labels=labels, rnn_type=supported_rnns[rnn_type], audio_conf=audio_conf, bidirectional=args.bidirectional) decoder = GreedyDecoder(labels) 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) if not args.distributed: train_sampler = BucketingSampler(train_dataset, batch_size=args.batch_size) else: train_sampler = DistributedBucketingSampler(train_dataset, batch_size=args.batch_size, num_replicas=args.world_size, rank=args.rank) train_loader = AudioDataLoader(train_dataset, num_workers=args.num_workers, batch_sampler=train_sampler) test_loader = AudioDataLoader(test_dataset, batch_size=args.batch_size, num_workers=args.num_workers) print("AudioDataLoader init done !") if (not args.no_shuffle and start_epoch != 0) or args.no_sorta_grad: print("Shuffling batches for the following epochs") train_sampler.shuffle(start_epoch) if not args.distributed: model = model.to(device) else : model.to(device) model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[device], output_device=[device]) print(model) print("Number of parameters: %d" % DeepSpeech.get_param_size(model)) # optimizer 实例化 需要 在 DistributedDataParallel 操作之后 parameters = model.parameters()
def main(): args = parser.parse_args() torch.manual_seed(args.seed) torch.cuda.manual_seed_all(args.seed) if params.rnn_type == 'gru' and params.rnn_act_type != 'tanh': print "ERROR: GRU does not currently support activations other than tanh" sys.exit() if params.rnn_type == 'rnn' and params.rnn_act_type != 'relu': print "ERROR: We should be using ReLU RNNs" sys.exit() print("=======================================================") for arg in vars(args): print "***%s = %s " % (arg.ljust(25), getattr(args, arg)) print("=======================================================") save_folder = args.save_folder loss_results, cer_results, wer_results = torch.Tensor( params.epochs), torch.Tensor(params.epochs), torch.Tensor( params.epochs) best_wer = None try: os.makedirs(save_folder) except OSError as e: if e.errno == errno.EEXIST: print('Directory already exists.') else: raise criterion = CTCLoss() with open(params.labels_path) as label_file: labels = str(''.join(json.load(label_file))) audio_conf = dict(sample_rate=params.sample_rate, window_size=params.window_size, window_stride=params.window_stride, window=params.window, noise_dir=params.noise_dir, noise_prob=params.noise_prob, noise_levels=(params.noise_min, params.noise_max)) train_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=params.train_manifest, labels=labels, normalize=True, augment=params.augment) test_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=params.val_manifest, labels=labels, normalize=True, augment=False) train_loader = AudioDataLoader(train_dataset, batch_size=params.batch_size, num_workers=1) test_loader = AudioDataLoader(test_dataset, batch_size=params.batch_size, num_workers=1) rnn_type = params.rnn_type.lower() assert rnn_type in supported_rnns, "rnn_type should be either lstm, rnn or gru" model = DeepSpeech(rnn_hidden_size=params.hidden_size, nb_layers=params.hidden_layers, labels=labels, rnn_type=supported_rnns[rnn_type], audio_conf=audio_conf, bidirectional=True, rnn_activation=params.rnn_act_type, bias=params.bias) parameters = model.parameters() optimizer = torch.optim.SGD(parameters, lr=params.lr, momentum=params.momentum, nesterov=True, weight_decay=params.l2) 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)) if args.start_epoch != -1: start_epoch = args.start_epoch loss_results[: start_epoch], cer_results[:start_epoch], wer_results[:start_epoch] = package[ 'loss_results'][:start_epoch], package[ 'cer_results'][:start_epoch], package[ 'wer_results'][:start_epoch] print loss_results epoch = start_epoch else: avg_loss = 0 start_epoch = 0 start_iter = 0 avg_training_loss = 0 if params.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() ctc_time = AverageMeter() for epoch in range(start_epoch, params.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 params.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) ctc_start_time = time.time() loss = criterion(out, targets, sizes, target_sizes) ctc_time.update(time.time() - ctc_start_time) 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(), params.max_norm) # SGD step optimizer.step() if params.cuda: torch.cuda.synchronize() # measure elapsed time batch_time.update(time.time() - end) end = time.time() 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' 'CTC Time {ctc_time.val:.3f} ({ctc_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, ctc_time=ctc_time, loss=losses)) 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() wer, cer = eval_model(model, test_loader, decoder) 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.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'] / params.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 set to exit at a given accuracy, exit if params.exit_at_acc and (best_wer <= args.acc): break print("=======================================================") print "***Best WER = ", best_wer for arg in vars(args): print "***%s = %s " % (arg.ljust(25), getattr(args, arg)) print("=======================================================")
cutoff_prob=args.cutoff_prob, beam_width=args.beam_width, num_processes=args.lm_workers) elif args.decoder == "greedy": decoder = GreedyDecoder(model.labels, blank_index=model.labels.index('_')) else: decoder = None target_decoder = GreedyDecoder(model.labels, blank_index=model.labels.index('_')) test_dataset = SpectrogramDataset(audio_conf=model.audio_conf, manifest_filepath=args.test_manifest, labels=model.labels, normalize=True) test_loader = AudioDataLoader(test_dataset, batch_size=args.batch_size, num_workers=args.num_workers) wer, cer, output_data = evaluate(test_loader=test_loader, device=device, model=model, decoder=decoder, target_decoder=target_decoder, save_output=args.save_output, verbose=args.verbose, half=args.half) print('Test Summary \t' 'Average WER {wer:.3f}\t' 'Average CER {cer:.3f}\t'.format(wer=wer, cer=cer)) if args.save_output is not None: np.save(args.save_output, output_data)
def decode_dataset(logits, test_dataset, batch_size, lm_alpha, lm_beta, mesh_x, mesh_y, index, labels, eval): print("Beginning decode for {}, {}".format(lm_alpha, lm_beta)) test_loader = AudioDataLoader(test_dataset, batch_size=batch_size, num_workers=0) target_decoder = GreedyDecoder(labels, blank_index=labels.index('_')) decoder = BeamCTCDecoder(labels, beam_width=args.beam_width, cutoff_top_n=args.cutoff_top_n, blank_index=labels.index('_'), lm_path=args.lm_path, alpha=lm_alpha, beta=lm_beta, num_processes=1) model_name = re.sub('.json.pth.tar', '', os.path.basename(args.model_path)) ref_file = None if eval == 'concept': eval_dir = "%s/%s/%s" % (os.path.dirname( args.output_path), model_name, index) if not os.path.exists(eval_dir): os.makedirs(eval_dir) ref_file = open( "%s/%s_reference.txt" % (eval_dir, re.sub('.csv', '', os.path.basename( args.test_manifest))), 'w') trans_file = open( "%s/%s_transcription.txt" % (eval_dir, re.sub('.csv', '', os.path.basename( args.test_manifest))), 'w') total_cer, total_wer = 0, 0 for i, (data) in enumerate(test_loader): inputs, targets, input_percentages, target_sizes, audio_ids = data # unflatten targets split_targets = [] offset = 0 for size in target_sizes: split_targets.append(targets[offset:offset + size]) offset += size out = torch.from_numpy(logits[i][0]) sizes = torch.from_numpy(logits[i][1]) decoded_output, _, _, _, _ = decoder.decode(out, sizes) target_strings = target_decoder.convert_to_strings(split_targets) wer, cer = 0, 0 for x in range(len(target_strings)): transcript, reference = decoded_output[x][0], target_strings[x][0] if eval == 'concept': ref_file.write( reference.encode('utf-8') + "(" + audio_ids[x] + ")\n") trans_file.write( transcript.encode('utf-8') + "(" + audio_ids[x] + ")\n") wer_inst = decoder.wer(transcript, reference) / float( len(reference.split())) cer_inst = decoder.cer(transcript, reference) / float( len(reference)) wer += wer_inst cer += cer_inst total_cer += cer total_wer += wer ref_file.close() trans_file.close() wer = total_wer / len(test_loader.dataset) cer = total_cer / len(test_loader.dataset) if eval == 'concept': # Concept error rate evaluation cmd = "perl /lium/buster1/ghannay/deepSpeech2/deepspeech.pytorch/data/eval.sclit_cer.pl %s" % ( eval_dir) print("cmd ", cmd) p = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, shell=True) coner, error = p.communicate() print(" coner ", coner) return [mesh_x, mesh_y, lm_alpha, lm_beta, float(coner) / 100, cer] else: return [mesh_x, mesh_y, lm_alpha, lm_beta, wer, cer]
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
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)) else: print("Must load model!") exit() train_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=args.train_manifest, labels=labels, normalize=True, augment=False) train_sampler = BucketingSampler(train_dataset, batch_size=1) train_loader = AudioDataLoader(train_dataset, batch_sampler=train_sampler) # get the previous output f* & have modified the forward function with torch.no_grad(): for i, (data) in enumerate(train_loader): # just once # inputs: Nx1xKxT inputs, targets, input_percentages, target_sizes = data input_sizes = input_percentages.mul_(int(inputs.size(3))).int() inputs = inputs.to(device) print('input size is:', inputs.size()) # initial M_Noise model M_model = M_Noise_Deepspeech(package, inputs.size()) M_model.to(device) # no update to these parameters for para in M_model.deepspeech_net.parameters(): para.requires_grad = False
if not args.distributed: train_sampler = BucketingSampler(train_dataset, batch_size=batch_size) val_sampler = BucketingSampler(val_dataset, batch_size=batch_size) else: train_sampler = DistributedBucketingSampler( train_dataset, batch_size=batch_size, num_replicas=args.world_size, rank=args.rank) val_sampler = DistributedBucketingSampler(val_dataset, batch_size=batch_size, num_replicas=args.world_size, rank=args.rank) train_loader = AudioDataLoader(train_dataset, num_workers=args.num_workers, batch_sampler=train_sampler) val_loader = AudioDataLoader(val_dataset, num_workers=args.num_workers, batch_sampler=val_sampler) if (shuffle and start_epoch != 0) or not sorta_grad: logger.info("Shuffling batches for the following epochs") train_sampler.shuffle(start_epoch) val_sampler.shuffle(1) # Optimizer. optim_name = train_conf['optimizer'] optim_conf = config[optim_name] parameters = model.parameters() learning_rate = train_conf.getfloat('learning_rate')
val_manifest = "data/dev-clean_manifest.csv" device = "cuda" model_path = args.model_path package = torch.load(model_path, map_location=lambda storage, loc: storage) model = DeepSpeech.load_model_package(package) labels = model.labels audio_conf = model.audio_conf model.to(device) test_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=val_manifest, labels=labels, normalize=True, speed_volume_perturb=False, spec_augment=False) test_loader = AudioDataLoader(test_dataset, batch_size=20, num_workers=4) decoder = GreedyDecoder(model.labels, blank_index=model.labels.index('_')) target_decoder = GreedyDecoder(model.labels, blank_index=model.labels.index('_')) with torch.no_grad(): evaluate(test_loader, device, model, decoder, target_decoder, verbose=True) print("made it this far")
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() 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 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 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'])) 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
batch_size=args.batch_size) train_sampler_adv = BucketingSampler(train_dataset_adv, batch_size=args.batch_size) else: train_sampler_clean = DistributedBucketingSampler( train_dataset_clean, batch_size=args.batch_size, num_replicas=args.world_size, rank=args.rank) train_sampler_adv = DistributedBucketingSampler( train_dataset_adv, batch_size=args.batch_size, num_replicas=args.world_size, rank=args.rank) train_loader_clean = AudioDataLoader(train_dataset_clean, num_workers=args.num_workers, batch_sampler=train_sampler_clean) train_loader_adv = AudioDataLoader(train_dataset_adv, num_workers=args.num_workers, batch_sampler=train_sampler_adv) test_loader = AudioDataLoader(test_dataset, batch_size=args.batch_size, num_workers=args.num_workers) ''' if (not args.no_shuffle and start_epoch != 0) or args.no_sorta_grad: print("Shuffling batches for the following epochs") train_sampler.shuffle(start_epoch) ''' model = model.to(device) denoiser = denoiser.to(device) if args.mixed_precision:
log_file = f'{save_folder}/{datetime.now().strftime("%Y%m%d-%H%M%S")}_{test_job}' logger = config_logger('test', log_file=log_file, console_level='ERROR') torch.set_grad_enabled(False) model, _ = load_model(args.model_path) device = torch.device("cuda" if args.cuda else "cpu") label_decoder = LabelDecoder(model.labels) model.eval() model = model.to(device) test_dataset = SpectrogramDataset(audio_conf=model.audio_conf, manifest_filepath=args.test_manifest, labels=model.labels) test_sampler = BucketingSampler(test_dataset, batch_size=args.batch_size) test_loader = AudioDataLoader(test_dataset, batch_sampler=test_sampler, num_workers=args.num_workers) test_sampler.shuffle(1) total_wer, total_cer, total_ler, num_words, num_chars, num_labels = 0, 0, 0, 0, 0, 0 output_data = [] for i, (data) in tqdm(enumerate(test_loader), total=len(test_loader), ascii=True): inputs, targets, input_sizes, target_sizes, filenames = data inputs = inputs.to(device) input_sizes = input_sizes.to(device) outputs = model.transcribe(inputs, input_sizes) for i, target in enumerate(targets):
def train(self, **kwargs): """ Run optimization to train the model. Parameters ---------- """ world_size = kwargs.pop('world_size', 1) gpu_rank = kwargs.pop('gpu_rank', 0) rank = kwargs.pop('rank', 0) dist_backend = kwargs.pop('dist_backend', 'nccl') dist_url = kwargs.pop('dist_url', None) os.environ['MASTER_ADDR'] = '127.0.0.1' os.environ['MASTER_PORT'] = '1234' main_proc = True self.distributed = world_size > 1 if self.distributed: if self.gpu_rank: torch.cuda.set_device(int(gpu_rank)) dist.init_process_group(backend=dist_backend, init_method=dist_url, world_size=world_size, rank=rank) print('Initiated process group') main_proc = rank == 0 # Only the first proc should save models if main_proc and self.tensorboard: tensorboard_logger = TensorBoardLogger(self.id, self.log_dir, self.log_params, comment=self.sufix) if self.distributed: train_sampler = DistributedBucketingSampler( self.data_train, batch_size=self.batch_size, num_replicas=world_size, rank=rank) else: if self.sampler_type == 'bucketing': train_sampler = BucketingSampler(self.data_train, batch_size=self.batch_size, shuffle=True) if self.sampler_type == 'random': train_sampler = RandomBucketingSampler( self.data_train, batch_size=self.batch_size) print("Shuffling batches for the following epochs..") train_sampler.shuffle(self.start_epoch) train_loader = AudioDataLoader(self.data_train, num_workers=self.num_workers, batch_sampler=train_sampler) val_loader = AudioDataLoader(self.data_val, batch_size=self.batch_size_val, num_workers=self.num_workers, shuffle=True) if self.tensorboard and self.generate_graph: # TO DO get some audios also with torch.no_grad(): inputs, targets, input_percentages, target_sizes = next( iter(train_loader)) input_sizes = input_percentages.mul_(int(inputs.size(3))).int() tensorboard_logger.add_image(inputs, input_sizes, targets, network=self.model) self.model = self.model.to(self.device) parameters = self.model.parameters() if self.update_rule == 'adam': optimizer = torch.optim.Adam(parameters, lr=self.lr, weight_decay=self.reg) if self.update_rule == 'sgd': optimizer = torch.optim.SGD(parameters, lr=self.lr, weight_decay=self.reg) self.model, self.optimizer = amp.initialize( self.model, optimizer, opt_level=self.opt_level, keep_batchnorm_fp32=self.keep_batchnorm_fp32, loss_scale=self.loss_scale) if self.optim_state is not None: self.optimizer.load_state_dict(self.optim_state) if self.amp_state is not None: amp.load_state_dict(self.amp_state) if self.distributed: self.model = DistributedDataParallel(self.model) print(self.model) if self.criterion_type == 'cross_entropy_loss': self.criterion = torch.nn.CrossEntropyLoss() # Useless for now because I don't save. accuracies_train_iters = [] losses_iters = [] avg_loss = 0 batch_time = AverageMeter() epoch_time = AverageMeter() losses = AverageMeter() start_training = time.time() for epoch in range(self.start_epoch, self.num_epochs): print("Start epoch..") # Put model in train mode self.model.train() y_true_train_epoch = np.array([]) y_pred_train_epoch = np.array([]) start_epoch = time.time() for i, (data) in enumerate(train_loader, start=0): start_batch = time.time() print('Start batch..') if i == len(train_sampler): # QUE pq isso deus break inputs, targets, input_percentages, _ = data input_sizes = input_percentages.mul_(int(inputs.size(3))).int() inputs = inputs.to(self.device) targets = targets.to(self.device) output, loss_value = self._step(inputs, input_sizes, targets) print('Step finished.') avg_loss += loss_value with torch.no_grad(): y_pred = self.decoder.decode(output.detach()).cpu().numpy() # import pdb; pdb.set_trace() y_true_train_epoch = np.concatenate( (y_true_train_epoch, targets.cpu().numpy() )) # maybe I should do it with tensors? y_pred_train_epoch = np.concatenate( (y_pred_train_epoch, y_pred)) inputs_size = inputs.size(0) del output, inputs, input_percentages if self.intra_epoch_sanity_check: with torch.no_grad(): acc, _ = self.check_accuracy(targets.cpu().numpy(), y_pred=y_pred) accuracies_train_iters.append(acc) losses_iters.append(loss_value) cm = confusion_matrix(targets.cpu().numpy(), y_pred, labels=self.labels) print('[it %i/%i] Confusion matrix train step:' % ((i + 1, len(train_sampler)))) print(pd.DataFrame(cm)) if self.tensorboard: tensorboard_logger.update( len(train_loader) * epoch + i + 1, { 'Loss/through_iterations': loss_value, 'Accuracy/train_through_iterations': acc }) del targets batch_time.update(time.time() - start_batch) epoch_time.update(time.time() - start_epoch) losses.update(loss_value, inputs_size) # Write elapsed time (and loss) to terminal print('Epoch: [{0}][{1}/{2}]\t' 'Batch {batch_time.val:.3f} ({batch_time.avg:.3f})\t' 'Epoch {data_time.val:.3f} ({data_time.avg:.3f})\t' 'Loss {loss.val:.4f} ({loss.avg:.4f})\t'.format( (epoch + 1), (i + 1), len(train_sampler), batch_time=batch_time, data_time=epoch_time, loss=losses)) # Loss log avg_loss /= len(train_sampler) self.loss_epochs.append(avg_loss) # Accuracy train log acc_train, _ = self.check_accuracy(y_true_train_epoch, y_pred=y_pred_train_epoch) self.accuracy_train_epochs.append(acc_train) # Accuracy val log with torch.no_grad(): y_pred_val = np.array([]) targets_val = np.array([]) for data in val_loader: inputs, targets, input_percentages, _ = data input_sizes = input_percentages.mul_(int( inputs.size(3))).int() _, y_pred_val_batch = self.check_accuracy( targets.cpu().numpy(), inputs=inputs, input_sizes=input_sizes) y_pred_val = np.concatenate((y_pred_val, y_pred_val_batch)) targets_val = np.concatenate( (targets_val, targets.cpu().numpy() )) # TO DO: think of a smarter way to do this later del inputs, targets, input_percentages # import pdb; pdb.set_trace() acc_val, y_pred_val = self.check_accuracy(targets_val, y_pred=y_pred_val) self.accuracy_val_epochs.append(acc_val) cm = confusion_matrix(targets_val, y_pred_val, labels=self.labels) print('Confusion matrix validation:') print(pd.DataFrame(cm)) # Write epoch stuff to tensorboard if self.tensorboard: tensorboard_logger.update( epoch + 1, {'Loss/through_epochs': avg_loss}, parameters=self.model.named_parameters) tensorboard_logger.update(epoch + 1, { 'train': acc_train, 'validation': acc_val }, together=True, name='Accuracy/through_epochs') # Keep track of the best model if acc_val > self.best_acc_val: self.best_acc_val = acc_val self.best_params = {} for k, v in self.model.named_parameters( ): # TO DO: actually copy model and save later? idk.. self.best_params[k] = v.clone() # Anneal learning rate. TO DO: find better way to this this specific to every parameter as cs231n does. for g in self.optimizer.param_groups: g['lr'] = g['lr'] / self.learning_anneal print('Learning rate annealed to: {lr:.6f}'.format(lr=g['lr'])) # Shuffle batches order print("Shuffling batches...") train_sampler.shuffle(epoch) # Rechoose batches elements if self.sampler_type == 'random': train_sampler.recompute_bins() end_training = time.time() if self.tensorboard: tensorboard_logger.close() print('Elapsed time in training: %.02f ' % ((end_training - start_training) / 60.0))
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
def main(): args = parser.parse_args() params.cuda = not bool(args.cpu) print("Use cuda: {}".format(params.cuda)) torch.manual_seed(args.seed) torch.cuda.manual_seed_all(args.seed) if params.rnn_type == 'gru' and params.rnn_act_type != 'tanh': print( "ERROR: GRU does not currently support activations other than tanh" ) sys.exit() if params.rnn_type == 'rnn' and params.rnn_act_type != 'relu': print("ERROR: We should be using ReLU RNNs") sys.exit() print("=======================================================") for arg in vars(args): print("***%s = %s " % (arg.ljust(25), getattr(args, arg))) print("=======================================================") save_folder = args.save_folder loss_results, cer_results, wer_results = torch.Tensor( params.epochs), torch.Tensor(params.epochs), torch.Tensor( params.epochs) best_wer = None try: os.makedirs(save_folder) except OSError as e: if e.errno == errno.EEXIST: print('Directory already exists.') else: raise criterion = CTCLoss() with open(params.labels_path) as label_file: labels = str(''.join(json.load(label_file))) audio_conf = dict(sample_rate=params.sample_rate, window_size=params.window_size, window_stride=params.window_stride, window=params.window, noise_dir=params.noise_dir, noise_prob=params.noise_prob, noise_levels=(params.noise_min, params.noise_max)) if args.use_set == 'libri': testing_manifest = params.val_manifest + ("_held" if args.hold_idx >= 0 else "") else: testing_manifest = params.test_manifest if args.batch_size_val > 0: params.batch_size_val = args.batch_size_val print("Testing on: {}".format(testing_manifest)) train_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=params.val_manifest, labels=labels, normalize=True, augment=params.augment) test_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=testing_manifest, labels=labels, normalize=True, augment=False) train_loader = AudioDataLoader(train_dataset, batch_size=params.batch_size, num_workers=1) test_loader = AudioDataLoader(test_dataset, batch_size=params.batch_size_val, num_workers=1) rnn_type = params.rnn_type.lower() assert rnn_type in supported_rnns, "rnn_type should be either lstm, rnn or gru" model = DeepSpeech(rnn_hidden_size=params.hidden_size, nb_layers=params.hidden_layers, labels=labels, rnn_type=supported_rnns[rnn_type], audio_conf=audio_conf, bidirectional=False, rnn_activation=params.rnn_act_type, bias=params.bias) parameters = model.parameters() optimizer = torch.optim.SGD(parameters, lr=params.lr, momentum=params.momentum, nesterov=True, weight_decay=params.l2) 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)) if args.start_epoch != -1: start_epoch = args.start_epoch avg_loss = 0 start_epoch = 0 start_iter = 0 avg_training_loss = 0 epoch = 1 else: avg_loss = 0 start_epoch = 0 start_iter = 0 avg_training_loss = 0 if params.cuda: model = torch.nn.DataParallel(model).cuda() # model = torch.nn.parallel.DistributedDataParallel(model).cuda() print(model) print("Number of parameters: %d" % DeepSpeech.get_param_size(model)) batch_time = AverageMeter() data_time = AverageMeter() losses = AverageMeter() ctc_time = AverageMeter() for epoch in range(start_epoch, params.epochs): ################################################################################################################# # The test script only really cares about this section. ################################################################################################################# model.eval() wer, cer, trials = eval_model_verbose(model, test_loader, decoder, params.cuda, args.n_trials) root = os.getcwd() outfile = osp.join( root, "inference_bs{}_i{}_gpu{}.csv".format(params.batch_size_val, args.hold_idx, params.cuda)) print("Exporting inference to: {}".format(outfile)) make_file(outfile) write_line( outfile, "batch times pre normalized by hold_sec =,{}\n".format( args.hold_sec)) write_line(outfile, "wer, {}\n".format(wer)) write_line(outfile, "cer, {}\n".format(cer)) write_line(outfile, "bs, {}\n".format(params.batch_size_val)) write_line(outfile, "hold_idx, {}\n".format(args.hold_idx)) write_line(outfile, "cuda, {}\n".format(params.cuda)) write_line(outfile, "avg batch time, {}\n".format(trials.avg / args.hold_sec)) percentile_50 = np.percentile( trials.array, 50) / params.batch_size_val / args.hold_sec write_line(outfile, "50%-tile latency, {}\n".format(percentile_50)) percentile_99 = np.percentile( trials.array, 99) / params.batch_size_val / args.hold_sec write_line(outfile, "99%-tile latency, {}\n".format(percentile_99)) write_line(outfile, "through put, {}\n".format(1 / percentile_50)) write_line(outfile, "data\n") for trial in trials.array: write_line(outfile, "{}\n".format(trial / args.hold_sec)) 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)) # anneal lr optim_state = optimizer.state_dict() optim_state['param_groups'][0]['lr'] = optim_state['param_groups'][0][ 'lr'] / params.learning_anneal optimizer.load_state_dict(optim_state) print('Learning rate annealed to: {lr:.6f}'.format( lr=optim_state['param_groups'][0]['lr'])) break print("=======================================================") print("***Best WER = ", best_wer) for arg in vars(args): print("***%s = %s " % (arg.ljust(25), getattr(args, arg))) print("=======================================================")
# elif args.decoder == "greedy": # decoder = GreedyDecoder(model.labels, blank_index=model.labels.index('_')) # else: # decoder = None # target_decoder = GreedyDecoder(model.labels, blank_index=model.labels.index('_')) decoder = MyDecoder(model.labels) target_decoder = MyDecoder(model.labels) test_dataset = SpectrogramDataset(audio_conf=model.audio_conf, manifest_filepath=args.test_manifest, metadata_file_path=metadata_path, labels=model.labels, normalize=True) test_loader = AudioDataLoader( test_dataset, batch_size=args.batch_size, num_workers=args.num_workers, shuffle=True ) # in train, the manifest will be already stratified with only train data so it's ok. accuracy_mean, acccuracy_std, output_data = evaluate( test_loader=test_loader, device=device, model=model, decoder=decoder, target_decoder=target_decoder, save_output=args.save_output, verbose=args.verbose, half=args.half) print('Test Summary \t' 'Average accuracy {acc_mean:.3f}\t' 'Standard deviation accuracy {acc_std:.3f}\t'.format(
def main(): args = parser.parse_args() save_folder = args.save_folder 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) train_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=args.train_manifest, labels=labels, normalize=True) test_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=args.val_manifest, labels=labels, normalize=True) 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) model = DeepSpeech(rnn_hidden_size=args.hidden_size, nb_layers=args.hidden_layers, num_classes=len(labels)) decoder = ArgMaxDecoder(labels) if args.cuda: model = torch.nn.DataParallel(model).cuda() print(model) parameters = model.parameters() optimizer = torch.optim.SGD(parameters, lr=args.lr, momentum=args.momentum, nesterov=True) batch_time = AverageMeter() data_time = AverageMeter() losses = AverageMeter() for epoch in range(args.epochs): model.train() end = time.time() avg_loss = 0 for i, (data) in enumerate(train_loader): inputs, targets, input_percentages, target_sizes = data # measure data loading time data_time.update(time.time() - end) inputs = Variable(inputs) target_sizes = Variable(target_sizes) targets = Variable(targets) 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()) 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() # 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)) avg_loss /= len(train_loader) print('Training Summary Epoch: [{0}]\t' 'Average Loss {loss:.3f}\t'.format((epoch + 1), loss=avg_loss)) total_cer, total_wer = 0, 0 for i, (data) in enumerate(test_loader): # test inputs, targets, input_percentages, target_sizes = data inputs = Variable(inputs) # 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 = Variable(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 wer = total_wer / len(test_loader.dataset) cer = total_cer / len(test_loader.dataset) print('Validation Summary Epoch: [{0}]\t' 'Average WER {wer:.0f}\t' 'Average CER {cer:.0f}\t'.format((epoch + 1), wer=wer * 100, cer=cer * 100)) if args.epoch_save: file_path = '%s/deepspeech_%d.pth.tar' % (save_folder, epoch) torch.save(checkpoint(model, args, len(labels), epoch), file_path) torch.save(checkpoint(model, args, len(labels)), args.final_model_path)