def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Image preprocessing
# For normalization, see https://github.com/pytorch/vision#models
transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
# Load vocabulary wrapper.
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
#Load vocab_list for uniskip
vocab_list = pd.read_csv("./data/vocab_list.csv", header=None)
vocab_list = vocab_list.values.tolist()[0]
#Build data loader
data_loader = get_loader(args.image_dir,
args.img_embeddings_dir,
args.data_path,
vocab,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
# Build the models
#im_encoder = preprocess_get_model.model()
attention = T_Att()
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers, args.dropout)
uniskip = UniSkip('./data/skip-thoughts', vocab_list)
decoder.eval()
if torch.cuda.is_available():
#im_encoder.cuda()
attention.cuda()
decoder.cuda()
uniskip.cuda()
attention.load_state_dict(torch.load(args.attention_path))
decoder.load_state_dict(torch.load(args.decoder_path))
for i, (images, captions, cap_lengths, qa, qa_lengths,
vocab_words) in enumerate(data_loader):
# # Set mini-batch dataset
img_embeddings = to_var(images.data, volatile=True)
captions = to_var(captions)
# qa = to_var(qa)
# targets = pack_padded_sequence(qa, qa_lengths, batch_first=True)[0]
# # Forward, Backward and Optimize
# decoder.zero_grad()
# attention.zero_grad()
# #features = encoder(images)
#img_embeddings = im_encoder(images)
#uniskip = UniSkip('/Users/tushar/Downloads/code/data/skip-thoughts', vocab_list)
cap_embeddings = uniskip(captions, cap_lengths)
cap_embeddings = cap_embeddings.data
img_embeddings = img_embeddings.data
ctx_vec = attention(img_embeddings, cap_embeddings)
outputs = decoder.sample(ctx_vec)
output_ids = outputs.cpu().data.numpy()
qa = qa.numpy()
qa = qa[0]
# predicted_q = []
# predicted_a = []
sample = []
# flag = -1
for word_id in output_ids:
word = vocab.idx2word[word_id]
sample.append(word)
# if word == '<end>':
# if flag == -1:
# predicted_q = sample
# sample = []
# flag = 0
# else:
# predicted_a = sample
# predicted_q = ' '.join(predicted_q[1:])
# predicted_a = ' '.join(predicted_a[1:])
sample = ' '.join(sample)
actual = []
# print("predicted q was : " + predicted_q)
for word_id in qa:
word = vocab.idx2word[word_id]
actual.append(word)
actual = ' '.join(actual)
#print(im_id)
print("actual_qa : " + actual + " | predicted_qa : " + sample)
def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Image preprocessing
# For normalization, see https://github.com/pytorch/vision#models
transform = transforms.Compose([
transforms.Resize((224,224)),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406),
(0.229, 0.224, 0.225))])
# Load vocabulary wrapper.
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
#Load vocab_list for uniskip
vocab_list = pd.read_csv("./data/vocab_list.csv", header=None)
vocab_list = vocab_list.values.tolist()[0]
# Build data loader
data_loader = get_loader(args.image_dir, args.img_embeddings_dir, args.data_path, vocab,
transform, args.batch_size,
shuffle=True, num_workers=args.num_workers)
# Build the models
#im_encoder = preprocess_get_model.model()
attention = T_Att()
decoder = DecoderRNN(args.embed_size, args.hidden_size,
len(vocab), args.num_layers, args.dropout)
uniskip = UniSkip('./data/skip-thoughts', vocab_list)
if torch.cuda.is_available():
#im_encoder.cuda()
attention.cuda()
decoder.cuda()
uniskip.cuda()
# Loss and Optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(attention.parameters())
optimizer = torch.optim.Adam(params, lr=args.learning_rate)
# Train the Models
total_step = len(data_loader)
for epoch in range(args.num_epochs):
for i, (images, captions, cap_lengths, qa, qa_lengths, vocab_words) in enumerate(tqdm(data_loader)):
#Re-initialize decoder hidden state
decoder.hidden = decoder.init_hidden()
# Set mini-batch dataset
img_embeddings = to_var(images.data, volatile=True)
captions = to_var(captions)
qa = to_var(qa)
targets = pack_padded_sequence(qa, qa_lengths, batch_first=True)[0]
# Forward, Backward and Optimize
decoder.zero_grad()
attention.zero_grad()
#features = encoder(images)
#img_embeddings = im_encoder(images)
cap_embeddings = uniskip(captions, cap_lengths)
cap_embeddings = cap_embeddings.data
img_embeddings = img_embeddings.data
# print(img_embeddings.size())
# print(type(img_embeddings))
# print(cap_embeddings.size())
#print(type(cap_embeddings))
ctx_vec = attention(img_embeddings,cap_embeddings)
outputs = decoder(ctx_vec, qa, qa_lengths)
predicted = outputs.max(1)[1]
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
#pred_ids = []
#print(predicted.size())
# pred_ids.append(predicted)
# Print log info
if i % args.log_step == 0:
print('Epoch [%d/%d], Step [%d/%d], Loss: %.4f, Perplexity: %5.4f'
%(epoch, args.num_epochs, i, total_step,
loss.data[0], np.exp(loss.data[0])))
#output_ids = predicted.cpu().data.numpy()
#sample = []
#for word_id in output_ids:
# word = vocab.idx2word[word_id]
# sample.append(word)
#sample = ' '.join(sample)
#print("predicted qa : " + sample)
# Save the models
if (i+1)%args.save_step == 0:
torch.save(decoder.state_dict(),
os.path.join(args.model_path,
'decoder-%d-%d.pkl' %(epoch+1, i+1)))
torch.save(attention.state_dict(),
os.path.join(args.model_path,
'attention-%d-%d.pkl' %(epoch+1, i+1)))