def main(): args = parser.parse_args() save_folder = args.save_folder ######## """ loss_results, cer_results, wer_results = torch.Tensor(args.epochs), torch.Tensor(args.epochs), torch.Tensor( args.epochs) best_wer = None if args.visdom: from visdom import Visdom viz = Visdom() opts = [dict(title=args.visdom_id + ' Loss', ylabel='Loss', xlabel='Epoch'), dict(title=args.visdom_id + ' WER', ylabel='WER', xlabel='Epoch'), dict(title=args.visdom_id + ' CER', ylabel='CER', xlabel='Epoch')] viz_windows = [None, None, None] epochs = torch.arange(1, args.epochs + 1) if args.tensorboard: from logger import TensorBoardLogger try: os.makedirs(args.log_dir) except OSError as e: if e.errno == errno.EEXIST: print('Directory already exists.') for file in os.listdir(args.log_dir): file_path = os.path.join(args.log_dir, file) try: if os.path.isfile(file_path): os.unlink(file_path) except Exception as e: raise else: raise logger = TensorBoardLogger(args.log_dir) try: os.makedirs(save_folder) except OSError as e: if e.errno == errno.EEXIST: print('Directory already exists.') else: raise """ ######## ######## """ criterion = CTCLoss() """ criterion = nn.CrossEntropyLoss() class_accu = tnt.meter.ClassErrorMeter(topk=[1], accuracy=True) class_accu_sum = tnt.meter.ClassErrorMeter(topk=[1], accuracy=True) class_accu_sum_120 = tnt.meter.ClassErrorMeter(topk=[1], accuracy=True) class_accu_sum_240 = tnt.meter.ClassErrorMeter(topk=[1], accuracy=True) class_accu_sum_360 = tnt.meter.ClassErrorMeter(topk=[1], accuracy=True) class_accu_sum_480 = tnt.meter.ClassErrorMeter(topk=[1], accuracy=True) ######## with open(args.labels_path) as label_file: labels = str(''.join(json.load(label_file))) audio_conf = dict(sample_rate=args.sample_rate, window_size=args.window_size, window_stride=args.window_stride, window=args.window, noise_dir=args.noise_dir, noise_prob=args.noise_prob, noise_levels=(args.noise_min, args.noise_max)) train_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=args.train_manifest, labels=labels, normalize=True, augment=args.augment) test_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=args.val_manifest, labels=labels, normalize=True, augment=False) train_loader = AudioDataLoader(train_dataset, batch_size=args.batch_size, num_workers=args.num_workers) test_loader = AudioDataLoader(test_dataset, batch_size=args.batch_size, num_workers=args.num_workers) rnn_type = args.rnn_type.lower() assert rnn_type in supported_rnns, "rnn_type should be either lstm, rnn or gru" ######## """ model = DeepSpeech(rnn_hidden_size=args.hidden_size, nb_layers=args.hidden_layers, labels=labels, rnn_type=supported_rnns[rnn_type], audio_conf=audio_conf, bidirectional=True, cnn_features=args.cnn_features) """ model = DeepSpeech(rnn_hidden_size=args.hidden_size, nb_layers=args.hidden_layers, labels=labels, rnn_type=supported_rnns[rnn_type], audio_conf=audio_conf, bidirectional=True, cnn_features=args.cnn_features, kernel=args.kernel, stride=args.stride) ######## ######## #print(list(model.rnns.modules())) #for rnn in model.rnns.modules(): # print(rnn)#.flatten_parameters() #def flat_model(model): # for m in model.modules(): # if isinstance(m, nn.LSTM): # m.flatten_parameters() ######## parameters = model.parameters() optimizer = torch.optim.SGD(parameters, lr=args.lr, momentum=args.momentum, nesterov=True) ######## #scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=args.learning_rate_decay_epochs, gamma=args.learning_rate_decay_rate) #scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.99) ######## ######## """ decoder = GreedyDecoder(labels) """ ######## ######## """ if args.continue_from: print("Loading checkpoint model %s" % args.continue_from) package = torch.load(args.continue_from) model.load_state_dict(package['state_dict']) optimizer.load_state_dict(package['optim_dict']) start_epoch = int(package.get('epoch', 1)) - 1 # Python index start at 0 for training start_iter = package.get('iteration', None) if start_iter is None: start_epoch += 1 # Assume that we saved a model after an epoch finished, so start at the next epoch. start_iter = 0 else: start_iter += 1 avg_loss = int(package.get('avg_loss', 0)) loss_results, cer_results, wer_results = package['loss_results'], package[ 'cer_results'], package['wer_results'] if args.visdom and \ package['loss_results'] is not None and start_epoch > 0: # Add previous scores to visdom graph x_axis = epochs[0:start_epoch] y_axis = [loss_results[0:start_epoch], wer_results[0:start_epoch], cer_results[0:start_epoch]] for x in range(len(viz_windows)): viz_windows[x] = viz.line( X=x_axis, Y=y_axis[x], opts=opts[x], ) if args.tensorboard and \ package['loss_results'] is not None and start_epoch > 0: # Previous scores to tensorboard logs for i in range(start_epoch): info = { 'Avg Train Loss': loss_results[i], 'Avg WER': wer_results[i], 'Avg CER': cer_results[i] } for tag, val in info.items(): logger.scalar_summary(tag, val, i + 1) if not args.no_bucketing and epoch != 0: print("Using bucketing sampler for the following epochs") train_dataset = SpectrogramDatasetWithLength(audio_conf=audio_conf, manifest_filepath=args.train_manifest, labels=labels, normalize=True, augment=args.augment) sampler = BucketingSampler(train_dataset) train_loader.sampler = sampler else: avg_loss = 0 start_epoch = 0 start_iter = 0 """ avg_loss = 0 start_epoch = 0 start_iter = 0 best_train_accu = 0 best_train_accu_sum = 0 best_train_accu_sum_120 = 0 best_train_accu_sum_240 = 0 best_train_accu_sum_360 = 0 best_train_accu_sum_480 = 0 best_test_accu = 0 best_test_accu_sum = 0 best_test_accu_sum_120 = 0 best_test_accu_sum_240 = 0 best_test_accu_sum_360 = 0 best_test_accu_sum_480 = 0 best_avg_loss = float("inf") # sys.float_info.max # 1000000 epoch_70 = None epoch_90 = None epoch_95 = None epoch_99 = None if args.stride == 1: multiplier = 6 if args.stride == 2: multiplier = 3 if args.stride == 3: multiplier = 2 if args.stride == 4: multiplier = 1 # (Should be 1.5...) #sample_time_steps = int(args.sample_miliseconds / 10) loss_begin = round(args.crop_begin / (10 * args.stride)) loss_end = -round(args.crop_end / (10 * args.stride)) or None print("LOSS BEGIN:", loss_begin) print("LOSS END:", loss_end) ######## if args.cuda: model = torch.nn.DataParallel(model).cuda() print(model) print("Number of parameters: %d" % DeepSpeech.get_param_size(model)) batch_time = AverageMeter() data_time = AverageMeter() losses = AverageMeter() for epoch in range(start_epoch, args.epochs): ######## #scheduler.step() optim_state_now = optimizer.state_dict() print('\nLEARNING RATE: {lr:.6f}'.format(lr=optim_state_now['param_groups'][0]['lr'])) class_accu.reset() class_accu_sum.reset() class_accu_sum_120.reset() class_accu_sum_240.reset() class_accu_sum_360.reset() class_accu_sum_480.reset() ######## model.train() end = time.time() for i, (data) in enumerate(train_loader, start=start_iter): if i == len(train_loader): break ######## """ inputs, targets, input_percentages, target_sizes = data """ inputs, targets, input_percentages, target_sizes, speaker_labels = data ######## # measure data loading time data_time.update(time.time() - end) inputs = Variable(inputs, requires_grad=False) ######## """ target_sizes = Variable(target_sizes, requires_grad=False) targets = Variable(targets, requires_grad=False) """ speaker_labels = Variable(speaker_labels, requires_grad=False) ######## if args.cuda: inputs = inputs.cuda() ######## """ out = model(inputs) """ #temp_random = random.randint(0, (inputs.size(3)-1)-sample_time_steps) #print("INPUT", inputs[...,temp_random:temp_random+sample_time_steps].size(),temp_random, temp_random+sample_time_steps) #out = model(inputs[...,temp_random:temp_random+sample_time_steps]) #print("OUTPUT", out.size()) start = random.randint(0, int((inputs.size(3)-1)*(1-args.sample_proportion))) print("INPUT", inputs.size(3), inputs[...,start:start+int((inputs.size(3))*(args.sample_proportion))].size(),start, start+int((inputs.size(3))*(args.sample_proportion))) out = model(inputs[...,start:start+int((inputs.size(3))*(args.sample_proportion))]) print("OUTPUT", out.size()) ######## out = out.transpose(0, 1) # TxNxH ######## speaker_labels = speaker_labels.cuda(async=True).long() # Prints the output of the model in a sequence of probabilities of char for each audio... torch.set_printoptions(profile="full") ####print("OUT: " + str(out.size()), "SPEAKER LABELS:" + str(speaker_labels.size()), "INPUT PERCENTAGES MEAN: " + str(input_percentages.mean())) """ seq_length = out.size(0) sizes = Variable(input_percentages.mul_(int(seq_length)).int(), requires_grad=False) loss = criterion(out, targets, sizes, target_sizes) """ #print(out[:,:,0]) #print("SPEAKER LABELS: " + str(speaker_labels)) #print(out[0][0]) #softmax_output = F.softmax(out).data # This DOES NOT what I want... #softmax_output_alt = flex_softmax(out, axis=2).data # This is FINE!!! <<<=== #print(softmax_output[0][0]) #print(softmax_output_alt[0][0]) ####new_out = torch.sum(out, 0) ####new_out = torch.sum(out[20:], 0) #print(out.size()) #print(new_out.size()) #print(out[-1].size()) ######## ######## if args.loss_type == "reg": #loss_out = out[-1]; loss_speaker_labels = speaker_labels loss_out = out[round(out.size(0)/2)]; loss_speaker_labels = speaker_labels #print("LOSS TYPE = REGULAR") elif args.loss_type == "sum": loss_out = torch.sum(out[loss_begin:loss_end], 0); loss_speaker_labels = speaker_labels #print("LOSS TYPE = SUM") elif args.loss_type == "full": # Don't know if is ok!!! Don't use!!! => loss_out = out.contiguous().view(-1,48); loss_speaker_labels = speaker_labels.repeat(out.size(0)) #speaker_labels = speaker_labels.expand(20, out.size(0)) # Don't know if is ok!!! Don't use!!! => loss_out = out.contiguous().view(-1,48); loss_speaker_labels = speaker_labels.repeat(1, out.size(0)).squeeze() #speaker_labels = speaker_labels.expand(20, out.size(0)) loss_out = out.contiguous()[loss_begin:loss_end].view(-1,48); loss_speaker_labels = speaker_labels.repeat(out.size(0),1)[loss_begin:loss_end].view(-1) #speaker_labels = speaker_labels.expand(20, out.size(0)) #print("LOSS TYPE = FULL") print("LOSS_OUT: " + str(loss_out.size()), "SPEAKER LABELS:" + str(loss_speaker_labels.size())) loss = criterion(loss_out, loss_speaker_labels) ######## loss = loss / inputs.size(0) # average the loss by minibatch loss_sum = loss.data.sum() inf = float("inf") if loss_sum == inf or loss_sum == -inf: print("WARNING: received an inf loss, setting loss value to 0") loss_value = 0 else: loss_value = loss.data[0] avg_loss += loss_value losses.update(loss_value, inputs.size(0)) ######## #if args.stride == 1: multiplier = 6 #if args.stride == 2: multiplier = 3 #if args.stride == 3: multiplier = 2 #if args.stride == 4: multiplier = 1 #(Should be 1.5...) #if args.stride == 5: multiplier = 1 #(Should be 1.25...) class_accu.add(out[round(out.size(0)/2)].data, speaker_labels.data) class_accu_sum.add(torch.sum(out, 0).data, speaker_labels.data) #class_accu_sum_120.add(torch.sum(out[1*multiplier:-1*multiplier], 0).data, speaker_labels.data) #class_accu_sum_240.add(torch.sum(out[2*multiplier:-2*multiplier], 0).data, speaker_labels.data) #class_accu_sum_360.add(torch.sum(out[3*multiplier:-3*multiplier], 0).data, speaker_labels.data) #class_accu_sum_480.add(torch.sum(out[4*multiplier:-4*multiplier], 0).data, speaker_labels.data) ####class_accu_sum_120.add(torch.sum(out[round(out.size(0)/2)-1*multiplier:round(out.size(0)/2)+1*multiplier], 0).data, speaker_labels.data) ####class_accu_sum_240.add(torch.sum(out[round(out.size(0)/2)-2*multiplier:round(out.size(0)/2)+2*multiplier], 0).data, speaker_labels.data) ####class_accu_sum_360.add(torch.sum(out[round(out.size(0)/2)-3*multiplier:round(out.size(0)/2)+3*multiplier], 0).data, speaker_labels.data) ####class_accu_sum_480.add(torch.sum(out[round(out.size(0)/2)-4*multiplier:round(out.size(0)/2)+4*multiplier], 0).data, speaker_labels.data) #accu_out3 = torch.sum(flex_softmax(out[20:], axis=2), 0) #print(classaccu.value()[0], classaccu.value()[1]) # Cross Entropy Loss for a Sequence (Time Series) of Output? #output = output.view(-1,29) #target = target.view(-1) #criterion = nn.CrossEntropyLoss() #loss = criterion(output,target) ######## # compute gradient optimizer.zero_grad() loss.backward() torch.nn.utils.clip_grad_norm(model.parameters(), args.max_norm) # SGD step optimizer.step() if args.cuda: torch.cuda.synchronize() # measure elapsed time batch_time.update(time.time() - end) end = time.time() if not args.silent: ######## """ print('Epoch: [{0}][{1}/{2}]\t' 'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t' 'Data {data_time.val:.3f} ({data_time.avg:.3f})\t' 'Loss {loss.val:.4f} ({loss.avg:.4f})\t'.format( (epoch + 1), (i + 1), len(train_loader), batch_time=batch_time, data_time=data_time, loss=losses)) """ print('Epoch: [{0}][{1}/{2}]\t' # 'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t' # 'Data {data_time.val:.3f} ({data_time.avg:.3f})\t' 'Loss {loss.val:.4f} ({loss.avg:.4f})\t' 'CAR {car:.3f}\t' 'CAR_SUM {car_sum:.3f}\t' #'CAR_SUM_120 {car_sum_120:.3f}\t' #'CAR_SUM_240 {car_sum_240:.3f}\t' #'CAR_SUM_360 {car_sum_360:.3f}\t' #'CAR_SUM_480 {car_sum_480:.3f}\t' .format((epoch + 1), (i + 1), len(train_loader), batch_time=batch_time, data_time=data_time, loss=losses, car=class_accu.value()[0], car_sum=class_accu_sum.value()[0], # car_sum_240=class_accu_sum_240.value()[0], car_sum_120=class_accu_sum_120.value()[0], # car_sum_360=class_accu_sum_360.value()[0], car_sum_480=class_accu_sum_480.value()[0] ) ) ######## ######## """ if args.checkpoint_per_batch > 0 and i > 0 and (i + 1) % args.checkpoint_per_batch == 0: file_path = '%s/deepspeech_checkpoint_epoch_%d_iter_%d.pth.tar' % (save_folder, epoch + 1, i + 1) print("Saving checkpoint model to %s" % file_path) torch.save(DeepSpeech.serialize(model, optimizer=optimizer, epoch=epoch, iteration=i, loss_results=loss_results, wer_results=wer_results, cer_results=cer_results, avg_loss=avg_loss), file_path) """ ######## del loss del out ######## del loss_out del speaker_labels del loss_speaker_labels ######## avg_loss /= len(train_loader) ######## """ print('Training Summary Epoch: [{0}]\t' 'Average Loss {loss:.3f}\t'.format( epoch + 1, loss=avg_loss)) """ if (best_avg_loss > avg_loss): best_avg_loss = avg_loss print("\nCURRENT EPOCH TRAINING RESULTS:\t", class_accu.value()[0], "\t", class_accu_sum.value()[0],"\t", #class_accu_sum_120.value()[0], class_accu_sum_240.value()[0], class_accu_sum_360.value()[0], "\t", class_accu_sum_480.value()[0], "\n" ) if (best_train_accu < class_accu.value()[0]): best_train_accu = class_accu.value()[0] if (best_train_accu_sum < class_accu_sum.value()[0]): best_train_accu_sum = class_accu_sum.value()[0] #if (best_train_accu_sum_120 < class_accu_sum_120.value()[0]): best_train_accu_sum_120 = class_accu_sum_120.value()[0] #if (best_train_accu_sum_240 < class_accu_sum_240.value()[0]): best_train_accu_sum_240 = class_accu_sum_240.value()[0] #if (best_train_accu_sum_360 < class_accu_sum_360.value()[0]): best_train_accu_sum_360 = class_accu_sum_360.value()[0] #if (best_train_accu_sum_480 < class_accu_sum_480.value()[0]): best_train_accu_sum_480 = class_accu_sum_480.value()[0] get_70 = ((class_accu.value()[0] > 70) or (class_accu_sum.value()[0] > 70) #or (class_accu_sum_120.value()[0] > 70) or (class_accu_sum_240.value()[0] > 70) #or (class_accu_sum_360.value()[0] > 70) or (class_accu_sum_480.value()[0] > 70) ) if ((epoch_70 is None) and (get_70 == True)): epoch_70 = epoch + 1 get_90 = ((class_accu.value()[0] > 90) or (class_accu_sum.value()[0] > 90) #or (class_accu_sum_120.value()[0] > 90) or (class_accu_sum_240.value()[0] > 90) #or (class_accu_sum_360.value()[0] > 90) or (class_accu_sum_480.value()[0] > 90) ) if ((epoch_90 is None) and (get_90 == True)): epoch_90 = epoch + 1 get_95 = ((class_accu.value()[0] > 95) or (class_accu_sum.value()[0] > 95) #or (class_accu_sum_120.value()[0] > 95) or (class_accu_sum_240.value()[0] > 95) #or (class_accu_sum_360.value()[0] > 95) or (class_accu_sum_480.value()[0] > 95) ) if ((epoch_95 is None) and (get_95 == True)): epoch_95 = epoch + 1 get_99 = ((class_accu.value()[0] > 99) or (class_accu_sum.value()[0] > 99) #or (class_accu_sum_120.value()[0] > 99) or (class_accu_sum_240.value()[0] > 99) #or (class_accu_sum_360.value()[0] > 99) or (class_accu_sum_480.value()[0] > 99) ) if ((epoch_99 is None) and (get_99 == True)): epoch_99 = epoch + 1 ######## start_iter = 0 # Reset start iteration for next epoch total_cer, total_wer = 0, 0 model.eval() ######## class_accu.reset() class_accu_sum.reset() class_accu_sum_120.reset() class_accu_sum_240.reset() class_accu_sum_360.reset() class_accu_sum_480.reset() ######## for i, (data) in enumerate(test_loader): # test ######## """ inputs, targets, input_percentages, target_sizes = data """ inputs, targets, input_percentages, target_sizes, speaker_labels = data ######## inputs = Variable(inputs, volatile=True) ######## speaker_labels = Variable(speaker_labels, requires_grad=False) speaker_labels = speaker_labels.cuda(async=True).long() """ # unflatten targets split_targets = [] offset = 0 for size in target_sizes: split_targets.append(targets[offset:offset + size]) offset += size """ ######## if args.cuda: inputs = inputs.cuda() out = model(inputs) out = out.transpose(0, 1) # TxNxH ######## speaker_labels = speaker_labels.cuda(async=True).long() # Prints the output of the model in a sequence of probabilities of char for each audio... torch.set_printoptions(profile="full") ########print("OUT: " + str(out.size()), "NEW OUT:" + str(new_out.size()), "SPEAKER LABELS:" + str(speaker_labels.size()), "INPUT PERCENTAGES MEAN: " + str(input_percentages.mean())) #print(out[:,:,0]) #print("SPEAKER LABELS: " + str(speaker_labels)) #print(out[0][0]) #softmax_output = F.softmax(out).data # This DOES NOT what I want... #softmax_output_alt = flex_softmax(out, axis=2).data # This is FINE!!! <<<=== #print(softmax_output[0][0]) #print(softmax_output_alt[0][0]) ######## ######## #if args.stride == 1: multiplier = 6 #if args.stride == 2: multiplier = 3 #if args.stride == 3: multiplier = 2 #if args.stride == 4: multiplier = 1 #(Should be 1.5...) #if args.stride == 5: multiplier = 1 #(Should be 1.25...) class_accu.add(out[round(out.size(0)/2)].data, speaker_labels.data) class_accu_sum.add(torch.sum(out, 0).data, speaker_labels.data) class_accu_sum_120.add(torch.sum(out[1*multiplier:-1*multiplier], 0).data, speaker_labels.data) class_accu_sum_240.add(torch.sum(out[2*multiplier:-2*multiplier], 0).data, speaker_labels.data) class_accu_sum_360.add(torch.sum(out[3*multiplier:-3*multiplier], 0).data, speaker_labels.data) class_accu_sum_480.add(torch.sum(out[4*multiplier:-4*multiplier], 0).data, speaker_labels.data) #class_accu_sum_120.add(torch.sum(out[round(out.size(0)/2)-1*multiplier:round(out.size(0)/2)+1*multiplier], 0).data, speaker_labels.data) #class_accu_sum_240.add(torch.sum(out[round(out.size(0)/2)-2*multiplier:round(out.size(0)/2)+2*multiplier], 0).data, speaker_labels.data) #class_accu_sum_360.add(torch.sum(out[round(out.size(0)/2)-3*multiplier:round(out.size(0)/2)+3*multiplier], 0).data, speaker_labels.data) #class_accu_sum_480.add(torch.sum(out[round(out.size(0)/2)-4*multiplier:round(out.size(0)/2)+4*multiplier], 0).data, speaker_labels.data) #accu_out3 = torch.sum(flex_softmax(out[20:], axis=2), 0) #print(classaccu.value()[0], classaccu.value()[1]) # Cross Entropy Loss for a Sequence (Time Series) of Output? #output = output.view(-1,29) #target = target.view(-1) #criterion = nn.CrossEntropyLoss() #loss = criterion(output,target) print('Validation Summary Epoch: [{0}]\t' 'CAR {car:.3f}\t' 'CAR_SUM {car_sum:.3f}\t' 'CAR_SUM_120 {car_sum_120:.3f}\t' 'CAR_SUM_240 {car_sum_240:.3f}\t' 'CAR_SUM_360 {car_sum_360:.3f}\t' 'CAR_SUM_480 {car_sum_480:.3f}\t' .format(epoch + 1, car=class_accu.value()[0], car_sum=class_accu_sum.value()[0], car_sum_240=class_accu_sum_240.value()[0], car_sum_120=class_accu_sum_120.value()[0], car_sum_360=class_accu_sum_360.value()[0], car_sum_480=class_accu_sum_480.value()[0] ) ) """ seq_length = out.size(0) sizes = input_percentages.mul_(int(seq_length)).int() decoded_output = decoder.decode(out.data, sizes) target_strings = decoder.process_strings(decoder.convert_to_strings(split_targets)) wer, cer = 0, 0 for x in range(len(target_strings)): wer += decoder.wer(decoded_output[x], target_strings[x]) / float(len(target_strings[x].split())) cer += decoder.cer(decoded_output[x], target_strings[x]) / float(len(target_strings[x])) total_cer += cer total_wer += wer """ ######## if args.cuda: torch.cuda.synchronize() del out ######## """ wer = total_wer / len(test_loader.dataset) cer = total_cer / len(test_loader.dataset) wer *= 100 cer *= 100 loss_results[epoch] = avg_loss wer_results[epoch] = wer cer_results[epoch] = cer print('Validation Summary Epoch: [{0}]\t' 'Average WER {wer:.3f}\t' 'Average CER {cer:.3f}\t'.format( epoch + 1, wer=wer, cer=cer)) """ ######## ######## print("\nCURRENT EPOCH TEST RESULTS:\t", class_accu.value()[0], "\t", class_accu_sum.value()[0], "\t", class_accu_sum_120.value()[0], "\t", class_accu_sum_240.value()[0], "\t", class_accu_sum_360.value()[0], "\t", class_accu_sum_480.value()[0], "\n") if (best_test_accu < class_accu.value()[0]): best_test_accu = class_accu.value()[0] if (best_test_accu_sum < class_accu_sum.value()[0]): best_test_accu_sum = class_accu_sum.value()[0] if (best_test_accu_sum_120 < class_accu_sum_120.value()[0]): best_test_accu_sum_120 = class_accu_sum_120.value()[0] if (best_test_accu_sum_240 < class_accu_sum_240.value()[0]): best_test_accu_sum_240 = class_accu_sum_240.value()[0] if (best_test_accu_sum_360 < class_accu_sum_360.value()[0]): best_test_accu_sum_360 = class_accu_sum_360.value()[0] if (best_test_accu_sum_480 < class_accu_sum_480.value()[0]): best_test_accu_sum_480 = class_accu_sum_480.value()[0] print("\nBEST EPOCH TRAINING RESULTS:\t", best_train_accu, "\t", best_train_accu_sum, "\t", best_train_accu_sum_120, "\t", best_train_accu_sum_240, "\t", best_train_accu_sum_360, "\t", best_train_accu_sum_480) print("\nBEST EPOCH TEST RESULTS:\t", best_test_accu, "\t", best_test_accu_sum, "\t", best_test_accu_sum_120, "\t", best_test_accu_sum_240, "\t", best_test_accu_sum_360, "\t", best_test_accu_sum_480) print("\nEPOCHS 70%, 90%, 95%, 99%:\t", epoch_70, "\t", epoch_90, "\t", epoch_95, "\t", epoch_99) print("\nBEST AVERAGE LOSS:\t", best_avg_loss, "\n") ######## ######## """ if args.visdom: # epoch += 1 x_axis = epochs[0:epoch + 1] y_axis = [loss_results[0:epoch + 1], wer_results[0:epoch + 1], cer_results[0:epoch + 1]] for x in range(len(viz_windows)): if viz_windows[x] is None: viz_windows[x] = viz.line( X=x_axis, Y=y_axis[x], opts=opts[x], ) else: viz.line( X=x_axis, Y=y_axis[x], win=viz_windows[x], update='replace', ) if args.tensorboard: info = { 'Avg Train Loss': avg_loss, 'Avg WER': wer, 'Avg CER': cer } for tag, val in info.items(): logger.scalar_summary(tag, val, epoch + 1) if args.log_params: for tag, value in model.named_parameters(): tag = tag.replace('.', '/') logger.histo_summary(tag, to_np(value), epoch + 1) logger.histo_summary(tag + '/grad', to_np(value.grad), epoch + 1) if args.checkpoint: file_path = '%s/deepspeech_%d.pth.tar' % (save_folder, epoch + 1) torch.save(DeepSpeech.serialize(model, optimizer=optimizer, epoch=epoch, loss_results=loss_results, wer_results=wer_results, cer_results=cer_results), file_path) # anneal lr optim_state = optimizer.state_dict() optim_state['param_groups'][0]['lr'] = optim_state['param_groups'][0]['lr'] / args.learning_anneal optimizer.load_state_dict(optim_state) print('Learning rate annealed to: {lr:.6f}'.format(lr=optim_state['param_groups'][0]['lr'])) if best_wer is None or best_wer > wer: print("Found better validated model, saving to %s" % args.model_path) torch.save(DeepSpeech.serialize(model, optimizer=optimizer, epoch=epoch, loss_results=loss_results, wer_results=wer_results, cer_results=cer_results) , args.model_path) best_wer = wer """ ######## avg_loss = 0 if not args.no_bucketing and epoch == 0: print("Switching to bucketing sampler for following epochs") train_dataset = SpectrogramDatasetWithLength(audio_conf=audio_conf, manifest_filepath=args.train_manifest, labels=labels, normalize=True, augment=args.augment) sampler = BucketingSampler(train_dataset) train_loader.sampler = sampler
def main(): args = parser.parse_args() save_folder = args.save_folder loss_results, cer_results, wer_results = torch.Tensor( args.epochs), torch.Tensor(args.epochs), torch.Tensor(args.epochs) if args.visdom: from visdom import Visdom viz = Visdom() opts = [ dict(title='Loss', ylabel='Loss', xlabel='Epoch'), dict(title='WER', ylabel='WER', xlabel='Epoch'), dict(title='CER', ylabel='CER', xlabel='Epoch') ] viz_windows = [None, None, None] epochs = torch.arange(1, args.epochs + 1) if args.tensorboard: from logger import TensorBoardLogger try: os.makedirs(args.log_dir) except OSError as e: if e.errno == errno.EEXIST: print('Directory already exists.') for file in os.listdir(args.log_dir): file_path = os.path.join(args.log_dir, file) try: if os.path.isfile(file_path): os.unlink(file_path) except Exception as e: raise else: raise logger = TensorBoardLogger(args.log_dir) try: os.makedirs(save_folder) except OSError as e: if e.errno == errno.EEXIST: print('Directory already exists.') else: raise criterion = CTCLoss() with open(args.labels_path) as label_file: labels = str(''.join(json.load(label_file))) audio_conf = dict(sample_rate=args.sample_rate, window_size=args.window_size, window_stride=args.window_stride, window=args.window, noise_dir=args.noise_dir, noise_prob=args.noise_prob, noise_levels=(args.noise_min, args.noise_max)) train_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=args.train_manifest, labels=labels, normalize=True, augment=args.augment) test_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=args.val_manifest, labels=labels, normalize=True, augment=False) train_loader = AudioDataLoader(train_dataset, batch_size=args.batch_size, num_workers=args.num_workers) test_loader = AudioDataLoader(test_dataset, batch_size=args.batch_size, num_workers=args.num_workers) rnn_type = args.rnn_type.lower() assert rnn_type in supported_rnns, "rnn_type should be either lstm, rnn or gru" model = DeepSpeech(rnn_hidden_size=args.hidden_size, nb_layers=args.hidden_layers, labels=labels, rnn_type=supported_rnns[rnn_type], audio_conf=audio_conf, bidirectional=True) parameters = model.parameters() optimizer = torch.optim.SGD(parameters, lr=args.lr, momentum=args.momentum, nesterov=True) decoder = GreedyDecoder(labels) if args.continue_from: print("Loading checkpoint model %s" % args.continue_from) package = torch.load(args.continue_from) model.load_state_dict(package['state_dict']) optimizer.load_state_dict(package['optim_dict']) start_epoch = int(package.get( 'epoch', 1)) - 1 # Python index start at 0 for training start_iter = package.get('iteration', None) if start_iter is None: start_epoch += 1 # Assume that we saved a model after an epoch finished, so start at the next epoch. start_iter = 0 else: start_iter += 1 avg_loss = int(package.get('avg_loss', 0)) loss_results, cer_results, wer_results = package[ 'loss_results'], package['cer_results'], package['wer_results'] if args.visdom and \ package['loss_results'] is not None and start_epoch > 0: # Add previous scores to visdom graph x_axis = epochs[0:start_epoch] y_axis = [ loss_results[0:start_epoch], wer_results[0:start_epoch], cer_results[0:start_epoch] ] for x in range(len(viz_windows)): viz_windows[x] = viz.line( X=x_axis, Y=y_axis[x], opts=opts[x], ) if args.tensorboard and \ package['loss_results'] is not None and start_epoch > 0: # Previous scores to tensorboard logs for i in range(start_epoch): info = { 'Avg Train Loss': loss_results[i], 'Avg WER': wer_results[i], 'Avg CER': cer_results[i] } for tag, val in info.items(): logger.scalar_summary(tag, val, i + 1) if not args.no_bucketing: print("Using bucketing sampler for the following epochs") train_dataset = SpectrogramDatasetWithLength( audio_conf=audio_conf, manifest_filepath=args.train_manifest, labels=labels, normalize=True, augment=args.augment) sampler = BucketingSampler(train_dataset) train_loader.sampler = sampler else: avg_loss = 0 start_epoch = 0 start_iter = 0 if args.cuda: model = torch.nn.DataParallel(model).cuda() print(model) print("Number of parameters: %d" % DeepSpeech.get_param_size(model)) batch_time = AverageMeter() data_time = AverageMeter() losses = AverageMeter() for epoch in range(start_epoch, args.epochs): model.train() end = time.time() for i, (data) in enumerate(train_loader, start=start_iter): if i == len(train_loader): break inputs, targets, input_percentages, target_sizes = data # measure data loading time data_time.update(time.time() - end) inputs = Variable(inputs, requires_grad=False) target_sizes = Variable(target_sizes, requires_grad=False) targets = Variable(targets, requires_grad=False) if args.cuda: inputs = inputs.cuda() out = model(inputs) out = out.transpose(0, 1) # TxNxH seq_length = out.size(0) sizes = Variable(input_percentages.mul_(int(seq_length)).int(), requires_grad=False) loss = criterion(out, targets, sizes, target_sizes) loss = loss / inputs.size(0) # average the loss by minibatch loss_sum = loss.data.sum() inf = float("inf") if loss_sum == inf or loss_sum == -inf: print("WARNING: received an inf loss, setting loss value to 0") loss_value = 0 else: loss_value = loss.data[0] avg_loss += loss_value losses.update(loss_value, inputs.size(0)) # compute gradient optimizer.zero_grad() loss.backward() torch.nn.utils.clip_grad_norm(model.parameters(), args.max_norm) # SGD step optimizer.step() if args.cuda: torch.cuda.synchronize() # measure elapsed time batch_time.update(time.time() - end) end = time.time() if not args.silent: print('Epoch: [{0}][{1}/{2}]\t' 'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t' 'Data {data_time.val:.3f} ({data_time.avg:.3f})\t' 'Loss {loss.val:.4f} ({loss.avg:.4f})\t'.format( (epoch + 1), (i + 1), len(train_loader), batch_time=batch_time, data_time=data_time, loss=losses)) if args.checkpoint_per_batch > 0 and i > 0 and ( i + 1) % args.checkpoint_per_batch == 0: file_path = '%s/deepspeech_checkpoint_epoch_%d_iter_%d.pth.tar' % ( save_folder, epoch + 1, i + 1) print("Saving checkpoint model to %s" % file_path) torch.save( DeepSpeech.serialize(model, optimizer=optimizer, epoch=epoch, iteration=i, loss_results=loss_results, wer_results=wer_results, cer_results=cer_results, avg_loss=avg_loss), file_path) del loss del out avg_loss /= len(train_loader) print('Training Summary Epoch: [{0}]\t' 'Average Loss {loss:.3f}\t'.format(epoch + 1, loss=avg_loss)) start_iter = 0 # Reset start iteration for next epoch total_cer, total_wer = 0, 0 model.eval() for i, (data) in enumerate(test_loader): # test inputs, targets, input_percentages, target_sizes = data inputs = Variable(inputs, volatile=True) # unflatten targets split_targets = [] offset = 0 for size in target_sizes: split_targets.append(targets[offset:offset + size]) offset += size if args.cuda: inputs = inputs.cuda() out = model(inputs) out = out.transpose(0, 1) # TxNxH seq_length = out.size(0) sizes = input_percentages.mul_(int(seq_length)).int() decoded_output = decoder.decode(out.data, sizes) target_strings = decoder.process_strings( decoder.convert_to_strings(split_targets)) wer, cer = 0, 0 for x in range(len(target_strings)): wer += decoder.wer(decoded_output[x], target_strings[x]) / float( len(target_strings[x].split())) cer += decoder.cer(decoded_output[x], target_strings[x]) / float( len(target_strings[x])) total_cer += cer total_wer += wer if args.cuda: torch.cuda.synchronize() del out wer = total_wer / len(test_loader.dataset) cer = total_cer / len(test_loader.dataset) wer *= 100 cer *= 100 loss_results[epoch] = avg_loss wer_results[epoch] = wer cer_results[epoch] = cer print('Validation Summary Epoch: [{0}]\t' 'Average WER {wer:.3f}\t' 'Average CER {cer:.3f}\t'.format(epoch + 1, wer=wer, cer=cer)) if args.visdom: # epoch += 1 x_axis = epochs[0:epoch + 1] y_axis = [ loss_results[0:epoch + 1], wer_results[0:epoch + 1], cer_results[0:epoch + 1] ] for x in range(len(viz_windows)): if viz_windows[x] is None: viz_windows[x] = viz.line( X=x_axis, Y=y_axis[x], opts=opts[x], ) else: viz.line( X=x_axis, Y=y_axis[x], win=viz_windows[x], update='replace', ) if args.tensorboard: info = {'Avg Train Loss': avg_loss, 'Avg WER': wer, 'Avg CER': cer} for tag, val in info.items(): logger.scalar_summary(tag, val, epoch + 1) if args.log_params: for tag, value in model.named_parameters(): tag = tag.replace('.', '/') logger.histo_summary(tag, to_np(value), epoch + 1) logger.histo_summary(tag + '/grad', to_np(value.grad), epoch + 1) if args.checkpoint: file_path = '%s/deepspeech_%d.pth.tar' % (save_folder, epoch + 1) torch.save( DeepSpeech.serialize(model, optimizer=optimizer, epoch=epoch, loss_results=loss_results, wer_results=wer_results, cer_results=cer_results), file_path) # anneal lr optim_state = optimizer.state_dict() optim_state['param_groups'][0][ 'lr'] = optim_state['param_groups'][0]['lr'] / args.learning_anneal optimizer.load_state_dict(optim_state) print('Learning rate annealed to: {lr:.6f}'.format( lr=optim_state['param_groups'][0]['lr'])) avg_loss = 0 if not args.no_bucketing and epoch == 0: print("Switching to bucketing sampler for following epochs") train_dataset = SpectrogramDatasetWithLength( audio_conf=audio_conf, manifest_filepath=args.train_manifest, labels=labels, normalize=True, augment=args.augment) sampler = BucketingSampler(train_dataset) train_loader.sampler = sampler torch.save(DeepSpeech.serialize(model, optimizer=optimizer), args.final_model_path)
def main(): args = parser.parse_args() torch.set_printoptions(profile="full") criterion = nn.CrossEntropyLoss() class_accu_reg = tnt.meter.ClassErrorMeter(topk=[1], accuracy=True) class_accu_sum = tnt.meter.ClassErrorMeter(topk=[1], accuracy=True) audio_conf = dict(sample_rate=args.sample_rate, window_size=args.window_size, window_stride=args.window_stride, window=args.window, noise_dir=args.noise_dir, noise_prob=args.noise_prob, noise_levels=(args.noise_min, args.noise_max)) train_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=args.train_manifest, normalize=True, augment=args.augment) test_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=args.val_manifest, normalize=True, augment=False) train_loader = AudioDataLoader(train_dataset, batch_size=args.batch_size, num_workers=args.num_workers) test_loader = AudioDataLoader(test_dataset, batch_size=args.batch_size, num_workers=args.num_workers) rnn_type = args.rnn_type.lower() assert rnn_type in supported_rnns, "rnn_type should be either lstm, rnn or gru" #print("FIRST LAYER TYPE:\t", args.first_layer_type) #print("MFCC TRANSFORM:\t\t", args.mfcc) model = DeepSpeech(rnn_hidden_size=args.hidden_size, nb_layers=args.hidden_layers, rnn_type=supported_rnns[rnn_type], audio_conf=audio_conf, bidirectional=True, cnn_features=args.cnn_features, kernel=args.kernel, first_layer_type=args.first_layer_type, stride=args.stride, mfcc=args.mfcc) ######## #print(list(model.rnns.modules())) #for rnn in model.rnns.modules(): # print(rnn)#.flatten_parameters() #def flat_model(model): # for m in model.modules(): # if isinstance(m, nn.LSTM): # m.flatten_parameters() ######## parameters = model.parameters() optimizer = torch.optim.SGD(parameters, lr=args.lr, momentum=args.momentum, nesterov=True) #scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=args.learning_rate_decay_epochs, gamma=args.learning_rate_decay_rate) scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.99) avg_loss = 0 start_epoch = 0 start_iter = 0 best_train_accu_reg = 0 best_train_accu_sum = 0 best_test_accu_reg = 0 best_test_accu_sum = 0 best_avg_loss = float("inf") # sys.float_info.max # 1000000 epoch_70 = None epoch_90 = None epoch_95 = None epoch_99 = None utterance_sequence_length = int(args.utterance_miliseconds / 10) loss_begin = round(args.crop_begin / (10 * args.stride)) loss_end = -round(args.crop_end / (10 * args.stride)) or None gap = loss_begin print("LOSS BEGIN:", loss_begin) print("LOSS END:", loss_end) if args.cuda: model = torch.nn.DataParallel(model).cuda() print(model) print("Number of parameters: ", DeepSpeech.get_param_size(model)) batch_time = AverageMeter() data_time = AverageMeter() #losses = AverageMeter() print(args, "\n") for epoch in range(start_epoch, args.epochs): losses = AverageMeter() scheduler.step() optim_state_now = optimizer.state_dict() print('\nLEARNING RATE: {lr:.6f}'.format( lr=optim_state_now['param_groups'][0]['lr'])) class_accu_reg.reset() class_accu_sum.reset() model.train() end = time.time() for i, (data) in enumerate(train_loader, start=start_iter): if i == len(train_loader): break inputs, input_percentages, speaker_labels, mfccs = data # measure data loading time data_time.update(time.time() - end) inputs = Variable(inputs, requires_grad=False) ######## mfccs = Variable(mfccs, requires_grad=False) if args.mfcc == "true": inputs = mfccs # <<-- This line makes us to use mfccs... #print("INPUTS SIZE:", inputs.size()) #print("MFCCS SIZE:", mfccs.size()) ######## speaker_labels = Variable(speaker_labels, requires_grad=False) speaker_labels = speaker_labels.cuda(async=True).long() if args.cuda: inputs = inputs.cuda() ######## ######## sizes = inputs.size() inputs = inputs.view(sizes[0], sizes[1] * sizes[2], sizes[3]) # Collapse feature dimension #print("INPUTS SIZE: ====>>>>>\t", inputs.size()) #start = 0 #duration = 100 start = random.randint( 0, int((inputs.size(2) - 1) * (1 - args.sample_proportion))) duration = int((inputs.size(2)) * (args.sample_proportion)) #start = random.randint(0, (inputs.size(3)-1)-utterance_sequence_length) #duration = utterance_sequence_length utterances = inputs[ ..., start:start + duration] # <<<<<<====== THIS IS THE MOST IMPORTANT CODE OF THE PROJECT #print("UTTERS SIZE: ====>>>>>\t", utterances.size(), start, start+duration) out = model(utterances) #print("OUTPUT SIZE: ====>>>>>\t", out.size()) out = out.transpose(0, 1) # TxNxH ######## ######## # Prints the output of the model in a sequence of probabilities of char for each audio... #torch.set_printoptions(profile="full") ####print("OUT: " + str(out.size()), "SPEAKER LABELS:" + str(speaker_labels.size()), "INPUT PERCENTAGES MEAN: " + str(input_percentages.mean())) #print(out[:,:,0]) #print("SPEAKER LABELS: " + str(speaker_labels)) #print(out[0][0]) #softmax_output = F.softmax(out).data # This DOES NOT what I want... #softmax_output_alt = flex_softmax(out, axis=2).data # This is FINE!!! <<<=== #print(softmax_output[0][0]) #print(softmax_output_alt[0][0]) ####new_out = torch.sum(out, 0) ####new_out = torch.sum(out[20:], 0) #print(out.size()) #print(new_out.size()) #print(out[-1].size()) class_accu_reg.add(out[round(out.size(0) / 2)].data, speaker_labels.data) class_accu_sum.add( torch.sum(out[loss_begin:loss_end], 0).data, speaker_labels.data) #class_accu_reg.add(processed_out.data, processed_speaker_labels.data) if args.loss_type == "reg": processed_out = out[round(out.size(0) / 2)] processed_speaker_labels = speaker_labels if args.loss_type == "mult": #indices = torch.LongTensor([0,2]) mult = (round(out.size(0) / 4), round(out.size(0) / 2), round(3 * out.size(0) / 4)) processed_out = out.contiguous()[mult, ...].view(-1, 48) processed_speaker_labels = speaker_labels.repeat( out.size(0), 1)[mult, ...].view(-1) #processed_out = out.contiguous()[(round(out.size(0)/4),round(out.size(0)/2),round(3*out.size(0)/4)),...].view(-1,48) #processed_speaker_labels = speaker_labels.repeat(out.size(0),1)[(round(out.size(0)/4),round(out.size(0)/2),round(3*out.size(0)/4)),...].view(-1) #processed_out = out.contiguous()[(loss_begin,round(out.size(0)/2),loss_end),...].view(-1,48) #processed_speaker_labels = speaker_labels.repeat(out.size(0),1)[(loss_begin,round(out.size(0)/2),loss_end),...].view(-1) ##speaker_labels = speaker_labels.expand(20, out.size(0)) elif args.loss_type == "sum": sum_begin = round(out.size(0) / 2) - round(out.size(0) / 4) sum_end = round(out.size(0) / 2) + round(out.size(0) / 4) processed_out = torch.sum(out[sum_begin:sum_end], 0) processed_speaker_labels = speaker_labels #processed_out = torch.sum(out[loss_begin:loss_end], 0) #processed_speaker_labels = speaker_labels #processed_out = torch.sum(out, 0) #processed_speaker_labels = speaker_labels elif args.loss_type == "full": full_begin = round(out.size(0) / 2) - round(out.size(0) / 4) full_end = round(out.size(0) / 2) + round(out.size(0) / 4) processed_out = out.contiguous()[full_begin:full_end].view( -1, 48) processed_speaker_labels = speaker_labels.repeat( out.size(0), 1)[full_begin:full_end].view(-1) ##speaker_labels = speaker_labels.expand(20, out.size(0)) #processed_out = out.contiguous()[loss_begin:loss_end].view(-1,48) #processed_speaker_labels = speaker_labels.repeat(out.size(0),1)[loss_begin:loss_end].view(-1) ##speaker_labels = speaker_labels.expand(20, out.size(0)) #processed_out = out.contiguous().view(-1, 48) #processed_speaker_labels = speaker_labels.repeat(out.size(0),1).view(-1) ##speaker_labels = speaker_labels.expand(20, out.size(0)) #print("PROC OUTPUT: ====>>>>>\t" + str(processed_out.size())) #print("PROC LABELS: ====>>>>>\t" + str(processed_speaker_labels.size())) loss = criterion(processed_out, processed_speaker_labels) loss = loss / inputs.size(0) # average the loss by minibatch loss_sum = loss.data.sum() inf = float("inf") if loss_sum == inf or loss_sum == -inf: print("WARNING: received an inf loss, setting loss value to 0") loss_value = 0 else: loss_value = loss.data[0] avg_loss += loss_value losses.update(loss_value, inputs.size(0)) #accu_out3 = torch.sum(flex_softmax(out[20:], axis=2), 0) #print(classaccu.value()[0], classaccu.value()[1]) # Cross Entropy Loss for a Sequence (Time Series) of Output? #output = output.view(-1,29) #target = target.view(-1) #criterion = nn.CrossEntropyLoss() #loss = criterion(output,target) # compute gradient optimizer.zero_grad() loss.backward() #torch.nn.utils.clip_grad_norm(model.parameters(), args.max_norm) # SGD step optimizer.step() # measure elapsed time batch_time.update(time.time() - end) end = time.time() if not args.silent: print('Epoch: [{0}][{1}/{2}]\t' 'Loss {loss.val:.8f} ({loss.avg:.8f})\t' 'CARR {carr:.2f}\t' 'CARS {cars:.2f}\t'.format( (epoch + 1), (i + 1), len(train_loader), batch_time=batch_time, data_time=data_time, loss=losses, carr=class_accu_reg.value()[0], cars=class_accu_sum.value()[0])) if args.cuda: torch.cuda.synchronize() del loss del out del processed_out del speaker_labels del processed_speaker_labels avg_loss /= len(train_loader) if (best_avg_loss > avg_loss): best_avg_loss = avg_loss print("\nCURRENT EPOCH AVERAGE LOSS:\t", avg_loss) print("\nCURRENT EPOCH TRAINING RESULTS:\t", class_accu_reg.value()[0], "\t", class_accu_sum.value()[0], "\n") if (best_train_accu_reg < class_accu_reg.value()[0]): best_train_accu_reg = class_accu_reg.value()[0] if (best_train_accu_sum < class_accu_sum.value()[0]): best_train_accu_sum = class_accu_sum.value()[0] get_70 = (class_accu_reg.value()[0] > 70) if ((epoch_70 is None) and (get_70 == True)): epoch_70 = epoch + 1 get_90 = (class_accu_reg.value()[0] > 90) if ((epoch_90 is None) and (get_90 == True)): epoch_90 = epoch + 1 get_95 = (class_accu_reg.value()[0] > 95) if ((epoch_95 is None) and (get_95 == True)): epoch_95 = epoch + 1 get_99 = (class_accu_reg.value()[0] > 99) if ((epoch_99 is None) and (get_99 == True)): epoch_99 = epoch + 1 start_iter = 0 # Reset start iteration for next epoch model.eval() class_accu_reg.reset() class_accu_sum.reset() for i, (data) in enumerate(test_loader): # test inputs, input_percentages, speaker_labels, mfccs = data inputs = Variable(inputs, volatile=True) ######## mfccs = Variable(mfccs, requires_grad=False) if args.mfcc == "true": inputs = mfccs # <<-- This line makes us to use mfccs... #print("INPUTS SIZE:", inputs.size()) #print("MFCCS SIZE:", mfccs.size()) ######## speaker_labels = Variable(speaker_labels, requires_grad=False) speaker_labels = speaker_labels.cuda(async=True).long() if args.cuda: inputs = inputs.cuda() ######## ######## sizes = inputs.size() inputs = inputs.view(sizes[0], sizes[1] * sizes[2], sizes[3]) # Collapse feature dimension #print("INPUTS SIZE: ====>>>>>\t", inputs.size()) #start = round(inputs.size(2)/2)-40 #duration = 80 #start = random.randint(0, int((inputs.size(3)-1)*(1-args.sample_proportion))) #duration = int((inputs.size(3))*(args.sample_proportion)) #start = random.randint(0, (inputs.size(3)-1)-utterance_sequence_length) #duration = utterance_sequence_length utterances = inputs #[...,start:start+duration] # <<<<<<====== THIS IS THE MOST IMPORTANT CODE OF THE PROJECT #print("UTTERS SIZE: ====>>>>>\t", utterances.size(), start, start+duration) out = model(utterances) #print("OUTPUT SIZE: ====>>>>>\t", out.size()) out = out.transpose(0, 1) # TxNxH ######## ######## # Prints the output of the model in a sequence of probabilities of char for each audio... #torch.set_printoptions(profile="full") ########print("OUT: " + str(out.size()), "NEW OUT:" + str(new_out.size()), "SPEAKER LABELS:" + str(speaker_labels.size()), "INPUT PERCENTAGES MEAN: " + str(input_percentages.mean())) #print(out[:,:,0]) #print("SPEAKER LABELS: " + str(speaker_labels)) #print(out[0][0]) #softmax_output = F.softmax(out).data # This DOES NOT what I want... #softmax_output_alt = flex_softmax(out, axis=2).data # This is FINE!!! <<<=== #print(softmax_output[0][0]) #print(softmax_output_alt[0][0]) ######## #if args.loss_type == "reg": # processed_out = out[round(out.size(0)/2)]; processed_speaker_labels = speaker_labels #elif args.loss_type == "sum" or "full": # #processed_out = torch.sum(out[loss_begin:loss_end], 0); processed_speaker_labels = speaker_labels # processed_out = torch.sum(out, 0); processed_speaker_labels = speaker_labels #elif args.loss_type == "full": # #processed_out = out.contiguous()[loss_begin:loss_end].view(-1,48); processed_speaker_labels = speaker_labels.repeat(out.size(0),1)[loss_begin:loss_end].view(-1) #speaker_labels = speaker_labels.expand(20, out.size(0)) # processed_out = out.contiguous().view(-1, 48); processed_speaker_labels = speaker_labels.repeat(out.size(0),1).view(-1) # speaker_labels = speaker_labels.expand(20, out.size(0)) #print("OUT: " + str(out.size()), "SPEAKER LABELS:" + str(speaker_labels.size())) #print("PROC OUTPUT: ====>>>>>\t" + str(processed_out.size())) #print("PROC LABELS: ====>>>>>\t" + str(processed_speaker_labels.size())) class_accu_reg.add(out[round(out.size(0) / 2)].data, speaker_labels.data) class_accu_sum.add( torch.sum(out[loss_begin:loss_end], 0).data, speaker_labels.data) #class_accu_reg.add(processed_out.data, processed_speaker_labels.data) print('Validation Summary Epoch: [{0}]\t' 'CARR {carr:.2f}\t' 'CARS {cars:.2f}\t'.format(epoch + 1, carr=class_accu_reg.value()[0], cars=class_accu_sum.value()[0])) if args.cuda: torch.cuda.synchronize() del out print("\nCURRENT EPOCH TEST RESULTS:\t", class_accu_reg.value()[0], "\t", class_accu_sum.value()[0], "\n") if (best_test_accu_reg < class_accu_reg.value()[0]): best_test_accu_reg = class_accu_reg.value()[0] if (best_test_accu_sum < class_accu_sum.value()[0]): best_test_accu_sum = class_accu_sum.value()[0] print("\nBEST AVERAGE LOSS:\t\t", best_avg_loss) print("\nBEST EPOCH TRAINING RESULTS:\t", best_train_accu_reg, "\t", best_train_accu_sum) print("\nBEST EPOCH TEST RESULTS:\t", best_test_accu_reg, "\t", best_test_accu_sum) print("\nEPOCHS 70%, 90%, 95%, 99%:\t", epoch_70, "\t", epoch_90, "\t", epoch_95, "\t", epoch_99, "\n") torch.save( DeepSpeech.serialize(model, optimizer=optimizer, epoch=epoch), args.model_path) avg_loss = 0 if not args.no_bucketing and epoch == 0: print("Switching to bucketing sampler for following epochs") train_dataset = SpectrogramDatasetWithLength( audio_conf=audio_conf, manifest_filepath=args.train_manifest, normalize=True, augment=args.augment) sampler = BucketingSampler(train_dataset) train_loader.sampler = sampler