def main(): script_dir = os.path.dirname(__file__) module_path = os.path.abspath(os.path.join(script_dir, '..', '..')) global msglogger # Parse arguments args = parser.get_parser().parse_args() if args.epochs is None: args.epochs = 90 if not os.path.exists(args.output_dir): os.makedirs(args.output_dir) msglogger = apputils.config_pylogger( os.path.join(script_dir, 'logging.conf'), args.name, args.output_dir) # Log various details about the execution environment. It is sometimes useful # to refer to past experiment executions and this information may be useful. apputils.log_execution_env_state( filter(None, [args.compress, args.qe_stats_file ]), # remove both None and empty strings msglogger.logdir, gitroot=module_path) msglogger.debug("Distiller: %s", distiller.__version__) if args.evaluate: args.deterministic = True if args.deterministic: distiller.set_deterministic( args.seed) # For experiment reproducability else: if args.seed is not None: distiller.set_seed(args.seed) # Turn on CUDNN benchmark mode for best performance. This is usually "safe" for image # classification models, as the input sizes don't change during the run # See here: https://discuss.pytorch.org/t/what-does-torch-backends-cudnn-benchmark-do/5936/3 cudnn.benchmark = True start_epoch = 0 ending_epoch = args.epochs perf_scores_history = [] if args.cpu or not torch.cuda.is_available(): # Set GPU index to -1 if using CPU args.device = 'cpu' args.gpus = -1 else: args.device = 'cuda' if args.gpus is not None: try: args.gpus = [int(s) for s in args.gpus.split(',')] except ValueError: raise ValueError( 'ERROR: Argument --gpus must be a comma-separated list of integers only' ) available_gpus = torch.cuda.device_count() for dev_id in args.gpus: if dev_id >= available_gpus: raise ValueError( 'ERROR: GPU device ID {0} requested, but only {1} devices available' .format(dev_id, available_gpus)) # Set default device in case the first one on the list != 0 torch.cuda.set_device(args.gpus[0]) # Infer the dataset from the model name args.dataset = distiller.apputils.classification_dataset_str_from_arch( args.arch) args.num_classes = distiller.apputils.classification_num_classes( args.dataset) if args.earlyexit_thresholds: args.num_exits = len(args.earlyexit_thresholds) + 1 args.loss_exits = [0] * args.num_exits args.losses_exits = [] args.exiterrors = [] # Create the model model = create_model(args.pretrained, args.dataset, args.arch, parallel=not args.load_serialized, device_ids=args.gpus) compression_scheduler = None # Create a couple of logging backends. TensorBoardLogger writes log files in a format # that can be read by Google's Tensor Board. PythonLogger writes to the Python logger. tflogger = TensorBoardLogger(msglogger.logdir) pylogger = PythonLogger(msglogger) # capture thresholds for early-exit training if args.earlyexit_thresholds: msglogger.info('=> using early-exit threshold values of %s', args.earlyexit_thresholds) # TODO(barrh): args.deprecated_resume is deprecated since v0.3.1 if args.deprecated_resume: msglogger.warning( 'The "--resume" flag is deprecated. Please use "--resume-from=YOUR_PATH" instead.' ) if not args.reset_optimizer: msglogger.warning( 'If you wish to also reset the optimizer, call with: --reset-optimizer' ) args.reset_optimizer = True args.resumed_checkpoint_path = args.deprecated_resume # We can optionally resume from a checkpoint optimizer = None if args.resumed_checkpoint_path: model, compression_scheduler, optimizer, start_epoch = apputils.load_checkpoint( model, args.resumed_checkpoint_path, model_device=args.device) elif args.load_model_path: model = apputils.load_lean_checkpoint(model, args.load_model_path, model_device=args.device) if args.reset_optimizer: start_epoch = 0 if optimizer is not None: optimizer = None msglogger.info( '\nreset_optimizer flag set: Overriding resumed optimizer and resetting epoch count to 0' ) # Define loss function (criterion) criterion = nn.CrossEntropyLoss().to(args.device) if optimizer is None: optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay, nesterov=args.nesterov) msglogger.info('Optimizer Type: %s', type(optimizer)) msglogger.info('Optimizer Args: %s', optimizer.defaults) if args.AMC: return automated_deep_compression(model, criterion, optimizer, pylogger, args) if args.greedy: return greedy(model, criterion, optimizer, pylogger, args) # This sample application can be invoked to produce various summary reports. if args.summary: for summary in args.summary: distiller.model_summary(model, summary, args.dataset) return if args.export_onnx is not None: return distiller.export_img_classifier_to_onnx(model, os.path.join( msglogger.logdir, args.export_onnx), args.dataset, add_softmax=True, verbose=False) if args.qe_calibration: return acts_quant_stats_collection(model, criterion, pylogger, args) if args.activation_histograms: return acts_histogram_collection(model, criterion, pylogger, args) activations_collectors = create_activation_stats_collectors( model, *args.activation_stats) # Load the datasets: the dataset to load is inferred from the model name passed # in args.arch. The default dataset is ImageNet, but if args.arch contains the # substring "_cifar", then cifar10 is used. train_loader, val_loader, test_loader, _ = load_data(args) msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d', len(train_loader.sampler), len(val_loader.sampler), len(test_loader.sampler)) if args.sensitivity is not None: sensitivities = np.arange(args.sensitivity_range[0], args.sensitivity_range[1], args.sensitivity_range[2]) return sensitivity_analysis(model, criterion, test_loader, pylogger, args, sensitivities) if args.evaluate: return evaluate_model(model, criterion, test_loader, pylogger, activations_collectors, args, compression_scheduler) if args.compress: # The main use-case for this sample application is CNN compression. Compression # requires a compression schedule configuration file in YAML. compression_scheduler = distiller.file_config( model, optimizer, args.compress, compression_scheduler, (start_epoch - 1) if args.resumed_checkpoint_path else None) # Model is re-transferred to GPU in case parameters were added (e.g. PACTQuantizer) model.to(args.device) elif compression_scheduler is None: compression_scheduler = distiller.CompressionScheduler(model) if args.thinnify: #zeros_mask_dict = distiller.create_model_masks_dict(model) assert args.resumed_checkpoint_path is not None, \ "You must use --resume-from to provide a checkpoint file to thinnify" distiller.remove_filters(model, compression_scheduler.zeros_mask_dict, args.arch, args.dataset, optimizer=None) apputils.save_checkpoint(0, args.arch, model, optimizer=None, scheduler=compression_scheduler, name="{}_thinned".format( args.resumed_checkpoint_path.replace( ".pth.tar", "")), dir=msglogger.logdir) print( "Note: your model may have collapsed to random inference, so you may want to fine-tune" ) return args.kd_policy = None if args.kd_teacher: teacher = create_model(args.kd_pretrained, args.dataset, args.kd_teacher, device_ids=args.gpus) if args.kd_resume: teacher = apputils.load_lean_checkpoint(teacher, args.kd_resume) dlw = distiller.DistillationLossWeights(args.kd_distill_wt, args.kd_student_wt, args.kd_teacher_wt) args.kd_policy = distiller.KnowledgeDistillationPolicy( model, teacher, args.kd_temp, dlw) compression_scheduler.add_policy(args.kd_policy, starting_epoch=args.kd_start_epoch, ending_epoch=args.epochs, frequency=1) msglogger.info('\nStudent-Teacher knowledge distillation enabled:') msglogger.info('\tTeacher Model: %s', args.kd_teacher) msglogger.info('\tTemperature: %s', args.kd_temp) msglogger.info('\tLoss Weights (distillation | student | teacher): %s', ' | '.join(['{:.2f}'.format(val) for val in dlw])) msglogger.info('\tStarting from Epoch: %s', args.kd_start_epoch) if start_epoch >= ending_epoch: msglogger.error( 'epoch count is too low, starting epoch is {} but total epochs set to {}' .format(start_epoch, ending_epoch)) raise ValueError('Epochs parameter is too low. Nothing to do.') for epoch in range(start_epoch, ending_epoch): # This is the main training loop. msglogger.info('\n') if compression_scheduler: compression_scheduler.on_epoch_begin(epoch) # Train for one epoch with collectors_context(activations_collectors["train"]) as collectors: train(train_loader, model, criterion, optimizer, epoch, compression_scheduler, loggers=[tflogger, pylogger], args=args) if args.print_sparsity: distiller.log_weights_sparsity(model, epoch, loggers=[tflogger, pylogger]) distiller.log_activation_statsitics( epoch, "train", loggers=[tflogger], collector=collectors["sparsity"]) if args.masks_sparsity: msglogger.info( distiller.masks_sparsity_tbl_summary( model, compression_scheduler)) # evaluate on validation set with collectors_context(activations_collectors["valid"]) as collectors: top1, top5, vloss = validate(val_loader, model, criterion, [pylogger], args, epoch) distiller.log_activation_statsitics( epoch, "valid", loggers=[tflogger], collector=collectors["sparsity"]) save_collectors_data(collectors, msglogger.logdir) stats = ('Performance/Validation/', OrderedDict([('Loss', vloss), ('Top1', top1), ('Top5', top5)])) distiller.log_training_progress(stats, None, epoch, steps_completed=0, total_steps=1, log_freq=1, loggers=[tflogger]) if compression_scheduler: compression_scheduler.on_epoch_end(epoch, optimizer, metrics={ 'min': vloss, 'max': top1 }) # Update the list of top scores achieved so far, and save the checkpoint update_training_scores_history(perf_scores_history, model, top1, top5, epoch, args.num_best_scores) is_best = epoch == perf_scores_history[0].epoch checkpoint_extras = { 'current_top1': top1, 'best_top1': perf_scores_history[0].top1, 'best_epoch': perf_scores_history[0].epoch } apputils.save_checkpoint(epoch, args.arch, model, optimizer=optimizer, scheduler=compression_scheduler, extras=checkpoint_extras, is_best=is_best, name=args.name, dir=msglogger.logdir) # ADDED: look at learning rate msglogger.info(f'\n==> Learning rate: {get_lr(optimizer)}') # Finally run results on the test set test(test_loader, model, criterion, [pylogger], activations_collectors, args=args)
def main(): global msglogger check_pytorch_version() args = parser.parse_args() if not os.path.exists(args.output_dir): os.makedirs(args.output_dir) msglogger = apputils.config_pylogger(os.path.join(script_dir, 'logging.conf'), args.name, args.output_dir) # Log various details about the execution environment. It is sometimes useful # to refer to past experiment executions and this information may be useful. apputils.log_execution_env_state(sys.argv, gitroot=module_path) msglogger.debug("Distiller: %s", distiller.__version__) start_epoch = 0 best_top1 = 0 if args.deterministic: # Experiment reproducibility is sometimes important. Pete Warden expounded about this # in his blog: https://petewarden.com/2018/03/19/the-machine-learning-reproducibility-crisis/ # In Pytorch, support for deterministic execution is still a bit clunky. if args.workers > 1: msglogger.error('ERROR: Setting --deterministic requires setting --workers/-j to 0 or 1') exit(1) # Use a well-known seed, for repeatability of experiments torch.manual_seed(0) random.seed(0) np.random.seed(0) cudnn.deterministic = True else: # This issue: https://github.com/pytorch/pytorch/issues/3659 # Implies that cudnn.benchmark should respect cudnn.deterministic, but empirically we see that # results are not re-produced when benchmark is set. So enabling only if deterministic mode disabled. cudnn.benchmark = True if args.gpus is not None: try: args.gpus = [int(s) for s in args.gpus.split(',')] except ValueError: msglogger.error('ERROR: Argument --gpus must be a comma-separated list of integers only') exit(1) available_gpus = torch.cuda.device_count() for dev_id in args.gpus: if dev_id >= available_gpus: msglogger.error('ERROR: GPU device ID {0} requested, but only {1} devices available' .format(dev_id, available_gpus)) exit(1) # Set default device in case the first one on the list != 0 torch.cuda.set_device(args.gpus[0]) # Infer the dataset from the model name if args.dataset.lower() == 'none': args.dataset = 'cifar10' if 'cifar' in args.arch else 'imagenet' # Create the model png_summary = args.summary is not None and args.summary.startswith('png') is_parallel = not png_summary and args.summary != 'compute' # For PNG summary, parallel graphs are illegible model = create_model(args.pretrained, args.dataset, args.arch, parallel=is_parallel, device_ids=args.gpus) #If the teacher is not None, create teacher model and load it if args.teacher_arch.lower() != 'none': try: teacher_model = create_model(True, args.dataset, args.teacher_arch, parallel=is_parallel, device_ids=args.gpus) print('The teacher model has been loaded from the torchvision pretrained model list. User input ignored') except: if args.teacher_weights.lower() == 'none': raise ValueError('If you pass a teacher model you also have to pass the path of its weights') teacher_model = create_model(False, args.dataset, args.teacher_arch, parallel=is_parallel, device_ids=args.gpus) try: teacher_weights = torch.load(args.teacher_weights) if 'state_dict' in teacher_weights: teacher_weights = teacher_weights['state_dict'] teacher_model.load_state_dict(teacher_weights) except: raise ValueError('Unable to load teacher weights. Loading path {} resulted in error'.format(args.teacher_weights)) else: teacher_model = None compression_scheduler = None # Create a couple of logging backends. TensorBoardLogger writes log files in a format # that can be read by Google's Tensor Board. PythonLogger writes to the Python logger. tflogger = TensorBoardLogger(msglogger.logdir) pylogger = PythonLogger(msglogger) # We can optionally resume from a checkpoint if args.resume: model, compression_scheduler, start_epoch = apputils.load_checkpoint( model, chkpt_file=args.resume) # Define loss function (criterion) and optimizer criterion = nn.CrossEntropyLoss().cuda() optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay) msglogger.info('Optimizer Type: %s', type(optimizer)) msglogger.info('Optimizer Args: %s', optimizer.defaults) # This sample application can be invoked to produce various summary reports. if args.summary: which_summary = args.summary if which_summary.startswith('png'): apputils.draw_img_classifier_to_file(model, 'model.png', args.dataset, which_summary == 'png_w_params') else: distiller.model_summary(model, which_summary, args.dataset) exit() # Load the datasets: the dataset to load is inferred from the model name passed # in args.arch. The default dataset is ImageNet, but if args.arch contains the # substring "_cifar", then cifar10 is used. train_loader, val_loader, test_loader, _ = apputils.load_data( args.dataset, os.path.expanduser(args.data), args.batch_size, args.workers, args.validation_size, args.deterministic) msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d', len(train_loader.sampler), len(val_loader.sampler), len(test_loader.sampler)) if teacher_model is not None: t1, t5, _ = _validate(test_loader, teacher_model, criterion, None, len(test_loader) + 10) print('Teacher model passed in input has Top-1 of {} and Top-5 of {}'.format(t1, t5)) activations_sparsity = None if args.activation_stats: # If your model has ReLU layers, then those layers have sparse activations. # ActivationSparsityCollector will collect information about this sparsity. # WARNING! Enabling activation sparsity collection will significantly slow down training! activations_sparsity = ActivationSparsityCollector(model) if args.sensitivity is not None: # This sample application can be invoked to execute Sensitivity Analysis on your # model. The ouptut is saved to CSV and PNG. msglogger.info("Running sensitivity tests") test_fnc = partial(test, test_loader=test_loader, criterion=criterion, loggers=[pylogger], print_freq=args.print_freq) which_params = [param_name for param_name, _ in model.named_parameters()] sensitivity = distiller.perform_sensitivity_analysis(model, net_params=which_params, sparsities=np.arange(0.0, 0.50, 0.05) if args.sensitivity == 'filter' else np.arange(0.0, 0.95, 0.05), test_func=test_fnc, group=args.sensitivity) distiller.sensitivities_to_png(sensitivity, 'sensitivity.png') distiller.sensitivities_to_csv(sensitivity, 'sensitivity.csv') exit() if args.evaluate: # This sample application can be invoked to evaluate the accuracy of your model on # the test dataset. # You can optionally quantize the model to 8-bit integer before evaluation. # For example: # python3 compress_classifier.py --arch resnet20_cifar ../data.cifar10 -p=50 --resume=checkpoint.pth.tar --evaluate if args.quantize: model.cpu() quantizer = quantization.SymmetricLinearQuantizer(model, 8, 8) quantizer.prepare_model() model.cuda() top1, _, _ = test(test_loader, model, criterion, [pylogger], args.print_freq) if args.quantize: checkpoint_name = 'quantized' apputils.save_checkpoint(0, args.arch, model, optimizer=None, best_top1=top1, name='_'.split(args.name, checkpoint_name) if args.name else checkpoint_name, dir=msglogger.logdir) exit() if args.compress: # The main use-case for this sample application is CNN compression. Compression # requires a compression schedule configuration file in YAML. compression_scheduler = distiller.file_config(model, optimizer, args.compress) for epoch in range(start_epoch, start_epoch + args.epochs): # This is the main training loop. msglogger.info('\n') if compression_scheduler: compression_scheduler.on_epoch_begin(epoch) # Train for one epoch if epoch >= args.start_distillation_from_epoch: train(train_loader, model, criterion, optimizer, epoch, compression_scheduler, loggers=[tflogger, pylogger], print_freq=args.print_freq, log_params_hist=args.log_params_histograms, teacher_model=teacher_model, temperature_distillation=args.temp_distillation, weight_distillation_loss=args.weight_distillation_loss) else: train(train_loader, model, criterion, optimizer, epoch, compression_scheduler, loggers=[tflogger, pylogger], print_freq=args.print_freq, log_params_hist=args.log_params_histograms, teacher_model=None) distiller.log_weights_sparsity(model, epoch, loggers=[tflogger, pylogger]) if args.activation_stats: distiller.log_activation_sparsity(epoch, loggers=[tflogger, pylogger], collector=activations_sparsity) # evaluate on validation set top1, top5, vloss = validate(val_loader, model, criterion, [pylogger], args.print_freq, epoch) stats = ('Peformance/Validation/', OrderedDict([('Loss', vloss), ('Top1', top1), ('Top5', top5)])) distiller.log_training_progress(stats, None, epoch, steps_completed=0, total_steps=1, log_freq=1, loggers=[tflogger]) if compression_scheduler: compression_scheduler.on_epoch_end(epoch, optimizer) # remember best top1 and save checkpoint is_best = top1 > best_top1 best_top1 = max(top1, best_top1) apputils.save_checkpoint(epoch, args.arch, model, optimizer, compression_scheduler, best_top1, is_best, args.name, msglogger.logdir) # Finally run results on the test set test(test_loader, model, criterion, [pylogger], args.print_freq)
def main(): global msglogger script_dir = os.path.dirname(__file__) args = parse_args() # Distiller loggers msglogger = apputils.config_pylogger('logging.conf', args.name, output_dir=args.output_dir) tflogger = TensorBoardLogger(msglogger.logdir) # tflogger.log_gradients = True # pylogger = PythonLogger(msglogger) if args.seed is not None: msglogger.info("Using seed = {}".format(args.seed)) torch.manual_seed(args.seed) np.random.seed(seed=args.seed) args.qe_mode = str(args.qe_mode).split('.')[1] args.qe_clip_acts = str(args.qe_clip_acts).split('.')[1] apputils.log_execution_env_state(sys.argv) if args.gpus is not None: try: args.gpus = [int(s) for s in args.gpus.split(',')] except ValueError: msglogger.error( 'ERROR: Argument --gpus must be a comma-separated list of integers only' ) exit(1) if len(args.gpus) > 1: msglogger.error('ERROR: Only single GPU supported for NCF') exit(1) available_gpus = torch.cuda.device_count() for dev_id in args.gpus: if dev_id >= available_gpus: msglogger.error( 'ERROR: GPU device ID {0} requested, but only {1} devices available' .format(dev_id, available_gpus)) exit(1) # Set default device in case the first one on the list != 0 torch.cuda.set_device(args.gpus[0]) # Save configuration to file config = {k: v for k, v in args.__dict__.items()} config['timestamp'] = "{:.0f}".format(datetime.utcnow().timestamp()) config['local_timestamp'] = str(datetime.now()) run_dir = msglogger.logdir msglogger.info("Saving config and results to {}".format(run_dir)) if not os.path.exists(run_dir) and run_dir != '': os.makedirs(run_dir) utils.save_config(config, run_dir) # Check that GPUs are actually available use_cuda = not args.no_cuda and torch.cuda.is_available() t1 = time.time() # Load Data training = not (args.eval or args.qe_calibration or args.activation_histograms) msglogger.info('Loading data') if training: train_dataset = CFTrainDataset( os.path.join(args.data, TRAIN_RATINGS_FILENAME), args.negative_samples) train_dataloader = torch.utils.data.DataLoader( dataset=train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, pin_memory=True) nb_users, nb_items = train_dataset.nb_users, train_dataset.nb_items else: train_dataset = None train_dataloader = None nb_users, nb_items = (138493, 26744) test_ratings = load_test_ratings( os.path.join(args.data, TEST_RATINGS_FILENAME)) # noqa: E501 test_negs = load_test_negs(os.path.join(args.data, TEST_NEG_FILENAME)) msglogger.info( 'Load data done [%.1f s]. #user=%d, #item=%d, #train=%s, #test=%d' % (time.time() - t1, nb_users, nb_items, str(train_dataset.mat.nnz) if training else 'N/A', len(test_ratings))) # Create model model = NeuMF(nb_users, nb_items, mf_dim=args.factors, mf_reg=0., mlp_layer_sizes=args.layers, mlp_layer_regs=[0. for i in args.layers], split_final=args.split_final) if use_cuda: model = model.cuda() msglogger.info(model) msglogger.info("{} parameters".format(utils.count_parameters(model))) # Save model text description with open(os.path.join(run_dir, 'model.txt'), 'w') as file: file.write(str(model)) compression_scheduler = None start_epoch = 0 optimizer = None if args.load: if training: model, compression_scheduler, optimizer, start_epoch = apputils.load_checkpoint( model, args.load) if args.reset_optimizer: start_epoch = 0 optimizer = None else: model = apputils.load_lean_checkpoint(model, args.load) # Add loss to graph criterion = nn.BCEWithLogitsLoss() if use_cuda: criterion = criterion.cuda() if training and optimizer is None: optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate) msglogger.info('Optimizer Type: %s', type(optimizer)) msglogger.info('Optimizer Args: %s', optimizer.defaults) if args.compress: compression_scheduler = distiller.file_config(model, optimizer, args.compress) model.cuda() # Create files for tracking training valid_results_file = os.path.join(run_dir, 'valid_results.csv') if args.qe_calibration or args.activation_histograms: calib = { 'portion': args.qe_calibration, 'desc_str': 'quantization calibration stats', 'collect_func': partial(distiller.data_loggers.collect_quant_stats, inplace_runtime_check=True, disable_inplace_attrs=True) } hists = { 'portion': args.activation_histograms, 'desc_str': 'activation histograms', 'collect_func': partial(distiller.data_loggers.collect_histograms, activation_stats=None, nbins=2048, save_hist_imgs=True) } d = calib if args.qe_calibration else hists distiller.utils.assign_layer_fq_names(model) num_users = int(np.floor(len(test_ratings) * d['portion'])) msglogger.info( "Generating {} based on {:.1%} of the test-set ({} users)".format( d['desc_str'], d['portion'], num_users)) test_fn = partial(val_epoch, ratings=test_ratings, negs=test_negs, K=args.topk, use_cuda=use_cuda, processes=args.processes, num_users=num_users) d['collect_func'](model=model, test_fn=test_fn, save_dir=run_dir, classes=None) return 0 if args.eval: if args.quantize_eval and args.qe_calibration is None: model.cpu() quantizer = quantization.PostTrainLinearQuantizer.from_args( model, args) dummy_input = (torch.tensor([1]), torch.tensor([1]), torch.tensor([True], dtype=torch.bool)) quantizer.prepare_model(dummy_input) model.cuda() distiller.utils.assign_layer_fq_names(model) if args.eval_fp16: model = model.half() # Calculate initial Hit Ratio and NDCG begin = time.time() hits, ndcgs = val_epoch(model, test_ratings, test_negs, args.topk, use_cuda=use_cuda, processes=args.processes) val_time = time.time() - begin hit_rate = np.mean(hits) msglogger.info( 'Initial HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f}, val_time = {val_time:.2f}' .format(K=args.topk, hit_rate=hit_rate, ndcg=np.mean(ndcgs), val_time=val_time)) hit_rate = 0 if args.quantize_eval: checkpoint_name = 'quantized' apputils.save_checkpoint(0, 'NCF', model, optimizer=None, extras={'quantized_hr@10': hit_rate}, name='_'.join([args.name, 'quantized']) if args.name else checkpoint_name, dir=msglogger.logdir) return 0 total_samples = len(train_dataloader.sampler) steps_per_epoch = math.ceil(total_samples / args.batch_size) best_hit_rate = 0 best_epoch = 0 for epoch in range(start_epoch, args.epochs): msglogger.info('') model.train() losses = utils.AverageMeter() begin = time.time() if compression_scheduler: compression_scheduler.on_epoch_begin(epoch, optimizer) loader = tqdm.tqdm(train_dataloader) for batch_index, (user, item, label) in enumerate(loader): user = torch.autograd.Variable(user, requires_grad=False) item = torch.autograd.Variable(item, requires_grad=False) label = torch.autograd.Variable(label, requires_grad=False) if use_cuda: user = user.cuda(async=True) item = item.cuda(async=True) label = label.cuda(async=True) if compression_scheduler: compression_scheduler.on_minibatch_begin( epoch, batch_index, steps_per_epoch, optimizer) outputs = model(user, item, torch.tensor([False], dtype=torch.bool)) loss = criterion(outputs, label) if compression_scheduler: compression_scheduler.before_backward_pass( epoch, batch_index, steps_per_epoch, loss, optimizer, return_loss_components=False) losses.update(loss.data.item(), user.size(0)) optimizer.zero_grad() loss.backward() optimizer.step() if compression_scheduler: compression_scheduler.on_minibatch_end(epoch, batch_index, steps_per_epoch, optimizer) # Save stats to file description = ( 'Epoch {} Loss {loss.val:.4f} ({loss.avg:.4f})'.format( epoch, loss=losses)) loader.set_description(description) steps_completed = batch_index + 1 if steps_completed % args.log_freq == 0: stats_dict = OrderedDict() stats_dict['Loss'] = losses.avg stats = ('Performance/Training/', stats_dict) params = model.named_parameters( ) if args.log_params_histograms else None distiller.log_training_progress(stats, params, epoch, steps_completed, steps_per_epoch, args.log_freq, [tflogger]) tflogger.log_model_buffers(model, ['tracked_min', 'tracked_max'], 'Quant/Train/Acts/TrackedMinMax', epoch, steps_completed, steps_per_epoch, args.log_freq) train_time = time.time() - begin begin = time.time() hits, ndcgs = val_epoch(model, test_ratings, test_negs, args.topk, use_cuda=use_cuda, output=valid_results_file, epoch=epoch, processes=args.processes) val_time = time.time() - begin if compression_scheduler: compression_scheduler.on_epoch_end(epoch, optimizer) hit_rate = np.mean(hits) mean_ndcgs = np.mean(ndcgs) stats_dict = OrderedDict() stats_dict['HR@{0}'.format(args.topk)] = hit_rate stats_dict['NDCG@{0}'.format(args.topk)] = mean_ndcgs stats = ('Performance/Validation/', stats_dict) distiller.log_training_progress(stats, None, epoch, steps_completed=0, total_steps=1, log_freq=1, loggers=[tflogger]) msglogger.info( 'Epoch {epoch}: HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f}, AvgTrainLoss = {loss.avg:.4f}, ' 'train_time = {train_time:.2f}, val_time = {val_time:.2f}'.format( epoch=epoch, K=args.topk, hit_rate=hit_rate, ndcg=mean_ndcgs, loss=losses, train_time=train_time, val_time=val_time)) is_best = False if hit_rate > best_hit_rate: best_hit_rate = hit_rate is_best = True best_epoch = epoch extras = { 'current_hr@10': hit_rate, 'best_hr@10': best_hit_rate, 'best_epoch': best_epoch } apputils.save_checkpoint(epoch, 'NCF', model, optimizer, compression_scheduler, extras, is_best, dir=run_dir) if args.threshold is not None: if np.mean(hits) >= args.threshold: msglogger.info("Hit threshold of {}".format(args.threshold)) break
def main(): script_dir = os.path.dirname(__file__) module_path = os.path.abspath(os.path.join(script_dir, '..', '..')) global msglogger # Parse arguments args = parser.get_parser().parse_args() if args.epochs is None: args.epochs = 90 if not os.path.exists(args.output_dir): os.makedirs(args.output_dir) msglogger = apputils.config_pylogger( os.path.join(script_dir, 'logging.conf'), args.name, args.output_dir) # Log various details about the execution environment. It is sometimes useful # to refer to past experiment executions and this information may be useful. apputils.log_execution_env_state(args.compress, msglogger.logdir, gitroot=module_path) msglogger.debug("Distiller: %s", distiller.__version__) start_epoch = 0 ending_epoch = args.epochs perf_scores_history = [] if args.deterministic: # Experiment reproducibility is sometimes important. Pete Warden expounded about this # in his blog: https://petewarden.com/2018/03/19/the-machine-learning-reproducibility-crisis/ # In Pytorch, support for deterministic execution is still a bit clunky. if args.workers > 1: raise ValueError( 'ERROR: Setting --deterministic requires setting --workers/-j to 0 or 1' ) # Use a well-known seed, for repeatability of experiments distiller.set_deterministic() else: # This issue: https://github.com/pytorch/pytorch/issues/3659 # Implies that cudnn.benchmark should respect cudnn.deterministic, but empirically we see that # results are not re-produced when benchmark is set. So enabling only if deterministic mode disabled. cudnn.benchmark = True if args.cpu or not torch.cuda.is_available(): # Set GPU index to -1 if using CPU args.device = 'cpu' args.gpus = -1 else: args.device = 'cuda' if args.gpus is not None: try: args.gpus = [int(s) for s in args.gpus.split(',')] except ValueError: raise ValueError( 'ERROR: Argument --gpus must be a comma-separated list of integers only' ) available_gpus = torch.cuda.device_count() for dev_id in args.gpus: if dev_id >= available_gpus: raise ValueError( 'ERROR: GPU device ID {0} requested, but only {1} devices available' .format(dev_id, available_gpus)) # Set default device in case the first one on the list != 0 torch.cuda.set_device(args.gpus[0]) # Infer the dataset from the model name args.dataset = 'cifar10' if 'cifar' in args.arch else 'imagenet' args.num_classes = 10 if args.dataset == 'cifar10' else 1000 if args.earlyexit_thresholds: args.num_exits = len(args.earlyexit_thresholds) + 1 args.loss_exits = [0] * args.num_exits args.losses_exits = [] args.exiterrors = [] # Create the model model = create_model(args.pretrained, args.dataset, args.arch, parallel=not args.load_serialized, device_ids=args.gpus) compression_scheduler = None # Create a couple of logging backends. TensorBoardLogger writes log files in a format # that can be read by Google's Tensor Board. PythonLogger writes to the Python logger. tflogger = TensorBoardLogger(msglogger.logdir) pylogger = PythonLogger(msglogger) # capture thresholds for early-exit training if args.earlyexit_thresholds: msglogger.info('=> using early-exit threshold values of %s', args.earlyexit_thresholds) # TODO(barrh): args.deprecated_resume is deprecated since v0.3.1 if args.deprecated_resume: msglogger.warning( 'The "--resume" flag is deprecated. Please use "--resume-from=YOUR_PATH" instead.' ) if not args.reset_optimizer: msglogger.warning( 'If you wish to also reset the optimizer, call with: --reset-optimizer' ) args.reset_optimizer = True args.resumed_checkpoint_path = args.deprecated_resume # We can optionally resume from a checkpoint optimizer = None if args.resumed_checkpoint_path: model, compression_scheduler, optimizer, start_epoch = apputils.load_checkpoint( model, args.resumed_checkpoint_path, model_device=args.device) elif args.load_model_path: model = apputils.load_lean_checkpoint(model, args.load_model_path, model_device=args.device) if args.reset_optimizer: start_epoch = 0 if optimizer is not None: optimizer = None msglogger.info( '\nreset_optimizer flag set: Overriding resumed optimizer and resetting epoch count to 0' ) # Define loss function (criterion) criterion = nn.CrossEntropyLoss().to(args.device) if optimizer is None: optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay) msglogger.info('Optimizer Type: %s', type(optimizer)) msglogger.info('Optimizer Args: %s', optimizer.defaults) if args.AMC: return automated_deep_compression(model, criterion, optimizer, pylogger, args) if args.greedy: return greedy(model, criterion, optimizer, pylogger, args) # This sample application can be invoked to produce various summary reports. if args.summary: return summarize_model(model, args.dataset, which_summary=args.summary) activations_collectors = create_activation_stats_collectors( model, *args.activation_stats) if args.qe_calibration: msglogger.info('Quantization calibration stats collection enabled:') msglogger.info( '\tStats will be collected for {:.1%} of test dataset'.format( args.qe_calibration)) msglogger.info( '\tSetting constant seeds and converting model to serialized execution' ) distiller.set_deterministic() model = distiller.make_non_parallel_copy(model) activations_collectors.update( create_quantization_stats_collector(model)) args.evaluate = True args.effective_test_size = args.qe_calibration # Load the datasets: the dataset to load is inferred from the model name passed # in args.arch. The default dataset is ImageNet, but if args.arch contains the # substring "_cifar", then cifar10 is used. train_loader, val_loader, test_loader, _ = apputils.load_data( args.dataset, os.path.expanduser(args.data), args.batch_size, args.workers, args.validation_split, args.deterministic, args.effective_train_size, args.effective_valid_size, args.effective_test_size) msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d', len(train_loader.sampler), len(val_loader.sampler), len(test_loader.sampler)) if args.sensitivity is not None: sensitivities = np.arange(args.sensitivity_range[0], args.sensitivity_range[1], args.sensitivity_range[2]) return sensitivity_analysis(model, criterion, test_loader, pylogger, args, sensitivities) if args.evaluate: return evaluate_model(model, criterion, test_loader, pylogger, activations_collectors, args, compression_scheduler) if args.compress: # The main use-case for this sample application is CNN compression. Compression # requires a compression schedule configuration file in YAML. compression_scheduler = distiller.file_config( model, optimizer, args.compress, compression_scheduler, (start_epoch - 1) if args.resumed_checkpoint_path else None) # Model is re-transferred to GPU in case parameters were added (e.g. PACTQuantizer) model.to(args.device) elif compression_scheduler is None: compression_scheduler = distiller.CompressionScheduler(model) if args.thinnify: #zeros_mask_dict = distiller.create_model_masks_dict(model) assert args.resumed_checkpoint_path is not None, \ "You must use --resume-from to provide a checkpoint file to thinnify" distiller.remove_filters(model, compression_scheduler.zeros_mask_dict, args.arch, args.dataset, optimizer=None) apputils.save_checkpoint(0, args.arch, model, optimizer=None, scheduler=compression_scheduler, name="{}_thinned".format( args.resumed_checkpoint_path.replace( ".pth.tar", "")), dir=msglogger.logdir) print( "Note: your model may have collapsed to random inference, so you may want to fine-tune" ) return args.kd_policy = None if args.kd_teacher: teacher = create_model(args.kd_pretrained, args.dataset, args.kd_teacher, device_ids=args.gpus) if args.kd_resume: teacher = apputils.load_lean_checkpoint(teacher, args.kd_resume) dlw = distiller.DistillationLossWeights(args.kd_distill_wt, args.kd_student_wt, args.kd_teacher_wt) args.kd_policy = distiller.KnowledgeDistillationPolicy( model, teacher, args.kd_temp, dlw) compression_scheduler.add_policy(args.kd_policy, starting_epoch=args.kd_start_epoch, ending_epoch=args.epochs, frequency=1) msglogger.info('\nStudent-Teacher knowledge distillation enabled:') msglogger.info('\tTeacher Model: %s', args.kd_teacher) msglogger.info('\tTemperature: %s', args.kd_temp) msglogger.info('\tLoss Weights (distillation | student | teacher): %s', ' | '.join(['{:.2f}'.format(val) for val in dlw])) msglogger.info('\tStarting from Epoch: %s', args.kd_start_epoch) if start_epoch >= ending_epoch: msglogger.error( 'epoch count is too low, starting epoch is {} but total epochs set to {}' .format(start_epoch, ending_epoch)) raise ValueError('Epochs parameter is too low. Nothing to do.') for epoch in range(start_epoch, ending_epoch): # This is the main training loop. msglogger.info('\n') if compression_scheduler: compression_scheduler.on_epoch_begin( epoch, metrics=(vloss if (epoch != start_epoch) else 10**6)) # Train for one epoch with collectors_context(activations_collectors["train"]) as collectors: train(train_loader, model, criterion, optimizer, epoch, compression_scheduler, loggers=[tflogger, pylogger], args=args) distiller.log_weights_sparsity(model, epoch, loggers=[tflogger, pylogger]) distiller.log_activation_statsitics( epoch, "train", loggers=[tflogger], collector=collectors["sparsity"]) if args.masks_sparsity: msglogger.info( distiller.masks_sparsity_tbl_summary( model, compression_scheduler)) # evaluate on validation set with collectors_context(activations_collectors["valid"]) as collectors: top1, top5, vloss = validate(val_loader, model, criterion, [pylogger], args, epoch) distiller.log_activation_statsitics( epoch, "valid", loggers=[tflogger], collector=collectors["sparsity"]) save_collectors_data(collectors, msglogger.logdir) stats = ('Performance/Validation/', OrderedDict([('Loss', vloss), ('Top1', top1), ('Top5', top5)])) distiller.log_training_progress(stats, None, epoch, steps_completed=0, total_steps=1, log_freq=1, loggers=[tflogger]) if compression_scheduler: compression_scheduler.on_epoch_end(epoch, optimizer) # Update the list of top scores achieved so far, and save the checkpoint update_training_scores_history(perf_scores_history, model, top1, top5, epoch, args.num_best_scores) is_best = epoch == perf_scores_history[0].epoch checkpoint_extras = { 'current_top1': top1, 'best_top1': perf_scores_history[0].top1, 'best_epoch': perf_scores_history[0].epoch } apputils.save_checkpoint(epoch, args.arch, model, optimizer=optimizer, scheduler=compression_scheduler, extras=checkpoint_extras, is_best=is_best, name=args.name, dir=msglogger.logdir) # Finally run results on the test set test(test_loader, model, criterion, [pylogger], activations_collectors, args=args)
def main(): global msglogger check_pytorch_version() args = parser.parse_args() if not os.path.exists(args.output_dir): os.makedirs(args.output_dir) msglogger = apputils.config_pylogger(os.path.join(script_dir, 'logging.conf'), args.name, args.output_dir) # Log various details about the execution environment. It is sometimes useful # to refer to past experiment executions and this information may be useful. apputils.log_execution_env_state(sys.argv, gitroot=module_path) msglogger.debug("Distiller: %s", distiller.__version__) start_epoch = 0 best_epochs = [distiller.MutableNamedTuple({'epoch': 0, 'top1': 0, 'sparsity': 0}) for i in range(args.num_best_scores)] if args.deterministic: # Experiment reproducibility is sometimes important. Pete Warden expounded about this # in his blog: https://petewarden.com/2018/03/19/the-machine-learning-reproducibility-crisis/ # In Pytorch, support for deterministic execution is still a bit clunky. if args.workers > 1: msglogger.error('ERROR: Setting --deterministic requires setting --workers/-j to 0 or 1') exit(1) # Use a well-known seed, for repeatability of experiments torch.manual_seed(0) random.seed(0) np.random.seed(0) cudnn.deterministic = True else: # This issue: https://github.com/pytorch/pytorch/issues/3659 # Implies that cudnn.benchmark should respect cudnn.deterministic, but empirically we see that # results are not re-produced when benchmark is set. So enabling only if deterministic mode disabled. cudnn.benchmark = True if args.gpus is not None: try: args.gpus = [int(s) for s in args.gpus.split(',')] except ValueError: msglogger.error('ERROR: Argument --gpus must be a comma-separated list of integers only') exit(1) available_gpus = torch.cuda.device_count() for dev_id in args.gpus: if dev_id >= available_gpus: msglogger.error('ERROR: GPU device ID {0} requested, but only {1} devices available' .format(dev_id, available_gpus)) exit(1) # Set default device in case the first one on the list != 0 torch.cuda.set_device(args.gpus[0]) # Infer the dataset from the model name if 'cifar' in args.arch: args.dataset = 'cifar10' elif 'mnist' in args.arch: args.dataset = 'mnist' elif 'svhn' in args.arch: args.dataset = 'svhn' else: args.dataset = 'imagenet' args.num_classes = 10 if args.dataset == 'cifar10' or args.dataset == 'svhn' else 1000 if args.earlyexit_thresholds: args.num_exits = len(args.earlyexit_thresholds) + 1 args.loss_exits = [0] * args.num_exits args.losses_exits = [] args.exiterrors = [] # Create the model model = create_model(args.pretrained, args.dataset, args.arch, parallel=not args.load_serialized, device_ids=args.gpus) #DEBUG print('++++++++++++++++++++++++++++++++++++++++++') print(model) print('++++++++++++++++++++++++++++++++++++++++++') compression_scheduler = None # Create a couple of logging backends. TensorBoardLogger writes log files in a format # that can be read by Google's Tensor Board. PythonLogger writes to the Python logger. tflogger = TensorBoardLogger(msglogger.logdir) pylogger = PythonLogger(msglogger) # capture thresholds for early-exit training if args.earlyexit_thresholds: msglogger.info('=> using early-exit threshold values of %s', args.earlyexit_thresholds) # We can optionally resume from a checkpoint if args.resume: model, compression_scheduler, start_epoch = apputils.load_checkpoint( model, chkpt_file=args.resume) model.cuda() # Define loss function (criterion) and optimizer criterion = nn.CrossEntropyLoss().cuda() optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay) msglogger.info('Optimizer Type: %s', type(optimizer)) msglogger.info('Optimizer Args: %s', optimizer.defaults) if args.ADC: return automated_deep_compression(model, criterion, pylogger, args) # This sample application can be invoked to produce various summary reports. if args.summary: return summarize_model(model, args.dataset, which_summary=args.summary) # Load the datasets: the dataset to load is inferred from the model name passed # in args.arch. The default dataset is ImageNet, but if args.arch contains the # substring "_cifar", then cifar10 is used. train_loader, val_loader, test_loader, _ = apputils.load_data( args.dataset, os.path.expanduser(args.data), args.batch_size, args.workers, args.validation_size, args.deterministic) msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d', len(train_loader.sampler), len(val_loader.sampler), len(test_loader.sampler)) activations_collectors = create_activation_stats_collectors(model, collection_phase=args.activation_stats) if args.sensitivity is not None: sensitivities = np.arange(args.sensitivity_range[0], args.sensitivity_range[1], args.sensitivity_range[2]) return sensitivity_analysis(model, criterion, test_loader, pylogger, args, sensitivities) if args.evaluate: return evaluate_model(model, criterion, test_loader, pylogger, activations_collectors, args) if args.compress: # The main use-case for this sample application is CNN compression. Compression # requires a compression schedule configuration file in YAML. compression_scheduler = distiller.file_config(model, optimizer, args.compress, compression_scheduler) # Model is re-transferred to GPU in case parameters were added (e.g. PACTQuantizer) model.cuda() elif compression_scheduler is None: compression_scheduler = distiller.CompressionScheduler(model) args.kd_policy = None if args.kd_teacher: teacher = create_model(args.kd_pretrained, args.dataset, args.kd_teacher, device_ids=args.gpus) if args.kd_resume: teacher, _, _ = apputils.load_checkpoint(teacher, chkpt_file=args.kd_resume) dlw = distiller.DistillationLossWeights(args.kd_distill_wt, args.kd_student_wt, args.kd_teacher_wt) args.kd_policy = distiller.KnowledgeDistillationPolicy(model, teacher, args.kd_temp, dlw) compression_scheduler.add_policy(args.kd_policy, starting_epoch=args.kd_start_epoch, ending_epoch=args.epochs, frequency=1) msglogger.info('\nStudent-Teacher knowledge distillation enabled:') msglogger.info('\tTeacher Model: %s', args.kd_teacher) msglogger.info('\tTemperature: %s', args.kd_temp) msglogger.info('\tLoss Weights (distillation | student | teacher): %s', ' | '.join(['{:.2f}'.format(val) for val in dlw])) msglogger.info('\tStarting from Epoch: %s', args.kd_start_epoch) for epoch in range(start_epoch, start_epoch + args.epochs): # This is the main training loop. msglogger.info('\n') if compression_scheduler: compression_scheduler.on_epoch_begin(epoch) # Train for one epoch with collectors_context(activations_collectors["train"]) as collectors: train(train_loader, model, criterion, optimizer, epoch, compression_scheduler, loggers=[tflogger, pylogger], args=args) distiller.log_weights_sparsity(model, epoch, loggers=[tflogger, pylogger]) distiller.log_activation_statsitics(epoch, "train", loggers=[tflogger], collector=collectors["sparsity"]) if args.masks_sparsity: msglogger.info(distiller.masks_sparsity_tbl_summary(model, compression_scheduler)) # evaluate on validation set with collectors_context(activations_collectors["valid"]) as collectors: top1, top5, vloss = validate(val_loader, model, criterion, [pylogger], args, epoch) distiller.log_activation_statsitics(epoch, "valid", loggers=[tflogger], collector=collectors["sparsity"]) save_collectors_data(collectors, msglogger.logdir) f= open("val_accuracy.txt","w+") f.write(str(top1)) #f.write(str(top5)) f.close() stats = ('Peformance/Validation/', OrderedDict([('Loss', vloss), ('Top1', top1), ('Top5', top5)])) distiller.log_training_progress(stats, None, epoch, steps_completed=0, total_steps=1, log_freq=1, loggers=[tflogger]) if compression_scheduler: compression_scheduler.on_epoch_end(epoch, optimizer) # Update the list of top scores achieved so far, and save the checkpoint is_best = top1 > best_epochs[-1].top1 if top1 > best_epochs[0].top1: best_epochs[0].epoch = epoch best_epochs[0].top1 = top1 # Keep best_epochs sorted such that best_epochs[0] is the lowest top1 in the best_epochs list best_epochs = sorted(best_epochs, key=lambda score: score.top1) for score in reversed(best_epochs): if score.top1 > 0: msglogger.info('==> Best Top1: %.3f on Epoch: %d', score.top1, score.epoch) apputils.save_checkpoint(epoch, args.arch, model, optimizer, compression_scheduler, best_epochs[-1].top1, is_best, args.name, msglogger.logdir)
def main(opt): # basic settings os.environ["CUDA_VISIBLE_DEVICES"] = str(opt.gpu_ids)[1:-1] if torch.cuda.is_available(): device = "cuda" torch.backends.cudnn.benchmark = True else: device = "cpu" ##################### Get Dataloader #################### dataloader_train, dataloader_val = custom_get_dataloaders(opt) # dummy_input is sample input of dataloaders if hasattr(dataloader_val, "dataset"): dummy_input = dataloader_val.dataset.__getitem__(0) dummy_input = dummy_input[0] dummy_input = dummy_input.unsqueeze(0) else: # for imagenet dali loader dummy_input = torch.rand(1, 3, 224, 224) ##################### Create Baseline Model #################### net = ModelWrapper(opt) net.load_checkpoint(opt.checkpoint) flops_before, params_before = model_summary(net.get_compress_part(), dummy_input) ##################### Load Pruning Strategy ############### compression_scheduler = distiller.file_config(net.get_compress_part(), net.optimizer, opt.compress_schedule_path) channel_config = get_channel_config(opt.search_result, opt.strategy_id) # pruning strategy compression_scheduler = random_compression_scheduler( compression_scheduler, channel_config) ###### Adaptive-BN-based Candidate Evaluation of Pruning Strategy ### thinning(net, compression_scheduler, input_tensor=dummy_input) flops_after, params_after = model_summary(net.get_compress_part(), dummy_input) ratio = flops_after / flops_before print("FLOPs ratio:", ratio) net = net.to(device) net.parallel(opt.gpu_ids) net.get_compress_part().train() with torch.no_grad(): for index, sample in enumerate(tqdm(dataloader_train, leave=False)): _ = net.get_loss(sample) if index > 100: break strategy_score = net.get_eval_scores(dataloader_val)["accuracy"] print("Result file:{}, Strategy ID:{}, Evaluation score:{}".format( opt.search_result, opt.strategy_id, strategy_score)) ##################### Fine-tuning ######################### lr_scheduler = optim.lr_scheduler.CosineAnnealingLR( net.optimizer, opt.epoch) reporter = Reporter(opt) best_acc = 0 net._net.train() for epoch in range(1, opt.epoch + 1): reporter.log_metric("lr", net.optimizer.param_groups[0]["lr"], epoch) train_loss = train_epoch( net, dataloader_train, net.optimizer, ) reporter.log_metric("train_loss", train_loss, epoch) lr_scheduler.step() scores = net.get_eval_scores(dataloader_val) print("==> Evaluation: Epoch={} Acc={}".format(epoch, str(scores))) reporter.log_metric("eval_acc", scores["accuracy"], epoch) if scores["accuracy"] > best_acc: best_acc = scores["accuracy"] reporter.log_metric("best_acc", best_acc, epoch) save_checkpoints( scores["accuracy"], net._net, reporter, opt.exp_name, epoch, ) print("==> Training epoch %d" % epoch)
def main(): global msglogger check_pytorch_version() args = parser.parse_args() if not os.path.exists(args.output_dir): os.makedirs(args.output_dir) msglogger = apputils.config_pylogger(os.path.join(script_dir, 'logging.conf'), args.name, args.output_dir) # Log various details about the execution environment. It is sometimes useful # to refer to past experiment executions and this information may be useful. apputils.log_execution_env_state(sys.argv, gitroot=module_path) msglogger.debug("Distiller: %s", distiller.__version__) start_epoch = 0 best_top1 = 0 if args.deterministic: # Experiment reproducibility is sometimes important. Pete Warden expounded about this # in his blog: https://petewarden.com/2018/03/19/the-machine-learning-reproducibility-crisis/ # In Pytorch, support for deterministic execution is still a bit clunky. if args.workers > 1: msglogger.error('ERROR: Setting --deterministic requires setting --workers/-j to 0 or 1') exit(1) # Use a well-known seed, for repeatability of experiments torch.manual_seed(0) random.seed(0) np.random.seed(0) cudnn.deterministic = True else: # This issue: https://github.com/pytorch/pytorch/issues/3659 # Implies that cudnn.benchmark should respect cudnn.deterministic, but empirically we see that # results are not re-produced when benchmark is set. So enabling only if deterministic mode disabled. cudnn.benchmark = True if args.gpus is not None: try: args.gpus = [int(s) for s in args.gpus.split(',')] except ValueError: msglogger.error('ERROR: Argument --gpus must be a comma-separated list of integers only') exit(1) available_gpus = torch.cuda.device_count() for dev_id in args.gpus: if dev_id >= available_gpus: msglogger.error('ERROR: GPU device ID {0} requested, but only {1} devices available' .format(dev_id, available_gpus)) exit(1) # Set default device in case the first one on the list != 0 torch.cuda.set_device(args.gpus[0]) # Infer the dataset from the model name args.dataset = 'cifar10' if 'cifar' in args.arch else 'imagenet' args.num_classes = 10 if args.dataset == 'cifar10' else 1000 # Create the model model = create_model(args.pretrained, args.dataset, args.arch, device_ids=args.gpus) compression_scheduler = None # Create a couple of logging backends. TensorBoardLogger writes log files in a format # that can be read by Google's Tensor Board. PythonLogger writes to the Python logger. tflogger = TensorBoardLogger(msglogger.logdir) pylogger = PythonLogger(msglogger) # We can optionally resume from a checkpoint if args.resume: model, compression_scheduler, start_epoch = apputils.load_checkpoint( model, chkpt_file=args.resume) # Define loss function (criterion) and optimizer criterion = nn.CrossEntropyLoss().cuda() optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay) msglogger.info('Optimizer Type: %s', type(optimizer)) msglogger.info('Optimizer Args: %s', optimizer.defaults) if args.ADC: return automated_deep_compression(model, criterion, pylogger, args) # This sample application can be invoked to produce various summary reports. if args.summary: return summarize_model(model, args.dataset, which_summary=args.summary) # Load the datasets: the dataset to load is inferred from the model name passed # in args.arch. The default dataset is ImageNet, but if args.arch contains the # substring "_cifar", then cifar10 is used. train_loader, val_loader, test_loader, _ = apputils.load_data( args.dataset, os.path.expanduser(args.data), args.batch_size, args.workers, args.validation_size, args.deterministic) msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d', len(train_loader.sampler), len(val_loader.sampler), len(test_loader.sampler)) activations_sparsity = None if args.activation_stats: # If your model has ReLU layers, then those layers have sparse activations. # ActivationSparsityCollector will collect information about this sparsity. # WARNING! Enabling activation sparsity collection will significantly slow down training! activations_sparsity = ActivationSparsityCollector(model) if args.sensitivity is not None: return sensitivity_analysis(model, criterion, test_loader, pylogger, args) if args.evaluate: return evaluate_model(model, criterion, test_loader, pylogger, args) if args.compress: # The main use-case for this sample application is CNN compression. Compression # requires a compression schedule configuration file in YAML. compression_scheduler = distiller.file_config(model, optimizer, args.compress) # Model is re-transferred to GPU in case parameters were added (e.g. PACTQuantizer) model.cuda() for epoch in range(start_epoch, start_epoch + args.epochs): # This is the main training loop. msglogger.info('\n') if compression_scheduler: compression_scheduler.on_epoch_begin(epoch) # Train for one epoch train(train_loader, model, criterion, optimizer, epoch, compression_scheduler, loggers=[tflogger, pylogger], args=args) distiller.log_weights_sparsity(model, epoch, loggers=[tflogger, pylogger]) if args.activation_stats: distiller.log_activation_sparsity(epoch, loggers=[tflogger, pylogger], collector=activations_sparsity) # evaluate on validation set top1, top5, vloss = validate(val_loader, model, criterion, [pylogger], args, epoch) stats = ('Peformance/Validation/', OrderedDict([('Loss', vloss), ('Top1', top1), ('Top5', top5)])) distiller.log_training_progress(stats, None, epoch, steps_completed=0, total_steps=1, log_freq=1, loggers=[tflogger]) if compression_scheduler: compression_scheduler.on_epoch_end(epoch, optimizer) # remember best top1 and save checkpoint is_best = top1 > best_top1 if is_best: best_epoch = epoch best_top1 = top1 msglogger.info('==> Best validation Top1: %.3f Epoch: %d', best_top1, best_epoch) apputils.save_checkpoint(epoch, args.arch, model, optimizer, compression_scheduler, best_top1, is_best, args.name, msglogger.logdir) # Finally run results on the test set test(test_loader, model, criterion, [pylogger], args=args)
def main(): script_dir = os.path.dirname(__file__) module_path = os.path.abspath(os.path.join(script_dir, '..', '..')) global msglogger # Parse arguments args = parser.get_parser().parse_args() if not os.path.exists(args.output_dir): os.makedirs(args.output_dir) msglogger = apputils.config_pylogger( os.path.join(script_dir, 'logging.conf'), args.name, args.output_dir) # Log various details about the execution environment. It is sometimes useful # to refer to past experiment executions and this information may be useful. apputils.log_execution_env_state(args.compress, msglogger.logdir, gitroot=module_path) msglogger.debug("Distiller: %s", distiller.__version__) start_epoch = 0 best_epochs = list() if args.deterministic: if args.loaders is None: args.loaders = 1 # Experiment reproducibility is sometimes important. Pete Warden expounded about this # in his blog: https://petewarden.com/2018/03/19/the-machine-learning-reproducibility-crisis/ # In Pytorch, support for deterministic execution is still a bit clunky. if args.loaders > 1: msglogger.error( 'ERROR: Setting --deterministic requires setting --loaders to 0 or 1' ) exit(1) # Use a well-known seed, for repeatability of experiments distiller.set_deterministic() else: # This issue: https://github.com/pytorch/pytorch/issues/3659 # Implies that cudnn.benchmark should respect cudnn.deterministic, but empirically we see that # results are not re-produced when benchmark is set. So enabling only if deterministic mode disabled. cudnn.benchmark = True if args.use_cpu or (args.gpus is None and not torch.cuda.is_available()) or (args.gpus == ''): # Set GPU index to -1 if using CPU args.device = 'cpu' args.gpus = -1 else: args.device = 'cuda' if args.gpus is not None: try: args.gpus = [int(s) for s in args.gpus.split(',')] except ValueError: msglogger.error( 'ERROR: Argument --gpus must be a comma-separated list of integers only' ) exit(1) available_gpus = torch.cuda.device_count() for dev_id in args.gpus: if dev_id >= available_gpus: msglogger.error( 'ERROR: GPU device ID {0} requested, but only {1} devices available' .format(dev_id, available_gpus)) exit(1) # Set default device in case the first one on the list != 0 torch.cuda.set_device(args.gpus[0]) if args.loaders is None: active_gpus = args.gpus if args.gpus is not None else torch.cuda.device_count( ) args.loaders = max(parser.DEFAULT_LOADERS_COUNT, parser.DEFAULT_LOADERS_COUNT * active_gpus) msglogger.debug('Number of data loaders set to: {}'.format(args.loaders)) # Infer the dataset from the model name args.dataset = 'cifar10' if 'cifar' in args.arch else 'imagenet' args.num_classes = 10 if args.dataset == 'cifar10' else 1000 if args.earlyexit_thresholds: args.num_exits = len(args.earlyexit_thresholds) + 1 args.loss_exits = [0] * args.num_exits args.losses_exits = [] args.exiterrors = [] # Create the model model = create_model(args.pretrained, args.dataset, args.arch, parallel=not args.load_serialized, device_ids=args.gpus) compression_scheduler = None # Create a couple of logging backends. TensorBoardLogger writes log files in a format # that can be read by Google's Tensor Board. PythonLogger writes to the Python logger. tflogger = TensorBoardLogger(msglogger.logdir) pylogger = PythonLogger(msglogger) # capture thresholds for early-exit training if args.earlyexit_thresholds: msglogger.info('=> using early-exit threshold values of %s', args.earlyexit_thresholds) # We can optionally resume from a checkpoint optimizer = None resumed_training_steps = None if args.resume or args.load_state_dict: if args.resume and not args.reset_optimizer: # initiate SGD with dummy lr optimizer = torch.optim.SGD(model.parameters(), lr=0.36787944117) model, compression_scheduler, optimizer, start_epoch, resumed_training_steps = apputils.load_checkpoint( model, args.resume or args.load_state_dict, optimizer=optimizer) model.to(args.device) # Define loss function (criterion) and optimizer criterion = nn.CrossEntropyLoss().to(args.device) if optimizer is not None: # optimizer was resumed from checkpoint # check if user has tried to set optimizer arguments # if so, ignore arguments with a warning. optimizer_group_args = [ 'lr', 'learning-rate', 'momentum', 'weight-decay', 'wd' ] user_optim_args = [ x for x in optimizer_group_args for arg in sys.argv if arg.startswith('--' + x) ] if user_optim_args: msglogger.warning( '{} optimizer arguments are ignored.'.format(user_optim_args)) msglogger.info( 'setting optimizer arguments when optimizer is resumed ' 'from checkpoint is forbidden.') else: optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay) msglogger.info('Optimizer Type: %s', type(optimizer)) msglogger.info('Optimizer Args: %s', optimizer.defaults) if args.AMC: return automated_deep_compression(model, criterion, optimizer, pylogger, args) if args.greedy: return greedy(model, criterion, optimizer, pylogger, args) # This sample application can be invoked to produce various summary reports. if args.summary: return summarize_model(model, args.dataset, which_summary=args.summary) activations_collectors = create_activation_stats_collectors( model, *args.activation_stats) if args.qe_calibration: msglogger.info('Quantization calibration stats collection enabled:') msglogger.info( '\tStats will be collected for {:.1%} of test dataset'.format( args.qe_calibration)) msglogger.info( '\tSetting constant seeds and converting model to serialized execution' ) distiller.set_deterministic() model = distiller.make_non_parallel_copy(model) activations_collectors.update( create_quantization_stats_collector(model)) args.evaluate = True args.effective_test_size = args.qe_calibration # Load the datasets: the dataset to load is inferred from the model name passed # in args.arch. The default dataset is ImageNet, but if args.arch contains the # substring "_cifar", then cifar10 is used. train_loader, val_loader, test_loader, _ = apputils.load_data( args.dataset, os.path.expanduser(args.data), args.batch_size, args.loaders, args.validation_split, args.deterministic, args.effective_train_size, args.effective_valid_size, args.effective_test_size) msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d', len(train_loader.sampler), len(val_loader.sampler), len(test_loader.sampler)) args.trainset_print_period = parser.getPrintPeriod( args, len(train_loader.sampler), args.batch_size) args.validset_print_period = parser.getPrintPeriod(args, len(val_loader.sampler), args.batch_size) args.testset_print_period = parser.getPrintPeriod(args, len(test_loader.sampler), args.batch_size) if args.sensitivity is not None: sensitivities = np.arange(args.sensitivity_range[0], args.sensitivity_range[1], args.sensitivity_range[2]) return sensitivity_analysis(model, criterion, test_loader, pylogger, args, sensitivities) if args.evaluate: return evaluate_model(model, criterion, test_loader, pylogger, activations_collectors, args, compression_scheduler) if args.compress: # The main use-case for this sample application is CNN compression. Compression # requires a compression schedule configuration file in YAML. compression_scheduler = distiller.file_config( model, optimizer, args.compress, compression_scheduler, (start_epoch - 1) if (args.resume and not args.reset_optimizer) else None) # Model is re-transferred to GPU in case parameters were added (e.g. PACTQuantizer) model.to(args.device) elif compression_scheduler is None: compression_scheduler = distiller.CompressionScheduler(model) if args.thinnify: #zeros_mask_dict = distiller.create_model_masks_dict(model) assert args.resume is not None, "You must use --resume to provide a checkpoint file to thinnify" distiller.remove_filters(model, compression_scheduler.zeros_mask_dict, args.arch, args.dataset, optimizer=None) apputils.save_checkpoint(0, args.arch, model, optimizer=None, scheduler=compression_scheduler, name="{}_thinned".format( args.resume.replace(".pth.tar", "")), dir=msglogger.logdir) print( "Note: your model may have collapsed to random inference, so you may want to fine-tune" ) return args.kd_policy = None if args.kd_teacher: teacher = create_model(args.kd_pretrained, args.dataset, args.kd_teacher, device_ids=args.gpus) if args.kd_resume: teacher = apputils.load_checkpoint(teacher, chkpt_file=args.kd_resume)[0] dlw = distiller.DistillationLossWeights(args.kd_distill_wt, args.kd_student_wt, args.kd_teacher_wt) args.kd_policy = distiller.KnowledgeDistillationPolicy( model, teacher, args.kd_temp, dlw) compression_scheduler.add_policy( args.kd_policy, range(args.kd_start_epoch, args.epochs, 1)) msglogger.info('\nStudent-Teacher knowledge distillation enabled:') msglogger.info('\tTeacher Model: %s', args.kd_teacher) msglogger.info('\tTemperature: %s', args.kd_temp) msglogger.info('\tLoss Weights (distillation | student | teacher): %s', ' | '.join(['{:.2f}'.format(val) for val in dlw])) msglogger.info('\tStarting from Epoch: %s', args.kd_start_epoch) if getattr(compression_scheduler, 'global_policy_end_epoch', None) is not None: if compression_scheduler.global_policy_end_epoch >= (start_epoch + args.epochs): msglogger.warning( 'scheduler requires at least {} epochs, but only {} are sanctioned' .format(compression_scheduler.global_policy_end_epoch, args.epochs)) accumulated_training_steps = resumed_training_steps if resumed_training_steps is not None else 0 for epoch in range(start_epoch, start_epoch + args.epochs): # This is the main training loop. msglogger.info('\n') if compression_scheduler: compression_scheduler.on_epoch_begin(epoch) # Train for one epoch with collectors_context(activations_collectors["train"]) as collectors: try: train(train_loader, model, criterion, optimizer, epoch, accumulated_training_steps, compression_scheduler, loggers=[tflogger, pylogger], args=args) except RuntimeError as e: if ('cuda out of memory' in str(e).lower()): msglogger.error( 'CUDA memory failure has been detected.\n' 'Sometimes it helps to decrease batch size.\n' 'e.g. Add the following flag to your call: --batch-size={}' .format(args.batch_size // 10)) raise distiller.log_weights_sparsity(model, epoch, loggers=[tflogger, pylogger]) distiller.log_activation_statsitics( epoch, "train", loggers=[tflogger], collector=collectors["sparsity"]) if args.masks_sparsity: msglogger.info( distiller.masks_sparsity_tbl_summary( model, compression_scheduler)) accumulated_training_steps += math.ceil( len(train_loader.sampler) / train_loader.batch_size) # evaluate on validation set with collectors_context(activations_collectors["valid"]) as collectors: top1, top5, vloss = validate(val_loader, model, criterion, [pylogger], args, epoch) distiller.log_activation_statsitics( epoch, "valid", loggers=[tflogger], collector=collectors["sparsity"]) save_collectors_data(collectors, msglogger.logdir) stats = ('Performance/Validation/', OrderedDict([('Loss', vloss), ('Top1', top1), ('Top5', top5)])) tflogger.log_training_progress(stats, epoch, None) if compression_scheduler: compression_scheduler.on_epoch_end(epoch, optimizer) if getattr(compression_scheduler, 'global_policy_end_epoch', None) is None or ( compression_scheduler.global_policy_end_epoch <= epoch): # Update the list of top scores achieved since all policies have concluded if top1 > 0: best_epochs.append( distiller.MutableNamedTuple({ 'top1': top1, 'top5': top5, 'epoch': epoch })) # Keep best_epochs sorted from best to worst # Sort by top1 first, secondary sort by top5, and so forth best_epochs.sort(key=operator.attrgetter('top1', 'top5', 'epoch'), reverse=True) for score in best_epochs[:args.num_best_scores]: msglogger.info('==> Best Top1: %.3f Top5: %.3f on epoch: %d', score.top1, score.top5, score.epoch) is_best = best_epochs and (epoch == best_epochs[0].epoch) apputils.save_checkpoint(epoch, args.arch, model, optimizer, compression_scheduler, best_epochs[0].top1 if best_epochs else None, is_best, args.name, msglogger.logdir, accumulated_training_steps) # Finally run results on the test set test(test_loader, model, criterion, [pylogger], activations_collectors, args=args)
target_transform=gt_transform) save_dir = os.path.join('pruning', model.name, pruner) if not os.path.exists(save_dir): os.makedirs(save_dir) stats_dir = os.path.join(save_dir, 'csv') save_model_stats(model, stats_dir, 'initial') train_loader = torch.utils.data.DataLoader(dataset, batch_size=args.batch_size, shuffle=True) writer = tensorboard.SummaryWriter(os.path.join(save_dir, 'logs')) print('Dataset loaded successfully') ## Load disitller scheduler compression_scheduler = distiller.file_config(model, optimizer, args.scheduler, None, None) for epoch in range(1, args.epoch + 1): compression_scheduler.on_epoch_begin( epoch) ## Call distiller on_epoch_begin print('Started epoch {:03d}/{:03d}'.format(epoch, args.epoch)) train(train_loader, model, optimizer, epoch, writer, compression_scheduler) compression_scheduler.on_epoch_end(epoch) ## Call distiller on_epoch_begin torch.save(model.state_dict(), '{}/{}.pruned.pth'.format(save_dir, model.name)) save_model_stats(model, stats_dir, epoch) save_model_stats(model, stats_dir, 'final')
def main(): script_dir = os.path.dirname(__file__) module_path = os.path.abspath(os.path.join(script_dir, '..', '..')) global msglogger # Parse arguments args = parser.get_parser().parse_args() if args.epochs is None: args.epochs = 200 if not os.path.exists(args.output_dir): os.makedirs(args.output_dir) msglogger = apputils.config_pylogger( os.path.join(script_dir, 'logging.conf'), args.name, args.output_dir) # Log various details about the execution environment. It is sometimes useful # to refer to past experiment executions and this information may be useful. apputils.log_execution_env_state( filter(None, [args.compress, args.qe_stats_file ]), # remove both None and empty strings msglogger.logdir, gitroot=module_path) msglogger.debug("Distiller: %s", distiller.__version__) if args.evaluate: args.deterministic = True if args.deterministic: distiller.set_deterministic( args.seed) # For experiment reproducability else: if args.seed is not None: distiller.set_seed(args.seed) # Turn on CUDNN benchmark mode for best performance. This is usually "safe" for image # classification models, as the input sizes don't change during the run # See here: https://discuss.pytorch.org/t/what-does-torch-backends-cudnn-benchmark-do/5936/3 cudnn.benchmark = True start_epoch = 0 ending_epoch = args.epochs perf_scores_history = [] if args.cpu or not torch.cuda.is_available(): # Set GPU index to -1 if using CPU args.device = 'cpu' args.gpus = -1 else: args.device = 'cuda' if args.gpus is not None: try: args.gpus = [int(s) for s in args.gpus.split(',')] except ValueError: raise ValueError( 'ERROR: Argument --gpus must be a comma-separated list of integers only' ) available_gpus = torch.cuda.device_count() for dev_id in args.gpus: if dev_id >= available_gpus: raise ValueError( 'ERROR: GPU device ID {0} requested, but only {1} devices available' .format(dev_id, available_gpus)) # Set default device in case the first one on the list != 0 torch.cuda.set_device(args.gpus[0]) # Infer the dataset from the model name # TODO args.dataset = 'coco' # args.num_classes = 21 # wc -l ~/data/VOC2012/voc-model-labels.txt if args.load_vgg19 and args.arch != 'vgg19': raise ValueError( '``load_vgg19`` should be set only when vgg19 is used') model = create_pose_estimation_model(args.pretrained, args.dataset, args.arch, load_vgg19=args.load_vgg19, parallel=not args.load_serialized, device_ids=args.gpus) compression_scheduler = None # Create a couple of logging backends. TensorBoardLogger writes log files in a format # that can be read by Google's Tensor Board. PythonLogger writes to the Python logger. tflogger = TensorBoardLogger(msglogger.logdir) pylogger = PythonLogger(msglogger) # <editor-fold desc=">>> Load Model"> # We can optionally resume from a checkpoint optimizer = None if args.resumed_checkpoint_path: model, compression_scheduler, optimizer, start_epoch = apputils.load_checkpoint( model, args.resumed_checkpoint_path, model_device=args.device) elif args.load_model_path: model = apputils.load_lean_checkpoint(model, args.load_model_path, model_device=args.device) if args.reset_optimizer: start_epoch = 0 if optimizer is not None: optimizer = None msglogger.info( '\nreset_optimizer flag set: Overriding resumed optimizer and resetting epoch count to 0' ) # </editor-fold> # Define loss function (criterion) # get_loss(saved_for_loss, heat_temp, heat_weight,vec_temp, vec_weight) criterion = { 'shufflenetv2': shufflenetv2_get_loss, 'vgg19': vgg19_get_loss, 'hourglass': hourglass_get_loss, }[args.arch] if optimizer is None: trainable_vars = [ param for param in model.parameters() if param.requires_grad ] optimizer = torch.optim.SGD(trainable_vars, lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay) msglogger.info('Optimizer Type: %s', type(optimizer)) msglogger.info('Optimizer Args: %s', optimizer.defaults) # TODO: load lr_scheduler lr_scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.8, patience=5, verbose=True, threshold=0.0001, threshold_mode='rel', cooldown=3, min_lr=0, eps=1e-08) if args.AMC: return automated_deep_compression(model, criterion, optimizer, pylogger, args) if args.greedy: return greedy(model, criterion, optimizer, pylogger, args) # This sample application can be invoked to produce various summary reports. if args.summary: for summary in args.summary: distiller.model_summary(model, summary, args.dataset) return if args.export_onnx is not None: return distiller.export_img_classifier_to_onnx(model, os.path.join( msglogger.logdir, args.export_onnx), args.dataset, add_softmax=True, verbose=False) if args.qe_calibration: return acts_quant_stats_collection(model, criterion, pylogger, args) if args.activation_histograms: return acts_histogram_collection(model, criterion, pylogger, args) print('Building activations_collectors...') activations_collectors = create_activation_stats_collectors( model, *args.activation_stats) # Load the datasets: the dataset to load is inferred from the model name passed # in args.arch. The default dataset is ImageNet, but if args.arch contains the # substring "_cifar", then cifar10 is used. print('Loading data...') train_loader, val_loader, test_loader, _ = load_data(args) msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d', len(train_loader.sampler), len(val_loader.sampler), len(test_loader.sampler)) if args.sensitivity is not None: sensitivities = np.arange(args.sensitivity_range[0], args.sensitivity_range[1], args.sensitivity_range[2]) return sensitivity_analysis(model, criterion, test_loader, pylogger, args, sensitivities) if args.evaluate: return evaluate_model(model, criterion, test_loader, pylogger, activations_collectors, args, compression_scheduler) if args.compress: # The main use-case for this sample application is CNN compression. Compression # requires a compression schedule configuration file in YAML. compression_scheduler = distiller.file_config( model, optimizer, args.compress, compression_scheduler, (start_epoch - 1) if args.resumed_checkpoint_path else None) # Model is re-transferred to GPU in case parameters were added (e.g. PACTQuantizer) model.to(args.device) elif compression_scheduler is None: compression_scheduler = distiller.CompressionScheduler(model) if args.thinnify: # zeros_mask_dict = distiller.create_model_masks_dict(model) assert args.resumed_checkpoint_path is not None, \ "You must use --resume-from to provide a checkpoint file to thinnify" distiller.remove_filters(model, compression_scheduler.zeros_mask_dict, args.arch, args.dataset, optimizer=None) apputils.save_checkpoint(0, args.arch, model, optimizer=None, scheduler=compression_scheduler, name="{}_thinned".format( args.resumed_checkpoint_path.replace( ".pth.tar", "")), dir=msglogger.logdir) print( "Note: your model may have collapsed to random inference, so you may want to fine-tune" ) return if start_epoch >= ending_epoch: msglogger.error( 'epoch count is too low, starting epoch is {} but total epochs set to {}' .format(start_epoch, ending_epoch)) raise ValueError('Epochs parameter is too low. Nothing to do.') for epoch in range(start_epoch, ending_epoch): # This is the main training loop. msglogger.info('\n') if compression_scheduler: compression_scheduler.on_epoch_begin( epoch, metrics=(total_loss if (epoch != start_epoch) else 10**6)) # Train for one epoch with collectors_context(activations_collectors["train"]) as collectors: train(train_loader, model, criterion, optimizer, epoch, compression_scheduler, loggers=[tflogger, pylogger], args=args) distiller.log_weights_sparsity(model, epoch, loggers=[tflogger, pylogger]) distiller.log_activation_statsitics( epoch, "train", loggers=[tflogger], collector=collectors["sparsity"]) if args.masks_sparsity: msglogger.info( distiller.masks_sparsity_tbl_summary( model, compression_scheduler)) # evaluate on validation set with collectors_context(activations_collectors["valid"]) as collectors: loss = validate(val_loader, model, criterion, [pylogger], args, epoch) distiller.log_activation_statsitics( epoch, "valid", loggers=[tflogger], collector=collectors["sparsity"]) save_collectors_data(collectors, msglogger.logdir) lr_scheduler.step(loss) stats = ('Performance/Validation/', OrderedDict([('Loss', loss)])) distiller.log_training_progress(stats, None, epoch, steps_completed=0, total_steps=1, log_freq=1, loggers=[tflogger]) if compression_scheduler: compression_scheduler.on_epoch_end(epoch, optimizer) # Update the list of top scores achieved so far, and save the checkpoint update_training_scores_history(perf_scores_history, model, loss, epoch, args.num_best_scores) is_best = epoch == perf_scores_history[0].epoch checkpoint_extras = { 'current_loss': loss, 'best_loss': perf_scores_history[0].loss, 'best_epoch': perf_scores_history[0].epoch } apputils.save_checkpoint(epoch, args.arch, model, optimizer=optimizer, scheduler=compression_scheduler, extras=checkpoint_extras, is_best=is_best, name=args.name, dir=msglogger.logdir) # Finally run results on the test set test(test_loader, model, criterion, [pylogger], activations_collectors, args=args)
def _init_learner(args): # Create the model model = create_model(args.pretrained, args.dataset, args.arch, parallel=not args.load_serialized, device_ids=args.gpus) compression_scheduler = None # TODO(barrh): args.deprecated_resume is deprecated since v0.3.1 if args.deprecated_resume: msglogger.warning( 'The "--resume" flag is deprecated. Please use "--resume-from=YOUR_PATH" instead.' ) if not args.reset_optimizer: msglogger.warning( 'If you wish to also reset the optimizer, call with: --reset-optimizer' ) args.reset_optimizer = True args.resumed_checkpoint_path = args.deprecated_resume optimizer = None start_epoch = 0 if args.resumed_checkpoint_path: model, compression_scheduler, optimizer, start_epoch = apputils.load_checkpoint( model, args.resumed_checkpoint_path, model_device=args.device) elif args.load_model_path: model = apputils.load_lean_checkpoint(model, args.load_model_path, model_device=args.device) if args.reset_optimizer: start_epoch = 0 if optimizer is not None: optimizer = None msglogger.info( '\nreset_optimizer flag set: Overriding resumed optimizer and resetting epoch count to 0' ) if optimizer is None: optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay) msglogger.debug('Optimizer Type: %s', type(optimizer)) msglogger.debug('Optimizer Args: %s', optimizer.defaults) if args.compress: # The main use-case for this sample application is CNN compression. Compression # requires a compression schedule configuration file in YAML. compression_scheduler = distiller.file_config( model, optimizer, args.compress, compression_scheduler, (start_epoch - 1) if args.resumed_checkpoint_path else None) # Model is re-transferred to GPU in case parameters were added (e.g. PACTQuantizer) model.to(args.device) elif compression_scheduler is None: compression_scheduler = distiller.CompressionScheduler(model) ending_epoch = args.epochs if start_epoch >= ending_epoch: msglogger.error( 'epoch count is too low, starting epoch is {} but total epochs set to {}' .format(start_epoch, ending_epoch)) raise ValueError('Epochs parameter is too low. Nothing to do.') return model, compression_scheduler, optimizer, start_epoch, ending_epoch
args.effective_train_size, args.effective_valid_size, args.effective_test_size) msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d', len(train_loader.sampler), len(val_loader.sampler), len(test_loader.sampler)) if args.sensitivity is not None: sensitivities = np.arange(args.sensitivity_range[0], args.sensitivity_range[1], args.sensitivity_range[2]) return sensitivity_analysis(model, criterion, test_loader, pylogger, args, sensitivities) if args.evaluate: return evaluate_model(model, criterion, test_loader, pylogger, activations_collectors, args, compression_scheduler) if args.compress: # The main use-case for this sample application is CNN compression. Compression # requires a compression schedule configuration file in YAML. compression_scheduler = distiller.file_config(model, optimizer, args.compress, compression_scheduler, (start_epoch-1) if args.resumed_checkpoint_path else None) # Model is re-transferred to GPU in case parameters were added (e.g. PACTQuantizer) model.to(args.device) elif compression_scheduler is None: compression_scheduler = distiller.CompressionScheduler(model) if args.thinnify: #zeros_mask_dict = distiller.create_model_masks_dict(model) assert args.resumed_checkpoint_path is not None, "You must use --resume-from to provide a checkpoint file to thinnify" distiller.remove_filters(model, compression_scheduler.zeros_mask_dict, args.arch, args.dataset, optimizer=None) apputils.save_checkpoint(0, args.arch, model, optimizer=None, scheduler=compression_scheduler, name="{}_thinned".format(args.resumed_checkpoint_path.replace(".pth.tar", "")), dir=msglogger.logdir) print("Note: your model may have collapsed to random inference, so you may want to fine-tune") return args.kd_policy = None
def main(opt, channel_config, dataloader_train, dataloader_val, path): # basic settings torch.backends.cudnn.enabled = False os.environ["CUDA_VISIBLE_DEVICES"] = str(opt.gpu_ids)[1:-1] if torch.cuda.is_available(): device = "cuda" torch.backends.cudnn.benchmark = False else: device = "cpu" ##################### Get Dataloader #################### # dummy_input is sample input of dataloaders if hasattr(dataloader_val, "dataset"): dummy_input = dataloader_val.dataset.__getitem__(0) dummy_input = dummy_input[0] dummy_input = dummy_input.unsqueeze(0) else: # for imagenet dali loader dummy_input = torch.rand(1, 3, 224, 224) ##################### Create Baseline Model #################### net = ModelWrapper(opt) load(net, path) #net.load_checkpoint(opt.checkpoint) ##################### Load Pruning Strategy ############### compression_scheduler = distiller.file_config(net.get_compress_part(), net.optimizer, opt.compress_schedule_path) compression_scheduler = setCompressionScheduler(compression_scheduler, channel_config) ###### Adaptive-BN-based Candidate Evaluation of Pruning Strategy ### thinning(net, compression_scheduler, input_tensor=dummy_input) flops_after, params_after = model_summary(net.get_compress_part(), dummy_input) net = net.to(device) net.parallel(opt.gpu_ids) net.get_compress_part().train() t = tqdm(dataloader_train, leave=False) with torch.no_grad(): for index, sample in enumerate(t): _ = net.get_loss(sample) if index > 100: break strategy_score = net.get_eval_scores(dataloader_val)["accuracy"] old = strategy_score print("Evaluation score:{}".format(strategy_score)) ##################### Fine-tuning ######################### lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(net.optimizer, 100, eta_min=5e-5) #lr_scheduler=optim.lr_scheduler.StepLR(net.optimizer,5,0.9) reporter = Reporter(opt) best_acc = strategy_score best_kappa = 0 net._net.train() for epoch in range(1, opt.epoch + 1): net.confusion_matrix.reset() reporter.log_metric("lr", net.optimizer.param_groups[0]["lr"], epoch) train_loss = train_epoch( net, dataloader_train, net.optimizer, ) reporter.log_metric("train_loss", train_loss, epoch) lr_scheduler.step() scores = net.get_eval_scores(dataloader_val) kappa = CaluKappa(net.confusion_matrix) print("==> Evaluation: Epoch={} Acc={}".format(epoch, str(scores))) reporter.log_metric("eval_acc", scores["accuracy"], epoch) reporter.log_metric("kappa", kappa, epoch) if scores["accuracy"] > best_acc: best_acc = scores["accuracy"] best_kappa = kappa save_checkpoints( scores["accuracy"], net._net, reporter, opt.exp_name, epoch, ) reporter.log_metric("best_acc", best_acc, epoch) save_checkpoints( scores["accuracy"], net._net, reporter, opt.exp_name, epoch, ) print("==> Training epoch %d" % epoch) """将模型转换为torch script保存""" ckpt_name = "{}_best.pth".format(opt.exp_name) load(net, os.path.join(reporter.ckpt_log_dir, ckpt_name)) net._net.eval() traced_script_module = torch.jit.trace(net._net, torch.rand(1, 3, 256, 256)) traced_script_module.save(os.path.join(reporter.log_dir, "model.pt")) del net return old, best_acc, best_kappa, flops_after, params_after
def main(): script_dir = os.path.dirname(__file__) module_path = os.path.abspath(os.path.join(script_dir, '..', '..')) global msglogger # Parse arguments args = parser.get_parser().parse_args() if not os.path.exists(args.output_dir): os.makedirs(args.output_dir) msglogger = apputils.config_pylogger( os.path.join(script_dir, 'logging.conf'), args.name, args.output_dir) # Log various details about the execution environment. It is sometimes useful # to refer to past experiment executions and this information may be useful. # 记录有关执行环境的各种详细信息。有时是有用的 # 参考过去的实验执行,这些信息可能有用。 apputils.log_execution_env_state(args.compress, msglogger.logdir, gitroot=module_path) msglogger.debug("Distiller: %s", distiller.__version__) start_epoch = 0 perf_scores_history = [] if args.deterministic: # Experiment reproducibility is sometimes important. Pete Warden expounded about this # in his blog: https://petewarden.com/2018/03/19/the-machine-learning-reproducibility-crisis/ # In Pytorch, support for deterministic execution is still a bit clunky. if args.workers > 1: msglogger.error( 'ERROR: Setting --deterministic requires setting --workers/-j to 0 or 1' ) # 错误:设置--确定性要求将--workers/-j设置为0或1 exit(1) # 正常退出程序 # Use a well-known seed, for repeatability of experiments 使用一种众所周知的种子,用于实验的重复性。 distiller.set_deterministic() else: # This issue: https://github.com/pytorch/pytorch/issues/3659 # Implies that cudnn.benchmark should respect cudnn.deterministic, but empirically we see that # results are not re-produced when benchmark is set. So enabling only if deterministic mode disabled. cudnn.benchmark = True if args.cpu or not torch.cuda.is_available(): # Set GPU index to -1 if using CPU args.device = 'cpu' args.gpus = -1 else: args.device = 'cuda' if args.gpus is not None: try: args.gpus = [int(s) for s in args.gpus.split(',')] except ValueError: msglogger.error( 'ERROR: Argument --gpus must be a comma-separated list of integers only' ) exit(1) available_gpus = torch.cuda.device_count() for dev_id in args.gpus: if dev_id >= available_gpus: msglogger.error( 'ERROR: GPU device ID {0} requested, but only {1} devices available' .format(dev_id, available_gpus)) exit(1) # Set default device in case the first one on the list != 0 torch.cuda.set_device(args.gpus[0]) # Infer the dataset from the model name args.dataset = 'cousm' if args.earlyexit_thresholds: args.num_exits = len(args.earlyexit_thresholds) + 1 args.loss_exits = [0] * args.num_exits args.losses_exits = [] args.exiterrors = [] # Create the model model = ResNet152() # model = torch.nn.DataParallel(model, device_ids=args.gpus) # 并行GPU model.to(args.device) compression_scheduler = None # 压缩调度 # Create a couple of logging backends. TensorBoardLogger writes log files in a format # that can be read by Google's Tensor Board. PythonLogger writes to the Python logger. # 创建两个日志后端 TensorBoardLogger以Google的Tensor板可以读取的格式写入日志文件。python logger将写入python记录器。 tflogger = TensorBoardLogger(msglogger.logdir) pylogger = PythonLogger(msglogger) # capture thresholds for early-exit training if args.earlyexit_thresholds: msglogger.info('=> using early-exit threshold values of %s', args.earlyexit_thresholds) # We can optionally resume from a checkpoint if args.resume: # 加载训练模型 # checkpoint = torch.load(args.resume) # model.load_state_dict(checkpoint['state_dict']) model, compression_scheduler, start_epoch = apputils.load_checkpoint( model, chkpt_file=args.resume) model.to(args.device) # Define loss function (criterion) and optimizer # 定义损失函数和优化器SGD criterion = nn.CrossEntropyLoss().to(args.device) # optimizer = torch.optim.SGD(model.fc.parameters(), lr=args.lr, # momentum=args.momentum, # weight_decay=args.weight_decay) optimizer = torch.optim.Adam(model.model.fc.parameters(), lr=args.lr, weight_decay=args.weight_decay) msglogger.info('Optimizer Type: %s', type(optimizer)) msglogger.info('Optimizer Args: %s', optimizer.defaults) if args.AMC: # 自动化的深层压缩 return automated_deep_compression(model, criterion, optimizer, pylogger, args) if args.greedy: # 贪婪的 return greedy(model, criterion, optimizer, pylogger, args) # This sample application can be invoked to produce various summary reports. # 可以调用此示例应用程序来生成各种摘要报告。 if args.summary: return summarize_model(model, args.dataset, which_summary=args.summary) # 激活统计收集器 activations_collectors = create_activation_stats_collectors( model, *args.activation_stats) if args.qe_calibration: msglogger.info('Quantization calibration stats collection enabled:') msglogger.info( '\tStats will be collected for {:.1%} of test dataset'.format( args.qe_calibration)) msglogger.info( '\tSetting constant seeds and converting model to serialized execution' ) distiller.set_deterministic() model = distiller.make_non_parallel_copy(model) activations_collectors.update( create_quantization_stats_collector(model)) # 量化统计收集器 args.evaluate = True args.effective_test_size = args.qe_calibration # Load the datasets: the dataset to load is inferred from the model name passed # in args.arch. The default dataset is ImageNet, but if args.arch contains the # substring "_cifar", then cifar10 is used. # 加载数据集:从传递的模型名称推断要加载的数据集 train_loader, val_loader, test_loader, _ = get_data_loaders( datasets_fn, r'/home/tian/Desktop/image_yasuo', args.batch_size, args.workers, args.validation_split, args.deterministic, args.effective_train_size, args.effective_valid_size, args.effective_test_size) msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d', len(train_loader.sampler), len(val_loader.sampler), len(test_loader.sampler)) # 可以调用此示例应用程序来对模型执行敏感性分析。输出保存到csv和png。 if args.sensitivity is not None: sensitivities = np.arange(args.sensitivity_range[0], args.sensitivity_range[1], args.sensitivity_range[2]) return sensitivity_analysis(model, criterion, test_loader, pylogger, args, sensitivities) if args.evaluate: return evaluate_model(model, criterion, test_loader, pylogger, activations_collectors, args, compression_scheduler) if args.compress: # The main use-case for this sample application is CNN compression. Compression # requires a compression schedule configuration file in YAML. # #这个示例应用程序的主要用例是CNN压缩 # #需要yaml中的压缩计划配置文件。 compression_scheduler = distiller.file_config(model, optimizer, args.compress, compression_scheduler) # Model is re-transferred to GPU in case parameters were added (e.g. PACTQuantizer) # 如果添加了参数(如PactQualifier),则模型会重新传输到GPU。 model.to(args.device) elif compression_scheduler is None: compression_scheduler = distiller.CompressionScheduler(model) # 压缩计划程序 if args.thinnify: # zeros_mask_dict = distiller.create_model_masks_dict(model) assert args.resume is not None, "You must use --resume to provide a checkpoint file to thinnify" # 必须使用--resume提供检查点文件以细化 distiller.remove_filters(model, compression_scheduler.zeros_mask_dict, args.arch, args.dataset, optimizer=None) apputils.save_checkpoint(0, args.arch, model, optimizer=None, scheduler=compression_scheduler, name="{}_thinned".format( args.resume.replace(".pth.tar", "")), dir=msglogger.logdir) print( "Note: your model may have collapsed to random inference, so you may want to fine-tune" ) # 注意:您的模型可能已折叠为随机推理,因此您可能需要对其进行微调。 return args.kd_policy = None # 蒸馏 if args.kd_teacher: teacher = create_model(args.kd_pretrained, args.dataset, args.kd_teacher, device_ids=args.gpus) if args.kd_resume: teacher, _, _ = apputils.load_checkpoint(teacher, chkpt_file=args.kd_resume) dlw = distiller.DistillationLossWeights(args.kd_distill_wt, args.kd_student_wt, args.kd_teacher_wt) args.kd_policy = distiller.KnowledgeDistillationPolicy( model, teacher, args.kd_temp, dlw) compression_scheduler.add_policy(args.kd_policy, starting_epoch=args.kd_start_epoch, ending_epoch=args.epochs, frequency=1) msglogger.info('\nStudent-Teacher knowledge distillation enabled:') msglogger.info('\tTeacher Model: %s', args.kd_teacher) msglogger.info('\tTemperature: %s', args.kd_temp) msglogger.info('\tLoss Weights (distillation | student | teacher): %s', ' | '.join(['{:.2f}'.format(val) for val in dlw])) msglogger.info('\tStarting from Epoch: %s', args.kd_start_epoch) lr = args.lr lr_decay = 0.5 for epoch in range(start_epoch, args.epochs): # This is the main training loop. msglogger.info('\n') if compression_scheduler: compression_scheduler.on_epoch_begin(epoch) # Train for one epoch with collectors_context(activations_collectors["train"]) as collectors: train(train_loader, model, criterion, optimizer, epoch, compression_scheduler, loggers=[tflogger, pylogger], args=args) distiller.log_weights_sparsity(model, epoch, loggers=[tflogger, pylogger]) distiller.log_activation_statsitics( epoch, "train", loggers=[tflogger], collector=collectors["sparsity"]) if args.masks_sparsity: # 打印掩盖稀疏表 在end of each epoch msglogger.info( distiller.masks_sparsity_tbl_summary( model, compression_scheduler)) # evaluate on validation set with collectors_context(activations_collectors["valid"]) as collectors: top1, top5, vloss = validate(val_loader, model, criterion, [pylogger], args, epoch) distiller.log_activation_statsitics( epoch, "valid", loggers=[tflogger], collector=collectors["sparsity"]) save_collectors_data(collectors, msglogger.logdir) stats = ('Peformance/Validation/', OrderedDict([('Loss', vloss), ('Top1', top1), ('Top5', top5)])) distiller.log_training_progress(stats, None, epoch, steps_completed=0, total_steps=1, log_freq=1, loggers=[tflogger]) if compression_scheduler: compression_scheduler.on_epoch_end(epoch, optimizer) # Update the list of top scores achieved so far, and save the checkpoint # 更新到目前为止获得的最高分数列表,并保存检查点 sparsity = distiller.model_sparsity(model) perf_scores_history.append( distiller.MutableNamedTuple({ 'sparsity': sparsity, 'top1': top1, 'top5': top5, 'epoch': epoch })) # Keep perf_scores_history sorted from best to worst # Sort by sparsity as main sort key, then sort by top1, top5 and epoch # 保持绩效分数历史记录从最好到最差的排序 # 按稀疏度排序为主排序键,然后按top1、top5、epoch排序 perf_scores_history.sort(key=operator.attrgetter( 'sparsity', 'top1', 'top5', 'epoch'), reverse=True) for score in perf_scores_history[:args.num_best_scores]: msglogger.info( '==> Best [Top1: %.3f Top5: %.3f Sparsity: %.2f on epoch: %d]', score.top1, score.top5, score.sparsity, score.epoch) is_best = epoch == perf_scores_history[0].epoch apputils.save_checkpoint(epoch, args.arch, model, optimizer, compression_scheduler, perf_scores_history[0].top1, is_best, args.name, msglogger.logdir) if not is_best: lr = lr * lr_decay # 当loss大于上一次loss,降低学习率 for param_group in optimizer.param_groups: param_group['lr'] = lr # Finally run results on the test set # 最后在测试集上运行结果 test(test_loader, model, criterion, [pylogger], activations_collectors, args=args)