# Load data
trainset = CSL_Phoenix_Openpose(skeleton_root=train_skeleton_root,annotation_file=train_annotation_file,dictionary=dictionary,
clip_length=clip_length,stride=stride)
devset = CSL_Phoenix_Openpose(skeleton_root=dev_skeleton_root,annotation_file=dev_annotation_file,dictionary=dictionary,
clip_length=clip_length,stride=stride)
testset = CSL_Phoenix_Openpose(skeleton_root=test_skeleton_root,annotation_file=test_annotation_file,dictionary=dictionary,
clip_length=clip_length,stride=stride)
print("Dataset samples: {}".format(len(trainset)+len(devset)))
trainloader = DataLoader(trainset, batch_size=batch_size, shuffle=True, num_workers=num_workers, pin_memory=True,
collate_fn=skeleton_collate)
devloader = DataLoader(devset, batch_size=batch_size, shuffle=True, num_workers=num_workers, pin_memory=True,
collate_fn=skeleton_collate)
testloader = DataLoader(testset, batch_size=batch_size, shuffle=True, num_workers=num_workers, pin_memory=True,
collate_fn=skeleton_collate)
# Create model
model = hcn_lstm(vocab_size,clip_length=clip_length,
num_classes=num_classes,hidden_dim=hidden_dim).to(device)
start_epoch, best_wer = resume_model(model, checkpoint)
# Run the model parallelly
if torch.cuda.device_count() > 1:
print("Using {} GPUs".format(torch.cuda.device_count()))
model = nn.DataParallel(model)
# Create loss criterion
criterion = nn.CTCLoss()
wer = 100.00
# Start training
print("Evaluation Started".center(60, '#'))
for epoch in range(start_epoch, start_epoch+1):
# Test the model
wer = test_hcn_lstm(model, criterion, trainloader, device, epoch, log_interval, writer, reverse_dict)
devloader = DataLoader(devset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
pin_memory=True,
collate_fn=skeleton_collate)
testloader = DataLoader(testset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
pin_memory=True,
collate_fn=skeleton_collate)
# Create model
model = hcn_lstm(vocab_size,
clip_length=clip_length,
num_classes=num_classes,
hidden_dim=hidden_dim,
dropout=0.8).to(device) # default dropout is 0.4
model = resume_hcn_module(model, hcn_checkpoint)
if hcn_lstm_ckpt is not None:
model = resume_hcn_lstm(model, hcn_lstm_ckpt)
if mainpart_ckpt is not None:
model = resume_main_part(model, mainpart_ckpt)
if checkpoint is not None:
start_epoch, best_wer = resume_model(model, checkpoint)
# Run the model parallelly
if torch.cuda.device_count() > 1:
print("Using {} GPUs".format(torch.cuda.device_count()))
model = nn.DataParallel(model)
# Create loss criterion & optimizer
# criterion = nn.CTCLoss(zero_infinity=True)