diff --git a/experiments/toy_exp/configs.py b/experiments/toy_exp/configs.py
index bd36ebd..c9703f1 100644
--- a/experiments/toy_exp/configs.py
+++ b/experiments/toy_exp/configs.py
@@ -1,344 +1,345 @@
#!/usr/bin/env python
# Copyright 2018 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
class configs(DefaultConfigs):
def __init__(self, server_env=None):
#########################
# Preprocessing #
#########################
- self.root_dir = '/media/gregor/HDD2TB/data/toy_mdt'
+ self.root_dir = '/home/gregor/datasets/toy_mdt'
#########################
# I/O #
#########################
# one out of [2, 3]. dimension the model operates in.
self.dim = 2
# one out of ['mrcnn', 'retina_net', 'retina_unet', 'detection_unet', 'ufrcnn', 'detection_unet'].
self.model = 'mrcnn'
DefaultConfigs.__init__(self, self.model, server_env, self.dim)
# int [0 < dataset_size]. select n patients from dataset for prototyping.
self.select_prototype_subset = None
self.hold_out_test_set = True
- self.n_train_data = 2500
+ # including val set. will be 3/4 train, 1/4 val.
+ self.n_train_val_data = 1500
# choose one of the 3 toy experiments described in https://arxiv.org/pdf/1811.08661.pdf
# one of ['donuts_shape', 'donuts_pattern', 'circles_scale'].
- toy_mode = 'donuts_shape_nonoise_dia12'
+ toy_mode = 'donuts_shape_noise'
# path to preprocessed data.
self.input_df_name = 'info_df.pickle'
self.pp_name = os.path.join(toy_mode, 'train')
self.pp_data_path = os.path.join(self.root_dir, self.pp_name)
self.pp_test_name = os.path.join(toy_mode, 'test')
self.pp_test_data_path = os.path.join(self.root_dir, self.pp_test_name)
# settings for deployment in cloud.
if server_env:
# path to preprocessed data.
pp_root_dir = '/datasets/datasets_ramien/toy_exp/data'
self.pp_name = os.path.join(toy_mode, 'train')
self.pp_data_path = os.path.join(pp_root_dir, self.pp_name)
self.pp_test_name = os.path.join(toy_mode, 'test')
self.pp_test_data_path = os.path.join(pp_root_dir, self.pp_test_name)
self.select_prototype_subset = None
#########################
# Data Loader #
#########################
# 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.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_sample_slack = 0.2
# set 2D network to operate in 3D images.
self.merge_2D_to_3D_preds = False
# 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)
#########################
# Architecture #
#########################
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'
self.weight_decay = 0
# one of 'xavier_uniform', 'xavier_normal', or 'kaiming_normal', None (=default = 'kaiming_uniform')
self.weight_init = None
#########################
# Schedule / Selection #
#########################
self.num_epochs = 24
self.num_train_batches = 100 if self.dim == 2 else 200
self.batch_size = 20 if self.dim == 2 else 8
self.do_validation = True
# 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_patient' # one of 'val_sampling' , 'val_patient'
if self.val_mode == 'val_patient':
self.max_val_patients = None # if 'None' iterates over entire val_set once.
if self.val_mode == 'val_sampling':
self.num_val_batches = 50
#########################
# Testing / Plotting #
#########################
# set the top-n-epochs to be saved for temporal averaging in testing.
self.save_n_models = 5
self.test_n_epochs = 5
# 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
self.report_score_level = ['patient', 'rois'] # choose list from 'patient', 'rois'
self.class_dict = {1: 'benign', 2: 'malignant'} # 0 is background.
self.patient_class_of_interest = 2 # patient metrics are only plotted for one class.
self.ap_match_ious = [0.1] # list of ious to be evaluated for ap-scoring.
self.model_selection_criteria = ['benign_ap', 'malignant_ap'] # criteria to average over for saving epochs.
self.min_det_thresh = 0.1 # minimum confidence value to select predictions for evaluation.
# threshold for clustering predictions together (wcs = weighted cluster scoring).
# needs to be >= the expected overlap of predictions coming from one model (typically NMS threshold).
# if too high, preds of the same object are separate clusters.
self.wcs_iou = 1e-5
self.plot_prediction_histograms = True
self.plot_stat_curves = False
#########################
# Data Augmentation #
#########################
self.da_kwargs={
'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)
#########################
# Add model specifics #
#########################
{'detection_unet': self.add_det_unet_configs,
'mrcnn': self.add_mrcnn_configs,
'ufrcnn': self.add_mrcnn_configs,
'ufrcnn_surrounding': self.add_mrcnn_configs,
'retina_net': self.add_mrcnn_configs,
'retina_unet': self.add_mrcnn_configs,
'prob_detector': self.add_mrcnn_configs,
}[self.model]()
def add_det_unet_configs(self):
self.learning_rate = [1e-4] * self.num_epochs
# aggregation from pixel perdiction to object scores (connected component). One of ['max', 'median']
self.aggregation_operation = 'max'
# max number of roi candidates to identify per image (slice in 2D, volume in 3D)
self.n_roi_candidates = 3 if self.dim == 2 else 8
# loss mode: either weighted cross entropy ('wce'), batch-wise dice loss ('dice), or the sum of both ('dice_wce')
self.seg_loss_mode = 'dice_wce'
# if <1, false positive predictions in foreground are penalized less.
self.fp_dice_weight = 1 if self.dim == 2 else 1
self.wce_weights = [1, 1, 1]
self.detection_min_confidence = self.min_det_thresh
# if 'True', loss distinguishes all classes, else only foreground vs. background (class agnostic).
self.class_specific_seg_flag = True
self.num_seg_classes = 3 if self.class_specific_seg_flag else 2
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
# disable mask head loss. (e.g. if no pixelwise annotations available)
self.frcnn_mode = 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_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 head networks: n_foreground_classes + 1 (background)
self.head_classes = 3
# seg_classes hier refers to the first stage classifier (RPN)
self.num_seg_classes = 2 # foreground vs. background
# 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 = 2 #per batch element
self.train_rois_per_image = 2 #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]])
self.scale = np.array([self.patch_size[0], self.patch_size[1], self.patch_size[0], self.patch_size[1]])
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 = 800 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 == 'ufrcnn':
self.operate_stride1 = True
self.class_specific_seg_flag = True
self.num_seg_classes = 3 if self.class_specific_seg_flag else 2
self.frcnn_mode = True
if self.model == 'retina_net' or self.model == 'retina_unet' or self.model == 'prob_detector':
# 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 64
# pre-selection of detections for NMS-speedup. per entire batch.
self.pre_nms_limit = 10000 if self.dim == 2 else 50000
# 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 'True', seg loss distinguishes all classes, else only foreground vs. background (class agnostic).
self.num_seg_classes = 3 if self.class_specific_seg_flag else 2
if self.model == 'retina_unet':
self.operate_stride1 = True
diff --git a/experiments/toy_exp/data_loader.py b/experiments/toy_exp/data_loader.py
index 7ceec4b..ae87d76 100644
--- a/experiments/toy_exp/data_loader.py
+++ b/experiments/toy_exp/data_loader.py
@@ -1,305 +1,308 @@
#!/usr/bin/env python
# Copyright 2018 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 numpy as np
import os
from collections import OrderedDict
import pandas as pd
import pickle
import time
import subprocess
import utils.dataloader_utils as dutils
# batch generator tools from https://github.com/MIC-DKFZ/batchgenerators
from batchgenerators.dataloading.data_loader import SlimDataLoaderBase
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.dataloading import SingleThreadedAugmenter
from batchgenerators.transforms.spatial_transforms import SpatialTransform
from batchgenerators.transforms.crop_and_pad_transforms import CenterCropTransform
from batchgenerators.transforms.utility_transforms import ConvertSegToBoundingBoxCoordinates
def get_train_generators(cf, logger):
"""
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.hold_out_test_set is True, adds the test split to the training data.
"""
all_data = load_dataset(cf, logger)
all_pids_list = np.unique([v['pid'] for (k, v) in all_data.items()])
- train_pids = all_pids_list[:cf.n_train_data]
- val_pids = all_pids_list[1000:1500]
+ assert cf.n_train_val_data <= len(all_pids_list), \
+ "requested {} train val samples, but dataset only has {} train val samples.".format(
+ cf.n_train_val_data, len(all_pids_list))
+ train_pids = all_pids_list[:int(2 * cf.n_train_val_data // 3)]
+ val_pids = all_pids_list[int(np.ceil(2 * cf.n_train_val_data // 3)):cf.n_train_val_data]
train_data = {k: v for (k, v) in all_data.items() if any(p == v['pid'] for p in train_pids)}
val_data = {k: v for (k, v) in all_data.items() if any(p == v['pid'] for p in val_pids)}
logger.info("data set loaded with: {} train / {} val patients".format(len(train_pids), len(val_pids)))
batch_gen = {}
batch_gen['train'] = create_data_gen_pipeline(train_data, cf=cf, do_aug=False)
batch_gen['val_sampling'] = create_data_gen_pipeline(val_data, cf=cf, do_aug=False)
if cf.val_mode == 'val_patient':
batch_gen['val_patient'] = PatientBatchIterator(val_data, cf=cf)
batch_gen['n_val'] = len(val_pids) if cf.max_val_patients is None else min(len(val_pids), 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.hold_out_test_set is True, gets the data from an external folder instead.
"""
if cf.hold_out_test_set:
pp_name = cf.pp_test_name
test_ix = None
else:
pp_name = None
with open(os.path.join(cf.exp_dir, 'fold_ids.pickle'), 'rb') as handle:
fold_list = pickle.load(handle)
_, _, test_ix, _ = fold_list[cf.fold]
# warnings.warn('WARNING: using validation set for testing!!!')
test_data = load_dataset(cf, logger, test_ix, pp_data_path=cf.pp_test_data_path, pp_name=pp_name)
logger.info("data set loaded with: {} test patients from {}".format(len(test_data.keys()), cf.pp_test_data_path))
batch_gen = {}
batch_gen['test'] = PatientBatchIterator(test_data, cf=cf)
batch_gen['n_test'] = len(test_data.keys()) if cf.max_test_patients=="all" else \
min(cf.max_test_patients, len(test_data.keys()))
return batch_gen
def load_dataset(cf, logger, subset_ixs=None, pp_data_path=None, pp_name=None):
"""
loads the dataset. if deployed in cloud also copies and unpacks the data to the working directory.
:param subset_ixs: subset indices to be loaded from the dataset. used e.g. for testing to only load the test folds.
:return: data: dictionary with one entry per patient (in this case per patient-breast, since they are treated as
individual images for training) each entry is a dictionary containing respective meta-info as well as paths to the preprocessed
numpy arrays to be loaded during batch-generation
"""
if pp_data_path is None:
pp_data_path = cf.pp_data_path
if pp_name is None:
pp_name = cf.pp_name
if cf.server_env:
copy_data = True
target_dir = os.path.join(cf.data_dest, pp_name)
if not os.path.exists(target_dir):
cf.data_source_dir = pp_data_path
os.makedirs(target_dir)
subprocess.call('rsync -av {} {}'.format(
os.path.join(cf.data_source_dir, cf.input_df_name), os.path.join(target_dir, cf.input_df_name)), shell=True)
logger.info('created target dir and info df at {}'.format(os.path.join(target_dir, cf.input_df_name)))
elif subset_ixs is None:
copy_data = False
pp_data_path = target_dir
p_df = pd.read_pickle(os.path.join(pp_data_path, cf.input_df_name))
if subset_ixs is not None:
subset_pids = [np.unique(p_df.pid.tolist())[ix] for ix in subset_ixs]
p_df = p_df[p_df.pid.isin(subset_pids)]
logger.info('subset: selected {} instances from df'.format(len(p_df)))
if cf.server_env:
if copy_data:
copy_and_unpack_data(logger, p_df.pid.tolist(), cf.fold_dir, cf.data_source_dir, target_dir)
class_targets = p_df['class_id'].tolist()
pids = p_df.pid.tolist()
imgs = [os.path.join(pp_data_path, '{}.npy'.format(pid)) for pid in pids]
segs = [os.path.join(pp_data_path,'{}.npy'.format(pid)) for pid in pids]
data = OrderedDict()
for ix, pid in enumerate(pids):
data[pid] = {'data': imgs[ix], 'seg': segs[ix], 'pid': pid, 'class_target': [class_targets[ix]]}
return data
def create_data_gen_pipeline(patient_data, cf, do_aug=True):
"""
create mutli-threaded train/val/test batch generation and augmentation pipeline.
: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
"""
# create instance of batch generator as first element in pipeline.
data_gen = BatchGenerator(patient_data, batch_size=cf.batch_size, cf=cf)
# add transformations to pipeline.
my_transforms = []
if do_aug:
mirror_transform = Mirror(axes=np.arange(2, cf.dim+2, 1))
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]))
my_transforms.append(ConvertSegToBoundingBoxCoordinates(cf.dim, get_rois_from_seg_flag=False, class_specific_seg_flag=cf.class_specific_seg_flag))
all_transforms = Compose(my_transforms)
# multithreaded_generator = SingleThreadedAugmenter(data_gen, all_transforms)
multithreaded_generator = MultiThreadedAugmenter(data_gen, all_transforms, num_processes=cf.n_workers, seeds=range(cf.n_workers))
return multithreaded_generator
class BatchGenerator(SlimDataLoaderBase):
"""
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, data, batch_size, cf):
super(BatchGenerator, self).__init__(data, batch_size)
self.cf = cf
def generate_train_batch(self):
batch_data, batch_segs, batch_pids, batch_targets = [], [], [], []
class_targets_list = [v['class_target'] for (k, v) in self._data.items()]
#samples patients towards equilibrium of foreground classes on a roi-level (after randomly sampling the ratio "batch_sample_slack).
batch_ixs = dutils.get_class_balanced_patients(
class_targets_list, self.batch_size, self.cf.head_classes - 1, slack_factor=self.cf.batch_sample_slack)
patients = list(self._data.items())
for b in batch_ixs:
patient = patients[b][1]
all_data = np.load(patient['data'], mmap_mode='r')
data = all_data[0]
seg = all_data[1].astype('uint8')
batch_pids.append(patient['pid'])
batch_targets.append(patient['class_target'])
batch_data.append(data[np.newaxis])
batch_segs.append(seg[np.newaxis])
data = np.array(batch_data)
seg = np.array(batch_segs).astype(np.uint8)
class_target = np.array(batch_targets)
return {'data': data, 'seg': seg, 'pid': batch_pids, 'class_target': class_target}
class PatientBatchIterator(SlimDataLoaderBase):
"""
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, data, cf): #threads in augmenter
super(PatientBatchIterator, self).__init__(data, 0)
self.cf = cf
self.patient_ix = 0
self.dataset_pids = [v['pid'] for (k, v) in data.items()]
self.patch_size = cf.patch_size
if len(self.patch_size) == 2:
self.patch_size = self.patch_size + [1]
def generate_train_batch(self):
pid = self.dataset_pids[self.patient_ix]
patient = self._data[pid]
all_data = np.load(patient['data'], mmap_mode='r')
data = all_data[0]
seg = all_data[1].astype('uint8')
batch_class_targets = np.array([patient['class_target']])
out_data = data[None, None]
out_seg = seg[None, None]
print('check patient data loader', out_data.shape, out_seg.shape)
batch_2D = {'data': out_data, 'seg': out_seg, 'class_target': batch_class_targets, 'pid': pid}
converter = ConvertSegToBoundingBoxCoordinates(dim=2, get_rois_from_seg_flag=False, class_specific_seg_flag=self.cf.class_specific_seg_flag)
batch_2D = converter(**batch_2D)
batch_2D.update({'patient_bb_target': batch_2D['bb_target'],
'patient_roi_labels': batch_2D['roi_labels'],
'original_img_shape': out_data.shape})
self.patient_ix += 1
if self.patient_ix == len(self.dataset_pids):
self.patient_ix = 0
return batch_2D
def copy_and_unpack_data(logger, pids, fold_dir, source_dir, target_dir):
start_time = time.time()
with open(os.path.join(fold_dir, 'file_list.txt'), 'w') as handle:
for pid in pids:
handle.write('{}.npy\n'.format(pid))
subprocess.call('rsync -av --files-from {} {} {}'.format(os.path.join(fold_dir, 'file_list.txt'),
source_dir, target_dir), shell=True)
# dutils.unpack_dataset(target_dir)
copied_files = os.listdir(target_dir)
logger.info("copying and unpacking data set finsihed : {} files in target dir: {}. took {} sec".format(
len(copied_files), target_dir, np.round(time.time() - start_time, 0)))
if __name__=="__main__":
import utils.exp_utils as utils
from .configs import Configs
total_stime = time.time()
cf = Configs()
logger = utils.get_logger(0)
batch_gen = get_train_generators(cf, logger)
train_batch = next(batch_gen["train"])
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))
\ No newline at end of file