diff --git a/default_configs.py b/default_configs.py index e68a3ed..fe4e0e0 100644 --- a/default_configs.py +++ b/default_configs.py @@ -1,203 +1,203 @@ #!/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. # ============================================================================== """Default Configurations script. Avoids changing configs of all experiments if general settings are to be changed.""" import os from collections import namedtuple boxLabel = namedtuple('boxLabel', ["name", "color"]) class DefaultConfigs: def __init__(self, server_env=None, dim=2): self.server_env = server_env self.cuda_benchmark = True - self.sysmetrics_interval = -1 # set > 0 to record system metrics to tboard with this time span in seconds. + self.sysmetrics_interval = 2 # set > 0 to record system metrics to tboard with this time span in seconds. ######################### # I/O # ######################### self.dim = dim # int [0 < dataset_size]. select n patients from dataset for prototyping. self.select_prototype_subset = None # some default paths. self.source_dir = os.path.dirname(os.path.realpath(__file__)) # current dir. self.backbone_path = os.path.join(self.source_dir, 'models/backbone.py') self.input_df_name = 'info_df.pickle' if server_env: self.select_prototype_subset = None ######################### # Colors/legends # ######################### # in part from solarized theme. self.black = (0.1, 0.05, 0.) self.gray = (0.514, 0.580, 0.588) self.beige = (1., 1., 0.85) self.white = (0.992, 0.965, 0.890) self.green = (0.659, 0.792, 0.251) # [168, 202, 64] self.dark_green = (0.522, 0.600, 0.000) # [133.11, 153. , 0. ] self.cyan = (0.165, 0.631, 0.596) # [ 42.075, 160.905, 151.98 ] self.bright_blue = (0.85, 0.95, 1.) self.blue = (0.149, 0.545, 0.824) # [ 37.995, 138.975, 210.12 ] self.dkfz_blue = (0, 75. / 255, 142. / 255) self.dark_blue = (0.027, 0.212, 0.259) # [ 6.885, 54.06 , 66.045] self.purple = (0.424, 0.443, 0.769) # [108.12 , 112.965, 196.095] self.aubergine = (0.62, 0.21, 0.44) # [ 157, 53 , 111] self.magenta = (0.827, 0.212, 0.510) # [210.885, 54.06 , 130.05 ] self.coral = (1., 0.251, 0.4) # [255,64,102] self.bright_red = (1., 0.15, 0.1) # [255, 38.25, 25.5] self.brighter_red = (0.863, 0.196, 0.184) # [220.065, 49.98 , 46.92 ] self.red = (0.87, 0.05, 0.01) # [ 223, 13, 2] self.dark_red = (0.6, 0.04, 0.005) self.orange = (0.91, 0.33, 0.125) # [ 232.05 , 84.15 , 31.875] self.dark_orange = (0.796, 0.294, 0.086) #[202.98, 74.97, 21.93] self.yellow = (0.95, 0.9, 0.02) # [ 242.25, 229.5 , 5.1 ] self.dark_yellow = (0.710, 0.537, 0.000) # [181.05 , 136.935, 0. ] self.color_palette = [self.blue, self.dark_blue, self.aubergine, self.green, self.yellow, self.orange, self.red, self.cyan, self.black] self.box_labels = [ # name color boxLabel("det", self.blue), boxLabel("prop", self.gray), boxLabel("pos_anchor", self.cyan), boxLabel("neg_anchor", self.cyan), boxLabel("neg_class", self.green), boxLabel("pos_class", self.aubergine), boxLabel("gt", self.red) ] # neg and pos in a medical sense, i.e., pos=positive diagnostic finding self.box_type2label = {label.name: label for label in self.box_labels} self.box_color_palette = {label.name: label.color for label in self.box_labels} # whether the input data is mono-channel or RGB/rgb self.has_colorchannels = False ######################### # Data Loader # ######################### #random seed for fold_generator and batch_generator. self.seed = 0 #number of threads for multithreaded tasks like batch generation, wcs, merge2dto3d self.n_workers = 16 if server_env else os.cpu_count() self.create_bounding_box_targets = True self.class_specific_seg = True # False if self.model=="mrcnn" else True ######################### # Architecture # ######################### self.prediction_tasks = ["class"] # 'class', 'regression_class', 'regression_kendall', 'regression_feindt' self.weight_decay = 0.0 # nonlinearity to be applied after convs with nonlinearity. one of 'relu' or 'leaky_relu' self.relu = 'relu' # if True initializes weights as specified in model script. else use default Pytorch init. self.weight_init = None # if True adds high-res decoder levels to feature pyramid: P1 + P0. (e.g. set to true in retina_unet configs) self.operate_stride1 = False ######################### # Optimization # ######################### self.optimizer = "ADAM" # "ADAM" or "SGD" or implemented additionals ######################### # Schedule # ######################### # number of folds in cross validation. self.n_cv_splits = 5 ######################### # Testing / Plotting # ######################### # perform mirroring at test time. (only XY. Z not done to not blow up predictions times). self.test_aug = True # if True, test data lies in a separate folder and is not part of the cross validation. self.held_out_test_set = False # if hold-out test set: eval each fold's parameters separately on the test set self.eval_test_fold_wise = True # if held_out_test_set provided, ensemble predictions over models of all trained cv-folds. self.ensemble_folds = False # what metrics to evaluate self.metrics = ['ap'] # whether to evaluate fold means when evaluating over more than one fold self.evaluate_fold_means = False # how often (in nr of epochs) to plot example batches during train/val self.plot_frequency = 1 # color specifications for all box_types in prediction_plot. self.box_color_palette = {'det': 'b', 'gt': 'r', 'neg_class': 'purple', 'prop': 'w', 'pos_class': 'g', 'pos_anchor': 'c', 'neg_anchor': 'c'} # scan over confidence score in evaluation to optimize it on the validation set. self.scan_det_thresh = False # plots roc-curves / prc-curves in evaluation. self.plot_stat_curves = False # if True: evaluate average precision per patient id and average over per-pid results, # instead of computing one ap over whole data set. self.per_patient_ap = False # threshold for clustering 2D box predictions to 3D Cubes. Overlap is computed in XY. self.merge_3D_iou = 0.1 ######################### # MRCNN # ######################### # if True, mask loss is not applied. used for data sets, where no pixel-wise annotations are provided. self.frcnn_mode = False self.return_masks_in_train = False # if True, unmolds masks in Mask R-CNN to full-res for plotting/monitoring. self.return_masks_in_val = False self.return_masks_in_test = False # needed if doing instance segmentation. evaluation not yet implemented. # add P6 to Feature Pyramid Network. self.sixth_pooling = False ######################### # RetinaNet # ######################### self.focal_loss = False self.focal_loss_gamma = 2. diff --git a/exec.py b/exec.py index 0903b48..74c10ba 100644 --- a/exec.py +++ b/exec.py @@ -1,344 +1,343 @@ #!/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. # ============================================================================== """ execution script. this where all routines come together and the only script you need to call. refer to parse args below to see options for execution. """ import plotting as plg import os import warnings import argparse import time import torch import utils.exp_utils as utils from evaluator import Evaluator from predictor import Predictor for msg in ["Attempting to set identical bottom==top results", "This figure includes Axes that are not compatible with tight_layout", "Data has no positive values, and therefore cannot be log-scaled.", ".*invalid value encountered in true_divide.*"]: warnings.filterwarnings("ignore", msg) def train(cf, logger): """ performs the training routine for a given fold. saves plots and selected parameters to the experiment dir specified in the configs. logs to file and tensorboard. """ logger.info('performing training in {}D over fold {} on experiment {} with model {}'.format( cf.dim, cf.fold, cf.exp_dir, cf.model)) logger.time("train_val") # -------------- inits and settings ----------------- net = model.net(cf, logger).cuda() if cf.optimizer == "ADAM": optimizer = torch.optim.Adam(net.parameters(), lr=cf.learning_rate[0], weight_decay=cf.weight_decay) elif cf.optimizer == "SGD": optimizer = torch.optim.SGD(net.parameters(), lr=cf.learning_rate[0], weight_decay=cf.weight_decay, momentum=0.3) if cf.dynamic_lr_scheduling: scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode=cf.scheduling_mode, factor=cf.lr_decay_factor, patience=cf.scheduling_patience) model_selector = utils.ModelSelector(cf, logger) starting_epoch = 1 if cf.resume_from_checkpoint: starting_epoch = utils.load_checkpoint(cf.resume_from_checkpoint, net, optimizer) logger.info('resumed from checkpoint {} at epoch {}'.format(cf.resume_from_checkpoint, starting_epoch)) # prepare monitoring monitor_metrics = utils.prepare_monitoring(cf) logger.info('loading dataset and initializing batch generators...') batch_gen = data_loader.get_train_generators(cf, logger) # -------------- training ----------------- for epoch in range(starting_epoch, cf.num_epochs + 1): logger.info('starting training epoch {}/{}'.format(epoch, cf.num_epochs)) logger.time("train_epoch") net.train() train_results_list = [] train_evaluator = Evaluator(cf, logger, mode='train') for i in range(cf.num_train_batches): logger.time("train_batch_loadfw") batch = next(batch_gen['train']) batch_gen['train'].generator.stats['roi_counts'] += batch['roi_counts'] batch_gen['train'].generator.stats['empty_samples_count'] += batch['empty_samples_count'] logger.time("train_batch_loadfw") logger.time("train_batch_netfw") results_dict = net.train_forward(batch) logger.time("train_batch_netfw") logger.time("train_batch_bw") optimizer.zero_grad() results_dict['torch_loss'].backward() if cf.clip_norm: torch.nn.utils.clip_grad_norm_(net.parameters(), cf.clip_norm, norm_type=2) # gradient clipping optimizer.step() train_results_list.append(({k:v for k,v in results_dict.items() if k != "seg_preds"}, batch["pid"])) # slim res dict if not cf.server_env: print("\rFinished training batch " + "{}/{} in {:.1f}s ({:.2f}/{:.2f} forw load/net, {:.2f} backw).".format(i+1, cf.num_train_batches, logger.get_time("train_batch_loadfw")+ logger.get_time("train_batch_netfw") +logger.time("train_batch_bw"), logger.get_time("train_batch_loadfw",reset=True), logger.get_time("train_batch_netfw", reset=True), logger.get_time("train_batch_bw", reset=True)), end="", flush=True) print() #--------------- train eval ---------------- if (epoch-1)%cf.plot_frequency==0: # view an example batch logger.time("train_plot") plg.view_batch(cf, batch, results_dict, has_colorchannels=cf.has_colorchannels, show_gt_labels=True, out_file=os.path.join(cf.plot_dir, 'batch_example_train_{}.png'.format(cf.fold))) logger.info("generated train-example plot in {:.2f}s".format(logger.time("train_plot"))) logger.time("evals") _, monitor_metrics['train'] = train_evaluator.evaluate_predictions(train_results_list, monitor_metrics['train']) logger.time("evals") logger.time("train_epoch", toggle=False) del train_results_list #----------- validation ------------ logger.info('starting validation in mode {}.'.format(cf.val_mode)) logger.time("val_epoch") with torch.no_grad(): net.eval() val_results_list = [] val_evaluator = Evaluator(cf, logger, mode=cf.val_mode) val_predictor = Predictor(cf, net, logger, mode='val') for i in range(batch_gen['n_val']): logger.time("val_batch") batch = next(batch_gen[cf.val_mode]) if cf.val_mode == 'val_patient': results_dict = val_predictor.predict_patient(batch) elif cf.val_mode == 'val_sampling': results_dict = net.train_forward(batch, is_validation=True) val_results_list.append([results_dict, batch["pid"]]) if not cf.server_env: print("\rFinished validation {} {}/{} in {:.1f}s.".format('patient' if cf.val_mode=='val_patient' else 'batch', i + 1, batch_gen['n_val'], logger.time("val_batch")), end="", flush=True) print() #------------ val eval ------------- if (epoch - 1) % cf.plot_frequency == 0: logger.time("val_plot") plg.view_batch(cf, batch, results_dict, has_colorchannels=cf.has_colorchannels, show_gt_labels=True, out_file=os.path.join(cf.plot_dir, 'batch_example_val_{}.png'.format(cf.fold))) logger.info("generated val plot in {:.2f}s".format(logger.time("val_plot"))) logger.time("evals") _, monitor_metrics['val'] = val_evaluator.evaluate_predictions(val_results_list, monitor_metrics['val']) model_selector.run_model_selection(net, optimizer, monitor_metrics, epoch) del val_results_list #----------- monitoring ------------- monitor_metrics.update({"lr": {str(g) : group['lr'] for (g, group) in enumerate(optimizer.param_groups)}}) logger.metrics2tboard(monitor_metrics, global_step=epoch) logger.time("evals") logger.info('finished epoch {}/{}, took {:.2f}s. train total: {:.2f}s, average: {:.2f}s. val total: {:.2f}s, average: {:.2f}s.'.format( epoch, cf.num_epochs, logger.get_time("train_epoch")+logger.time("val_epoch"), logger.get_time("train_epoch"), logger.get_time("train_epoch", reset=True)/cf.num_train_batches, logger.get_time("val_epoch"), logger.get_time("val_epoch", reset=True)/batch_gen["n_val"])) logger.info("time for evals: {:.2f}s".format(logger.get_time("evals", reset=True))) #-------------- scheduling ----------------- if not cf.dynamic_lr_scheduling: for param_group in optimizer.param_groups: param_group['lr'] = cf.learning_rate[epoch-1] else: scheduler.step(monitor_metrics["val"][cf.scheduling_criterion][-1]) logger.time("train_val") logger.info("Training and validating over {} epochs took {}".format(cf.num_epochs, logger.get_time("train_val", format="hms", reset=True))) batch_gen['train'].generator.print_stats(logger, plot=True) def test(cf, logger, max_fold=None): """performs testing for a given fold (or held out set). saves stats in evaluator. """ logger.time("test_fold") logger.info('starting testing model of fold {} in exp {}'.format(cf.fold, cf.exp_dir)) net = model.net(cf, logger).cuda() batch_gen = data_loader.get_test_generator(cf, logger) test_predictor = Predictor(cf, net, logger, mode='test') test_results_list = test_predictor.predict_test_set(batch_gen, return_results = not hasattr( cf, "eval_test_separately") or not cf.eval_test_separately) if test_results_list is not None: test_evaluator = Evaluator(cf, logger, mode='test') test_evaluator.evaluate_predictions(test_results_list) test_evaluator.score_test_df(max_fold=max_fold) logger.info('Testing of fold {} took {}.'.format(cf.fold, logger.get_time("test_fold", reset=True, format="hms"))) - if __name__ == '__main__': stime = time.time() parser = argparse.ArgumentParser() parser.add_argument('-m', '--mode', type=str, default='train_test', help='one out of: create_exp, analysis, train, train_test, or test') parser.add_argument('-f', '--folds', nargs='+', type=int, default=None, help='None runs over all folds in CV. otherwise specify list of folds.') parser.add_argument('--exp_dir', type=str, default='/home/gregor/Documents/regrcnn/datasets/toy/experiments/dev', help='path to experiment dir. will be created if non existent.') parser.add_argument('--server_env', default=False, action='store_true', help='change IO settings to deploy models on a cluster.') parser.add_argument('--data_dest', type=str, default=None, help="path to final data folder if different from config") parser.add_argument('--use_stored_settings', default=False, action='store_true', help='load configs from existing exp_dir instead of source dir. always done for testing, ' 'but can be set to true to do the same for training. useful in job scheduler environment, ' 'where source code might change before the job actually runs.') parser.add_argument('--resume_from_checkpoint', type=str, default=None, help='path to checkpoint. if resuming from checkpoint, the desired fold still needs to be parsed via --folds.') parser.add_argument('--dataset_name', type=str, default='toy', help="path to the dataset-specific code in source_dir/datasets") parser.add_argument('-d', '--dev', default=False, action='store_true', help="development mode: shorten everything") args = parser.parse_args() args.dataset_name = os.path.join("datasets", args.dataset_name) if not "datasets" in args.dataset_name else args.dataset_name folds = args.folds resume_from_checkpoint = None if args.resume_from_checkpoint in ['None', 'none'] else args.resume_from_checkpoint if args.mode == 'create_exp': cf = utils.prep_exp(args.dataset_name, args.exp_dir, args.server_env, use_stored_settings=False) logger = utils.get_logger(cf.exp_dir, cf.server_env, -1) logger.info('created experiment directory at {}'.format(args.exp_dir)) elif args.mode == 'train' or args.mode == 'train_test': cf = utils.prep_exp(args.dataset_name, args.exp_dir, args.server_env, args.use_stored_settings) if args.dev: folds = [0,1] cf.batch_size, cf.num_epochs, cf.min_save_thresh, cf.save_n_models = 3 if cf.dim==2 else 1, 1, 0, 1 cf.num_train_batches, cf.num_val_batches, cf.max_val_patients = 5, 1, 1 cf.test_n_epochs = cf.save_n_models cf.max_test_patients = 1 torch.backends.cudnn.benchmark = cf.dim==3 else: torch.backends.cudnn.benchmark = cf.cuda_benchmark if args.data_dest is not None: cf.data_dest = args.data_dest logger = utils.get_logger(cf.exp_dir, cf.server_env, cf.sysmetrics_interval) data_loader = utils.import_module('data_loader', os.path.join(args.dataset_name, 'data_loader.py')) model = utils.import_module('model', cf.model_path) logger.info("loaded model from {}".format(cf.model_path)) if folds is None: folds = range(cf.n_cv_splits) for fold in folds: """k-fold cross-validation: the dataset is split into k equally-sized folds, one used for validation, one for testing, the rest for training. This loop iterates k-times over the dataset, cyclically moving the splits. k==folds, fold in [0,folds) says which split is used for testing. """ cf.fold_dir = os.path.join(cf.exp_dir, 'fold_{}'.format(fold)) cf.fold, logger.fold = fold, fold cf.resume_from_checkpoint = resume_from_checkpoint if not os.path.exists(cf.fold_dir): os.mkdir(cf.fold_dir) train(cf, logger) cf.resume_from_checkpoint = None if args.mode == 'train_test': test(cf, logger) elif args.mode == 'test': cf = utils.prep_exp(args.dataset_name, args.exp_dir, args.server_env, use_stored_settings=True, is_training=False) if args.data_dest is not None: cf.data_dest = args.data_dest logger = utils.get_logger(cf.exp_dir, cf.server_env, cf.sysmetrics_interval) data_loader = utils.import_module('data_loader', os.path.join(args.dataset_name, 'data_loader.py')) model = utils.import_module('model', cf.model_path) logger.info("loaded model from {}".format(cf.model_path)) fold_dirs = sorted([os.path.join(cf.exp_dir, f) for f in os.listdir(cf.exp_dir) if os.path.isdir(os.path.join(cf.exp_dir, f)) and f.startswith("fold")]) if folds is None: folds = range(cf.n_cv_splits) if args.dev: folds = folds[:2] cf.batch_size, cf.max_test_patients, cf.test_n_epochs = 1 if cf.dim==2 else 1, 2, 2 else: torch.backends.cudnn.benchmark = cf.cuda_benchmark for fold in folds: cf.fold = fold cf.fold_dir = os.path.join(cf.exp_dir, 'fold_{}'.format(cf.fold)) if cf.fold_dir in fold_dirs: test(cf, logger, max_fold=max([int(f[-1]) for f in fold_dirs])) else: logger.info("Skipping fold {} since no model parameters found.".format(fold)) # load raw predictions saved by predictor during testing, run aggregation algorithms and evaluation. elif args.mode == 'analysis': """ analyse already saved predictions. """ cf = utils.prep_exp(args.dataset_name, args.exp_dir, args.server_env, use_stored_settings=True, is_training=False) - logger = utils.get_logger(cf.exp_dir, cf.server_env, -1) + logger = utils.get_logger(cf.exp_dir, cf.server_env, cf.sysmetrics_interval) if cf.held_out_test_set and not cf.eval_test_fold_wise: predictor = Predictor(cf, net=None, logger=logger, mode='analysis') results_list = predictor.load_saved_predictions() logger.info('starting evaluation...') cf.fold = 0 evaluator = Evaluator(cf, logger, mode='test') evaluator.evaluate_predictions(results_list) evaluator.score_test_df(max_fold=0) else: fold_dirs = sorted([os.path.join(cf.exp_dir, f) for f in os.listdir(cf.exp_dir) if os.path.isdir(os.path.join(cf.exp_dir, f)) and f.startswith("fold")]) if args.dev: fold_dirs = fold_dirs[:1] if folds is None: folds = range(cf.n_cv_splits) for fold in folds: cf.fold = fold cf.fold_dir = os.path.join(cf.exp_dir, 'fold_{}'.format(cf.fold)) if cf.fold_dir in fold_dirs: predictor = Predictor(cf, net=None, logger=logger, mode='analysis') results_list = predictor.load_saved_predictions() # results_list[x][1] is pid, results_list[x][0] is list of len samples-per-patient, each entry hlds # list of boxes per that sample, i.e., len(results_list[x][y][0]) would be nr of boxes in sample y of patient x logger.info('starting evaluation...') evaluator = Evaluator(cf, logger, mode='test') evaluator.evaluate_predictions(results_list) max_fold = max([int(f[-1]) for f in fold_dirs]) evaluator.score_test_df(max_fold=max_fold) else: logger.info("Skipping fold {} since no model parameters found.".format(fold)) else: raise ValueError('mode "{}" specified in args is not implemented.'.format(args.mode)) mins, secs = divmod((time.time() - stime), 60) h, mins = divmod(mins, 60) t = "{:d}h:{:02d}m:{:02d}s".format(int(h), int(mins), int(secs)) logger.info("{} total runtime: {}".format(os.path.split(__file__)[1], t)) del logger torch.cuda.empty_cache() diff --git a/requirements.txt b/requirements.txt index 88d3644..127a70d 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,66 +1,18 @@ -absl-py==0.8.1 -backcall==0.1.0 batchgenerators==0.19.3 -cachetools==3.1.1 -certifi==2019.11.28 -chardet==3.0.4 -cycler==0.10.0 Cython==0.29.14 -decorator==4.4.1 -future==0.18.2 -google-auth==1.7.1 -google-auth-oauthlib==0.4.1 -grpcio==1.25.0 -idna==2.8 -imageio==2.6.1 ipython==7.9.0 ipython-genutils==0.2.0 -jedi==0.15.1 -joblib==0.14.0 -kiwisolver==1.1.0 -linecache2==1.0.0 Markdown==3.1.1 matplotlib==3.1.2 -networkx==2.4 -ninja==1.9.0.post1 -nms-extension==0.0.0 -numpy==1.17.3 -oauthlib==3.1.0 pandas==0.25.3 -parso==0.5.1 -pexpect==4.7.0 -pickleshare==0.7.5 -Pillow==6.2.1 -prompt-toolkit==2.0.10 -protobuf==3.10.0 psutil==5.6.5 -ptyprocess==0.6.0 -pyasn1==0.4.8 -pyasn1-modules==0.2.7 -Pygments==2.4.2 -pyparsing==2.4.4 -python-dateutil==2.8.1 -pytz==2019.3 -PyWavelets==1.1.1 -RegRCNN==0.0.2 -requests==2.22.0 -requests-oauthlib==1.3.0 -RoIAlign-extension-2D==0.0.0 -RoIAlign-extension-3D==0.0.0 -rsa==4.0 scikit-image==0.16.2 scikit-learn==0.21.3 scipy==1.3.1 SimpleITK==1.2.3 six==1.13.0 tensorboard==2.0.2 -threadpoolctl==1.1.0 torch==1.3.1 torchvision==0.4.2 tqdm==4.39.0 -traceback2==1.4.0 -traitlets==4.3.3 -unittest2==1.1.0 -urllib3==1.25.7 -wcwidth==0.1.7 -Werkzeug==0.16.0 + diff --git a/utils/exp_utils.py b/utils/exp_utils.py index 5bf4b05..8625761 100644 --- a/utils/exp_utils.py +++ b/utils/exp_utils.py @@ -1,659 +1,652 @@ #!/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 sys import os import subprocess import threading import pickle import importlib.util import psutil import time import logging from torch.utils.tensorboard import SummaryWriter from collections import OrderedDict import numpy as np import pandas as pd import torch def import_module(name, path): """ correct way of importing a module dynamically in python 3. :param name: name given to module instance. :param path: path to module. :return: module: returned module instance. """ spec = importlib.util.spec_from_file_location(name, path) module = importlib.util.module_from_spec(spec) spec.loader.exec_module(module) return module def save_obj(obj, name): """Pickle a python object.""" with open(name + '.pkl', 'wb') as f: pickle.dump(obj, f, pickle.HIGHEST_PROTOCOL) def load_obj(file_path): with open(file_path, 'rb') as handle: return pickle.load(handle) def IO_safe(func, *args, _tries=5, _raise=True, **kwargs): """ Wrapper calling function func with arguments args and keyword arguments kwargs to catch input/output errors on cluster. :param func: function to execute (intended to be read/write operation to a problematic cluster drive, but can be any function). :param args: positional args of func. :param kwargs: kw args of func. :param _tries: how many attempts to make executing func. """ for _try in range(_tries): try: return func(*args, **kwargs) except OSError as e: # to catch cluster issues with network drives if _raise: raise e else: print("After attempting execution {} time{}, following error occurred:\n{}".format(_try + 1, "" if _try == 0 else "s", e)) continue def query_nvidia_gpu(device_id, d_keyword=None, no_units=False): """ :param device_id: :param d_keyword: -d, --display argument (keyword(s) for selective display), all are selected if None :return: dict of gpu-info items """ cmd = ['nvidia-smi', '-i', str(device_id), '-q'] if d_keyword is not None: cmd += ['-d', d_keyword] outp = subprocess.check_output(cmd).strip().decode('utf-8').split("\n") outp = [x for x in outp if len(x) > 0] headers = [ix for ix, item in enumerate(outp) if len(item.split(":")) == 1] + [len(outp)] out_dict = {} for lix, hix in enumerate(headers[:-1]): head = outp[hix].strip().replace(" ", "_").lower() out_dict[head] = {} for lix2 in range(hix, headers[lix + 1]): try: key, val = [x.strip().lower() for x in outp[lix2].split(":")] if no_units: val = val.split()[0] out_dict[head][key] = val except: pass return out_dict class CombinedPrinter(object): """combined print function. prints to logger and/or file if given, to normal print if non given. """ def __init__(self, logger=None, file=None): if logger is None and file is None: self.out = [print] elif logger is None: self.out = [file.write] elif file is None: self.out = [logger.info] else: self.out = [logger.info, file.write] def __call__(self, string): for fct in self.out: fct(string) class Nvidia_GPU_Logger(object): def __init__(self): self.count = None def get_vals(self): cmd = ['nvidia-settings', '-t', '-q', 'GPUUtilization'] gpu_util = subprocess.check_output(cmd).strip().decode('utf-8').split(",") gpu_util = dict([f.strip().split("=") for f in gpu_util]) cmd[-1] = 'UsedDedicatedGPUMemory' gpu_used_mem = subprocess.check_output(cmd).strip().decode('utf-8') current_vals = {"gpu_mem_alloc": gpu_used_mem, "gpu_graphics_util": int(gpu_util['graphics']), "gpu_mem_util": gpu_util['memory'], "time": time.time()} return current_vals def loop(self, interval): i = 0 while True: self.get_vals() self.log["time"].append(time.time()) self.log["gpu_util"].append(self.current_vals["gpu_graphics_util"]) if self.count is not None: i += 1 if i == self.count: exit(0) time.sleep(self.interval) def start(self, interval=1.): self.interval = interval self.start_time = time.time() self.log = {"time": [], "gpu_util": []} if self.interval is not None: thread = threading.Thread(target=self.loop) thread.daemon = True thread.start() -class DummyLogger(): - def __init__(self): - pass - def info(self, *args): - print(*args) - return None - class CombinedLogger(object): """Combine console and tensorboard logger and record system metrics. """ - def __init__(self, name, log_dir, server_env=True, fold="", sysmetrics_interval=-1): - self.pylogger = DummyLogger()#logging.getLogger(name) + def __init__(self, name, log_dir, server_env=True, fold="", sysmetrics_interval=2): + self.pylogger = logging.getLogger(name) self.tboard = SummaryWriter(log_dir=log_dir) self.times = {} self.fold = fold - # self.pylogger.setLevel(logging.DEBUG) - # self.log_file = os.path.join(log_dir, 'exec.log') - # self.pylogger.addHandler(logging.FileHandler(self.log_file)) - # if not server_env: - # self.pylogger.addHandler(ColorHandler()) - # else: - # self.pylogger.addHandler(logging.StreamHandler()) - # self.pylogger.propagate = False + self.pylogger.setLevel(logging.DEBUG) + self.log_file = os.path.join(log_dir, 'exec.log') + self.pylogger.addHandler(logging.FileHandler(self.log_file)) + if not server_env: + self.pylogger.addHandler(ColorHandler()) + else: + self.pylogger.addHandler(logging.StreamHandler()) + self.pylogger.propagate = False # monitor system metrics (cpu, mem, ...) if not server_env and sysmetrics_interval > 0: self.sysmetrics = pd.DataFrame( columns=["global_step", "rel_time", r"CPU (%)", "mem_used (GB)", r"mem_used (%)", r"swap_used (GB)", r"gpu_utilization (%)"], dtype="float16") for device in range(torch.cuda.device_count()): self.sysmetrics[ "mem_allocd (GB) by torch on {:10s}".format(torch.cuda.get_device_name(device))] = np.nan self.sysmetrics[ "mem_cached (GB) by torch on {:10s}".format(torch.cuda.get_device_name(device))] = np.nan self.sysmetrics_start(sysmetrics_interval) pass else: print("NOT logging sysmetrics") def __getattr__(self, attr): """delegate all undefined method requests to objects of this class in order pylogger, tboard (first find first serve). E.g., combinedlogger.add_scalars(...) should trigger self.tboard.add_scalars(...) """ for obj in [self.pylogger, self.tboard]: if attr in dir(obj): return getattr(obj, attr) print("logger attr not found") #raise AttributeError("CombinedLogger has no attribute {}".format(attr)) def time(self, name, toggle=None): """record time-spans as with a stopwatch. :param name: :param toggle: True^=On: start time recording, False^=Off: halt rec. if None determine from current status. :return: either start-time or last recorded interval """ if toggle is None: if name in self.times.keys(): toggle = not self.times[name]["toggle"] else: toggle = True if toggle: if not name in self.times.keys(): self.times[name] = {"total": 0, "last": 0} elif self.times[name]["toggle"] == toggle: self.info("restarting running stopwatch") self.times[name]["last"] = time.time() self.times[name]["toggle"] = toggle return time.time() else: if toggle == self.times[name]["toggle"]: self.info("WARNING: tried to stop stopped stop watch: {}.".format(name)) self.times[name]["last"] = time.time() - self.times[name]["last"] self.times[name]["total"] += self.times[name]["last"] self.times[name]["toggle"] = toggle return self.times[name]["last"] def get_time(self, name=None, kind="total", format=None, reset=False): """ :param name: :param kind: 'total' or 'last' :param format: None for float, "hms"/"ms" for (hours), mins, secs as string :param reset: reset time after retrieving :return: """ if name is None: times = self.times if reset: self.reset_time() return times else: if self.times[name]["toggle"]: self.time(name, toggle=False) time = self.times[name][kind] if format == "hms": m, s = divmod(time, 60) h, m = divmod(m, 60) time = "{:d}h:{:02d}m:{:02d}s".format(int(h), int(m), int(s)) elif format == "ms": m, s = divmod(time, 60) time = "{:02d}m:{:02d}s".format(int(m), int(s)) if reset: self.reset_time(name) return time def reset_time(self, name=None): if name is None: self.times = {} else: del self.times[name] def sysmetrics_update(self, global_step=None): if global_step is None: global_step = time.strftime("%x_%X") mem = psutil.virtual_memory() mem_used = (mem.total - mem.available) gpu_vals = self.gpu_logger.get_vals() rel_time = time.time() - self.sysmetrics_start_time self.sysmetrics.loc[len(self.sysmetrics)] = [global_step, rel_time, psutil.cpu_percent(), mem_used / 1024 ** 3, mem_used / mem.total * 100, psutil.swap_memory().used / 1024 ** 3, int(gpu_vals['gpu_graphics_util']), *[torch.cuda.memory_allocated(d) / 1024 ** 3 for d in range(torch.cuda.device_count())], *[torch.cuda.memory_cached(d) / 1024 ** 3 for d in range(torch.cuda.device_count())] ] return self.sysmetrics.loc[len(self.sysmetrics) - 1].to_dict() def sysmetrics2tboard(self, metrics=None, global_step=None, suptitle=None): tag = "per_time" if metrics is None: metrics = self.sysmetrics_update(global_step=global_step) tag = "per_epoch" if suptitle is not None: suptitle = str(suptitle) elif self.fold != "": suptitle = "Fold_" + str(self.fold) if suptitle is not None: self.tboard.add_scalars(suptitle + "/System_Metrics/" + tag, {k: v for (k, v) in metrics.items() if (k != "global_step" and k != "rel_time")}, global_step) def sysmetrics_loop(self): try: os.nice(-19) self.info("Logging system metrics with superior process priority.") except: self.info("Logging system metrics WITHOUT superior process priority.") while True: metrics = self.sysmetrics_update() self.sysmetrics2tboard(metrics, global_step=metrics["rel_time"]) # print("thread alive", self.thread.is_alive()) time.sleep(self.sysmetrics_interval) def sysmetrics_start(self, interval): if interval is not None and interval > 0: self.sysmetrics_interval = interval self.gpu_logger = Nvidia_GPU_Logger() self.sysmetrics_start_time = time.time() self.thread = threading.Thread(target=self.sysmetrics_loop) self.thread.daemon = True self.thread.start() def sysmetrics_save(self, out_file): self.sysmetrics.to_pickle(out_file) def metrics2tboard(self, metrics, global_step=None, suptitle=None): """ :param metrics: {'train': dataframe, 'val':df}, df as produced in evaluator.py.evaluate_predictions """ # print("metrics", metrics) if global_step is None: global_step = len(metrics['train'][list(metrics['train'].keys())[0]]) - 1 if suptitle is not None: suptitle = str(suptitle) else: suptitle = "Fold_" + str(self.fold) for key in ['train', 'val']: # series = {k:np.array(v[-1]) for (k,v) in metrics[key].items() if not np.isnan(v[-1]) and not 'Bin_Stats' in k} loss_series = {} unc_series = {} bin_stat_series = {} mon_met_series = {} for tag, val in metrics[key].items(): val = val[-1] # maybe remove list wrapping, recording in evaluator? if 'bin_stats' in tag.lower() and not np.isnan(val): bin_stat_series["{}".format(tag.split("/")[-1])] = val elif 'uncertainty' in tag.lower() and not np.isnan(val): unc_series["{}".format(tag)] = val elif 'loss' in tag.lower() and not np.isnan(val): loss_series["{}".format(tag)] = val elif not np.isnan(val): mon_met_series["{}".format(tag)] = val self.tboard.add_scalars(suptitle + "/Binary_Statistics/{}".format(key), bin_stat_series, global_step) self.tboard.add_scalars(suptitle + "/Uncertainties/{}".format(key), unc_series, global_step) self.tboard.add_scalars(suptitle + "/Losses/{}".format(key), loss_series, global_step) self.tboard.add_scalars(suptitle + "/Monitor_Metrics/{}".format(key), mon_met_series, global_step) self.tboard.add_scalars(suptitle + "/Learning_Rate", metrics["lr"], global_step) return def batchImgs2tboard(self, batch, results_dict, cmap, boxtype2color, img_bg=False, global_step=None): raise NotImplementedError("not up-to-date, problem with importing plotting-file, torchvision dependency.") if len(batch["seg"].shape) == 5: # 3D imgs slice_ix = np.random.randint(batch["seg"].shape[-1]) seg_gt = plg.to_rgb(batch['seg'][:, 0, :, :, slice_ix], cmap) seg_pred = plg.to_rgb(results_dict['seg_preds'][:, 0, :, :, slice_ix], cmap) mod_img = plg.mod_to_rgb(batch["data"][:, 0, :, :, slice_ix]) if img_bg else None elif len(batch["seg"].shape) == 4: seg_gt = plg.to_rgb(batch['seg'][:, 0, :, :], cmap) seg_pred = plg.to_rgb(results_dict['seg_preds'][:, 0, :, :], cmap) mod_img = plg.mod_to_rgb(batch["data"][:, 0]) if img_bg else None else: raise Exception("batch content has wrong format: {}".format(batch["seg"].shape)) # from here on only works in 2D seg_gt = np.transpose(seg_gt, axes=(0, 3, 1, 2)) # previous shp: b,x,y,c seg_pred = np.transpose(seg_pred, axes=(0, 3, 1, 2)) seg = np.concatenate((seg_gt, seg_pred), axis=0) # todo replace torchvision (tv) dependency seg = tv.utils.make_grid(torch.from_numpy(seg), nrow=2) self.tboard.add_image("Batch seg, 1st col: gt, 2nd: pred.", seg, global_step=global_step) if img_bg: bg_img = np.transpose(mod_img, axes=(0, 3, 1, 2)) else: bg_img = seg_gt box_imgs = plg.draw_boxes_into_batch(bg_img, results_dict["boxes"], boxtype2color) box_imgs = tv.utils.make_grid(torch.from_numpy(box_imgs), nrow=4) self.tboard.add_image("Batch bboxes", box_imgs, global_step=global_step) return def __del__(self): # otherwise might produce multiple prints e.g. in ipython console - # for hdlr in self.pylogger.handlers: - # hdlr.close() - # #self.pylogger.handlers = [] - # del self.pylogger + for hdlr in self.pylogger.handlers: + hdlr.close() + self.pylogger.handlers = [] + del self.pylogger self.tboard.close() -def get_logger(exp_dir, server_env=False, sysmetrics_interval=-1): +def get_logger(exp_dir, server_env=False, sysmetrics_interval=2): log_dir = os.path.join(exp_dir, "logs") logger = CombinedLogger('Reg R-CNN', os.path.join(log_dir, "tboard"), server_env=server_env, sysmetrics_interval=sysmetrics_interval) print("logging to {}".format(logger.log_file)) return logger def prep_exp(dataset_path, exp_path, server_env, use_stored_settings=True, is_training=True): """ I/O handling, creating of experiment folder structure. Also creates a snapshot of configs/model scripts and copies them to the exp_dir. This way the exp_dir contains all info needed to conduct an experiment, independent to changes in actual source code. Thus, training/inference of this experiment can be started at anytime. Therefore, the model script is copied back to the source code dir as tmp_model (tmp_backbone). Provides robust structure for cloud deployment. :param dataset_path: path to source code for specific data set. (e.g. medicaldetectiontoolkit/lidc_exp) :param exp_path: path to experiment directory. :param server_env: boolean flag. pass to configs script for cloud deployment. :param use_stored_settings: boolean flag. When starting training: If True, starts training from snapshot in existing experiment directory, else creates experiment directory on the fly using configs/model scripts from source code. :param is_training: boolean flag. distinguishes train vs. inference mode. :return: configs object. """ if is_training: if use_stored_settings: cf_file = import_module('cf', os.path.join(exp_path, 'configs.py')) cf = cf_file.Configs(server_env) # in this mode, previously saved model and backbone need to be found in exp dir. if not os.path.isfile(os.path.join(exp_path, 'model.py')) or \ not os.path.isfile(os.path.join(exp_path, 'backbone.py')): raise Exception( "Selected use_stored_settings option but no model and/or backbone source files exist in exp dir.") cf.model_path = os.path.join(exp_path, 'model.py') cf.backbone_path = os.path.join(exp_path, 'backbone.py') else: # this case overwrites settings files in exp dir, i.e., default_configs, configs, backbone, model if not os.path.exists(exp_path): os.mkdir(exp_path) # run training with source code info and copy snapshot of model to exp_dir for later testing (overwrite scripts if exp_dir already exists.) subprocess.call('cp {} {}'.format('default_configs.py', os.path.join(exp_path, 'default_configs.py')), shell=True) subprocess.call( 'cp {} {}'.format(os.path.join(dataset_path, 'configs.py'), os.path.join(exp_path, 'configs.py')), shell=True) cf_file = import_module('cf_file', os.path.join(dataset_path, 'configs.py')) cf = cf_file.Configs(server_env) subprocess.call('cp {} {}'.format(cf.model_path, os.path.join(exp_path, 'model.py')), shell=True) subprocess.call('cp {} {}'.format(cf.backbone_path, os.path.join(exp_path, 'backbone.py')), shell=True) if os.path.isfile(os.path.join(exp_path, "fold_ids.pickle")): subprocess.call('rm {}'.format(os.path.join(exp_path, "fold_ids.pickle")), shell=True) else: # testing, use model and backbone stored in exp dir. cf_file = import_module('cf', os.path.join(exp_path, 'configs.py')) cf = cf_file.Configs(server_env) cf.model_path = os.path.join(exp_path, 'model.py') cf.backbone_path = os.path.join(exp_path, 'backbone.py') cf.exp_dir = exp_path cf.test_dir = os.path.join(cf.exp_dir, 'test') cf.plot_dir = os.path.join(cf.exp_dir, 'plots') if not os.path.exists(cf.test_dir): os.mkdir(cf.test_dir) if not os.path.exists(cf.plot_dir): os.mkdir(cf.plot_dir) cf.experiment_name = exp_path.split("/")[-1] cf.dataset_name = dataset_path cf.server_env = server_env cf.created_fold_id_pickle = False return cf class ModelSelector: ''' saves a checkpoint after each epoch as 'last_state' (can be loaded to continue interrupted training). saves the top-k (k=cf.save_n_models) ranked epochs. In inference, predictions of multiple epochs can be ensembled to improve performance. ''' def __init__(self, cf, logger): self.cf = cf self.saved_epochs = [-1] * cf.save_n_models self.logger = logger def run_model_selection(self, net, optimizer, monitor_metrics, epoch): """rank epoch via weighted mean from self.cf.model_selection_criteria: {criterion : weight} :param net: :param optimizer: :param monitor_metrics: :param epoch: :return: """ crita = self.cf.model_selection_criteria # shorter alias non_nan_scores = {} for criterion in crita.keys(): # exclude first entry bc its dummy None entry non_nan_scores[criterion] = [0 if (ii is None or np.isnan(ii)) else ii for ii in monitor_metrics['val'][criterion]][1:] n_epochs = len(non_nan_scores[criterion]) epochs_scores = [] for e_ix in range(n_epochs): epochs_scores.append(np.sum([weight * non_nan_scores[criterion][e_ix] for criterion, weight in crita.items()]) / len(crita.keys())) # ranking of epochs according to model_selection_criterion epoch_ranking = np.argsort(epochs_scores)[::-1] + 1 # epochs start at 1 # if set in configs, epochs < min_save_thresh are discarded from saving process. epoch_ranking = epoch_ranking[epoch_ranking >= self.cf.min_save_thresh] # check if current epoch is among the top-k epchs. if epoch in epoch_ranking[:self.cf.save_n_models]: if self.cf.server_env: IO_safe(torch.save, net.state_dict(), os.path.join(self.cf.fold_dir, '{}_best_params.pth'.format(epoch))) # save epoch_ranking to keep info for inference. IO_safe(np.save, os.path.join(self.cf.fold_dir, 'epoch_ranking'), epoch_ranking[:self.cf.save_n_models]) else: torch.save(net.state_dict(), os.path.join(self.cf.fold_dir, '{}_best_params.pth'.format(epoch))) np.save(os.path.join(self.cf.fold_dir, 'epoch_ranking'), epoch_ranking[:self.cf.save_n_models]) self.logger.info( "saving current epoch {} at rank {}".format(epoch, np.argwhere(epoch_ranking == epoch))) # delete params of the epoch that just fell out of the top-k epochs. for se in [int(ii.split('_')[0]) for ii in os.listdir(self.cf.fold_dir) if 'best_params' in ii]: if se in epoch_ranking[self.cf.save_n_models:]: subprocess.call('rm {}'.format(os.path.join(self.cf.fold_dir, '{}_best_params.pth'.format(se))), shell=True) self.logger.info('deleting epoch {} at rank {}'.format(se, np.argwhere(epoch_ranking == se))) state = { 'epoch': epoch, 'state_dict': net.state_dict(), 'optimizer': optimizer.state_dict(), } if self.cf.server_env: IO_safe(torch.save, state, os.path.join(self.cf.fold_dir, 'last_state.pth')) else: torch.save(state, os.path.join(self.cf.fold_dir, 'last_state.pth')) def load_checkpoint(checkpoint_path, net, optimizer): checkpoint = torch.load(checkpoint_path) net.load_state_dict(checkpoint['state_dict']) optimizer.load_state_dict(checkpoint['optimizer']) return checkpoint['epoch'] def prepare_monitoring(cf): """ creates dictionaries, where train/val metrics are stored. """ metrics = {} # first entry for loss dict accounts for epoch starting at 1. metrics['train'] = OrderedDict() # [(l_name, [np.nan]) for l_name in cf.losses_to_monitor] ) metrics['val'] = OrderedDict() # [(l_name, [np.nan]) for l_name in cf.losses_to_monitor] ) metric_classes = [] if 'rois' in cf.report_score_level: metric_classes.extend([v for k, v in cf.class_dict.items()]) if hasattr(cf, "eval_bins_separately") and cf.eval_bins_separately: metric_classes.extend([v for k, v in cf.bin_dict.items()]) if 'patient' in cf.report_score_level: metric_classes.extend(['patient_' + cf.class_dict[cf.patient_class_of_interest]]) if hasattr(cf, "eval_bins_separately") and cf.eval_bins_separately: metric_classes.extend(['patient_' + cf.bin_dict[cf.patient_bin_of_interest]]) for cl in metric_classes: for m in cf.metrics: metrics['train'][cl + '_' + m] = [np.nan] metrics['val'][cl + '_' + m] = [np.nan] return metrics class _AnsiColorizer(object): """ A colorizer is an object that loosely wraps around a stream, allowing callers to write text to the stream in a particular color. Colorizer classes must implement C{supported()} and C{write(text, color)}. """ _colors = dict(black=30, red=31, green=32, yellow=33, blue=34, magenta=35, cyan=36, white=37, default=39) def __init__(self, stream): self.stream = stream @classmethod def supported(cls, stream=sys.stdout): """ A class method that returns True if the current platform supports coloring terminal output using this method. Returns False otherwise. """ if not stream.isatty(): return False # auto color only on TTYs try: import curses except ImportError: return False else: try: try: return curses.tigetnum("colors") > 2 except curses.error: curses.setupterm() return curses.tigetnum("colors") > 2 except: raise # guess false in case of error return False def write(self, text, color): """ Write the given text to the stream in the given color. @param text: Text to be written to the stream. @param color: A string label for a color. e.g. 'red', 'white'. """ color = self._colors[color] self.stream.write('\x1b[%sm%s\x1b[0m' % (color, text)) class ColorHandler(logging.StreamHandler): def __init__(self, stream=sys.stdout): super(ColorHandler, self).__init__(_AnsiColorizer(stream)) def emit(self, record): msg_colors = { logging.DEBUG: "green", logging.INFO: "default", logging.WARNING: "red", logging.ERROR: "red" } color = msg_colors.get(record.levelno, "blue") self.stream.write(record.msg + "\n", color)