diff --git a/datasets/lidc/configs.py b/datasets/lidc/configs.py index 413ce8f..e037756 100644 --- a/datasets/lidc/configs.py +++ b/datasets/lidc/configs.py @@ -1,445 +1,445 @@ #!/usr/bin/env python # Copyright 2019 Division of Medical Image Computing, German Cancer Research Center (DKFZ). # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== import sys import os from collections import namedtuple sys.path.append(os.path.dirname(os.path.realpath(__file__))) import numpy as np sys.path.append(os.path.dirname(os.path.realpath(__file__))+"/../..") from default_configs import DefaultConfigs # legends, nested classes are not handled well in multiprocessing! hence, Label class def in outer scope Label = namedtuple("Label", ['id', 'name', 'color', 'm_scores']) # m_scores = malignancy scores binLabel = namedtuple("binLabel", ['id', 'name', 'color', 'm_scores', 'bin_vals']) class Configs(DefaultConfigs): def __init__(self, server_env=None): super(Configs, self).__init__(server_env) ######################### # Preprocessing # ######################### self.root_dir = '/home/gregor/networkdrives/E130-Personal/Goetz/Datenkollektive/Lungendaten/Nodules_LIDC_IDRI' self.raw_data_dir = '{}/new_nrrd'.format(self.root_dir) self.pp_dir = '/media/gregor/HDD2TB/data/lidc/pp_20200309_dev' # 'merged' for one gt per image, 'single_annotator' for four gts per image. self.gts_to_produce = ["single_annotator", "merged"] self.target_spacing = (0.7, 0.7, 1.25) ######################### # I/O # ######################### # path to preprocessed data. #self.pp_name = 'pp_20190318' self.pp_name = 'pp_20200309_dev' self.input_df_name = 'info_df.pickle' self.data_sourcedir = '/media/gregor/HDD2TB/data/lidc/{}/'.format(self.pp_name) # settings for deployment on cluster. if server_env: # path to preprocessed data. self.data_sourcedir = '/datasets/data_ramien/lidc/{}_npz/'.format(self.pp_name) # one out of ['mrcnn', 'retina_net', 'retina_unet', 'detection_fpn']. self.model = 'mrcnn' self.model_path = 'models/{}.py'.format(self.model if not 'retina' in self.model else 'retina_net') self.model_path = os.path.join(self.source_dir, self.model_path) ######################### # Architecture # ######################### # dimension the model operates in. one out of [2, 3]. self.dim = 2 # 'class': standard object classification per roi, pairwise combinable with each of below tasks. # if 'class' is omitted from tasks, object classes will be fg/bg (1/0) from RPN. # 'regression': regress some vector per each roi # 'regression_ken_gal': use kendall-gal uncertainty sigma # 'regression_bin': classify each roi into a bin related to a regression scale - self.prediction_tasks = ['class'] + self.prediction_tasks = ['regression'] self.start_filts = 48 if self.dim == 2 else 18 self.end_filts = self.start_filts * 4 if self.dim == 2 else self.start_filts * 2 self.res_architecture = 'resnet50' # 'resnet101' , 'resnet50' self.norm = None # one of None, 'instance_norm', 'batch_norm' # one of 'xavier_uniform', 'xavier_normal', or 'kaiming_normal', None (=default = 'kaiming_uniform') self.weight_init = None self.regression_n_features = 1 ######################### # Data Loader # ######################### # distorted gt experiments: train on single-annotator gts in a random fashion to investigate network's # handling of noisy gts. # choose 'merged' for single, merged gt per image, or 'single_annotator' for four gts per image. # validation is always performed on same gt kind as training, testing always on merged gt. - self.training_gts = "merged" + self.training_gts = "sa" # select modalities from preprocessed data self.channels = [0] self.n_channels = len(self.channels) # patch_size to be used for training. pre_crop_size is the patch_size before data augmentation. self.pre_crop_size_2D = [320, 320] self.patch_size_2D = [320, 320] self.pre_crop_size_3D = [160, 160, 96] self.patch_size_3D = [160, 160, 96] self.patch_size = self.patch_size_2D if self.dim == 2 else self.patch_size_3D self.pre_crop_size = self.pre_crop_size_2D if self.dim == 2 else self.pre_crop_size_3D # ratio of free sampled batch elements before class balancing is triggered # (>0 to include "empty"/background patches.) self.batch_random_ratio = 0.3 self.balance_target = "class_targets" if 'class' in self.prediction_tasks else 'rg_bin_targets' # set 2D network to match 3D gt boxes. self.merge_2D_to_3D_preds = self.dim==2 self.observables_rois = [] #self.rg_map = {1:1, 2:2, 3:3, 4:4, 5:5} ######################### # Colors and Legends # ######################### self.plot_frequency = 5 binary_cl_labels = [Label(1, 'benign', (*self.dark_green, 1.), (1, 2)), Label(2, 'malignant', (*self.red, 1.), (3, 4, 5))] quintuple_cl_labels = [Label(1, 'MS1', (*self.dark_green, 1.), (1,)), Label(2, 'MS2', (*self.dark_yellow, 1.), (2,)), Label(3, 'MS3', (*self.orange, 1.), (3,)), Label(4, 'MS4', (*self.bright_red, 1.), (4,)), Label(5, 'MS5', (*self.red, 1.), (5,))] # choose here if to do 2-way or 5-way regression-bin classification task_spec_cl_labels = quintuple_cl_labels self.class_labels = [ # #id #name #color #malignancy score Label( 0, 'bg', (*self.gray, 0.), (0,))] if "class" in self.prediction_tasks: self.class_labels += task_spec_cl_labels else: self.class_labels += [Label(1, 'lesion', (*self.orange, 1.), (1,2,3,4,5))] if any(['regression' in task for task in self.prediction_tasks]): self.bin_labels = [binLabel(0, 'MS0', (*self.gray, 1.), (0,), (0,))] self.bin_labels += [binLabel(cll.id, cll.name, cll.color, cll.m_scores, tuple([ms for ms in cll.m_scores])) for cll in task_spec_cl_labels] self.bin_id2label = {label.id: label for label in self.bin_labels} self.ms2bin_label = {ms: label for label in self.bin_labels for ms in label.m_scores} bins = [(min(label.bin_vals), max(label.bin_vals)) for label in self.bin_labels] self.bin_id2rg_val = {ix: [np.mean(bin)] for ix, bin in enumerate(bins)} self.bin_edges = [(bins[i][1] + bins[i + 1][0]) / 2 for i in range(len(bins) - 1)] if self.class_specific_seg: self.seg_labels = self.class_labels else: self.seg_labels = [ # id #name #color Label(0, 'bg', (*self.gray, 0.)), Label(1, 'fg', (*self.orange, 1.)) ] self.class_id2label = {label.id: label for label in self.class_labels} self.class_dict = {label.id: label.name for label in self.class_labels if label.id != 0} # class_dict is used in evaluator / ap, auc, etc. statistics, and class 0 (bg) only needs to be # evaluated in debugging self.class_cmap = {label.id: label.color for label in self.class_labels} self.seg_id2label = {label.id: label for label in self.seg_labels} self.cmap = {label.id: label.color for label in self.seg_labels} self.plot_prediction_histograms = True self.plot_stat_curves = False self.has_colorchannels = False self.plot_class_ids = True self.num_classes = len(self.class_dict) # for instance classification (excl background) self.num_seg_classes = len(self.seg_labels) # incl background ######################### # Data Augmentation # ######################### self.da_kwargs={ 'mirror': True, 'mirror_axes': tuple(np.arange(0, self.dim, 1)), 'do_elastic_deform': True, 'alpha':(0., 1500.), 'sigma':(30., 50.), 'do_rotation':True, 'angle_x': (0., 2 * np.pi), 'angle_y': (0., 0), 'angle_z': (0., 0), 'do_scale': True, 'scale':(0.8, 1.1), 'random_crop':False, 'rand_crop_dist': (self.patch_size[0] / 2. - 3, self.patch_size[1] / 2. - 3), 'border_mode_data': 'constant', 'border_cval_data': 0, 'order_data': 1} if self.dim == 3: self.da_kwargs['do_elastic_deform'] = False self.da_kwargs['angle_x'] = (0, 0.0) self.da_kwargs['angle_y'] = (0, 0.0) #must be 0!! self.da_kwargs['angle_z'] = (0., 2 * np.pi) ################################# # Schedule / Selection / Optim # ################################# self.num_epochs = 130 if self.dim == 2 else 150 self.num_train_batches = 200 if self.dim == 2 else 200 self.batch_size = 20 if self.dim == 2 else 8 # decide whether to validate on entire patient volumes (like testing) or sampled patches (like training) # the former is morge accurate, while the latter is faster (depending on volume size) self.val_mode = 'val_sampling' # only 'val_sampling', 'val_patient' not implemented if self.val_mode == 'val_patient': raise NotImplementedError if self.val_mode == 'val_sampling': self.num_val_batches = 70 self.save_n_models = 4 # set a minimum epoch number for saving in case of instabilities in the first phase of training. self.min_save_thresh = 0 if self.dim == 2 else 0 # criteria to average over for saving epochs, 'criterion':weight. if "class" in self.prediction_tasks: # 'criterion': weight if len(self.class_labels)==3: self.model_selection_criteria = {"benign_ap": 0.5, "malignant_ap": 0.5} elif len(self.class_labels)==6: self.model_selection_criteria = {str(label.name)+"_ap": 1./5 for label in self.class_labels if label.id!=0} elif any("regression" in task for task in self.prediction_tasks): self.model_selection_criteria = {"lesion_ap": 0.2, "lesion_avp": 0.8} self.weight_decay = 0 self.clip_norm = 200 if 'regression_ken_gal' in self.prediction_tasks else None # number or None # int in [0, dataset_size]. select n patients from dataset for prototyping. If None, all data is used. self.select_prototype_subset = None #self.batch_size ######################### # Testing # ######################### # set the top-n-epochs to be saved for temporal averaging in testing. self.test_n_epochs = self.save_n_models self.test_aug_axes = (0,1,(0,1)) # None or list: choices are 0,1,(0,1) (0==spatial y, 1== spatial x). self.held_out_test_set = False self.max_test_patients = "all" # "all" or number self.report_score_level = ['rois', 'patient'] # choose list from 'patient', 'rois' self.patient_class_of_interest = 2 if 'class' in self.prediction_tasks else 1 self.metrics = ['ap', 'auc'] if any(['regression' in task for task in self.prediction_tasks]): self.metrics += ['avp', 'rg_MAE_weighted', 'rg_MAE_weighted_tp', 'rg_bin_accuracy_weighted', 'rg_bin_accuracy_weighted_tp'] if 'aleatoric' in self.model: self.metrics += ['rg_uncertainty', 'rg_uncertainty_tp', 'rg_uncertainty_tp_weighted'] self.evaluate_fold_means = True self.ap_match_ious = [0.1] # list of ious to be evaluated for ap-scoring. self.min_det_thresh = 0.1 # minimum confidence value to select predictions for evaluation. # aggregation method for test and val_patient predictions. # wbc = weighted box clustering as in https://arxiv.org/pdf/1811.08661.pdf, # nms = standard non-maximum suppression, or None = no clustering self.clustering = 'wbc' # iou thresh (exclusive!) for regarding two preds as concerning the same ROI self.clustering_iou = 0.1 # has to be larger than desired possible overlap iou of model predictions self.plot_prediction_histograms = True self.plot_stat_curves = False self.n_test_plots = 1 ######################### # Assertions # ######################### if not 'class' in self.prediction_tasks: assert self.num_classes == 1 ######################### # Add model specifics # ######################### {'detection_fpn': self.add_det_fpn_configs, 'mrcnn': self.add_mrcnn_configs, 'mrcnn_aleatoric': self.add_mrcnn_configs, 'retina_net': self.add_mrcnn_configs, 'retina_unet': self.add_mrcnn_configs, }[self.model]() def rg_val_to_bin_id(self, rg_val): return float(np.digitize(np.mean(rg_val), self.bin_edges)) def add_det_fpn_configs(self): self.learning_rate = [1e-4] * self.num_epochs self.dynamic_lr_scheduling = False # RoI score assigned to aggregation from pixel prediction (connected component). One of ['max', 'median']. self.score_det = 'max' # max number of roi candidates to identify per batch element and class. self.n_roi_candidates = 10 if self.dim == 2 else 30 # loss mode: either weighted cross entropy ('wce'), batch-wise dice loss ('dice), or the sum of both ('dice_wce') self.seg_loss_mode = 'wce' # if <1, false positive predictions in foreground are penalized less. self.fp_dice_weight = 1 if self.dim == 2 else 1 if len(self.class_labels)==3: self.wce_weights = [1., 1., 1.] if self.seg_loss_mode=="dice_wce" else [0.1, 1., 1.] elif len(self.class_labels)==6: self.wce_weights = [1., 1., 1., 1., 1., 1.] if self.seg_loss_mode == "dice_wce" else [0.1, 1., 1., 1., 1., 1.] else: raise Exception("mismatch loss weights & nr of classes") self.detection_min_confidence = self.min_det_thresh self.head_classes = self.num_seg_classes def add_mrcnn_configs(self): # learning rate is a list with one entry per epoch. self.learning_rate = [1e-4] * self.num_epochs self.dynamic_lr_scheduling = False # disable the re-sampling of mask proposals to original size for speed-up. # since evaluation is detection-driven (box-matching) and not instance segmentation-driven (iou-matching), # mask-outputs are optional. self.return_masks_in_train = False self.return_masks_in_val = True self.return_masks_in_test = False # set number of proposal boxes to plot after each epoch. self.n_plot_rpn_props = 5 if self.dim == 2 else 30 # number of classes for network heads: n_foreground_classes + 1 (background) self.head_classes = self.num_classes + 1 self.frcnn_mode = False # feature map strides per pyramid level are inferred from architecture. self.backbone_strides = {'xy': [4, 8, 16, 32], 'z': [1, 2, 4, 8]} # anchor scales are chosen according to expected object sizes in data set. Default uses only one anchor scale # per pyramid level. (outer list are pyramid levels (corresponding to BACKBONE_STRIDES), inner list are scales per level.) self.rpn_anchor_scales = {'xy': [[8], [16], [32], [64]], 'z': [[2], [4], [8], [16]]} # choose which pyramid levels to extract features from: P2: 0, P3: 1, P4: 2, P5: 3. self.pyramid_levels = [0, 1, 2, 3] # number of feature maps in rpn. typically lowered in 3D to save gpu-memory. self.n_rpn_features = 512 if self.dim == 2 else 128 # anchor ratios and strides per position in feature maps. self.rpn_anchor_ratios = [0.5, 1, 2] self.rpn_anchor_stride = 1 # Threshold for first stage (RPN) non-maximum suppression (NMS): LOWER == HARDER SELECTION self.rpn_nms_threshold = 0.7 if self.dim == 2 else 0.7 # loss sampling settings. self.rpn_train_anchors_per_image = 6 #per batch element self.train_rois_per_image = 6 #per batch element self.roi_positive_ratio = 0.5 self.anchor_matching_iou = 0.7 # factor of top-k candidates to draw from per negative sample (stochastic-hard-example-mining). # poolsize to draw top-k candidates from will be shem_poolsize * n_negative_samples. self.shem_poolsize = 10 self.pool_size = (7, 7) if self.dim == 2 else (7, 7, 3) self.mask_pool_size = (14, 14) if self.dim == 2 else (14, 14, 5) self.mask_shape = (28, 28) if self.dim == 2 else (28, 28, 10) self.rpn_bbox_std_dev = np.array([0.1, 0.1, 0.1, 0.2, 0.2, 0.2]) self.bbox_std_dev = np.array([0.1, 0.1, 0.1, 0.2, 0.2, 0.2]) self.window = np.array([0, 0, self.patch_size[0], self.patch_size[1], 0, self.patch_size_3D[2]]) self.scale = np.array([self.patch_size[0], self.patch_size[1], self.patch_size[0], self.patch_size[1], self.patch_size_3D[2], self.patch_size_3D[2]]) if self.dim == 2: self.rpn_bbox_std_dev = self.rpn_bbox_std_dev[:4] self.bbox_std_dev = self.bbox_std_dev[:4] self.window = self.window[:4] self.scale = self.scale[:4] # pre-selection in proposal-layer (stage 1) for NMS-speedup. applied per batch element. self.pre_nms_limit = 3000 if self.dim == 2 else 6000 # n_proposals to be selected after NMS per batch element. too high numbers blow up memory if "detect_while_training" is True, # since proposals of the entire batch are forwarded through second stage in as one "batch". self.roi_chunk_size = 2500 if self.dim == 2 else 600 self.post_nms_rois_training = 500 if self.dim == 2 else 75 self.post_nms_rois_inference = 500 # Final selection of detections (refine_detections) self.model_max_instances_per_batch_element = 10 if self.dim == 2 else 30 # per batch element and class. self.detection_nms_threshold = 1e-5 # needs to be > 0, otherwise all predictions are one cluster. self.model_min_confidence = 0.1 if self.dim == 2: self.backbone_shapes = np.array( [[int(np.ceil(self.patch_size[0] / stride)), int(np.ceil(self.patch_size[1] / stride))] for stride in self.backbone_strides['xy']]) else: self.backbone_shapes = np.array( [[int(np.ceil(self.patch_size[0] / stride)), int(np.ceil(self.patch_size[1] / stride)), int(np.ceil(self.patch_size[2] / stride_z))] for stride, stride_z in zip(self.backbone_strides['xy'], self.backbone_strides['z'] )]) if self.model == 'retina_net' or self.model == 'retina_unet': self.focal_loss = True # implement extra anchor-scales according to retina-net publication. self.rpn_anchor_scales['xy'] = [[ii[0], ii[0] * (2 ** (1 / 3)), ii[0] * (2 ** (2 / 3))] for ii in self.rpn_anchor_scales['xy']] self.rpn_anchor_scales['z'] = [[ii[0], ii[0] * (2 ** (1 / 3)), ii[0] * (2 ** (2 / 3))] for ii in self.rpn_anchor_scales['z']] self.n_anchors_per_pos = len(self.rpn_anchor_ratios) * 3 self.n_rpn_features = 256 if self.dim == 2 else 128 # pre-selection of detections for NMS-speedup. per entire batch. self.pre_nms_limit = (500 if self.dim == 2 else 6250) * self.batch_size # anchor matching iou is lower than in Mask R-CNN according to https://arxiv.org/abs/1708.02002 self.anchor_matching_iou = 0.5 if self.model == 'retina_unet': self.operate_stride1 = True diff --git a/datasets/lidc/data_loader.py b/datasets/lidc/data_loader.py index 222364d..fad15fc 100644 --- a/datasets/lidc/data_loader.py +++ b/datasets/lidc/data_loader.py @@ -1,1014 +1,1024 @@ # Copyright 2019 Division of Medical Image Computing, German Cancer Research Center (DKFZ). # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== ''' Data Loader for the LIDC data set. This dataloader expects preprocessed data in .npy or .npz files per patient and a pandas dataframe containing the meta info e.g. file paths, and some ground-truth info like labels, foreground slice ids. LIDC 4-fold annotations storage capacity problem: keep segmentation gts compressed (npz), unpack at each batch generation. ''' import plotting as plg import os import pickle import time from multiprocessing import Pool import numpy as np import pandas as pd from collections import OrderedDict # batch generator tools from https://github.com/MIC-DKFZ/batchgenerators from batchgenerators.transforms.spatial_transforms import MirrorTransform as Mirror from batchgenerators.transforms.abstract_transforms import Compose from batchgenerators.dataloading.multi_threaded_augmenter import MultiThreadedAugmenter from batchgenerators.transforms.spatial_transforms import SpatialTransform from batchgenerators.transforms.crop_and_pad_transforms import CenterCropTransform + import utils.dataloader_utils as dutils from utils.dataloader_utils import ConvertSegToBoundingBoxCoordinates - +from utils.dataloader_utils import BatchGenerator as BatchGeneratorParent def save_obj(obj, name): """Pickle a python object.""" with open(name + '.pkl', 'wb') as f: pickle.dump(obj, f, pickle.HIGHEST_PROTOCOL) def vector(item): """ensure item is vector-like (list or array or tuple) :param item: anything """ if not isinstance(item, (list, tuple, np.ndarray)): item = [item] return item class Dataset(dutils.Dataset): r"""Load a dict holding memmapped arrays and clinical parameters for each patient, evtly subset of those. If server_env: copy and evtly unpack (npz->npy) data in cf.data_rootdir to cf.data_dest. :param cf: config object. :param logger: logger. :param subset_ids: subset of patient/sample identifiers to load from whole set. :param data_sourcedir: directory in which to find data, defaults to cf.data_sourcedir if None. :return: dict with imgs, segs, pids, class_labels, observables """ def __init__(self, cf, logger=None, subset_ids=None, data_sourcedir=None, mode='train'): super(Dataset,self).__init__(cf, data_sourcedir) if mode == 'train' and not cf.training_gts == "merged": self.gt_dir = "patient_gts_sa" self.gt_kind = cf.training_gts else: self.gt_dir = "patient_gts_merged" self.gt_kind = "merged" if logger is not None: logger.info("loading {} ground truths for {}".format(self.gt_kind, 'training and validation' if mode=='train' else 'testing')) p_df = pd.read_pickle(os.path.join(self.data_sourcedir, self.gt_dir, cf.input_df_name)) #exclude_pids = ["0305a", "0447a"] # due to non-bg segmentation but bg mal label in nodules 5728, 8840 #p_df = p_df[~p_df.pid.isin(exclude_pids)] if subset_ids is not None: p_df = p_df[p_df.pid.isin(subset_ids)] if logger is not None: logger.info('subset: selected {} instances from df'.format(len(p_df))) if cf.select_prototype_subset is not None: prototype_pids = p_df.pid.tolist()[:cf.select_prototype_subset] p_df = p_df[p_df.pid.isin(prototype_pids)] if logger is not None: logger.warning('WARNING: using prototyping data subset of length {}!!!'.format(len(p_df))) pids = p_df.pid.tolist() # evtly copy data from data_sourcedir to data_dest if cf.server_env and not hasattr(cf, 'data_dir') and hasattr(cf, "data_dest"): # copy and unpack images file_subset = ["{}_img.npz".format(pid) for pid in pids if not os.path.isfile(os.path.join(cf.data_dest,'{}_img.npy'.format(pid)))] file_subset += [os.path.join(self.data_sourcedir, self.gt_dir, cf.input_df_name)] self.copy_data(cf, file_subset=file_subset, keep_packed=False, del_after_unpack=True) # copy and do not unpack segmentations file_subset = [os.path.join(self.gt_dir, "{}_rois.np*".format(pid)) for pid in pids] keep_packed = not cf.training_gts == "merged" self.copy_data(cf, file_subset=file_subset, keep_packed=keep_packed, del_after_unpack=(not keep_packed)) else: cf.data_dir = self.data_sourcedir ext = 'npy' if self.gt_kind == "merged" else 'npz' imgs = [os.path.join(self.data_dir, '{}_img.npy'.format(pid)) for pid in pids] segs = [os.path.join(self.data_dir, self.gt_dir, '{}_rois.{}'.format(pid, ext)) for pid in pids] orig_class_targets = p_df['class_target'].tolist() data = OrderedDict() if self.gt_kind == 'merged': for ix, pid in enumerate(pids): data[pid] = {'data': imgs[ix], 'seg': segs[ix], 'pid': pid} data[pid]['fg_slices'] = np.array(p_df['fg_slices'].tolist()[ix]) if 'class' in cf.prediction_tasks: if len(cf.class_labels)==3: # malignancy scores are binarized: (benign: 1-2 --> cl 1, malignant: 3-5 --> cl 2) data[pid]['class_targets'] = np.array([2 if ii >= 3 else 1 for ii in orig_class_targets[ix]], dtype='uint8') elif len(cf.class_labels)==6: # classify each malignancy score data[pid]['class_targets'] = np.array([1 if ii==0.5 else np.round(ii) for ii in orig_class_targets[ix]], dtype='uint8') else: raise Exception("mismatch class labels and data-loading implementations.") else: data[pid]['class_targets'] = np.ones_like(np.array(orig_class_targets[ix]), dtype='uint8') if any(['regression' in task for task in cf.prediction_tasks]): data[pid]["regression_targets"] = np.array([vector(v) for v in orig_class_targets[ix]], dtype='float16') data[pid]["rg_bin_targets"] = np.array( [cf.rg_val_to_bin_id(v) for v in data[pid]["regression_targets"]], dtype='uint8') else: for ix, pid in enumerate(pids): data[pid] = {'data': imgs[ix], 'seg': segs[ix], 'pid': pid} data[pid]['fg_slices'] = np.array(p_df['fg_slices'].values[ix]) if 'class' in cf.prediction_tasks: # malignancy scores are binarized: (benign: 1-2 --> cl 1, malignant: 3-5 --> cl 2) raise NotImplementedError # todo need to consider bg # data[pid]['class_targets'] = np.array( # [[2 if ii >= 3 else 1 for ii in four_fold_targs] for four_fold_targs in orig_class_targets[ix]]) else: data[pid]['class_targets'] = np.array( [[1 if ii > 0 else 0 for ii in four_fold_targs] for four_fold_targs in orig_class_targets[ix]], dtype='uint8') if any(['regression' in task for task in cf.prediction_tasks]): data[pid]["regression_targets"] = np.array( [[vector(v) for v in four_fold_targs] for four_fold_targs in orig_class_targets[ix]], dtype='float16') data[pid]["rg_bin_targets"] = np.array( [[cf.rg_val_to_bin_id(v) for v in four_fold_targs] for four_fold_targs in data[pid]["regression_targets"]], dtype='uint8') cf.roi_items = cf.observables_rois[:] cf.roi_items += ['class_targets'] if any(['regression' in task for task in cf.prediction_tasks]): cf.roi_items += ['regression_targets'] cf.roi_items += ['rg_bin_targets'] self.data = data self.set_ids = np.array(list(self.data.keys())) self.df = None # merged GTs class BatchGenerator_merged(dutils.BatchGenerator): """ creates the training/validation batch generator. Samples n_batch_size patients (draws a slice from each patient if 2D) from the data set while maintaining foreground-class balance. Returned patches are cropped/padded to pre_crop_size. Actual patch_size is obtained after data augmentation. :param data: data dictionary as provided by 'load_dataset'. :param batch_size: number of patients to sample for the batch :return dictionary containing the batch data (b, c, x, y, (z)) / seg (b, 1, x, y, (z)) / pids / class_target """ def __init__(self, cf, data, name="train"): super(BatchGenerator_merged, self).__init__(cf, data) self.crop_margin = np.array(self.cf.patch_size)/8. #min distance of ROI center to edge of cropped_patch. self.p_fg = 0.5 self.empty_samples_max_ratio = 0.6 self.random_count = int(cf.batch_random_ratio * cf.batch_size) self.class_targets = {k: v["class_targets"] for (k, v) in self._data.items()} self.balance_target_distribution(plot=name=="train") def generate_train_batch(self): # samples patients towards equilibrium of foreground classes on a roi-level after sampling a random ratio # fully random patients batch_patient_ids = list(np.random.choice(self.dataset_pids, size=self.random_count, replace=False)) # target-balanced patients batch_patient_ids += list(np.random.choice(self.dataset_pids, size=self.batch_size-self.random_count, replace=False, p=self.p_probs)) batch_data, batch_segs, batch_pids, batch_patient_labels = [], [], [], [] batch_roi_items = {name: [] for name in self.cf.roi_items} # record roi count of classes in batch batch_roi_counts = np.zeros((len(self.unique_ts),), dtype='uint32') batch_empty_counts = np.zeros((len(self.unique_ts),), dtype='uint32') # empty count for full bg samples (empty slices in 2D/patients in 3D) per class for sample in range(self.batch_size): patient = self._data[batch_patient_ids[sample]] data = np.transpose(np.load(patient['data'], mmap_mode='r'), axes=(1, 2, 0))[np.newaxis] seg = np.transpose(np.load(patient['seg'], mmap_mode='r'), axes=(1, 2, 0)) batch_pids.append(patient['pid']) (c, y, x, z) = data.shape if self.cf.dim == 2: elig_slices, choose_fg = [], False if len(patient['fg_slices']) > 0: if np.all(batch_empty_counts / self.batch_size >= self.empty_samples_max_ratio) or \ np.random.rand(1)<=self.p_fg: # fg is to be picked for tix in np.argsort(batch_roi_counts): # pick slices of patient that have roi of sought-for target # np.unique(seg[...,sl_ix][seg[...,sl_ix]>0]) gives roi_ids (numbering) of rois in slice sl_ix elig_slices = [sl_ix for sl_ix in np.arange(z) if np.count_nonzero( patient[self.balance_target][np.unique(seg[..., sl_ix][seg[..., sl_ix] > 0])-1] == self.unique_ts[tix]) > 0] if len(elig_slices) > 0: choose_fg = True break else: # pick bg elig_slices = np.setdiff1d(np.arange(z), patient['fg_slices']) if len(elig_slices)>0: sl_pick_ix = np.random.choice(elig_slices, size=None) else: sl_pick_ix = np.random.choice(z, size=None) data = data[..., sl_pick_ix] seg = seg[..., sl_pick_ix] # pad data if smaller than pre_crop_size. if np.any([data.shape[dim + 1] < ps for dim, ps in enumerate(self.cf.pre_crop_size)]): new_shape = [np.max([data.shape[dim + 1], ps]) for dim, ps in enumerate(self.cf.pre_crop_size)] data = dutils.pad_nd_image(data, new_shape, mode='constant') seg = dutils.pad_nd_image(seg, new_shape, mode='constant') # crop patches of size pre_crop_size, while sampling patches containing foreground with p_fg. crop_dims = [dim for dim, ps in enumerate(self.cf.pre_crop_size) if data.shape[dim + 1] > ps] if len(crop_dims) > 0: if self.cf.dim == 3: choose_fg = np.all(batch_empty_counts / self.batch_size >= self.empty_samples_max_ratio)\ or np.random.rand(1) <= self.p_fg if choose_fg and np.any(seg): available_roi_ids = np.unique(seg)[1:] for tix in np.argsort(batch_roi_counts): elig_roi_ids = available_roi_ids[patient[self.balance_target][available_roi_ids-1] == self.unique_ts[tix]] if len(elig_roi_ids)>0: seg_ics = np.argwhere(seg == np.random.choice(elig_roi_ids, size=None)) break roi_anchor_pixel = seg_ics[np.random.choice(seg_ics.shape[0], size=None)] assert seg[tuple(roi_anchor_pixel)] > 0 # sample the patch center coords. constrained by edges of images - pre_crop_size /2. And by # distance to the desired ROI < patch_size /2. # (here final patch size to account for center_crop after data augmentation). sample_seg_center = {} for ii in crop_dims: low = np.max((self.cf.pre_crop_size[ii]//2, roi_anchor_pixel[ii] - (self.cf.patch_size[ii]//2 - self.crop_margin[ii]))) high = np.min((data.shape[ii + 1] - self.cf.pre_crop_size[ii]//2, roi_anchor_pixel[ii] + (self.cf.patch_size[ii]//2 - self.crop_margin[ii]))) # happens if lesion on the edge of the image. dont care about roi anymore, # just make sure pre-crop is inside image. if low >= high: low = data.shape[ii + 1] // 2 - (data.shape[ii + 1] // 2 - self.cf.pre_crop_size[ii] // 2) high = data.shape[ii + 1] // 2 + (data.shape[ii + 1] // 2 - self.cf.pre_crop_size[ii] // 2) sample_seg_center[ii] = np.random.randint(low=low, high=high) else: # not guaranteed to be empty. probability of emptiness depends on the data. sample_seg_center = {ii: np.random.randint(low=self.cf.pre_crop_size[ii]//2, high=data.shape[ii + 1] - self.cf.pre_crop_size[ii]//2) for ii in crop_dims} for ii in crop_dims: min_crop = int(sample_seg_center[ii] - self.cf.pre_crop_size[ii] // 2) max_crop = int(sample_seg_center[ii] + self.cf.pre_crop_size[ii] // 2) data = np.take(data, indices=range(min_crop, max_crop), axis=ii + 1) seg = np.take(seg, indices=range(min_crop, max_crop), axis=ii) batch_data.append(data) batch_segs.append(seg[np.newaxis]) for o in batch_roi_items: #after loop, holds every entry of every batchpatient per roi-item batch_roi_items[o].append(patient[o]) if self.cf.dim == 3: for tix in range(len(self.unique_ts)): non_zero = np.count_nonzero(patient[self.balance_target] == self.unique_ts[tix]) batch_roi_counts[tix] += non_zero batch_empty_counts[tix] += int(non_zero==0) # todo remove assert when checked if not np.any(seg): assert non_zero==0 elif self.cf.dim == 2: for tix in range(len(self.unique_ts)): non_zero = np.count_nonzero(patient[self.balance_target][np.unique(seg[seg>0]) - 1] == self.unique_ts[tix]) batch_roi_counts[tix] += non_zero batch_empty_counts[tix] += int(non_zero == 0) # todo remove assert when checked if not np.any(seg): assert non_zero==0 data = np.array(batch_data).astype(np.float16) seg = np.array(batch_segs).astype(np.uint8) batch = {'data': data, 'seg': seg, 'pid': batch_pids, 'roi_counts':batch_roi_counts, 'empty_counts': batch_empty_counts} for key,val in batch_roi_items.items(): #extend batch dic by roi-wise items (obs, class ids, regression vectors...) batch[key] = np.array(val) return batch class PatientBatchIterator_merged(dutils.PatientBatchIterator): """ creates a test generator that iterates over entire given dataset returning 1 patient per batch. Can be used for monitoring if cf.val_mode = 'patient_val' for a monitoring closer to actualy evaluation (done in 3D), if willing to accept speed-loss during training. :return: out_batch: dictionary containing one patient with batch_size = n_3D_patches in 3D or batch_size = n_2D_patches in 2D . """ def __init__(self, cf, data): # threads in augmenter super(PatientBatchIterator_merged, self).__init__(cf, data) self.patient_ix = 0 self.patch_size = cf.patch_size + [1] if cf.dim == 2 else cf.patch_size def generate_train_batch(self, pid=None): if pid is None: pid = self.dataset_pids[self.patient_ix] patient = self._data[pid] data = np.transpose(np.load(patient['data'], mmap_mode='r'), axes=(1, 2, 0)) seg = np.transpose(np.load(patient['seg'], mmap_mode='r'), axes=(1, 2, 0)) # pad data if smaller than patch_size seen during training. if np.any([data.shape[dim] < ps for dim, ps in enumerate(self.patch_size)]): new_shape = [np.max([data.shape[dim], self.patch_size[dim]]) for dim, ps in enumerate(self.patch_size)] data = dutils.pad_nd_image(data, new_shape) # use 'return_slicer' to crop image back to original shape. seg = dutils.pad_nd_image(seg, new_shape) # get 3D targets for evaluation, even if network operates in 2D. 2D predictions will be merged to 3D in predictor. if self.cf.dim == 3 or self.cf.merge_2D_to_3D_preds: out_data = data[np.newaxis, np.newaxis] out_seg = seg[np.newaxis, np.newaxis] batch_3D = {'data': out_data, 'seg': out_seg} for o in self.cf.roi_items: batch_3D[o] = np.array([patient[o]]) converter = ConvertSegToBoundingBoxCoordinates(3, self.cf.roi_items, False, self.cf.class_specific_seg) batch_3D = converter(**batch_3D) batch_3D.update({'patient_bb_target': batch_3D['bb_target'], 'original_img_shape': out_data.shape}) for o in self.cf.roi_items: batch_3D["patient_" + o] = batch_3D[o] if self.cf.dim == 2: out_data = np.transpose(data, axes=(2, 0, 1))[:, np.newaxis] # (z, c, x, y ) out_seg = np.transpose(seg, axes=(2, 0, 1))[:, np.newaxis] batch_2D = {'data': out_data, 'seg': out_seg} for o in self.cf.roi_items: batch_2D[o] = np.repeat(np.array([patient[o]]), out_data.shape[0], axis=0) converter = ConvertSegToBoundingBoxCoordinates(2, self.cf.roi_items, False, self.cf.class_specific_seg) batch_2D = converter(**batch_2D) if self.cf.merge_2D_to_3D_preds: batch_2D.update({'patient_bb_target': batch_3D['patient_bb_target'], 'original_img_shape': out_data.shape}) for o in self.cf.roi_items: batch_2D["patient_" + o] = batch_3D[o] else: batch_2D.update({'patient_bb_target': batch_2D['bb_target'], 'original_img_shape': out_data.shape}) for o in self.cf.roi_items: batch_2D["patient_" + o] = batch_2D[o] out_batch = batch_3D if self.cf.dim == 3 else batch_2D out_batch.update({'pid': np.array([patient['pid']] * len(out_data))}) # crop patient-volume to patches of patch_size used during training. stack patches up in batch dimension. # in this case, 2D is treated as a special case of 3D with patch_size[z] = 1. if np.any([data.shape[dim] > self.patch_size[dim] for dim in range(3)]): patient_batch = out_batch patch_crop_coords_list = dutils.get_patch_crop_coords(data, self.patch_size) new_img_batch, new_seg_batch = [], [] for cix, c in enumerate(patch_crop_coords_list): seg_patch = seg[c[0]:c[1], c[2]: c[3], c[4]:c[5]] new_seg_batch.append(seg_patch) tmp_c_5 = c[5] new_img_batch.append(data[c[0]:c[1], c[2]:c[3], c[4]:tmp_c_5]) data = np.array(new_img_batch)[:, np.newaxis] # (n_patches, c, x, y, z) seg = np.array(new_seg_batch)[:, np.newaxis] # (n_patches, 1, x, y, z) if self.cf.dim == 2: # all patches have z dimension 1 (slices). discard dimension data = data[..., 0] seg = seg[..., 0] patch_batch = {'data': data.astype('float32'), 'seg': seg.astype('uint8'), 'pid': np.array([patient['pid']] * data.shape[0])} for o in self.cf.roi_items: patch_batch[o] = np.repeat(np.array([patient[o]]), len(patch_crop_coords_list), axis=0) # patient-wise (orig) batch info for putting the patches back together after prediction for o in self.cf.roi_items: patch_batch["patient_" + o] = patient_batch['patient_' + o] if self.cf.dim == 2: # this could also be named "unpatched_2d_roi_items" patch_batch["patient_" + o + "_2d"] = patient_batch[o] # adding patient-wise data and seg adds about 2 GB of additional RAM consumption to a batch 20x288x288 # and enables calculating test-dice/viewing patient-wise results in test # remove, but also remove dice from metrics, when like to save memory patch_batch['patient_data'] = patient_batch['data'] patch_batch['patient_seg'] = patient_batch['seg'] patch_batch['patch_crop_coords'] = np.array(patch_crop_coords_list) patch_batch['patient_bb_target'] = patient_batch['patient_bb_target'] if self.cf.dim == 2: patch_batch['patient_bb_target_2d'] = patient_batch['bb_target'] patch_batch['original_img_shape'] = patient_batch['original_img_shape'] converter = ConvertSegToBoundingBoxCoordinates(self.cf.dim, self.cf.roi_items, False, self.cf.class_specific_seg) patch_batch = converter(**patch_batch) out_batch = patch_batch self.patient_ix += 1 if self.patient_ix == len(self.dataset_pids): self.patient_ix = 0 return out_batch # single-annotator GTs -class BatchGenerator_sa(dutils.BatchGenerator): +class BatchGenerator_sa(BatchGeneratorParent): """ creates the training/validation batch generator. Samples n_batch_size patients (draws a slice from each patient if 2D) from the data set while maintaining foreground-class balance. Returned patches are cropped/padded to pre_crop_size. Actual patch_size is obtained after data augmentation. :param data: data dictionary as provided by 'load_dataset'. :param batch_size: number of patients to sample for the batch :return dictionary containing the batch data (b, c, x, y, (z)) / seg (b, 1, x, y, (z)) / pids / class_target """ # noinspection PyMethodOverriding def balance_target_distribution(self, rater, plot=False): """ :param rater: for which rater slot to generate the distribution :param self.targets: dic holding {patient_specifier : patient-wise-unique ROI targets} :param plot: whether to plot the generated patient distributions :return: probability distribution over all pids. draw without replace from this. """ unique_ts = np.unique([v[rater] for pat in self.targets.values() for v in pat]) sample_stats = pd.DataFrame(columns=[str(ix) + suffix for ix in unique_ts for suffix in ["", "_bg"]], index=list(self.targets.keys())) for pid in sample_stats.index: for targ in unique_ts: fg_count = 0 if len(self.targets[pid]) == 0 else np.count_nonzero(self.targets[pid][:, rater] == targ) sample_stats.loc[pid, str(targ)] = int(fg_count > 0) sample_stats.loc[pid, str(targ) + "_bg"] = int(fg_count == 0) target_stats = sample_stats.agg( ("sum", lambda col: col.sum() / len(self._data)), axis=0, sort=False).rename({"": "relative"}) anchor = 1. - target_stats.loc["relative"].iloc[0] fg_bg_weights = anchor / target_stats.loc["relative"] cum_weights = anchor * len(fg_bg_weights) fg_bg_weights /= cum_weights p_probs = sample_stats.apply(self.sample_targets_to_weights, args=(fg_bg_weights,), axis=1).sum(axis=1) p_probs = p_probs / p_probs.sum() if plot: print("Rater: {}. Applying class-weights:\n {}".format(rater, fg_bg_weights)) if len(sample_stats.columns) == 2: # assert that probs are calc'd correctly: # (p_probs * sample_stats["1"]).sum() == (p_probs * sample_stats["1_bg"]).sum() # only works if one label per patient (multi-label expectations depend on multi-label occurences). for rater in range(self.rater_bsize): expectations = [] for targ in sample_stats.columns: expectations.append((p_probs[rater] * sample_stats[targ]).sum()) assert np.allclose(expectations, expectations[0], atol=1e-4), "expectation values for fgs/bgs: {}".format( expectations) if plot: plg.plot_batchgen_distribution(self.cf, self.dataset_pids, p_probs, self.balance_target, out_file=os.path.join(self.plot_dir, "train_gen_distr_"+str(self.cf.fold)+"_rater"+str(rater)+".png")) return p_probs, unique_ts, sample_stats def __init__(self, cf, data, name="train"): super(BatchGenerator_sa, self).__init__(cf, data) self.name = name self.crop_margin = np.array(self.cf.patch_size) / 8. # min distance of ROI center to edge of cropped_patch. self.p_fg = 0.5 self.empty_samples_max_ratio = 0.6 self.random_count = int(cf.batch_random_ratio * cf.batch_size) self.rater_bsize = 4 unique_ts_total = set() self.p_probs = [] self.sample_stats = [] - p = Pool(processes=cf.n_workers) - mp_res = p.starmap(self.balance_target_distribution, [(r, name=="train") for r in range(self.rater_bsize)]) - p.close() - p.join() - for r, res in enumerate(mp_res): - p_probs, unique_ts, sample_stats = res + # todo resolve pickling error + # p = Pool(processes=min(self.rater_bsize, cf.n_workers)) + # mp_res = p.starmap(self.balance_target_distribution, [(r, name=="train") for r in range(self.rater_bsize)]) + # p.close() + # p.join() + # for r, res in enumerate(mp_res): + # p_probs, unique_ts, sample_stats = res + # self.p_probs.append(p_probs) + # self.sample_stats.append(sample_stats) + # unique_ts_total.update(unique_ts) + + for r in range(self.rater_bsize): + # todo multiprocess. takes forever + p_probs, unique_ts, sample_stats = self.balance_target_distribution(r, plot=name == "train") self.p_probs.append(p_probs) self.sample_stats.append(sample_stats) unique_ts_total.update(unique_ts) self.unique_ts = sorted(list(unique_ts_total)) - + self.stats = {"roi_counts": np.zeros(len(self.unique_ts,), dtype='uint32'), + "empty_counts": np.zeros(len(self.unique_ts,), dtype='uint32')} def generate_train_batch(self): rater = np.random.randint(self.rater_bsize) # samples patients towards equilibrium of foreground classes on a roi-level (after randomly sampling the ratio batch_random_ratio). # random patients batch_patient_ids = list(np.random.choice(self.dataset_pids, size=self.random_count, replace=False)) # target-balanced patients batch_patient_ids += list(np.random.choice(self.dataset_pids, size=self.batch_size-self.random_count, replace=False, p=self.p_probs[rater])) batch_data, batch_segs, batch_pids, batch_patient_labels = [], [], [], [] batch_roi_items = {name: [] for name in self.cf.roi_items} # record roi count of classes in batch batch_roi_counts = np.zeros((len(self.unique_ts),), dtype='uint32') batch_empty_counts = np.zeros((len(self.unique_ts),), dtype='uint32') # empty count for full bg samples (empty slices in 2D/patients in 3D) for sample in range(self.batch_size): patient = self._data[batch_patient_ids[sample]] patient_balance_ts = np.array([roi[rater] for roi in patient[self.balance_target]]) data = np.transpose(np.load(patient['data'], mmap_mode='r'), axes=(1, 2, 0))[np.newaxis] seg = np.load(patient['seg'], mmap_mode='r') seg = np.transpose(seg[list(seg.keys())[0]][rater], axes=(1, 2, 0)) batch_pids.append(patient['pid']) (c, y, x, z) = data.shape if self.cf.dim == 2: elig_slices, choose_fg = [], False if len(patient['fg_slices']) > 0: if np.all(batch_empty_counts / self.batch_size >= self.empty_samples_max_ratio) or \ np.random.rand(1) <= self.p_fg: # fg is to be picked for tix in np.argsort(batch_roi_counts): # pick slices of patient that have roi of sought-for target # np.unique(seg[...,sl_ix][seg[...,sl_ix]>0]) gives roi_ids (numbering) of rois in slice sl_ix elig_slices = [sl_ix for sl_ix in np.arange(z) if np.count_nonzero( patient_balance_ts[np.unique(seg[..., sl_ix][seg[..., sl_ix] > 0]) - 1] == self.unique_ts[tix]) > 0] if len(elig_slices) > 0: choose_fg = True break else: # pick bg elig_slices = np.setdiff1d(np.arange(z), patient['fg_slices'][rater]) if len(elig_slices) > 0: sl_pick_ix = np.random.choice(elig_slices, size=None) else: sl_pick_ix = np.random.choice(z, size=None) data = data[..., sl_pick_ix] seg = seg[..., sl_pick_ix] # pad data if smaller than pre_crop_size. if np.any([data.shape[dim + 1] < ps for dim, ps in enumerate(self.cf.pre_crop_size)]): new_shape = [np.max([data.shape[dim + 1], ps]) for dim, ps in enumerate(self.cf.pre_crop_size)] data = dutils.pad_nd_image(data, new_shape, mode='constant') seg = dutils.pad_nd_image(seg, new_shape, mode='constant') # crop patches of size pre_crop_size, while sampling patches containing foreground with p_fg. crop_dims = [dim for dim, ps in enumerate(self.cf.pre_crop_size) if data.shape[dim + 1] > ps] if len(crop_dims) > 0: if self.cf.dim == 3: choose_fg = np.all(batch_empty_counts / self.batch_size >= self.empty_samples_max_ratio) or \ np.random.rand(1) <= self.p_fg if choose_fg and np.any(seg): available_roi_ids = np.unique(seg[seg>0]) assert np.all(patient_balance_ts[available_roi_ids-1]>0), "trying to choose roi with rating 0" for tix in np.argsort(batch_roi_counts): elig_roi_ids = available_roi_ids[ patient_balance_ts[available_roi_ids-1] == self.unique_ts[tix] ] if len(elig_roi_ids)>0: seg_ics = np.argwhere(seg == np.random.choice(elig_roi_ids, size=None)) roi_anchor_pixel = seg_ics[np.random.choice(seg_ics.shape[0], size=None)] break assert seg[tuple(roi_anchor_pixel)] > 0, "roi_anchor_pixel not inside roi: {}, pb_ts {}, elig ids {}".format(tuple(roi_anchor_pixel), patient_balance_ts, elig_roi_ids) # sample the patch center coords. constrained by edges of images - pre_crop_size /2. And by # distance to the desired ROI < patch_size /2. # (here final patch size to account for center_crop after data augmentation). sample_seg_center = {} for ii in crop_dims: low = np.max((self.cf.pre_crop_size[ii]//2, roi_anchor_pixel[ii] - (self.cf.patch_size[ii]//2 - self.crop_margin[ii]))) high = np.min((data.shape[ii + 1] - self.cf.pre_crop_size[ii]//2, roi_anchor_pixel[ii] + (self.cf.patch_size[ii]//2 - self.crop_margin[ii]))) # happens if lesion on the edge of the image. dont care about roi anymore, # just make sure pre-crop is inside image. if low >= high: low = data.shape[ii + 1] // 2 - (data.shape[ii + 1] // 2 - self.cf.pre_crop_size[ii] // 2) high = data.shape[ii + 1] // 2 + (data.shape[ii + 1] // 2 - self.cf.pre_crop_size[ii] // 2) sample_seg_center[ii] = np.random.randint(low=low, high=high) else: # not guaranteed to be empty. probability of emptiness depends on the data. sample_seg_center = {ii: np.random.randint(low=self.cf.pre_crop_size[ii]//2, high=data.shape[ii + 1] - self.cf.pre_crop_size[ii]//2) for ii in crop_dims} for ii in crop_dims: min_crop = int(sample_seg_center[ii] - self.cf.pre_crop_size[ii] // 2) max_crop = int(sample_seg_center[ii] + self.cf.pre_crop_size[ii] // 2) data = np.take(data, indices=range(min_crop, max_crop), axis=ii + 1) seg = np.take(seg, indices=range(min_crop, max_crop), axis=ii) batch_data.append(data) batch_segs.append(seg[np.newaxis]) for o in batch_roi_items: #after loop, holds every entry of every batchpatient per roi-item batch_roi_items[o].append([roi[rater] for roi in patient[o]]) if self.cf.dim == 3: for tix in range(len(self.unique_ts)): non_zero = np.count_nonzero(patient[self.balance_target] == self.unique_ts[tix]) batch_roi_counts[tix] += non_zero batch_empty_counts[tix] += int(non_zero==0) # todo remove assert when checked if not np.any(seg): assert non_zero==0 elif self.cf.dim == 2: for tix in range(len(self.unique_ts)): non_zero = np.count_nonzero(patient[self.balance_target][np.unique(seg[seg>0]) - 1] == self.unique_ts[tix]) batch_roi_counts[tix] += non_zero batch_empty_counts[tix] += int(non_zero == 0) # todo remove assert when checked if not np.any(seg): assert non_zero==0 data = np.array(batch_data).astype('float16') seg = np.array(batch_segs).astype('uint8') batch = {'data': data, 'seg': seg, 'pid': batch_pids, 'rater_id': rater, - 'roi_counts':batch_roi_counts, 'empty_counts': batch_empty_counts} + 'roi_counts': batch_roi_counts, 'empty_counts': batch_empty_counts} for key,val in batch_roi_items.items(): #extend batch dic by roi-wise items (obs, class ids, regression vectors...) batch[key] = np.array(val) return batch class PatientBatchIterator_sa(dutils.PatientBatchIterator): """ creates a test generator that iterates over entire given dataset returning 1 patient per batch. Can be used for monitoring if cf.val_mode = 'patient_val' for a monitoring closer to actual evaluation (done in 3D), if willing to accept speed loss during training. :return: out_batch: dictionary containing one patient with batch_size = n_3D_patches in 3D or batch_size = n_2D_patches in 2D . This is the data & gt loader for the 4-fold single-annotator GTs: each data input has separate annotations of 4 annotators. the way the pipeline is currently setup, the single-annotator GTs are only used if training with validation mode val_patient; during testing the Iterator with the merged GTs is used. # todo mode val_patient not implemented yet (since very slow). would need to sample from all available rater GTs. """ def __init__(self, cf, data): #threads in augmenter super(PatientBatchIterator_sa, self).__init__(cf, data) self.cf = cf self.patient_ix = 0 self.dataset_pids = list(self._data.keys()) self.patch_size = cf.patch_size+[1] if cf.dim==2 else cf.patch_size self.rater_bsize = 4 def generate_train_batch(self, pid=None): if pid is None: pid = self.dataset_pids[self.patient_ix] patient = self._data[pid] data = np.transpose(np.load(patient['data'], mmap_mode='r'), axes=(1, 2, 0)) # all gts are 4-fold and npz! seg = np.load(patient['seg'], mmap_mode='r') seg = np.transpose(seg[list(seg.keys())[0]], axes=(0, 2, 3, 1)) # pad data if smaller than patch_size seen during training. if np.any([data.shape[dim] < ps for dim, ps in enumerate(self.patch_size)]): new_shape = [np.max([data.shape[dim], self.patch_size[dim]]) for dim, ps in enumerate(self.patch_size)] data = dutils.pad_nd_image(data, new_shape) # use 'return_slicer' to crop image back to original shape. seg = dutils.pad_nd_image(seg, new_shape) # get 3D targets for evaluation, even if network operates in 2D. 2D predictions will be merged to 3D in predictor. if self.cf.dim == 3 or self.cf.merge_2D_to_3D_preds: out_data = data[np.newaxis, np.newaxis] out_seg = seg[:, np.newaxis] batch_3D = {'data': out_data, 'seg': out_seg} for item in self.cf.roi_items: batch_3D[item] = [] for r in range(self.rater_bsize): for item in self.cf.roi_items: batch_3D[item].append(np.array([roi[r] for roi in patient[item]])) converter = ConvertSegToBoundingBoxCoordinates(3, self.cf.roi_items, False, self.cf.class_specific_seg) batch_3D = converter(**batch_3D) batch_3D.update({'patient_bb_target': batch_3D['bb_target'], 'original_img_shape': out_data.shape}) for o in self.cf.roi_items: batch_3D["patient_" + o] = batch_3D[o] if self.cf.dim == 2: out_data = np.transpose(data, axes=(2, 0, 1))[:, np.newaxis] # (z, c, y, x ) out_seg = np.transpose(seg, axes=(0, 3, 1, 2))[:, :, np.newaxis] # (n_raters, z, 1, y,x) batch_2D = {'data': out_data} for item in ["seg", "bb_target"]+self.cf.roi_items: batch_2D[item] = [] converter = ConvertSegToBoundingBoxCoordinates(2, self.cf.roi_items, False, self.cf.class_specific_seg) for r in range(self.rater_bsize): tmp_batch = {"seg": out_seg[r]} for item in self.cf.roi_items: tmp_batch[item] = np.repeat(np.array([[roi[r] for roi in patient[item]]]), out_data.shape[0], axis=0) tmp_batch = converter(**tmp_batch) for item in ["seg", "bb_target"]+self.cf.roi_items: batch_2D[item].append(tmp_batch[item]) # for item in ["seg", "bb_target"]+self.cf.roi_items: # batch_2D[item] = np.array(batch_2D[item]) if self.cf.merge_2D_to_3D_preds: batch_2D.update({'patient_bb_target': batch_3D['patient_bb_target'], 'original_img_shape': out_data.shape}) for o in self.cf.roi_items: batch_2D["patient_" + o] = batch_3D[o] else: batch_2D.update({'patient_bb_target': batch_2D['bb_target'], 'original_img_shape': out_data.shape}) for o in self.cf.roi_items: batch_2D["patient_" + o] = batch_2D[o] out_batch = batch_3D if self.cf.dim == 3 else batch_2D out_batch.update({'pid': np.array([patient['pid']] * out_data.shape[0])}) # crop patient-volume to patches of patch_size used during training. stack patches up in batch dimension. # in this case, 2D is treated as a special case of 3D with patch_size[z] = 1. if np.any([data.shape[dim] > self.patch_size[dim] for dim in range(3)]): patient_batch = out_batch patch_crop_coords_list = dutils.get_patch_crop_coords(data, self.patch_size) new_img_batch = [] new_seg_batch = [] for cix, c in enumerate(patch_crop_coords_list): seg_patch = seg[:, c[0]:c[1], c[2]: c[3], c[4]:c[5]] new_seg_batch.append(seg_patch) tmp_c_5 = c[5] new_img_batch.append(data[c[0]:c[1], c[2]:c[3], c[4]:tmp_c_5]) data = np.array(new_img_batch)[:, np.newaxis] # (n_patches, c, x, y, z) seg = np.transpose(np.array(new_seg_batch), axes=(1,0,2,3,4))[:,:,np.newaxis] # (n_raters, n_patches, x, y, z) if self.cf.dim == 2: # all patches have z dimension 1 (slices). discard dimension data = data[..., 0] seg = seg[..., 0] patch_batch = {'data': data.astype('float32'), 'pid': np.array([patient['pid']] * data.shape[0])} # for o in self.cf.roi_items: # patch_batch[o] = np.repeat(np.array([patient[o]]), len(patch_crop_coords_list), axis=0) converter = ConvertSegToBoundingBoxCoordinates(self.cf.dim, self.cf.roi_items, False, self.cf.class_specific_seg) for item in ["seg", "bb_target"]+self.cf.roi_items: patch_batch[item] = [] # coord_list = [np.min(seg_ixs[:, 1]) - 1, np.min(seg_ixs[:, 2]) - 1, np.max(seg_ixs[:, 1]) + 1, # IndexError: index 2 is out of bounds for axis 1 with size 2 for r in range(self.rater_bsize): tmp_batch = {"seg": seg[r]} for item in self.cf.roi_items: tmp_batch[item] = np.repeat(np.array([[roi[r] for roi in patient[item]]]), len(patch_crop_coords_list), axis=0) tmp_batch = converter(**tmp_batch) for item in ["seg", "bb_target"]+self.cf.roi_items: patch_batch[item].append(tmp_batch[item]) # patient-wise (orig) batch info for putting the patches back together after prediction for o in self.cf.roi_items: patch_batch["patient_" + o] = patient_batch['patient_'+o] if self.cf.dim==2: # this could also be named "unpatched_2d_roi_items" patch_batch["patient_"+o+"_2d"] = patient_batch[o] # adding patient-wise data and seg adds about 2 GB of additional RAM consumption to a batch 20x288x288 # and enables calculating test-dice/viewing patient-wise results in test # remove, but also remove dice from metrics, if you like to save memory patch_batch['patient_data'] = patient_batch['data'] patch_batch['patient_seg'] = patient_batch['seg'] patch_batch['patch_crop_coords'] = np.array(patch_crop_coords_list) patch_batch['patient_bb_target'] = patient_batch['patient_bb_target'] if self.cf.dim==2: patch_batch['patient_bb_target_2d'] = patient_batch['bb_target'] patch_batch['original_img_shape'] = patient_batch['original_img_shape'] out_batch = patch_batch self.patient_ix += 1 if self.patient_ix == len(self.dataset_pids): self.patient_ix = 0 return out_batch def create_data_gen_pipeline(cf, patient_data, is_training=True): """ create multi-threaded train/val/test batch generation and augmentation pipeline. :param cf: configs object. :param patient_data: dictionary containing one dictionary per patient in the train/test subset. :param is_training: (optional) whether to perform data augmentation (training) or not (validation/testing) :return: multithreaded_generator """ BG_name = "train" if is_training else "val" data_gen = BatchGenerator_merged(cf, patient_data, name=BG_name) if cf.training_gts=='merged' else \ BatchGenerator_sa(cf, patient_data, name=BG_name) # add transformations to pipeline. my_transforms = [] if is_training: if cf.da_kwargs["mirror"]: mirror_transform = Mirror(axes=cf.da_kwargs['mirror_axes']) my_transforms.append(mirror_transform) spatial_transform = SpatialTransform(patch_size=cf.patch_size[:cf.dim], patch_center_dist_from_border=cf.da_kwargs['rand_crop_dist'], do_elastic_deform=cf.da_kwargs['do_elastic_deform'], alpha=cf.da_kwargs['alpha'], sigma=cf.da_kwargs['sigma'], do_rotation=cf.da_kwargs['do_rotation'], angle_x=cf.da_kwargs['angle_x'], angle_y=cf.da_kwargs['angle_y'], angle_z=cf.da_kwargs['angle_z'], do_scale=cf.da_kwargs['do_scale'], scale=cf.da_kwargs['scale'], random_crop=cf.da_kwargs['random_crop']) my_transforms.append(spatial_transform) else: my_transforms.append(CenterCropTransform(crop_size=cf.patch_size[:cf.dim])) if cf.create_bounding_box_targets: my_transforms.append(ConvertSegToBoundingBoxCoordinates(cf.dim, cf.roi_items, False, cf.class_specific_seg)) all_transforms = Compose(my_transforms) multithreaded_generator = MultiThreadedAugmenter(data_gen, all_transforms, num_processes=cf.n_workers, seeds=range(cf.n_workers)) return multithreaded_generator def get_train_generators(cf, logger, data_statistics=True): """ wrapper function for creating the training batch generator pipeline. returns the train/val generators. selects patients according to cv folds (generated by first run/fold of experiment): splits the data into n-folds, where 1 split is used for val, 1 split for testing and the rest for training. (inner loop test set) If cf.held_out_test_set is True, adds the test split to the training data. """ dataset = Dataset(cf, logger) dataset.init_FoldGenerator(cf.seed, cf.n_cv_splits) dataset.generate_splits(check_file=os.path.join(cf.exp_dir, 'fold_ids.pickle')) set_splits = dataset.fg.splits test_ids, val_ids = set_splits.pop(cf.fold), set_splits.pop(cf.fold - 1) train_ids = np.concatenate(set_splits, axis=0) if cf.held_out_test_set: train_ids = np.concatenate((train_ids, test_ids), axis=0) test_ids = [] train_data = {k: v for (k, v) in dataset.data.items() if k in train_ids} val_data = {k: v for (k, v) in dataset.data.items() if k in val_ids} logger.info("data set loaded with: {} train / {} val / {} test patients".format(len(train_ids), len(val_ids), len(test_ids))) if data_statistics: dataset.calc_statistics(subsets={"train": train_ids, "val": val_ids, "test": test_ids}, plot_dir=os.path.join(cf.plot_dir,"dataset")) batch_gen = {} batch_gen['train'] = create_data_gen_pipeline(cf, train_data, is_training=True) batch_gen['val_sampling'] = create_data_gen_pipeline(cf, val_data, is_training=False) if cf.val_mode == 'val_patient': assert cf.training_gts == 'merged', 'val_patient not yet implemented for sa gts' batch_gen['val_patient'] = PatientBatchIterator_merged(cf, val_data) if cf.training_gts=='merged' \ else PatientBatchIterator_sa(cf, val_data) batch_gen['n_val'] = len(val_data) if cf.max_val_patients=="all" else min(len(val_data), cf.max_val_patients) else: batch_gen['n_val'] = cf.num_val_batches return batch_gen def get_test_generator(cf, logger): """ wrapper function for creating the test batch generator pipeline. selects patients according to cv folds (generated by first run/fold of experiment) If cf.held_out_test_set is True, gets the data from an external folder instead. """ if cf.held_out_test_set: sourcedir = cf.test_data_sourcedir test_ids = None else: sourcedir = None with open(os.path.join(cf.exp_dir, 'fold_ids.pickle'), 'rb') as handle: set_splits = pickle.load(handle) test_ids = set_splits[cf.fold] test_data = Dataset(cf, logger, subset_ids=test_ids, data_sourcedir=sourcedir, mode="test").data logger.info("data set loaded with: {} test patients".format(len(test_ids))) batch_gen = {} batch_gen['test'] = PatientBatchIterator_merged(cf, test_data) batch_gen['n_test'] = len(test_ids) if cf.max_test_patients == "all" else min(cf.max_test_patients, len(test_ids)) return batch_gen if __name__ == "__main__": import sys sys.path.append('../') import plotting as plg import utils.exp_utils as utils from configs import Configs cf = Configs() cf.batch_size = 3 #dataset_path = os.path.dirname(os.path.realpath(__file__)) #exp_path = os.path.join(dataset_path, "experiments/dev") #cf = utils.prep_exp(dataset_path, exp_path, server_env=False, use_stored_settings=False, is_training=True) cf.created_fold_id_pickle = False total_stime = time.time() times = {} # cf.server_env = True # cf.data_dir = "experiments/dev_data" # dataset = Dataset(cf) # patient = dataset['Master_00018'] cf.exp_dir = "experiments/dev/" cf.plot_dir = cf.exp_dir + "plots" os.makedirs(cf.exp_dir, exist_ok=True) cf.fold = 0 logger = utils.get_logger(cf.exp_dir) gens = get_train_generators(cf, logger) train_loader = gens['train'] for i in range(1): stime = time.time() #ex_batch = next(train_loader) print("train batch", i) times["train_batch"] = time.time() - stime #plg.view_batch(cf, ex_batch, out_file="experiments/dev/dev_exbatch.png", show_gt_labels=True) # # # with open(os.path.join(cf.exp_dir, "fold_"+str(cf.fold), "BatchGenerator_stats.txt"), mode="w") as file: # # train_loader.generator.print_stats(logger, file) # val_loader = gens['val_sampling'] stime = time.time() ex_batch = next(val_loader) times["val_batch"] = time.time() - stime stime = time.time() #plg.view_batch(cf, ex_batch, out_file="experiments/dev/dev_exvalbatch.png", show_gt_labels=True, plot_mods=False, # show_info=False) times["val_plot"] = time.time() - stime # test_loader = get_test_generator(cf, logger)["test"] stime = time.time() ex_batch = test_loader.generate_train_batch() times["test_batch"] = time.time() - stime stime = time.time() plg.view_batch(cf, ex_batch, show_gt_labels=True, out_file="experiments/dev/dev_expatchbatch.png", get_time=False)#, sample_picks=[0,1,2,3]) times["test_patchbatch_plot"] = time.time() - stime # ex_batch['data'] = ex_batch['patient_data'] # ex_batch['seg'] = ex_batch['patient_seg'] # ex_batch['bb_target'] = ex_batch['patient_bb_target'] # for item in cf.roi_items: # ex_batch[] # stime = time.time() # #ex_batch = next(test_loader) # ex_batch = next(test_loader) # plg.view_batch(cf, ex_batch, show_gt_labels=False, show_gt_boxes=True, patient_items=True,# vol_slice_picks=[146,148, 218,220], # out_file="experiments/dev/dev_expatientbatch.png") # , sample_picks=[0,1,2,3]) # times["test_patient_batch_plot"] = time.time() - stime print("Times recorded throughout:") for (k, v) in times.items(): print(k, "{:.2f}".format(v)) mins, secs = divmod((time.time() - total_stime), 60) h, mins = divmod(mins, 60) t = "{:d}h:{:02d}m:{:02d}s".format(int(h), int(mins), int(secs)) print("{} total runtime: {}".format(os.path.split(__file__)[1], t)) diff --git a/datasets/toy/configs.py b/datasets/toy/configs.py index 6cb5859..74f5927 100644 --- a/datasets/toy/configs.py +++ b/datasets/toy/configs.py @@ -1,490 +1,490 @@ #!/usr/bin/env python # Copyright 2019 Division of Medical Image Computing, German Cancer Research Center (DKFZ). # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== import sys import os sys.path.append(os.path.dirname(os.path.realpath(__file__))) import numpy as np from default_configs import DefaultConfigs from collections import namedtuple boxLabel = namedtuple('boxLabel', ["name", "color"]) Label = namedtuple("Label", ['id', 'name', 'shape', 'radius', 'color', 'regression', 'ambiguities', 'gt_distortion']) binLabel = namedtuple("binLabel", ['id', 'name', 'color', 'bin_vals']) class Configs(DefaultConfigs): def __init__(self, server_env=None): super(Configs, self).__init__(server_env) ######################### # Prepro # ######################### self.pp_rootdir = os.path.join('/media/gregor/HDD2TB/data/toy', "cyl1ps_dev") self.pp_npz_dir = self.pp_rootdir+"_npz" self.pre_crop_size = [320,320,8] #y,x,z; determines pp data shape (2D easily implementable, but only 3D for now) self.min_2d_radius = 6 #in pixels self.n_train_samples, self.n_test_samples = 1200, 1000 # not actually real one-hot encoding (ohe) but contains more info: roi-overlap only within classes. self.pp_create_ohe_seg = False self.pp_empty_samples_ratio = 0.1 self.pp_place_radii_mid_bin = True self.pp_only_distort_2d = True # outer-most intensity of blurred radii, relative to inner-object intensity. <1 for decreasing, > 1 for increasing. # e.g.: setting 0.1 means blurred edge has min intensity 10% as large as inner-object intensity. self.pp_blur_min_intensity = 0.2 self.max_instances_per_sample = 1 #how many max instances over all classes per sample (img if 2d, vol if 3d) self.max_instances_per_class = self.max_instances_per_sample # how many max instances per image per class self.noise_scale = 0. # std-dev of gaussian noise self.ambigs_sampling = "gaussian" #"gaussian" or "uniform" """ radius_calib: gt distort for calibrating uncertainty. Range of gt distortion is inferable from image by distinguishing it from the rest of the object. blurring width around edge will be shifted so that symmetric rel to orig radius. blurring scale: if self.ambigs_sampling is uniform, distribution's non-zero range (b-a) will be sqrt(12)*scale since uniform dist has variance (b-a)²/12. b,a will be placed symmetrically around unperturbed radius. if sampling is gaussian, then scale parameter sets one std dev, i.e., blurring width will be orig_radius * std_dev * 2. """ self.ambiguities = { #set which classes to apply which ambs to below in class labels #choose out of: 'outer_radius', 'inner_radius', 'radii_relations'. #kind #probability #scale (gaussian std, relative to unperturbed value) #"outer_radius": (1., 0.5), #"outer_radius_xy": (1., 0.5), #"inner_radius": (0.5, 0.1), #"radii_relations": (0.5, 0.1), "radius_calib": (1., 1./6) } # shape choices: 'cylinder', 'block' # id, name, shape, radius, color, regression, ambiguities, gt_distortion self.pp_classes = [Label(1, 'cylinder', 'cylinder', ((6,6,1),(40,40,8)), (*self.blue, 1.), "radius_2d", (), ()), #Label(2, 'block', 'block', ((6,6,1),(40,40,8)), (*self.aubergine,1.), "radii_2d", (), ('radius_calib',)) ] ######################### # I/O # ######################### self.data_sourcedir = '/home/gregor/data/toy/cyl1ps_dev' if server_env: self.data_sourcedir = '/datasets/data_ramien/toy/cyl1ps_dev_npz' self.test_data_sourcedir = os.path.join(self.data_sourcedir, 'test') self.data_sourcedir = os.path.join(self.data_sourcedir, "train") self.info_df_name = 'info_df.pickle' # one out of ['mrcnn', 'retina_net', 'retina_unet', 'detection_unet', 'ufrcnn', 'detection_fpn']. self.model = 'mrcnn' self.model_path = 'models/{}.py'.format(self.model if not 'retina' in self.model else 'retina_net') self.model_path = os.path.join(self.source_dir, self.model_path) ######################### # Architecture # ######################### # one out of [2, 3]. dimension the model operates in. - self.dim = 2 + self.dim = 3 # 'class', 'regression', 'regression_bin', 'regression_ken_gal' # currently only tested mode is a single-task at a time (i.e., only one task in below list) # but, in principle, tasks could be combined (e.g., object classes and regression per class) self.prediction_tasks = ['class', ] self.start_filts = 48 if self.dim == 2 else 18 self.end_filts = self.start_filts * 4 if self.dim == 2 else self.start_filts * 2 self.res_architecture = 'resnet50' # 'resnet101' , 'resnet50' self.norm = 'instance_norm' # one of None, 'instance_norm', 'batch_norm' self.relu = 'relu' # one of 'xavier_uniform', 'xavier_normal', or 'kaiming_normal', None (=default = 'kaiming_uniform') self.weight_init = None self.regression_n_features = 1 # length of regressor target vector ######################### # Data Loader # ######################### self.num_epochs = 32 self.num_train_batches = 120 if self.dim == 2 else 80 self.batch_size = 16 if self.dim == 2 else 8 self.n_cv_splits = 4 # select modalities from preprocessed data self.channels = [0] self.n_channels = len(self.channels) # which channel (mod) to show as bg in plotting, will be extra added to batch if not in self.channels self.plot_bg_chan = 0 self.crop_margin = [20, 20, 1] # has to be smaller than respective patch_size//2 self.patch_size_2D = self.pre_crop_size[:2] self.patch_size_3D = self.pre_crop_size[:2]+[8] # patch_size to be used for training. pre_crop_size is the patch_size before data augmentation. self.patch_size = self.patch_size_2D if self.dim == 2 else self.patch_size_3D # ratio of free sampled batch elements before class balancing is triggered # (>0 to include "empty"/background patches.) self.batch_random_ratio = 0.2 self.balance_target = "class_targets" if 'class' in self.prediction_tasks else "rg_bin_targets" self.observables_patient = [] self.observables_rois = [] self.seed = 3 #for generating folds ############################# # Colors, Classes, Legends # ############################# self.plot_frequency = 1 binary_bin_labels = [binLabel(1, 'r<=25', (*self.green, 1.), (1,25)), binLabel(2, 'r>25', (*self.red, 1.), (25,))] quintuple_bin_labels = [binLabel(1, 'r2-10', (*self.green, 1.), (2,10)), binLabel(2, 'r10-20', (*self.yellow, 1.), (10,20)), binLabel(3, 'r20-30', (*self.orange, 1.), (20,30)), binLabel(4, 'r30-40', (*self.bright_red, 1.), (30,40)), binLabel(5, 'r>40', (*self.red, 1.), (40,))] # choose here if to do 2-way or 5-way regression-bin classification task_spec_bin_labels = quintuple_bin_labels self.class_labels = [ # regression: regression-task label, either value or "(x,y,z)_radius" or "radii". # ambiguities: name of above defined ambig to apply to image data (not gt); need to be iterables! # gt_distortion: name of ambig to apply to gt only; needs to be iterable! # #id #name #shape #radius #color #regression #ambiguities #gt_distortion Label( 0, 'bg', None, (0, 0, 0), (*self.white, 0.), (0, 0, 0), (), ())] if "class" in self.prediction_tasks: self.class_labels += self.pp_classes else: self.class_labels += [Label(1, 'object', 'object', ('various',), (*self.orange, 1.), ('radius_2d',), ("various",), ('various',))] if any(['regression' in task for task in self.prediction_tasks]): self.bin_labels = [binLabel(0, 'bg', (*self.white, 1.), (0,))] self.bin_labels += task_spec_bin_labels self.bin_id2label = {label.id: label for label in self.bin_labels} bins = [(min(label.bin_vals), max(label.bin_vals)) for label in self.bin_labels] self.bin_id2rg_val = {ix: [np.mean(bin)] for ix, bin in enumerate(bins)} self.bin_edges = [(bins[i][1] + bins[i + 1][0]) / 2 for i in range(len(bins) - 1)] self.bin_dict = {label.id: label.name for label in self.bin_labels if label.id != 0} if self.class_specific_seg: self.seg_labels = self.class_labels self.box_type2label = {label.name: label for label in self.box_labels} self.class_id2label = {label.id: label for label in self.class_labels} self.class_dict = {label.id: label.name for label in self.class_labels if label.id != 0} self.seg_id2label = {label.id: label for label in self.seg_labels} self.cmap = {label.id: label.color for label in self.seg_labels} self.plot_prediction_histograms = True self.plot_stat_curves = False self.has_colorchannels = False self.plot_class_ids = True self.num_classes = len(self.class_dict) self.num_seg_classes = len(self.seg_labels) ######################### # Data Augmentation # ######################### self.do_aug = True self.da_kwargs = { 'mirror': True, 'mirror_axes': tuple(np.arange(0, self.dim, 1)), 'do_elastic_deform': False, 'alpha': (500., 1500.), 'sigma': (40., 45.), 'do_rotation': False, 'angle_x': (0., 2 * np.pi), 'angle_y': (0., 0), 'angle_z': (0., 0), 'do_scale': False, 'scale': (0.8, 1.1), 'random_crop': False, 'rand_crop_dist': (self.patch_size[0] / 2. - 3, self.patch_size[1] / 2. - 3), 'border_mode_data': 'constant', 'border_cval_data': 0, 'order_data': 1 } if self.dim == 3: self.da_kwargs['do_elastic_deform'] = False self.da_kwargs['angle_x'] = (0, 0.0) self.da_kwargs['angle_y'] = (0, 0.0) # must be 0!! self.da_kwargs['angle_z'] = (0., 2 * np.pi) ######################### # Schedule / Selection # ######################### # decide whether to validate on entire patient volumes (like testing) or sampled patches (like training) # the former is morge accurate, while the latter is faster (depending on volume size) self.val_mode = 'val_sampling' # one of 'val_sampling' , 'val_patient' if self.val_mode == 'val_patient': self.max_val_patients = 220 # if 'all' iterates over entire val_set once. if self.val_mode == 'val_sampling': self.num_val_batches = 35 if self.dim==2 else 25 self.save_n_models = 2 self.min_save_thresh = 1 if self.dim == 2 else 1 # =wait time in epochs if "class" in self.prediction_tasks: self.model_selection_criteria = {name + "_ap": 1. for name in self.class_dict.values()} elif any("regression" in task for task in self.prediction_tasks): self.model_selection_criteria = {name + "_ap": 0.2 for name in self.class_dict.values()} self.model_selection_criteria.update({name + "_avp": 0.8 for name in self.class_dict.values()}) self.lr_decay_factor = 0.5 self.scheduling_patience = int(self.num_epochs / 5) self.weight_decay = 1e-5 self.clip_norm = None # number or None ######################### # Testing / Plotting # ######################### self.test_aug_axes = (0,1,(0,1)) # None or list: choices are 0,1,(0,1) self.held_out_test_set = True self.max_test_patients = "all" # number or "all" for all self.test_against_exact_gt = True # only True implemented self.val_against_exact_gt = False # True is an unrealistic --> irrelevant scenario. self.report_score_level = ['rois'] # 'patient' or 'rois' (incl) self.patient_class_of_interest = 1 self.patient_bin_of_interest = 2 self.eval_bins_separately = False#"additionally" if not 'class' in self.prediction_tasks else False self.metrics = ['ap', 'auc', 'dice'] if any(['regression' in task for task in self.prediction_tasks]): self.metrics += ['avp', 'rg_MAE_weighted', 'rg_MAE_weighted_tp', 'rg_bin_accuracy_weighted', 'rg_bin_accuracy_weighted_tp'] if 'aleatoric' in self.model: self.metrics += ['rg_uncertainty', 'rg_uncertainty_tp', 'rg_uncertainty_tp_weighted'] self.evaluate_fold_means = True self.ap_match_ious = [0.5] # threshold(s) for considering a prediction as true positive self.min_det_thresh = 0.3 self.model_max_iou_resolution = 0.2 # aggregation method for test and val_patient predictions. # wbc = weighted box clustering as in https://arxiv.org/pdf/1811.08661.pdf, # nms = standard non-maximum suppression, or None = no clustering self.clustering = 'wbc' # iou thresh (exclusive!) for regarding two preds as concerning the same ROI self.clustering_iou = self.model_max_iou_resolution # has to be larger than desired possible overlap iou of model predictions self.merge_2D_to_3D_preds = False self.merge_3D_iou = self.model_max_iou_resolution self.n_test_plots = 1 # per fold and rank self.test_n_epochs = self.save_n_models # should be called n_test_ens, since is number of models to ensemble over during testing # is multiplied by (1 + nr of test augs) ######################### # Assertions # ######################### if not 'class' in self.prediction_tasks: assert self.num_classes == 1 ######################### # Add model specifics # ######################### {'mrcnn': self.add_mrcnn_configs, 'mrcnn_aleatoric': self.add_mrcnn_configs, 'retina_net': self.add_mrcnn_configs, 'retina_unet': self.add_mrcnn_configs, 'detection_unet': self.add_det_unet_configs, 'detection_fpn': self.add_det_fpn_configs }[self.model]() def rg_val_to_bin_id(self, rg_val): #only meant for isotropic radii!! # only 2D radii (x and y dims) or 1D (x or y) are expected return np.round(np.digitize(rg_val, self.bin_edges).mean()) def add_det_fpn_configs(self): self.learning_rate = [1 * 1e-4] * self.num_epochs self.dynamic_lr_scheduling = True self.scheduling_criterion = 'torch_loss' self.scheduling_mode = 'min' if "loss" in self.scheduling_criterion else 'max' self.n_roi_candidates = 4 if self.dim == 2 else 6 # max number of roi candidates to identify per image (slice in 2D, volume in 3D) # loss mode: either weighted cross entropy ('wce'), batch-wise dice loss ('dice), or the sum of both ('dice_wce') self.seg_loss_mode = 'wce' self.wce_weights = [1] * self.num_seg_classes if 'dice' in self.seg_loss_mode else [0.1, 1] self.fp_dice_weight = 1 if self.dim == 2 else 1 # if <1, false positive predictions in foreground are penalized less. self.detection_min_confidence = 0.05 # how to determine score of roi: 'max' or 'median' self.score_det = 'max' def add_det_unet_configs(self): self.learning_rate = [1 * 1e-4] * self.num_epochs self.dynamic_lr_scheduling = True self.scheduling_criterion = "torch_loss" self.scheduling_mode = 'min' if "loss" in self.scheduling_criterion else 'max' # max number of roi candidates to identify per image (slice in 2D, volume in 3D) self.n_roi_candidates = 4 if self.dim == 2 else 6 # loss mode: either weighted cross entropy ('wce'), batch-wise dice loss ('dice), or the sum of both ('dice_wce') self.seg_loss_mode = 'wce' self.wce_weights = [1] * self.num_seg_classes if 'dice' in self.seg_loss_mode else [0.1, 1] # if <1, false positive predictions in foreground are penalized less. self.fp_dice_weight = 1 if self.dim == 2 else 1 self.detection_min_confidence = 0.05 # how to determine score of roi: 'max' or 'median' self.score_det = 'max' self.init_filts = 32 self.kernel_size = 3 # ks for horizontal, normal convs self.kernel_size_m = 2 # ks for max pool self.pad = "same" # "same" or integer, padding of horizontal convs def add_mrcnn_configs(self): self.learning_rate = [1e-4] * self.num_epochs self.dynamic_lr_scheduling = True # with scheduler set in exec self.scheduling_criterion = max(self.model_selection_criteria, key=self.model_selection_criteria.get) self.scheduling_mode = 'min' if "loss" in self.scheduling_criterion else 'max' # number of classes for network heads: n_foreground_classes + 1 (background) self.head_classes = self.num_classes + 1 if 'class' in self.prediction_tasks else 2 # feed +/- n neighbouring slices into channel dimension. set to None for no context. self.n_3D_context = None if self.n_3D_context is not None and self.dim == 2: self.n_channels *= (self.n_3D_context * 2 + 1) self.detect_while_training = True # disable the re-sampling of mask proposals to original size for speed-up. # since evaluation is detection-driven (box-matching) and not instance segmentation-driven (iou-matching), # mask outputs are optional. self.return_masks_in_train = True self.return_masks_in_val = True self.return_masks_in_test = True # feature map strides per pyramid level are inferred from architecture. anchor scales are set accordingly. self.backbone_strides = {'xy': [4, 8, 16, 32], 'z': [1, 2, 4, 8]} # anchor scales are chosen according to expected object sizes in data set. Default uses only one anchor scale # per pyramid level. (outer list are pyramid levels (corresponding to BACKBONE_STRIDES), inner list are scales per level.) self.rpn_anchor_scales = {'xy': [[4], [8], [16], [32]], 'z': [[1], [2], [4], [8]]} # choose which pyramid levels to extract features from: P2: 0, P3: 1, P4: 2, P5: 3. self.pyramid_levels = [0, 1, 2, 3] # number of feature maps in rpn. typically lowered in 3D to save gpu-memory. self.n_rpn_features = 512 if self.dim == 2 else 64 # anchor ratios and strides per position in feature maps. self.rpn_anchor_ratios = [0.5, 1., 2.] self.rpn_anchor_stride = 1 # Threshold for first stage (RPN) non-maximum suppression (NMS): LOWER == HARDER SELECTION self.rpn_nms_threshold = max(0.8, self.model_max_iou_resolution) # loss sampling settings. self.rpn_train_anchors_per_image = 4 self.train_rois_per_image = 6 # per batch_instance self.roi_positive_ratio = 0.5 self.anchor_matching_iou = 0.8 # k negative example candidates are drawn from a pool of size k*shem_poolsize (stochastic hard-example mining), # where k<=#positive examples. self.shem_poolsize = 2 self.pool_size = (7, 7) if self.dim == 2 else (7, 7, 3) self.mask_pool_size = (14, 14) if self.dim == 2 else (14, 14, 5) self.mask_shape = (28, 28) if self.dim == 2 else (28, 28, 10) self.rpn_bbox_std_dev = np.array([0.1, 0.1, 0.1, 0.2, 0.2, 0.2]) self.bbox_std_dev = np.array([0.1, 0.1, 0.1, 0.2, 0.2, 0.2]) self.window = np.array([0, 0, self.patch_size[0], self.patch_size[1], 0, self.patch_size_3D[2]]) self.scale = np.array([self.patch_size[0], self.patch_size[1], self.patch_size[0], self.patch_size[1], self.patch_size_3D[2], self.patch_size_3D[2]]) # y1,x1,y2,x2,z1,z2 if self.dim == 2: self.rpn_bbox_std_dev = self.rpn_bbox_std_dev[:4] self.bbox_std_dev = self.bbox_std_dev[:4] self.window = self.window[:4] self.scale = self.scale[:4] self.plot_y_max = 1.5 self.n_plot_rpn_props = 5 if self.dim == 2 else 30 # per batch_instance (slice in 2D / patient in 3D) # pre-selection in proposal-layer (stage 1) for NMS-speedup. applied per batch element. self.pre_nms_limit = 2000 if self.dim == 2 else 4000 # n_proposals to be selected after NMS per batch element. too high numbers blow up memory if "detect_while_training" is True, # since proposals of the entire batch are forwarded through second stage as one "batch". self.roi_chunk_size = 1300 if self.dim == 2 else 500 self.post_nms_rois_training = 200 * (self.head_classes-1) if self.dim == 2 else 400 self.post_nms_rois_inference = 200 * (self.head_classes-1) # Final selection of detections (refine_detections) self.model_max_instances_per_batch_element = 9 if self.dim == 2 else 18 # per batch element and class. self.detection_nms_threshold = self.model_max_iou_resolution # needs to be > 0, otherwise all predictions are one cluster. self.model_min_confidence = 0.2 # iou for nms in box refining (directly after heads), should be >0 since ths>=x in mrcnn.py if self.dim == 2: self.backbone_shapes = np.array( [[int(np.ceil(self.patch_size[0] / stride)), int(np.ceil(self.patch_size[1] / stride))] for stride in self.backbone_strides['xy']]) else: self.backbone_shapes = np.array( [[int(np.ceil(self.patch_size[0] / stride)), int(np.ceil(self.patch_size[1] / stride)), int(np.ceil(self.patch_size[2] / stride_z))] for stride, stride_z in zip(self.backbone_strides['xy'], self.backbone_strides['z'] )]) if self.model == 'retina_net' or self.model == 'retina_unet': # whether to use focal loss or SHEM for loss-sample selection self.focal_loss = False # implement extra anchor-scales according to https://arxiv.org/abs/1708.02002 self.rpn_anchor_scales['xy'] = [[ii[0], ii[0] * (2 ** (1 / 3)), ii[0] * (2 ** (2 / 3))] for ii in self.rpn_anchor_scales['xy']] self.rpn_anchor_scales['z'] = [[ii[0], ii[0] * (2 ** (1 / 3)), ii[0] * (2 ** (2 / 3))] for ii in self.rpn_anchor_scales['z']] self.n_anchors_per_pos = len(self.rpn_anchor_ratios) * 3 # pre-selection of detections for NMS-speedup. per entire batch. self.pre_nms_limit = (500 if self.dim == 2 else 6250) * self.batch_size # anchor matching iou is lower than in Mask R-CNN according to https://arxiv.org/abs/1708.02002 self.anchor_matching_iou = 0.7 if self.model == 'retina_unet': self.operate_stride1 = True diff --git a/utils/dataloader_utils.py b/utils/dataloader_utils.py index cdd92e1..0724b28 100644 --- a/utils/dataloader_utils.py +++ b/utils/dataloader_utils.py @@ -1,743 +1,723 @@ #!/usr/bin/env python # Copyright 2019 Division of Medical Image Computing, German Cancer Research Center (DKFZ). # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== import plotting as plg import os from multiprocessing import Pool, Lock import pickle import warnings import numpy as np import pandas as pd from batchgenerators.transforms.abstract_transforms import AbstractTransform from scipy.ndimage.measurements import label as lb from torch.utils.data import Dataset as torchDataset from batchgenerators.dataloading.data_loader import SlimDataLoaderBase import utils.exp_utils as utils import data_manager as dmanager for msg in ["This figure includes Axes that are not compatible with tight_layout", "Data has no positive values, and therefore cannot be log-scaled."]: warnings.filterwarnings("ignore", msg) class AttributeDict(dict): __getattr__ = dict.__getitem__ __setattr__ = dict.__setitem__ ################################## # data loading, organisation # ################################## class fold_generator: """ generates splits of indices for a given length of a dataset to perform n-fold cross-validation. splits each fold into 3 subsets for training, validation and testing. This form of cross validation uses an inner loop test set, which is useful if test scores shall be reported on a statistically reliable amount of patients, despite limited size of a dataset. If hold out test set is provided and hence no inner loop test set needed, just add test_idxs to the training data in the dataloader. This creates straight-forward train-val splits. :returns names list: list of len n_splits. each element is a list of len 3 for train_ix, val_ix, test_ix. """ def __init__(self, seed, n_splits, len_data): """ :param seed: Random seed for splits. :param n_splits: number of splits, e.g. 5 splits for 5-fold cross-validation :param len_data: number of elements in the dataset. """ self.tr_ix = [] self.val_ix = [] self.te_ix = [] self.slicer = None self.missing = 0 self.fold = 0 self.len_data = len_data self.n_splits = n_splits self.myseed = seed self.boost_val = 0 def init_indices(self): t = list(np.arange(self.l)) # round up to next splittable data amount. split_length = int(np.ceil(len(t) / float(self.n_splits))) self.slicer = split_length self.mod = len(t) % self.n_splits if self.mod > 0: # missing is the number of folds, in which the new splits are reduced to account for missing data. self.missing = self.n_splits - self.mod self.te_ix = t[:self.slicer] self.tr_ix = t[self.slicer:] self.val_ix = self.tr_ix[:self.slicer] self.tr_ix = self.tr_ix[self.slicer:] def new_fold(self): slicer = self.slicer if self.fold < self.missing : slicer = self.slicer - 1 temp = self.te_ix # catch exception mod == 1: test set collects 1+ data since walk through both roudned up splits. # account for by reducing last fold split by 1. if self.fold == self.n_splits-2 and self.mod ==1: temp += self.val_ix[-1:] self.val_ix = self.val_ix[:-1] self.te_ix = self.val_ix self.val_ix = self.tr_ix[:slicer] self.tr_ix = self.tr_ix[slicer:] + temp def get_fold_names(self): names_list = [] rgen = np.random.RandomState(self.myseed) cv_names = np.arange(self.len_data) rgen.shuffle(cv_names) self.l = len(cv_names) self.init_indices() for split in range(self.n_splits): train_names, val_names, test_names = cv_names[self.tr_ix], cv_names[self.val_ix], cv_names[self.te_ix] names_list.append([train_names, val_names, test_names, self.fold]) self.new_fold() self.fold += 1 return names_list class FoldGenerator(): r"""takes a set of elements (identifiers) and randomly splits them into the specified amt of subsets. """ def __init__(self, identifiers, seed, n_splits=5): self.ids = np.array(identifiers) self.n_splits = n_splits self.seed = seed def generate_splits(self, n_splits=None): if n_splits is None: n_splits = self.n_splits rgen = np.random.RandomState(self.seed) rgen.shuffle(self.ids) self.splits = list(np.array_split(self.ids, n_splits, axis=0)) # already returns list, but to be sure return self.splits class Dataset(torchDataset): r"""Parent Class for actual Dataset classes to inherit from! """ def __init__(self, cf, data_sourcedir=None): super(Dataset, self).__init__() self.cf = cf self.data_sourcedir = cf.data_sourcedir if data_sourcedir is None else data_sourcedir self.data_dir = cf.data_dir if hasattr(cf, 'data_dir') else self.data_sourcedir self.data_dest = cf.data_dest if hasattr(cf, "data_dest") else self.data_sourcedir self.data = {} self.set_ids = [] def copy_data(self, cf, file_subset, keep_packed=False, del_after_unpack=False): if os.path.normpath(self.data_sourcedir) != os.path.normpath(self.data_dest): self.data_sourcedir = os.path.join(self.data_sourcedir, '') args = AttributeDict({ "source" : self.data_sourcedir, "destination" : self.data_dest, "recursive" : True, "cp_only_npz" : False, "keep_packed" : keep_packed, "del_after_unpack" : del_after_unpack, "threads" : 16 if self.cf.server_env else os.cpu_count() }) dmanager.copy(args, file_subset=file_subset) self.data_dir = self.data_dest def __len__(self): return len(self.data) def __getitem__(self, id): """Return a sample of the dataset, i.e.,the dict of the id """ return self.data[id] def __iter__(self): return self.data.__iter__() def init_FoldGenerator(self, seed, n_splits): self.fg = FoldGenerator(self.set_ids, seed=seed, n_splits=n_splits) def generate_splits(self, check_file): if not os.path.exists(check_file): self.fg.generate_splits() with open(check_file, 'wb') as handle: pickle.dump(self.fg.splits, handle) else: with open(check_file, 'rb') as handle: self.fg.splits = pickle.load(handle) def calc_statistics(self, subsets=None, plot_dir=None, overall_stats=True): if self.df is None: self.df = pd.DataFrame() balance_t = self.cf.balance_target if hasattr(self.cf, "balance_target") else "class_targets" self.df._metadata.append(balance_t) if balance_t=="class_targets": mapper = lambda cl_id: self.cf.class_id2label[cl_id] labels = self.cf.class_id2label.values() elif balance_t=="rg_bin_targets": mapper = lambda rg_bin: self.cf.bin_id2label[rg_bin] labels = self.cf.bin_id2label.values() # elif balance_t=="regression_targets": # # todo this wont work # mapper = lambda rg_val: AttributeDict({"name":rg_val}) #self.cf.bin_id2label[self.cf.rg_val_to_bin_id(rg_val)] # labels = self.cf.bin_id2label.values() elif balance_t=="lesion_gleasons": mapper = lambda gs: self.cf.gs2label[gs] labels = self.cf.gs2label.values() else: mapper = lambda x: AttributeDict({"name":x}) labels = None for pid, subj_data in self.data.items(): unique_ts, counts = np.unique(subj_data[balance_t], return_counts=True) self.df = self.df.append(pd.DataFrame({"pid": [pid], **{mapper(unique_ts[i]).name: [counts[i]] for i in range(len(unique_ts))}}), ignore_index=True, sort=True) self.df = self.df.fillna(0) if overall_stats: df = self.df.drop("pid", axis=1) df = df.reindex(sorted(df.columns), axis=1).astype('uint32') print("Overall dataset roi counts per target kind:"); print(df.sum()) if subsets is not None: self.df["subset"] = np.nan self.df["display_order"] = np.nan for ix, (subset, pids) in enumerate(subsets.items()): self.df.loc[self.df.pid.isin(pids), "subset"] = subset self.df.loc[self.df.pid.isin(pids), "display_order"] = ix df = self.df.groupby("subset").agg("sum").drop("pid", axis=1, errors='ignore').astype('int64') df = df.sort_values(by=['display_order']).drop('display_order', axis=1) df = df.reindex(sorted(df.columns), axis=1) print("Fold {} dataset roi counts per target kind:".format(self.cf.fold)); print(df) if plot_dir is not None: os.makedirs(plot_dir, exist_ok=True) if subsets is not None: plg.plot_fold_stats(self.cf, df, labels, os.path.join(plot_dir, "data_stats_fold_" + str(self.cf.fold))+".pdf") if overall_stats: plg.plot_data_stats(self.cf, df, labels, os.path.join(plot_dir, 'data_stats_overall.pdf')) return df, labels def get_class_balanced_patients(all_pids, class_targets, batch_size, num_classes, random_ratio=0): ''' samples towards equilibrium of classes (on basis of total RoI counts). for highly imbalanced dataset, this might be a too strong requirement. :param class_targets: dic holding {patient_specifier : ROI class targets}, list position of ROI target corresponds to respective seg label - 1 :param batch_size: :param num_classes: :return: ''' # assert len(all_pids)>=batch_size, "not enough eligible pids {} to form a single batch of size {}".format(len(all_pids), batch_size) class_counts = {k: 0 for k in range(1,num_classes+1)} not_picked = np.array(all_pids) batch_patients = np.empty((batch_size,), dtype=not_picked.dtype) rarest_class = np.random.randint(1,num_classes+1) for ix in range(batch_size): if len(not_picked) == 0: warnings.warn("Dataset too small to generate batch with unique samples; => recycling.") not_picked = np.array(all_pids) np.random.shuffle(not_picked) #this could actually go outside(above) the loop. pick = not_picked[0] for cand in not_picked: if np.count_nonzero(class_targets[cand] == rarest_class) > 0: pick = cand cand_rarest_class = np.argmin([np.count_nonzero(class_targets[cand] == cl) for cl in range(1,num_classes+1)])+1 # if current batch already bigger than the batch random ratio, then # check that weakest class in this patient is not the weakest in current batch (since needs to be boosted) # also that at least one roi of this patient belongs to weakest class. If True, keep patient, else keep looking. if (cand_rarest_class != rarest_class and np.count_nonzero(class_targets[cand] == rarest_class) > 0) \ or ix < int(batch_size * random_ratio): break for c in range(1,num_classes+1): class_counts[c] += np.count_nonzero(class_targets[pick] == c) if not ix < int(batch_size * random_ratio) and class_counts[rarest_class] == 0: # means searched thru whole set without finding rarest class print("Class {} not represented in current dataset.".format(rarest_class)) rarest_class = np.argmin(([class_counts[c] for c in range(1,num_classes+1)]))+1 batch_patients[ix] = pick not_picked = not_picked[not_picked != pick] # removes pick return batch_patients class BatchGenerator(SlimDataLoaderBase): """ create the training/validation batch generator. Randomly sample batch_size patients from the data set, (draw a random slice if 2D), pad-crop them to equal sizes and merge to an array. :param data: data dictionary as provided by 'load_dataset' :param img_modalities: list of strings ['adc', 'b1500'] from config :param batch_size: number of patients to sample for the batch :param pre_crop_size: equal size for merging the patients to a single array (before the final random-crop in data aug.) :return dictionary containing the batch data / seg / pids as lists; the augmenter will later concatenate them into an array. """ def __init__(self, cf, data, sample_pids_w_replace=True, max_batches=None, raise_stop_iteration=False, n_threads=None, seed=0): if n_threads is None: n_threads = cf.n_workers super(BatchGenerator, self).__init__(data, cf.batch_size, number_of_threads_in_multithreaded=n_threads) self.cf = cf self.random_count = int(cf.batch_random_ratio * cf.batch_size) self.plot_dir = os.path.join(self.cf.plot_dir, 'train_generator') + os.makedirs(self.plot_dir, exist_ok=True) self.max_batches = max_batches self.raise_stop = raise_stop_iteration self.thread_id = 0 self.batches_produced = 0 self.dataset_length = len(self._data) self.dataset_pids = list(self._data.keys()) self.rgen = np.random.RandomState(seed=seed) self.eligible_pids = self.rgen.permutation(self.dataset_pids.copy()) self.eligible_pids = np.array_split(self.eligible_pids, self.number_of_threads_in_multithreaded) self.eligible_pids = sorted(self.eligible_pids, key=len, reverse=True) self.sample_pids_w_replace = sample_pids_w_replace if not self.sample_pids_w_replace: assert len(self.dataset_pids) / self.number_of_threads_in_multithreaded >= self.batch_size, \ "at least one batch needed per thread. dataset size: {}, n_threads: {}, batch_size: {}.".format( len(self.dataset_pids), self.number_of_threads_in_multithreaded, self.batch_size) self.lock = Lock() - self.stats = {"roi_counts": np.zeros((self.cf.num_classes,), dtype='uint32'), - "empty_counts": np.zeros((self.cf.num_classes,), dtype='uint32')} - if hasattr(cf, "balance_target"): # WARNING: "balance targets are only implemented for 1-d targets (or 1-component vectors)" self.balance_target = cf.balance_target else: self.balance_target = "class_targets" self.targets = {k:v[self.balance_target] for (k,v) in self._data.items()} def set_thread_id(self, thread_id): self.thread_ids = self.eligible_pids[thread_id] self.thread_id = thread_id def reset(self): self.batches_produced = 0 self.thread_ids = self.rgen.permutation(self.eligible_pids[self.thread_id]) @staticmethod def sample_targets_to_weights(targets, fg_bg_weights): weights = targets * fg_bg_weights return weights def balance_target_distribution(self, plot=False): """Impose a drawing distribution over samples. Distribution should be designed so that classes' fg and bg examples are (as good as possible) shown in equal frequency. Since we are dealing with rois, fg/bg weights count a sample (e.g., a patient) with **at least** one occurrence as fg, otherwise bg. For fg weights among classes, each RoI counts. :param all_pids: :param self.targets: dic holding {patient_specifier : patient-wise-unique ROI targets} :return: probability distribution over all pids. draw without replace from this. """ self.unique_ts = np.unique([v for pat in self.targets.values() for v in pat]) self.sample_stats = pd.DataFrame(columns=[str(ix)+suffix for ix in self.unique_ts for suffix in ["", "_bg"]], index=list(self.targets.keys())) for pid in self.sample_stats.index: for targ in self.unique_ts: fg_count = np.count_nonzero(self.targets[pid] == targ) self.sample_stats.loc[pid, str(targ)] = int(fg_count > 0) self.sample_stats.loc[pid, str(targ)+"_bg"] = int(fg_count == 0) self.targ_stats = self.sample_stats.agg( ("sum", lambda col: col.sum() / len(self._data)), axis=0, sort=False).rename({"": "relative"}) anchor = 1. - self.targ_stats.loc["relative"].iloc[0] self.fg_bg_weights = anchor / self.targ_stats.loc["relative"] cum_weights = anchor * len(self.fg_bg_weights) self.fg_bg_weights /= cum_weights self.p_probs = self.sample_stats.apply(self.sample_targets_to_weights, args=(self.fg_bg_weights,), axis=1).sum(axis=1) self.p_probs = self.p_probs / self.p_probs.sum() if plot: print("Applying class-weights:\n {}".format(self.fg_bg_weights)) if len(self.sample_stats.columns) == 2: # assert that probs are calc'd correctly: # (self.p_probs * self.sample_stats["1"]).sum() == (self.p_probs * self.sample_stats["1_bg"]).sum() # only works if one label per patient (multi-label expectations depend on multi-label occurences). expectations = [] for targ in self.sample_stats.columns: expectations.append((self.p_probs * self.sample_stats[targ]).sum()) assert np.allclose(expectations, expectations[0], atol=1e-4), "expectation values for fgs/bgs: {}".format(expectations) - # get unique foreground targets per patient, assign -1 to an "empty" patient (has no foreground) - #patient_ts = [np.unique(lst) if len([t for t in lst if np.any(t>0)])>0 else [-1] for lst in self.targets.values()] - #bg_mask = np.array([np.all(lst == [-1]) for lst in patient_ts]) - #unique_ts, t_counts = np.unique([t for lst in patient_ts for t in lst if t!=-1], return_counts=True) - # t_probs = t_counts.sum() / t_counts - # t_probs /= t_probs.sum() - # t_probs = {t : t_probs[ix] for ix, t in enumerate(unique_ts)} - # t_probs[-1] = 0. - # # fail if balance target is not a number (i.e., a vector) - # self.p_probs = np.array([ max([t_probs[t] for t in lst]) for lst in patient_ts ]) - # #normalize - # self.p_probs /= self.p_probs.sum() - # rescale probs of empty samples - # if not 0 == self.p_probs[bg_mask].shape[0]: - # #rescale_f = (1 - self.cf.empty_samples_ratio) / self.p_probs[~bg_mask].sum() - # rescale_f = 1 / self.p_probs[~bg_mask].sum() - # self.p_probs *= rescale_f - # self.p_probs[bg_mask] = 0. #self.cf.empty_samples_ratio/self.p_probs[bg_mask].shape[0] - - #self.unique_ts = unique_ts + self.stats = {"roi_counts": np.zeros(len(self.unique_ts,), dtype='uint32'), + "empty_counts": np.zeros(len(self.unique_ts,), dtype='uint32')} if plot: os.makedirs(self.plot_dir, exist_ok=True) plg.plot_batchgen_distribution(self.cf, self.dataset_pids, self.p_probs, self.balance_target, out_file=os.path.join(self.plot_dir, "train_gen_distr_"+str(self.cf.fold)+".png")) return self.p_probs def get_batch_pids(self): if self.max_batches is not None and self.batches_produced * self.number_of_threads_in_multithreaded \ + self.thread_id >= self.max_batches: self.reset() raise StopIteration if self.sample_pids_w_replace: # fully random patients batch_pids = list(np.random.choice(self.dataset_pids, size=self.random_count, replace=False)) # target-balanced patients batch_pids += list(np.random.choice( self.dataset_pids, size=self.batch_size - self.random_count, replace=False, p=self.p_probs)) else: with self.lock: if len(self.thread_ids) == 0: if self.raise_stop: self.reset() raise StopIteration else: self.thread_ids = self.rgen.permutation(self.eligible_pids[self.thread_id]) batch_pids = self.thread_ids[:self.batch_size] # batch_pids = np.random.choice(self.thread_ids, size=self.batch_size, replace=False) self.thread_ids = [pid for pid in self.thread_ids if pid not in batch_pids] self.batches_produced += 1 return batch_pids def generate_train_batch(self): # to be overriden by child # everything done in here is per batch # print statements in here get confusing due to multithreading raise NotImplementedError def print_stats(self, logger=None, file=None, plot_file=None, plot=True): print_f = utils.CombinedPrinter(logger, file) print_f('\n***Final Training Stats***') total_count = np.sum(self.stats['roi_counts']) for tix, count in enumerate(self.stats['roi_counts']): #name = self.cf.class_dict[tix] if self.balance_target=="class_targets" else str(self.unique_ts[tix]) name=str(self.unique_ts[tix]) print_f('{}: {} rois seen ({:.1f}%).'.format(name, count, count / total_count * 100)) total_samples = self.cf.num_epochs*self.cf.num_train_batches*self.cf.batch_size empties = [ '{}: {} ({:.1f}%)'.format(str(name), self.stats['empty_counts'][tix], self.stats['empty_counts'][tix]/total_samples*100) for tix, name in enumerate(self.unique_ts) ] empties = ", ".join(empties) print_f('empty samples seen: {}\n'.format(empties)) if plot: if plot_file is None: plot_file = os.path.join(self.plot_dir, "train_gen_stats_{}.png".format(self.cf.fold)) os.makedirs(self.plot_dir, exist_ok=True) plg.plot_batchgen_stats(self.cf, self.stats, empties, self.balance_target, self.unique_ts, plot_file) class PatientBatchIterator(SlimDataLoaderBase): """ creates a val/test generator. Step through the dataset and return dictionaries per patient. 2D is a special case of 3D patching with patch_size[2] == 1 (slices) Creates whole Patient batch and targets, and - if necessary - patchwise batch and targets. Appends patient targets anyway for evaluation. For Patching, shifts all patches into batch dimension. batch_tiling_forward will take care of exceeding batch dimensions. This iterator/these batches are not intended to go through MTaugmenter afterwards """ def __init__(self, cf, data): super(PatientBatchIterator, self).__init__(data, 0) self.cf = cf self.dataset_length = len(self._data) self.dataset_pids = list(self._data.keys()) def generate_train_batch(self, pid=None): # to be overriden by child return ################################### # transforms, image manipulation # ################################### def get_patch_crop_coords(img, patch_size, min_overlap=30): """ _:param img (y, x, (z)) _:param patch_size: list of len 2 (2D) or 3 (3D). _:param min_overlap: minimum required overlap of patches. If too small, some areas are poorly represented only at edges of single patches. _:return ndarray: shape (n_patches, 2*dim). crop coordinates for each patch. """ crop_coords = [] for dim in range(len(img.shape)): n_patches = int(np.ceil(img.shape[dim] / patch_size[dim])) # no crops required in this dimension, add image shape as coordinates. if n_patches == 1: crop_coords.append([(0, img.shape[dim])]) continue # fix the two outside patches to coords patchsize/2 and interpolate. center_dists = (img.shape[dim] - patch_size[dim]) / (n_patches - 1) if (patch_size[dim] - center_dists) < min_overlap: n_patches += 1 center_dists = (img.shape[dim] - patch_size[dim]) / (n_patches - 1) patch_centers = np.round([(patch_size[dim] / 2 + (center_dists * ii)) for ii in range(n_patches)]) dim_crop_coords = [(center - patch_size[dim] / 2, center + patch_size[dim] / 2) for center in patch_centers] crop_coords.append(dim_crop_coords) coords_mesh_grid = [] for ymin, ymax in crop_coords[0]: for xmin, xmax in crop_coords[1]: if len(crop_coords) == 3 and patch_size[2] > 1: for zmin, zmax in crop_coords[2]: coords_mesh_grid.append([ymin, ymax, xmin, xmax, zmin, zmax]) elif len(crop_coords) == 3 and patch_size[2] == 1: for zmin in range(img.shape[2]): coords_mesh_grid.append([ymin, ymax, xmin, xmax, zmin, zmin + 1]) else: coords_mesh_grid.append([ymin, ymax, xmin, xmax]) return np.array(coords_mesh_grid).astype(int) def pad_nd_image(image, new_shape=None, mode="edge", kwargs=None, return_slicer=False, shape_must_be_divisible_by=None): """ one padder to pad them all. Documentation? Well okay. A little bit. by Fabian Isensee :param image: nd image. can be anything :param new_shape: what shape do you want? new_shape does not have to have the same dimensionality as image. If len(new_shape) < len(image.shape) then the last axes of image will be padded. If new_shape < image.shape in any of the axes then we will not pad that axis, but also not crop! (interpret new_shape as new_min_shape) Example: image.shape = (10, 1, 512, 512); new_shape = (768, 768) -> result: (10, 1, 768, 768). Cool, huh? image.shape = (10, 1, 512, 512); new_shape = (364, 768) -> result: (10, 1, 512, 768). :param mode: see np.pad for documentation :param return_slicer: if True then this function will also return what coords you will need to use when cropping back to original shape :param shape_must_be_divisible_by: for network prediction. After applying new_shape, make sure the new shape is divisibly by that number (can also be a list with an entry for each axis). Whatever is missing to match that will be padded (so the result may be larger than new_shape if shape_must_be_divisible_by is not None) :param kwargs: see np.pad for documentation """ if kwargs is None: kwargs = {} if new_shape is not None: old_shape = np.array(image.shape[-len(new_shape):]) else: assert shape_must_be_divisible_by is not None assert isinstance(shape_must_be_divisible_by, (list, tuple, np.ndarray)) new_shape = image.shape[-len(shape_must_be_divisible_by):] old_shape = new_shape num_axes_nopad = len(image.shape) - len(new_shape) new_shape = [max(new_shape[i], old_shape[i]) for i in range(len(new_shape))] if not isinstance(new_shape, np.ndarray): new_shape = np.array(new_shape) if shape_must_be_divisible_by is not None: if not isinstance(shape_must_be_divisible_by, (list, tuple, np.ndarray)): shape_must_be_divisible_by = [shape_must_be_divisible_by] * len(new_shape) else: assert len(shape_must_be_divisible_by) == len(new_shape) for i in range(len(new_shape)): if new_shape[i] % shape_must_be_divisible_by[i] == 0: new_shape[i] -= shape_must_be_divisible_by[i] new_shape = np.array([new_shape[i] + shape_must_be_divisible_by[i] - new_shape[i] % shape_must_be_divisible_by[i] for i in range(len(new_shape))]) difference = new_shape - old_shape pad_below = difference // 2 pad_above = difference // 2 + difference % 2 pad_list = [[0, 0]]*num_axes_nopad + list([list(i) for i in zip(pad_below, pad_above)]) res = np.pad(image, pad_list, mode, **kwargs) if not return_slicer: return res else: pad_list = np.array(pad_list) pad_list[:, 1] = np.array(res.shape) - pad_list[:, 1] slicer = list(slice(*i) for i in pad_list) return res, slicer def convert_seg_to_bounding_box_coordinates(data_dict, dim, roi_item_keys, get_rois_from_seg=False, class_specific_seg=False): '''adapted from batchgenerators :param data_dict: seg: segmentation with labels indicating roi_count (get_rois_from_seg=False) or classes (get_rois_from_seg=True), class_targets: list where list index corresponds to roi id (roi_count) :param dim: :param roi_item_keys: keys of the roi-wise items in data_dict to process :param n_rg_feats: nr of regression vector features :param get_rois_from_seg: :return: coords (y1,x1,y2,x2 (,z1,z2)) where the segmentation GT is framed by +1 voxel, i.e., for an object with z-extensions z1=0 through z2=5, bbox target coords will be z1=-1, z2=6. (analogically for x,y). data_dict['roi_masks']: (b, n(b), c, h(n), w(n) (z(n))) list like roi_labels but with arrays (masks) inplace of integers. c==1 if segmentation not one-hot encoded. ''' bb_target = [] roi_masks = [] roi_items = {name:[] for name in roi_item_keys} out_seg = np.copy(data_dict['seg']) for b in range(data_dict['seg'].shape[0]): p_coords_list = [] #p for patient? p_roi_masks_list = [] p_roi_items_lists = {name:[] for name in roi_item_keys} if np.sum(data_dict['seg'][b] != 0) > 0: if get_rois_from_seg: clusters, n_cands = lb(data_dict['seg'][b]) data_dict['class_targets'][b] = [data_dict['class_targets'][b]] * n_cands else: n_cands = int(np.max(data_dict['seg'][b])) rois = np.array( [(data_dict['seg'][b] == ii) * 1 for ii in range(1, n_cands + 1)], dtype='uint8') # separate clusters for rix, r in enumerate(rois): if np.sum(r != 0) > 0: # check if the roi survived slicing (3D->2D) and data augmentation (cropping etc.) seg_ixs = np.argwhere(r != 0) coord_list = [np.min(seg_ixs[:, 1]) - 1, np.min(seg_ixs[:, 2]) - 1, np.max(seg_ixs[:, 1]) + 1, np.max(seg_ixs[:, 2]) + 1] if dim == 3: coord_list.extend([np.min(seg_ixs[:, 3]) - 1, np.max(seg_ixs[:, 3]) + 1]) p_coords_list.append(coord_list) p_roi_masks_list.append(r) # add background class = 0. rix is a patient wide index of lesions. since 'class_targets' is # also patient wide, this assignment is not dependent on patch occurrences. for name in roi_item_keys: p_roi_items_lists[name].append(data_dict[name][b][rix]) assert data_dict["class_targets"][b][rix]>=1, "convertsegtobbox produced bg roi w cl targ {} and unique roi seg {}".format(data_dict["class_targets"][b][rix], np.unique(r)) if class_specific_seg: out_seg[b][data_dict['seg'][b] == rix + 1] = data_dict['class_targets'][b][rix] if not class_specific_seg: out_seg[b][data_dict['seg'][b] > 0] = 1 bb_target.append(np.array(p_coords_list)) roi_masks.append(np.array(p_roi_masks_list)) for name in roi_item_keys: roi_items[name].append(np.array(p_roi_items_lists[name])) else: bb_target.append([]) roi_masks.append(np.zeros_like(data_dict['seg'][b], dtype='uint8')[None]) for name in roi_item_keys: roi_items[name].append(np.array([])) if get_rois_from_seg: data_dict.pop('class_targets', None) data_dict['bb_target'] = np.array(bb_target) data_dict['roi_masks'] = np.array(roi_masks) data_dict['seg'] = out_seg for name in roi_item_keys: data_dict[name] = np.array(roi_items[name]) return data_dict class ConvertSegToBoundingBoxCoordinates(AbstractTransform): """ Converts segmentation masks into bounding box coordinates. """ def __init__(self, dim, roi_item_keys, get_rois_from_seg=False, class_specific_seg=False): self.dim = dim self.roi_item_keys = roi_item_keys self.get_rois_from_seg = get_rois_from_seg self.class_specific_seg = class_specific_seg def __call__(self, **data_dict): return convert_seg_to_bounding_box_coordinates(data_dict, self.dim, self.roi_item_keys, self.get_rois_from_seg, self.class_specific_seg) ############################# # data packing / unpacking # not used, data_manager.py used instead ############################# def get_case_identifiers(folder): case_identifiers = [i[:-4] for i in os.listdir(folder) if i.endswith("npz")] return case_identifiers def convert_to_npy(npz_file): if not os.path.isfile(npz_file[:-3] + "npy"): a = np.load(npz_file)['data'] np.save(npz_file[:-3] + "npy", a) def unpack_dataset(folder, threads=8): case_identifiers = get_case_identifiers(folder) p = Pool(threads) npz_files = [os.path.join(folder, i + ".npz") for i in case_identifiers] p.map(convert_to_npy, npz_files) p.close() p.join() def delete_npy(folder): case_identifiers = get_case_identifiers(folder) npy_files = [os.path.join(folder, i + ".npy") for i in case_identifiers] npy_files = [i for i in npy_files if os.path.isfile(i)] for n in npy_files: os.remove(n) \ No newline at end of file