def main():
# Load vocabulary wrapper.
with open(vocab_path) as f:
vocab = pickle.load(f)
encoder = EncoderCNN(4096, embed_dim)
decoder = DecoderRNN(embed_dim, hidden_size, len(vocab), num_layers_rnn)
if torch.cuda.is_available():
encoder.cuda()
decoder.cuda()
# Loss and Optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(encoder.linear.parameters())
optimizer = torch.optim.Adam(params, lr=0.001)
#load data
with open(image_data_file) as f:
image_data = pickle.load(f)
image_features = si.loadmat(image_feature_file)
img_features = image_features['fc7'][0]
img_features = np.concatenate(img_features)
print 'here'
iteration = 0
save_loss = []
for i in range(10): # epoch
use_caption = i % 5
print 'Epoch', i
for x, y in make_mini_batch(img_features,
image_data,
use_caption=use_caption):
word_padding, lengths = make_word_padding(y, vocab)
x = Variable(torch.from_numpy(x).cuda())
word_index = Variable(torch.from_numpy(word_padding).cuda())
encoder.zero_grad()
decoder.zero_grad()
features = encoder(x)
targets = pack_padded_sequence(word_index,
lengths,
batch_first=True)[0]
outputs = decoder(features, word_index, lengths)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
if iteration % 100 == 0:
print 'loss', loss.data[0]
save_loss.append(loss.data[0])
iteration += 1
torch.save(decoder.state_dict(), 'decoder.pkl')
torch.save(encoder.state_dict(), 'encoder.pkl')
with open('losses.txt', 'w') as f:
print >> f, losses
def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Build the models
encoder = EncoderCNN(args.embed_size).to(device)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers).to(device)
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(
encoder.linear.parameters()) + list(encoder.bn.parameters())
optimizer = torch.optim.Adam(params, lr=args.learning_rate)
# Train the models
total_step = len(data_loader)
# For each TSP problem
for epoch in range(args.num_epochs):
for i, (images, captions, lengths) in enumerate(data_loader):
# Set mini-batch dataset
images = images.to(device)
captions = captions.to(device)
targets = pack_padded_sequence(captions, lengths,
batch_first=True)[0]
# Forward, backward and optimize
features = encoder(images)
outputs = decoder(features, captions, lengths)
loss = criterion(outputs, targets)
decoder.zero_grad()
encoder.zero_grad()
loss.backward()
optimizer.step()
# Print log info
if i % args.log_step == 0:
print(
'Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, Perplexity: {:5.4f}'
.format(epoch, args.num_epochs, i, total_step, loss.item(),
np.exp(loss.item())))
# Save the model checkpoints
if (i + 1) % args.save_step == 0:
torch.save(
decoder.state_dict(),
os.path.join(args.model_path,
'decoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
torch.save(
encoder.state_dict(),
os.path.join(args.model_path,
'encoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Image preprocessing
# Composea all processing together, to a tensor with (C,H,W) and value in range (0 - 1)
transform = transforms.Compose([
transforms.RandomCrop(args.crop_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
# Load vocabulary wrapper.
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
# Build data loader
data_loader = get_loader(args.image_dir,
args.caption_path,
vocab,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
# Build the models
encoder = EncoderCNN(args.embed_size)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers)
if torch.cuda.is_available():
encoder.cuda()
decoder.cuda()
# Loss and Optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(encoder.resnet.fc.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, lengths) in enumerate(data_loader):
# Set mini-batch dataset
images = Variable(images)
captions = Variable(captions)
print("cap size %s" % str(captions.size()))
if torch.cuda.is_available():
images = images.cuda()
captions = captions.cuda()
targets = pack_padded_sequence(captions, lengths,
batch_first=True)[0]
print(targets)
# Forward, Backward and Optimize
decoder.zero_grad()
encoder.zero_grad()
features = encoder(images)
print("cnn feats %s" % str(features.size()))
outputs = decoder(features, captions, lengths)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
# 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])))
# 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(
encoder.state_dict(),
os.path.join(args.model_path,
'encoder-%d-%d.pkl' % (epoch + 1, i + 1)))
for epoch in range(num_epochs):
tic = time.time()
for i, (image, captions, lengths) in enumerate(dataset_loader):
image = image.to(device)
captions = captions.to(device)
targets = pack_padded_sequence(captions, lengths, batch_first=True)[0]
cnn.zero_grad()
rnn.zero_grad()
cnn_out = cnn.forward(image)
lstm_out = rnn.forward(cnn_out, captions, lengths)
loss = criterion(lstm_out, targets)
loss.backward()
optimizer.step()
if i % 1000 == 0:
print(
'Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, Perplexity: {:5.4f}'
.format(epoch, num_epochs, i, len(dataset_loader), loss.item(),
np.exp(loss.item())))
toc = time.time()
print('epoch %d time %.2f mins' % (epoch, (toc - tic) / 60))
torch.save(cnn.state_dict(), 'cnn.pkl')
torch.save(rnn.state_dict(), 'rnn.pkl')
# Get training statistics.
stats = 'Epoch [%d/%d], Step [%d/%d], Loss: %.4f, Perplexity: %5.4f' % (epoch, num_epochs, i_step, total_step, loss.item(), np.exp(loss.item()))
# Print training statistics (on same line).
print('\r' + stats, end="")
sys.stdout.flush()
# Print training statistics to file.
f.write(stats + '\n')
f.flush()
# Print training statistics (on different line).
if i_step % print_every == 0:
print('\r' + stats)
epoch_loss += loss.item()
epoch_loss /= total_step
# Save the weights.
if save_every == -1:
# Only save the best one so far!
if epoch_loss <= smallest_loss:
torch.save(decoder.state_dict(), os.path.join('./models', "{:02d}-decoder-{:.4f}.pkl".format(epoch, epoch_loss)))
torch.save(encoder.state_dict(), os.path.join('./models', "{:02d}-encoder-{:.4f}.pkl".format(epoch, epoch_loss)))
smallest_loss = epoch_loss
elif epoch % save_every == 0:
torch.save(decoder.state_dict(), os.path.join('./models', "{:02d}-decoder-{:.4f}.pkl".format(epoch, epoch_loss)))
torch.save(encoder.state_dict(), os.path.join('./models', "{:02d}-encoder-{:.4f}.pkl".format(epoch, epoch_loss)))
# Close the training log file.
f.close()
def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Image preprocessing, normalization for the pretrained resnet
transform = transforms.Compose([
transforms.RandomCrop(args.crop_size),
transforms.RandomHorizontalFlip(),
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)
# Build data loader
data_loader = get_loader(args.image_dir,
args.caption_path,
vocab,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
# Build the models
encoder = EncoderCNN(args.embed_size).to(device)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers).to(device)
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(
encoder.linear.parameters()) + list(encoder.bn.parameters())
optimizer = torch.optim.Adam(params, lr=args.learning_rate)
# Train the models
save_in_file_loss = open(
'/media/raid6/shivam/imagecaption/simple_cnn_img_attention/mod_epoch_loss402.txt',
"w")
save_in_file_perplex = open(
'/media/raid6/shivam/imagecaption/simple_cnn_img_attention/mod_epoch_perplex402.txt',
"w")
save_in_file = open(
'/media/raid6/shivam/imagecaption/simple_cnn_img_attention/mod_step_loss402.txt',
"w")
loss_per_epoch = {}
perplex_per_epoch = {}
total_step = len(data_loader)
print('\ntotal-step\n')
print(total_step)
for epoch in range(args.num_epochs):
total_loss = 0
for i, (images, captions, lengths) in enumerate(data_loader):
# Set mini-batch dataset
images = images.to(device)
captions = captions.to(device)
targets = pack_padded_sequence(captions, lengths,
batch_first=True)[0]
# Forward, backward and optimize
features = encoder(images)
outputs = decoder(features, captions, lengths)
loss = criterion(outputs, targets)
decoder.zero_grad()
encoder.zero_grad()
loss.backward()
optimizer.step()
# Print log info
if i % args.log_step == 0:
print(
'Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, Perplexity: {:5.4f}'
.format(epoch, args.num_epochs, i, total_step, loss.item(),
np.exp(loss.item())))
total_loss += loss.item()
text = 'Epoch : ' + str(epoch) + '\nStep : ' + str(
i) + '\nLoss : ' + str(
loss.item()) + '\nPerplexity : ' + str(
np.exp(loss.item()))
print('\ntext\n')
print(text)
save_in_file.write(text)
# Save the model checkpoints
if (i + 1) % args.save_step == 0:
print('saving')
torch.save(
decoder.state_dict(),
os.path.join(args.model_path,
'decoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
torch.save(
encoder.state_dict(),
os.path.join(args.model_path,
'encoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
loss_per_epoch[epoch + 1] = total_loss / (total_step * args.batch_size)
loss_text = str(epoch + 1) + ' : ' + str(loss_per_epoch[epoch + 1])
save_in_file_loss.write(loss_text)
save_in_file_loss.write('\n')
print('\nloss_text : ' + loss_text)
perplex_per_epoch[epoch + 1] = np.exp(loss_per_epoch[epoch + 1])
perplex_text = str(epoch + 1) + ' : ' + str(
perplex_per_epoch[epoch + 1])
save_in_file_perplex.write(perplex_text)
save_in_file_perplex.write('\n')
print('\nperplex_text : ' + perplex_text)
save_in_file.close()
# Update the parameters in the optimizer.
optimizer.step()
# Get training statistics.
stats = 'Epoch [%d/%d], Step [%d/%d], Loss: %.4f, Perplexity: %5.4f' % (epoch, num_epochs, i_step, total_step, loss.item(), np.exp(loss.item()))
# Print training statistics (on same line).
print('\r' + stats, end="")
sys.stdout.flush()
# Print training statistics to file.
f.write(stats + '\n')
f.flush()
# Print training statistics (on different line).
if i_step % print_every == 0:
print('\r' + stats)
# if epoch == 3 and i_step % 5000 == 0:
# torch.save(decoder.state_dict(), os.path.join('./models', 'decoder-%d-%d.pkl' % epoch, i_step))
# torch.save(encoder.state_dict(), os.path.join('./models', 'encoder-%d-%d.pkl' % epoch, i_step))
# Save the weights.
if epoch % save_every == 0:
torch.save(decoder.state_dict(), os.path.join(output_path, 'decoder-%d.pkl' % epoch))
torch.save(encoder.state_dict(), os.path.join(output_path, 'encoder-%d.pkl' % epoch))
scheduler.step()
# Close the training log file.
f.close()
def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Load vocabulary wrapper
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
# Build data loader
data_loader = get_caption_loader(args.caption_path,
vocab,
75,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
# Build the models
encoder = EncoderRNN(len(vocab), args.embed_size,
args.hidden_size).to(device)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers).to(device)
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(
encoder.embedding.parameters()) + list(encoder.rnn.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, (captions_src, captions_tgt, lengths) in enumerate(data_loader):
# Set mini-batch dataset
captions_src = captions_src.to(device)
captions_tgt = captions_tgt.to(device)
targets = pack_padded_sequence(captions_tgt,
lengths,
batch_first=True)[0]
# Forward, backward and optimize
enc_output, enc_hidden = encoder(captions_src)
outputs = decoder(enc_hidden[:, -1:, :], captions_tgt, lengths)
loss = criterion(outputs, targets)
decoder.zero_grad()
encoder.zero_grad()
loss.backward()
optimizer.step()
# Print log info
if i % args.log_step == 0:
print(
'Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, Perplexity: {:5.4f}'
.format(epoch, args.num_epochs, i, total_step, loss.item(),
np.exp(loss.item())))
# Save the model checkpoints
if (i + 1) % args.save_step == 0:
torch.save(
decoder.state_dict(),
os.path.join(args.model_path,
'decoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
torch.save(
encoder.state_dict(),
os.path.join(args.model_path,
'encoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
def main(args):
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Build the models, can use a feedforward/convolutional encoder and an RNN decoder
encoder = EncoderCNN(args.embed_size).to(
device) #can be sequential or convolutional
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers).to(device)
# Loss and optimizer
criterion1 = nn.CrossEntropyLoss()
criterion2 = nn.NLLLoss()
softmax = nn.LogSoftmax(dim=1)
params = list(decoder.parameters()) + list(encoder.parameters())
optimizer = torch.optim.Adam(params, lr=args.learning_rate)
total_training_steps = args.num_iters
losses = []
perplexity = []
for epoch in range(args.num_epochs):
for i in range(total_training_steps):
prog_data = generate_training_data(args.batch_size)
images = [im[0] for im in prog_data]
transforms = [transform[1] for transform in prog_data]
[ele.insert(0, '<start>')
for ele in transforms] #start token for each sequence
[ele.append('<end>')
for ele in transforms] #end token for each sequence
lengths = [len(trans) for trans in transforms]
maximum_len = max(lengths)
for trans in transforms:
if len(trans) != maximum_len:
trans.extend(['pad'] * (maximum_len - len(trans)))
padded_lengths = [len(trans) for trans in transforms]
transforms = [[word_to_int(word) for word in transform]
for transform in transforms]
transforms = torch.tensor(transforms, device=device)
images = torch.tensor(images, device=device)
images = images.unsqueeze(
1) #Uncomment this line when training using EncoderCNN
lengths = torch.tensor(lengths, device=device)
padded_lengths = torch.tensor(padded_lengths, device=device)
targets = pack_padded_sequence(transforms,
padded_lengths,
batch_first=True)[0]
features = encoder(images)
outputs = decoder(features, transforms, padded_lengths)
#print(outputs)
loss = criterion1(outputs, targets)
losses.append(loss.item())
perplexity.append(np.exp(loss.item()))
decoder.zero_grad()
encoder.zero_grad()
loss.backward()
optimizer.step()
# Print log info
if i % args.log_step == 0:
print(
'Epoch [{}/{}], Step [{}/{}], Loss: {:.4f},Perplexity: {:5.4f}'
.format(epoch, args.num_epochs, i, total_training_steps,
loss.item(), np.exp(loss.item())))
# Save the model checkpoints
if (i + 1) % args.save_step == 0:
torch.save(
decoder.state_dict(),
os.path.join(args.model_path,
'decoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
torch.save(
encoder.state_dict(),
os.path.join(args.model_path,
'encoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
y = losses
z = perplexity
x = np.arange(len(losses))
plt.plot(x, y, label='Cross Entropy Loss')
plt.plot(x, z, label='Perplexity')
plt.xlabel('Iterations')
plt.ylabel('Cross Entropy Loss and Perplexity')
plt.title("Cross Entropy Loss and Model Perplexity During Training")
plt.legend()
plt.savefig('plots/plots_cnn/cnn4_gpu', dpi=100)
class Worker:
def __init__(self, args):
# Initialize MPI/NCCL and set topology variables
self.init_dist(args.gpu_only)
self.rank = self.dist.get_rank()
self.world_size = self.dist.get_world_size()
self.local_rank = self.dist.get_local_rank()
self.local_size = self.dist.get_local_size()
self.n_gpus = self.dist.get_n_gpus()
self.n_nodes = self.world_size / self.local_size
self.node = self.rank // self.local_size
self.n_cpu_workers = (self.local_size - self.n_gpus) * self.n_nodes
self.n_gpu_workers = self.n_gpus * self.n_nodes
# Set RNG seed for reproducibility, can be left on
torch.manual_seed(1234)
# CuDNN reproducibility
if args.reproducible:
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# Set number of threads
if self.dist.is_cpu_rank():
#torch.set_num_threads(args.num_threads)
print("[Rank {}] Setting number of OMP threads to {}".format(
self.rank, args.num_threads),
flush=True)
# Calculate batch sizes
self.total_batch_size = args.batch_size
self.cpu_batch_size = args.cpu_batch_size
assert ((self.total_batch_size - self.cpu_batch_size * self.n_cpu_workers * self.n_nodes) \
% (self.n_gpus * self.n_nodes) == 0), "GPU batch size is not an integer"
self.gpu_batch_size = int((self.total_batch_size - self.cpu_batch_size * self.n_cpu_workers * self.n_nodes) \
/ (self.n_gpus * self.n_nodes))
self.batch_size = self.cpu_batch_size if self.dist.is_cpu_rank(
) else self.gpu_batch_size
print("[Rank {}] Current CUDA device: {}".format(
self.rank, torch.cuda.current_device()),
flush=True)
def init_dist(self, gpu_only):
# C++ extension module with JIT compilation
dist_module = load(
name="dist",
sources=["dist.cu"],
verbose=True,
with_cuda=True,
extra_cuda_cflags=[
'-ccbin',
'g++',
'-std=c++11',
'-O3',
#'-I/usr/mpi/gcc/openmpi-2.1.2-hfi/include',
#'-I/usr/mpi/gcc/mvapich2-2.3b-hfi/include',
'-I/opt/intel/compilers_and_libraries_2017.4.196/linux/mpi/intel64/include',
#'-I/opt/intel/compilers_and_libraries_2017.4.196/linux/mpi/include64',
'-I/pylon5/ac7k4vp/jchoi157/pytorch/build/nccl/include'
],
extra_ldflags=[
'-L/opt/packages/cuda/9.2/lib64',
'-lcudart',
'-lrt',
#'-L/usr/mpi/gcc/openmpi-2.1.2-hfi/lib64', '-lmpi',
#'-L/usr/mpi/gcc/mvapich2-2.3b-hfi/lib', '-lmpi',
'-L/opt/intel/compilers_and_libraries_2017.4.196/linux/mpi/intel64/lib',
'-lmpi',
#'-L/opt/intel/compilers_and_libraries_2017.4.196/linux/mpi/lib64', '-lmpi',
'-L/pylon5/ac7k4vp/jchoi157/pytorch/build/nccl/lib',
'-lnccl'
],
build_directory="/home/jchoi157/torch_extensions")
self.dist = dist_module.DistManager(gpu_only, False)
def average_gradients(self):
# Only all-reduce decoder parameters since encoder is pre-trained
for param in self.decoder.parameters():
if self.dist.is_cpu_rank():
param.grad.data = param.grad.data.cuda(0, non_blocking=True)
param.grad.data *= (self.cpu_batch_size /
self.total_batch_size)
else:
param.grad.data *= (self.gpu_batch_size /
self.total_batch_size)
self.dist.hetero_allreduce(param.grad.data)
if self.dist.is_cpu_rank():
param.grad.data = param.grad.data.cpu()
def train(self, args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Image preprocessing, normalization for the pretrained resnet
transform = transforms.Compose([
transforms.RandomCrop(args.crop_size),
transforms.RandomHorizontalFlip(),
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)
# Build data loader
data_loader = get_loader(
args.image_dir,
args.caption_path,
vocab,
transform,
self.rank,
self.world_size,
self.local_size,
self.n_gpus,
self.total_batch_size,
self.cpu_batch_size,
self.gpu_batch_size,
self.batch_size,
shuffle=(False if args.reproducible else True),
no_partition=args.no_partition)
self.num_batches = len(data_loader)
print("[Rank {}] batch size {}, num batches {}".format(
self.rank,
self.total_batch_size if args.no_partition else self.batch_size,
self.num_batches),
flush=True)
# Build the models
self.encoder = EncoderCNN(args.embed_size)
self.decoder = DecoderRNN(args.embed_size, args.hidden_size,
len(vocab), args.num_layers)
if self.dist.is_gpu_rank():
self.encoder = self.encoder.cuda(self.local_rank)
self.decoder = self.decoder.cuda(self.local_rank)
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
params = list(self.decoder.parameters()) + list(
self.encoder.linear.parameters()) + list(
self.encoder.bn.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):
epoch_start_time = time.time()
batch_time_sum = 0
batch_time_total = 0
processed_batches = 0
processed_batches_total = 0
batch_start_time = time.time()
for i, (images, captions, lengths) in enumerate(data_loader):
# Set mini-batch dataset
if self.dist.is_gpu_rank():
images = images.cuda(self.local_rank)
captions = captions.cuda(self.local_rank)
targets = pack_padded_sequence(captions,
lengths,
batch_first=True)[0]
# Forward, backward, all-reduce and optimize
features = self.encoder(images)
outputs = self.decoder(features, captions, lengths)
loss = criterion(outputs, targets)
self.decoder.zero_grad()
self.encoder.zero_grad()
loss.backward()
if not args.no_partition:
self.average_gradients()
optimizer.step()
batch_time = time.time() - batch_start_time
batch_time_sum += batch_time
batch_time_total += batch_time
processed_batches += 1
processed_batches_total += 1
saved_loss = loss.item()
# Print log info
if i % args.log_step == 0 and i != 0:
print(
'Rank [{}], Epoch [{}/{}], Step [{}/{}], Average time: {:.6f}, Loss: {:.4f}, Perplexity: {:5.4f}'
.format(self.rank, epoch, args.num_epochs, i,
total_step, batch_time_sum / processed_batches,
saved_loss, np.exp(saved_loss)),
flush=True)
batch_time_sum = 0
processed_batches = 0
# Save the model checkpoints
if (i + 1) % args.save_step == 0:
torch.save(
self.decoder.state_dict(),
os.path.join(
args.model_path,
'decoder-{}-{}.ckpt'.format(epoch + 1, i + 1)))
torch.save(
self.encoder.state_dict(),
os.path.join(
args.model_path,
'encoder-{}-{}.ckpt'.format(epoch + 1, i + 1)))
batch_start_time = time.time()
epoch_time = time.time() - epoch_start_time
print(
'!!! Rank [{}], Epoch [{}], Time: {:.6f}, Average batch time: {:.6f}, Loss: {:.4f}, Perplexity: {:5.4f}'
.format(self.rank, epoch, epoch_time,
batch_time_total / processed_batches_total, saved_loss,
np.exp(saved_loss)),
flush=True)
def main(args):
# random set
manualSeed = random.randint(1, 100)
# print("Random Seed: ", manualSeed)
random.seed(manualSeed)
torch.manual_seed(manualSeed)
torch.cuda.manual_seed_all(manualSeed)
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Load vocabulary wrapper
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
audio_len, comment_len, mfcc_dim = caculate_max_len(args.audio_dir,args.text_path, vocab)
# mfcc_features = audio_preprocess(args.audio_dir, N, AUDIO_LEN, MFCC_DIM).astype(np.float32)
# Build data loader
data_loader = data_get(args.audio_dir,audio_len, args.text_path, comment_len, vocab )
# Build the models
encoder = EncoderRNN(mfcc_dim, args.embed_size, args.hidden_size).to(device)
decoder = DecoderRNN(args.embed_size+Z_DIM, args.hidden_size, len(vocab), args.num_layers).to(device)
# decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab), args.num_layers).to(device)
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(encoder.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, ((audio,audio_len), (comment,comment_len)) in enumerate(data_loader):
audio = audio.to(device)
audio = audio.unsqueeze(0)
comment = comment.to(device)
comment = comment.unsqueeze(0)
targets = pack_padded_sequence(comment, [comment_len], batch_first=True)[0]
# Forward, backward and optimize
audio_features = encoder(audio, [audio_len])
if(Z_DIM>0):
z = Variable(torch.randn(audio_features.shape[0], Z_DIM)).cuda()
audio_features = torch.cat([z,audio_features],1)
outputs = decoder(audio_features, comment, [comment_len])
loss = criterion(outputs, targets)
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Print log info
if i % args.log_step == 0:
print('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, Perplexity: {:5.4f}'.format(epoch, args.num_epochs, i, total_step, loss.item(), np.exp(loss.item())))
# Save the model checkpoints
if (epoch+1) % args.save_step == 0:
torch.save(decoder.state_dict(), os.path.join(
args.model_path, 'decoder-{}-{}.ckpt'.format(epoch+1, i+1)))
torch.save(encoder.state_dict(), os.path.join(
args.model_path, 'encoder-{}-{}.ckpt'.format(epoch+1, i+1)))
def main():
# Configuration for hyper-parameters
config = Config()
# Create model directory
if not os.path.exists(config.model_path):
os.makedirs(config.model_path)
# Image preprocessing
transform = config.train_transform
# Load vocabulary wrapper
with open(os.path.join(config.vocab_path, 'vocab.pkl'), 'rb') as f:
vocab = pickle.load(f)
# Build data loader
image_path = os.path.join(config.image_path, 'train2014')
json_path = os.path.join(config.caption_path, 'captions_train2014.json')
train_loader = get_data_loader(image_path, json_path, vocab,
transform, config.batch_size,
shuffle=True, num_workers=config.num_threads)
total_step = len(train_loader)
# Build Models
encoder = EncoderCNN(config.embed_size)
decoder = DecoderRNN(config.embed_size, config.hidden_size,
len(vocab), config.num_layers)
if torch.cuda.is_available()
encoder.cuda()
decoder.cuda()
# Loss and Optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(encoder.resnet.fc.parameters())
optimizer = torch.optim.Adam(params, lr=config.learning_rate)
# Train the Models
for epoch in range(config.num_epochs):
for i, (images, captions, lengths) in enumerate(train_loader):
# Set mini-batch dataset
images = Variable(images)
captions = Variable(captions)
if torch.cuda.is_available():
images = images.cuda()
captions = captions.cuda()
targets = pack_padded_sequence(captions, lengths, batch_first=True)[0]
# Forward, Backward and Optimize
decoder.zero_grad()
encoder.zero_grad()
features = encoder(images)
outputs = decoder(features, captions, lengths)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
# Print log info
if i % config.log_step == 0:
print('Epoch [%d/%d], Step [%d/%d], Loss: %.4f, Perplexity: %5.4f'
%(epoch, config.num_epochs, i, total_step,
loss.data[0], np.exp(loss.data[0])))
# Save the Model
if (i+1) % config.save_step == 0:
torch.save(decoder.state_dict(),
os.path.join(config.model_path,
'decoder-%d-%d.pkl' %(epoch+1, i+1)))
torch.save(encoder.state_dict(),
os.path.join(config.model_path,
'encoder-%d-%d.pkl' %(epoch+1, i+1)))
best_val_accuracy = 0
for epoch in range(args.num_epoch):
# optimizer.step()
train(args, trainloader, epoch)
val_acc = validation(args, valloader, epoch)
if val_acc > best_val_accuracy:
print("Saving the models")
torch.save(force_encoder_model.state_dict(), os.path.join(
args.model_dir, 'force_encoder.ckpt'))
torch.save(rgb_mask_encoder_model.state_dict(), os.path.join(
args.model_dir, 'rgb_mask_encoder.ckpt'))
torch.save(decoder_model.state_dict(), os.path.join(
args.model_dir, 'decoder.ckpt'))
best_val_accuracy = val_acc
def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
worker_thread_count = 1
retry_for_failed = 2
# Image preprocessing
# For normalization, see https://github.com/pytorch/vision#models
transform = transforms.Compose([
# transforms.RandomCrop(args.crop_size),
# transforms.RandomHorizontalFlip(),
transforms.Scale(args.crop_size),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
#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)
# Build data loader
data_loader = get_loader(args.image_dir,
vocab,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
# Build the models
encoder = EncoderCNN(args.embed_size)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers)
if torch.cuda.is_available():
encoder.cuda()
decoder.cuda()
# Loss and Optimizer
criterion = nn.L1Loss()
params = list(decoder.parameters()) + list(
encoder.linear.parameters()) + list(encoder.bn.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, lengths) in enumerate(data_loader):
processed_items = []
threads = []
has_data_to_process = True
def do_request(item):
position = item['position']
#print(position)
#print(item)
retry = retry_for_failed
while retry:
r = requests.post('http://localhost:4567/', data=item)
if r.status_code == 200:
pil = Image.open(io.BytesIO(r.content)).convert('RGB')
processed_items[position] = transform(pil)
#print(position, processed_items[position])
break
else:
print("shouldb be here")
time.sleep(2)
retry -= 1
# Set mini-batch dataset
image_tensors = to_var(images, volatile=True)
captions = to_var(captions)
targets = pack_padded_sequence(captions, lengths,
batch_first=True)[0]
#print(images.size())
#print(torch.equal(images[0] ,images[1]))
# Forward, Backward and Optimize
decoder.zero_grad()
encoder.zero_grad()
features = encoder(image_tensors)
outputs = decoder(features, captions, lengths)
codes = []
def worker():
while items_to_process.qsize() > 0 or has_data_to_process:
item = items_to_process.get()
if item is None:
break
do_request(item)
items_to_process.task_done()
print("ended thread processing")
for j in range(worker_thread_count):
t = threading.Thread(target=worker)
t.daemon = True # thread dies when main thread (only non-daemon thread) exits.
t.start()
threads.append(t)
for ii, image in enumerate(images):
image_tensor = to_var(image.unsqueeze(0), volatile=True)
feature = encoder(image_tensor)
sampled_ids = decoder.sample(feature)
sampled_ids = sampled_ids.cpu().data.numpy()
sampled_caption = []
for word_id in sampled_ids:
word = vocab.idx2word[word_id]
sampled_caption.append(word)
if word == '<end>':
break
sentence = ' '.join(sampled_caption)
payload = {'code': sentence}
data = {'position': ii, 'code': sentence}
items_to_process.put(data)
processed_items.append('failed')
codes.append(sentence)
has_data_to_process = False
print(codes)
print(items_to_process.qsize())
print(image.size())
print("waiting for threads")
for t in threads:
t.join()
print("done reassembling images")
for t in threads:
t.shutdown = True
t.join()
bad_value = False
for pi in processed_items:
if isinstance(pi, str) and pi == "failed":
bad_value = True
if bad_value == True:
print("failed conversion,skipping batch")
continue
output_tensor = torch.FloatTensor(len(processed_items), 3,
images.size()[2],
images.size()[3])
for ii, image_tensor in enumerate(processed_items):
output_tensor[ii] = processed_items[ii]
output_var = to_var(output_tensor, False)
target_var = to_var(images, False)
#loss = criterion(output_var,target_var)
print("loss")
print(loss)
loss.backward()
optimizer.step()
# 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])))
# 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(
encoder.state_dict(),
os.path.join(args.model_path,
'encoder-%d-%d.pkl' % (epoch + 1, i + 1)))
def main(args):
if not os.path.exists(
args.model_path
): # # create model folder to keep model setting pickle files
os.makedirs(args.model_path)
# image preprocessing and normailzation
transform = transforms.Compose([
transforms.RandomCrop(args.crop_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f) # load vocabulary wrapper file
# get data loader
data_loader = get_loader(args.image_dir,
args.caption_path,
vocab,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
encoder = EncoderCNN(args.embed_size) # build encoder
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers) # build decoder
if torch.cuda.is_available(): # load GPU
encoder.cuda()
decoder.cuda()
criterion = nn.CrossEntropyLoss() # get loss
params = list(decoder.parameters()) + list(
encoder.linear.parameters()) + list(encoder.bn.parameters())
optimizer = torch.optim.Adam(params,
lr=args.learning_rate) # get optimization
# train the Models
total_step = len(data_loader)
for epoch in range(args.num_epochs):
for i, (images, captions, lengths) in enumerate(data_loader):
# set mini batch dataset
images = to_var(images, volatile=True)
captions = to_var(captions)
targets = pack_padded_sequence(captions, lengths,
batch_first=True)[0]
# forward and backward
decoder.zero_grad()
encoder.zero_grad()
features = encoder(images)
outputs = decoder(features, captions, lengths)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step() # optimization
# Print loss and perplexity
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])))
# save the models pickle file settings
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(
encoder.state_dict(),
os.path.join(args.model_path,
'encoder-%d-%d.pkl' % (epoch + 1, i + 1)))
def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Image preprocessing, normalization for the pretrained resnet
transform = transforms.Compose([
transforms.RandomCrop(args.crop_size),
transforms.RandomHorizontalFlip(),
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)
# Build data loader
data_loader = get_loader(args.image_dir,
args.caption_path,
vocab,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
val_loader = get_loader(args.val_image_dir,
args.val_caption_path,
vocab,
transform,
args.batch_size,
shuffle=False,
num_workers=args.num_workers)
# Build the models
encoder = EncoderCNN(args.embed_size).to(device)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers).to(device)
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(
encoder.linear.parameters()) + list(encoder.bn.parameters())
optimizer = torch.optim.Adam(params, lr=args.learning_rate)
# Train the models
total_step = len(data_loader)
train_loss_arr = []
val_loss_arr = []
train_bleu_arr = []
val_bleu_arr = []
for epoch in range(1, args.num_epochs + 1, 1):
iteration_loss = []
iteration_bleu = []
for i, (images, captions, lengths) in enumerate(data_loader):
# Set mini-batch dataset
images = images.to(device)
captions = captions.to(device)
targets = pack_padded_sequence(captions, lengths,
batch_first=True)[0]
# Forward, backward and optimize
features = encoder(images)
outputs = decoder(features, captions, lengths)
#print(outputs.shape, targets.shape)
loss = criterion(outputs, targets)
iteration_loss.append(loss.item())
decoder.zero_grad()
encoder.zero_grad()
loss.backward()
optimizer.step()
#get BLEU score for corresponding batch
sampled_ids = decoder.sample(features)
sampled_ids = sampled_ids.cpu().numpy()
bleu_score_batch = get_bleu(captions, sampled_ids, vocab)
iteration_bleu.append(bleu_score_batch)
# Print log info
if i % args.log_step == 0:
print(
'Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, Bleu: '.format(
epoch, args.num_epochs, i, total_step, loss.item()) +
str(bleu_score_batch))
f_log = open(os.path.join(args.model_path, "log.txt"), "a+")
f_log.write("Epoch: " + str(epoch) + "/" +
str(args.num_epochs) + " Step: " + str(i) + "/" +
str(total_step) + " loss: " + str(loss.item()) +
" Bleu: " + str(bleu_score_batch) + "\n")
f_log.close()
# Save the model checkpoints
if (i + 1) % args.save_step == 0:
torch.save(
decoder.state_dict(),
os.path.join(args.model_path,
'decoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
torch.save(
encoder.state_dict(),
os.path.join(args.model_path,
'encoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
train_loss_arr.append(np.array(iteration_loss))
train_bleu_arr.append(np.array(iteration_bleu))
val_loss = 0
val_steps = 0
val_iteration_loss = []
val_iteration_bleu = []
for j, (images_val, captions_val,
lengths_val) in enumerate(val_loader):
# Set mini-batch dataset
images_val = images_val.to(device)
captions_val = captions_val.to(device)
targets = pack_padded_sequence(captions_val,
lengths_val,
batch_first=True)[0]
# Forward, backward and optimize
features = encoder(images_val)
outputs = decoder(features, captions_val, lengths_val)
#print(outputs.shape, targets.shape)
loss = criterion(outputs, targets).item()
val_loss += loss
val_iteration_loss.append(loss)
val_steps += 1
#get BLEU score for corresponding batch
sampled_ids = decoder.sample(features)
sampled_ids = sampled_ids.cpu().numpy()
bleu_score_batch = get_bleu(captions_val, sampled_ids, vocab)
val_iteration_bleu.append(bleu_score_batch)
val_loss /= val_steps
print('Epoch [{}/{}], Val Loss: {:.4f}, Bleu: '.format(
epoch, args.num_epochs, val_loss) + str(bleu_score_batch))
f_log = open(os.path.join(args.model_path, "log.txt"), "a+")
f_log.write("Epoch: " + str(epoch) + "/" + str(args.num_epochs) +
" val loss: " + str(val_loss) + " Bleu: " +
str(bleu_score_batch) + "\n\n")
f_log.close()
val_loss_arr.append(np.array(val_iteration_loss))
val_bleu_arr.append(np.array(val_iteration_bleu))
np.save(os.path.join(args.model_path, "train_loss.npy"),
np.array(train_loss_arr))
np.save(os.path.join(args.model_path, "val_loss.npy"),
np.array(val_loss_arr))
np.save(os.path.join(args.model_path, "train_bleu.npy"),
np.array(train_bleu_arr))
np.save(os.path.join(args.model_path, "val_bleu.npy"),
np.array(val_bleu_arr))
def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Load vocabulary wrapper
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
# Build the models
encoder = EncoderCNN(args.embed_size).to(device)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers).to(device)
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(encoder.linear.parameters())
optimizer = torch.optim.Adam(params, lr=args.learning_rate)
obj = data_loader.MsvdDataset()
datas = obj.getAll()
#print(len(datas))
os.chdir(r'E:/jupyterNotebook/our_project/')
# Train the models
total_step = len(datas)
for epoch in range(args.num_epochs):
for i, (images, captions, lengths) in enumerate(datas):
#print(epoch,i,images.shape)
# Set mini-batch dataset
images = images.to(device)
# Forward, backward and optimize
features = encoder(images)
features = features.cpu().detach().numpy()
features = features.mean(axis=0)
features = torch.from_numpy(features).view(1, -1).to(device)
#print(features.shape)
for j in range(1):
#for j in range(len(captions)):
captions[j] = captions[j].long()
captions[j] = captions[j].view(1, -1).to(device)
targets = pack_padded_sequence(captions[j],
lengths[j],
batch_first=True)[0]
outputs = decoder(features, captions[j], lengths[j])
#print(targets.shape)
#print(outputs.shape)
loss = criterion(outputs, targets)
decoder.zero_grad()
#encoder.zero_grad()
loss.backward()
optimizer.step()
print(
'Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, Perplexity: {:5.4f}'
.format(epoch + 1, args.num_epochs, i, total_step, loss.item(),
np.exp(loss.item())))
#print(os.path)
if (i + 1) % 25 == 0: #args.save_step == 0:
torch.save(
decoder.state_dict(),
os.path.join('E:\jupyterNotebook\our_project\models',
'decoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
torch.save(
encoder.state_dict(),
os.path.join('E:\jupyterNotebook\our_project\models',
'encoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
def main(args):
train_losses = []
train_acc = []
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Image preprocessing, normalization for the pretrained resnet
transform = transforms.Compose([
transforms.RandomCrop(args.crop_size),
transforms.RandomHorizontalFlip(),
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)
# Build data loader
data_loader = get_loader(args.image_dir,
args.caption_path,
vocab,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
# Build the models
encoder = EncoderCNN(args.embed_size).to(device)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers).to(device)
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(
encoder.linear.parameters()) + list(encoder.bn.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):
losses = []
accuracy = 0.0
for i, (images, captions, lengths) in enumerate(data_loader):
# Set mini-batch dataset
images = images.to(device)
captions = captions.to(device)
targets = pack_padded_sequence(captions, lengths,
batch_first=True)[0]
# Forward, backward and optimize
features = encoder(images)
outputs = decoder(features, captions, lengths)
loss = criterion(outputs, targets)
# record accuracy and loss
losses.append(loss.item())
topv, topi = outputs.topk(1, dim=1)
targets = targets.unsqueeze(-1)
accuracy += float((topi == targets).sum()) / targets.shape[0]
# update params
decoder.zero_grad()
encoder.zero_grad()
loss.backward()
optimizer.step()
# Print log info
if i % args.log_step == 0:
print(
'Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, Perplexity: {:5.4f}, Accuracy: {:.4f}'
.format(epoch + 1, args.num_epochs, i, total_step,
loss.item(), np.exp(loss.item()),
accuracy / float(i + 1)))
with open('my_train_loss_t4_resnext.txt', 'a') as fi:
fi.write('\n' +
'epoch = {}, i = {}, tr_loss = {}, acc = {}'.
format(epoch + 1, i + 1, loss.item(), accuracy /
float(i + 1)))
# Save the model checkpoints
if (i + 1) % args.save_step == 0:
torch.save(
decoder.state_dict(),
os.path.join(
args.model_path,
'my-decoder-{}-{}-t4-resnext.ckpt'.format(
epoch + 1, i + 1)))
torch.save(
encoder.state_dict(),
os.path.join(
args.model_path,
'my-encoder-{}-{}-t4-resnext.ckpt'.format(
epoch + 1, i + 1)))
train_losses.append(sum(losses) / total_step)
train_acc.append(accuracy / total_step)
# save losses over epoch
f = open("train_loss.txt", "a")
f.write(str(train_losses))
f.close()
# save accuracies over epoch
f = open("train_acc.txt", "a")
f.write(str(train_acc))
f.close()
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.RandomCrop(args.crop_size),
transforms.RandomHorizontalFlip(),
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)
# Build data loader
data_loader = get_loader(args.image_dir, args.caption_path, vocab,
transform, args.batch_size,
shuffle=True, num_workers=args.num_workers)
# Build the models
encoder = EncoderCNN(args.embed_size)
decoder = DecoderRNN(args.embed_size, args.hidden_size,
len(vocab), args.num_layers)
if torch.cuda.is_available():
encoder.cuda()
decoder.cuda()
# Loss and Optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(encoder.linear.parameters()) + list(encoder.bn.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, lengths) in enumerate(data_loader):
# Set mini-batch dataset
images = to_var(images, volatile=True)
captions = to_var(captions)
targets = pack_padded_sequence(captions, lengths, batch_first=True)[0]
# Forward, Backward and Optimize
decoder.zero_grad()
encoder.zero_grad()
features = encoder(images)
outputs = decoder(features, captions, lengths)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
# 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])))
# 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(encoder.state_dict(),
os.path.join(args.model_path,
'encoder-%d-%d.pkl' %(epoch+1, i+1)))
def main(args):
torch.manual_seed(args.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(args.seed)
# 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.RandomCrop(args.crop_size),
# transforms.RandomHorizontalFlip(),
transforms.Scale(args.crop_size),
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)
# Build data loader
data_loader = get_loader(args.image_dir,
args.caption_path,
vocab,
args.MSCOCO_result,
args.coco_detection_result,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers,
dummy_object=99,
yolo=False)
# Build the models
encoder = EncoderCNN(args.embed_size)
# the layout encoder hidden state size must be the same with decoder input size
layout_encoder = LayoutEncoder(args.layout_embed_size, args.embed_size,
100, args.num_layers)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers)
if torch.cuda.is_available():
encoder.cuda()
layout_encoder.cuda()
decoder.cuda()
# Loss and Optimizer
criterion = nn.CrossEntropyLoss()
params = list(layout_encoder.parameters()) + list(decoder.parameters()) + \
list(encoder.linear.parameters()) + list(encoder.bn.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, lengths, label_seqs, location_seqs,
visual_seqs, layout_lengths) in enumerate(data_loader):
# Set mini-batch dataset
images = to_var(images)
captions = to_var(captions)
targets = pack_padded_sequence(captions, lengths,
batch_first=True)[0]
# Forward, Backward and Optimize
# decoder.zero_grad()
# layout_encoder.zero_grad()
# encoder.zero_grad()
# Modify This part for using visual features or not
# features = encoder(images)
layout_encoding = layout_encoder(label_seqs, location_seqs,
layout_lengths)
# comb_features = features + layout_encoding
comb_features = layout_encoding
outputs = decoder(comb_features, captions, lengths)
loss = criterion(outputs, targets)
optimizer.zero_grad()
loss.backward()
optimizer.step()
# 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])))
# 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(
layout_encoder.state_dict(),
os.path.join(
args.model_path,
'layout_encoding-%d-%d.pkl' % (epoch + 1, i + 1)))
class ImageDescriptor():
def __init__(self, args, encoder):
assert(args.mode == 'train' or 'val' or 'test')
self.__args = args
self.__mode = args.mode
self.__attention_mechanism = args.attention
self.__stats_manager = ImageDescriptorStatsManager()
self.__validate_when_training = args.validate_when_training
self.__history = []
if not os.path.exists(args.model_dir):
os.makedirs(args.model_dir)
self.__config_path = os.path.join(
args.model_dir, f'config-{args.encoder}{args.encoder_ver}.txt')
# Device configuration
self.__device = torch.device(
'cuda' if torch.cuda.is_available() else 'cpu')
# training set vocab
with open(args.vocab_path, 'rb') as f:
self.__vocab = pickle.load(f)
# validation set vocab
with open(args.vocab_path.replace('train', 'val'), 'rb') as f:
self.__vocab_val = pickle.load(f)
# coco dataset
self.__coco_train = CocoDataset(
args.image_dir, args.caption_path, self.__vocab, args.crop_size)
self.__coco_val = CocoDataset(
args.image_dir, args.caption_path.replace('train', 'val'), self.__vocab_val, args.crop_size)
# data loader
self.__train_loader = torch.utils.data.DataLoader(dataset=self.__coco_train,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
self.__val_loader = torch.utils.data.DataLoader(dataset=self.__coco_val,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Build the models
self.__encoder = encoder.to(self.__device)
self.__decoder = DecoderRNN(args.embed_size, args.hidden_size,
len(self.__vocab), args.num_layers, attention_mechanism=self.__attention_mechanism).to(self.__device)
# Loss and optimizer
self.__criterion = nn.CrossEntropyLoss()
self.__params = list(self.__decoder.parameters(
)) + list(self.__encoder.linear.parameters()) + list(self.__encoder.bn.parameters())
self.__optimizer = torch.optim.Adam(
self.__params, lr=args.learning_rate)
# Load checkpoint and check compatibility
if os.path.isfile(self.__config_path):
with open(self.__config_path, 'r') as f:
content = f.read()[:-1]
if content != repr(self):
# save the error info
with open('config.err', 'w') as f:
print(f'f.read():\n{content}', file=f)
print(f'repr(self):\n{repr(self)}', file=f)
raise ValueError(
"Cannot create this experiment: "
"I found a checkpoint conflicting with the current setting.")
self.load(file_name=args.checkpoint)
else:
self.save()
def setting(self):
'''
Return the setting of the experiment.
'''
return {'Net': (self.__encoder, self.__decoder),
'Optimizer': self.__optimizer,
'BatchSize': self.__args.batch_size}
@property
def epoch(self):
return len(self.__history)
@property
def history(self):
return self.__history
# @property
# def mode(self):
# return self.__args.mode
# @mode.setter
# def mode(self, m):
# self.__args.mode = m
def __repr__(self):
'''
Pretty printer showing the setting of the experiment. This is what
is displayed when doing `print(experiment). This is also what is
saved in the `config.txt file.
'''
string = ''
for key, val in self.setting().items():
string += '{}({})\n'.format(key, val)
return string
def state_dict(self):
'''
Returns the current state of the model.
'''
return {'Net': (self.__encoder.state_dict(), self.__decoder.state_dict()),
'Optimizer': self.__optimizer.state_dict(),
'History': self.__history}
def save(self):
'''
Saves the model on disk, i.e, create/update the last checkpoint.
'''
file_name = os.path.join(
self.__args.model_dir, '{}{}-epoch-{}.ckpt'.format(self.__args.encoder, self.__args.encoder_ver, self.epoch))
torch.save(self.state_dict(), file_name)
with open(self.__config_path, 'w') as f:
print(self, file=f)
print(f'Save to {file_name}.')
def load(self, file_name=None):
'''
Loads the model from the last checkpoint saved on disk.
Args:
file_name (str): path to the checkpoint file
'''
if not file_name:
# find the latest .ckpt file
try:
file_name = max(
glob.iglob(os.path.join(self.__args.model_dir, '*.ckpt')), key=os.path.getctime)
print(f'Load from {file_name}.')
except:
raise FileNotFoundError(
'No checkpoint file in the model directory.')
else:
file_name = os.path.join(self.__args.model_dir, file_name)
print(f'Load from {file_name}.')
try:
checkpoint = torch.load(file_name, map_location=self.__device)
except:
raise FileNotFoundError(
'Please check --checkpoint, the name of the file')
self.load_state_dict(checkpoint)
del checkpoint
def load_state_dict(self, checkpoint):
'''
Loads the model from the input checkpoint.
Args:
checkpoint: an object saved with torch.save() from a file.
'''
self.__encoder.load_state_dict(checkpoint['Net'][0])
self.__decoder.load_state_dict(checkpoint['Net'][1])
self.__optimizer.load_state_dict(checkpoint['Optimizer'])
self.__history = checkpoint['History']
# The following loops are used to fix a bug that was
# discussed here: https://github.com/pytorch/pytorch/issues/2830
# (it is supposed to be fixed in recent PyTorch version)
for state in self.__optimizer.state.values():
for k, v in state.items():
if isinstance(v, torch.Tensor):
state[k] = v.to(self.__device)
def train(self, plot_loss=None):
'''
Train the network using backpropagation based
on the optimizer and the training set.
Args:
plot_loss (func, optional): if not None, should be a function taking a
single argument being an experiment (meant to be `self`).
Similar to a visitor pattern, this function is meant to inspect
the current state of the experiment and display/plot/save
statistics. For example, if the experiment is run from a
Jupyter notebook, `plot` can be used to display the evolution
of the loss with `matplotlib`. If the experiment is run on a
server without display, `plot` can be used to show statistics
on `stdout` or save statistics in a log file. (default: None)
'''
self.__encoder.train()
self.__decoder.train()
self.__stats_manager.init()
total_step = len(self.__train_loader)
start_epoch = self.epoch
print("Start/Continue training from epoch {}".format(start_epoch))
if plot_loss is not None:
plot_loss(self)
for epoch in range(start_epoch, self.__args.num_epochs):
t_start = time.time()
self.__stats_manager.init()
for i, (images, captions, lengths) in enumerate(self.__train_loader):
# Set mini-batch dataset
if not self.__attention_mechanism:
images = images.to(self.__device)
captions = captions.to(self.__device)
else:
with torch.no_grad():
images = images.to(self.__device)
captions = captions.to(self.__device)
targets = pack_padded_sequence(
captions, lengths, batch_first=True)[0]
# Forward, backward and optimize
if not self.__attention_mechanism:
features = self.__encoder(images)
outputs = self.__decoder(features, captions, lengths)
self.__decoder.zero_grad()
self.__encoder.zero_grad()
else:
self.__encoder.zero_grad()
self.__decoder.zero_grad()
features, cnn_features = self.__encoder(images)
outputs = self.__decoder(
features, captions, lengths, cnn_features=cnn_features)
loss = self.__criterion(outputs, targets)
loss.backward()
self.__optimizer.step()
with torch.no_grad():
self.__stats_manager.accumulate(
loss=loss.item(), perplexity=np.exp(loss.item()))
# Print log info each iteration
if i % self.__args.log_step == 0:
print('[Training] Epoch: {}/{} | Step: {}/{} | Loss: {:.4f} | Perplexity: {:5.4f}'
.format(epoch+1, self.__args.num_epochs, i, total_step, loss.item(), np.exp(loss.item())))
if not self.__validate_when_training:
self.__history.append(self.__stats_manager.summarize())
print("Epoch {} | Time: {:.2f}s\nTraining Loss: {:.6f} | Training Perplexity: {:.6f}".format(
self.epoch, time.time() - t_start, self.__history[-1]['loss'], self.__history[-1]['perplexity']))
else:
self.__history.append(
(self.__stats_manager.summarize(), self.evaluate()))
print("Epoch {} | Time: {:.2f}s\nTraining Loss: {:.6f} | Training Perplexity: {:.6f}\nEvaluation Loss: {:.6f} | Evaluation Perplexity: {:.6f}".format(
self.epoch, time.time() - t_start,
self.__history[-1][0]['loss'], self.__history[-1][0]['perplexity'],
self.__history[-1][1]['loss'], self.__history[-1][1]['perplexity']))
# Save the model checkpoints
self.save()
if plot_loss is not None:
plot_loss(self)
print("Finish training for {} epochs".format(self.__args.num_epochs))
def evaluate(self, print_info=False):
'''
Evaluates the experiment, i.e., forward propagates the validation set
through the network and returns the statistics computed by the stats
manager.
Args:
print_info (bool): print the results of loss and perplexity
'''
self.__stats_manager.init()
self.__encoder.eval()
self.__decoder.eval()
total_step = len(self.__val_loader)
with torch.no_grad():
for i, (images, captions, lengths) in enumerate(self.__val_loader):
images = images.to(self.__device)
captions = captions.to(self.__device)
targets = pack_padded_sequence(
captions, lengths, batch_first=True)[0]
# Forward
if not self.__attention_mechanism:
features = self.__encoder(images)
outputs = self.__decoder(features, captions, lengths)
else:
features, cnn_features = self.__encoder(images)
outputs = self.__decoder(
features, captions, lengths, cnn_features=cnn_features)
loss = self.__criterion(outputs, targets)
self.__stats_manager.accumulate(
loss=loss.item(), perplexity=np.exp(loss.item()))
if i % self.__args.log_step == 0:
print('[Validation] Step: {}/{} | Loss: {:.4f} | Perplexity: {:5.4f}'
.format(i, total_step, loss.item(), np.exp(loss.item())))
summarize = self.__stats_manager.summarize()
if print_info:
print(
f'[Validation] Average loss for this epoch is {summarize["loss"]:.6f}')
print(
f'[Validation] Average perplexity for this epoch is {summarize["perplexity"]:.6f}\n')
self.__encoder.train()
self.__decoder.train()
return summarize
def mode(self, mode=None):
'''
Get the current mode or change mode.
Args:
mode (str): 'train' or 'eval' mode
'''
if not mode:
return self.__mode
self.__mode = mode
def __load_image(self, image):
'''
Load image at `image_path` for evaluation.
Args:
image (PIL Image): image
'''
image = image.resize([224, 224], Image.LANCZOS)
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406),
(0.229, 0.224, 0.225))])
image = transform(image).unsqueeze(0)
return image
def test(self, image_path=None, plot=False):
'''
Evaluate the model by generating the caption for the
corresponding image at `image_path`.
Note: This function will not provide BLEU socre.
Args:
image_path (str): file path of the evaluation image
plot (bool): plot or not
'''
self.__encoder.eval()
self.__decoder.eval()
with torch.no_grad():
if not image_path:
image_path = self.__args.image_path
image = Image.open(image_path)
# only process with RGB image
if np.array(image).ndim == 3:
img = self.__load_image(image).to(self.__device)
# generate an caption
if not self.__attention_mechanism:
feature = self.__encoder(img)
sampled_ids = self.__decoder.sample(feature)
sampled_ids = sampled_ids[0].cpu().numpy()
else:
feature, cnn_features = self.__encoder(img)
sampled_ids = self.__decoder.sample(feature, cnn_features)
sampled_ids = sampled_ids.cpu().data.numpy()
# Convert word_ids to words
sampled_caption = []
for word_id in sampled_ids:
word = self.__vocab.idx2word[word_id]
sampled_caption.append(word)
if word == '<end>':
break
sentence = ' '.join(sampled_caption[1:-1])
# Print out the image and the generated caption
print(sentence)
if plot:
image = Image.open(image_path)
plt.imshow(np.asarray(image))
else:
print('Not support for non-RGB image.')
self.__encoder.train()
self.__decoder.train()
def coco_image(self, idx, ds='val'):
'''
Access iamge_id (which is part of the file name)
and corresponding image caption of index `idx` in COCO dataset.
Note: For jupyter notebook
Args:
idx (int): index of COCO dataset
Returns:
(dict)
'''
assert(ds == 'train' or 'val')
if ds == 'train':
ann_id = self.__coco_train.ids[idx]
return self.__coco_train.coco.anns[ann_id]
else:
ann_id = self.__coco_val.ids[idx]
return self.__coco_val.coco.anns[ann_id]
@property
def len_of_train_set(self):
'''
Number of training
'''
return len(self.__coco_train)
@property
def len_of_val_set(self):
return len(self.__coco_val)
def bleu_score(self, idx, ds='val', plot=False, show_caption=False):
'''
Evaluate the BLEU score for index `idx` in COCO dataset.
Note: For jupyter notebook
Args:
idx (int): index
ds (str): training or validation dataset
plot (bool): plot the image or not
Returns:
score (float): bleu score
'''
assert(ds == 'train' or 'val')
self.__encoder.eval()
self.__decoder.eval()
with torch.no_grad():
try:
if ds == 'train':
ann_id = self.__coco_train.ids[idx]
coco_ann = self.__coco_train.coco.anns[ann_id]
else:
ann_id = self.__coco_val.ids[idx]
coco_ann = self.__coco_val.coco.anns[ann_id]
except:
raise IndexError('Invalid index')
image_id = coco_ann['image_id']
image_id = str(image_id)
if len(image_id) != 6:
for _ in range(6 - len(image_id)):
image_id = '0' + image_id
image_path = f'{self.__args.image_dir}/COCO_train2014_000000{image_id}.jpg'
if ds == 'val':
image_path = image_path.replace('train', 'val')
coco_list = coco_ann['caption'].split()
image = Image.open(image_path)
if np.array(image).ndim == 3:
img = self.__load_image(image).to(self.__device)
# generate an caption
if not self.__attention_mechanism:
feature = self.__encoder(img)
sampled_ids = self.__decoder.sample(feature)
sampled_ids = sampled_ids[0].cpu().numpy()
else:
feature, cnn_features = self.__encoder(img)
sampled_ids = self.__decoder.sample(feature, cnn_features)
sampled_ids = sampled_ids.cpu().data.numpy()
# Convert word_ids to words
sampled_caption = []
for word_id in sampled_ids:
word = self.__vocab.idx2word[word_id]
sampled_caption.append(word)
if word == '<end>':
break
# strip punctuations and spacing
sampled_list = [c for c in sampled_caption[1:-1]
if c not in punctuation]
score = sentence_bleu(coco_list, sampled_list,
smoothing_function=SmoothingFunction().method4)
if plot:
plt.figure()
image = Image.open(image_path)
plt.imshow(np.asarray(image))
plt.title(f'score: {score}')
plt.xlabel(f'file: {image_path}')
# Print out the generated caption
if show_caption:
print(f'Sampled caption:\n{sampled_list}')
print(f'COCO caption:\n{coco_list}')
else:
print('Not support for non-RGB image.')
return
return score
def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Load vocabulary wrapper
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
# Build data loader
data_loader = get_loader(args.image_dir,
args.caption_path,
vocab,
args.dictionary,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
# Build the models
#encoder = EncoderCNN(args.embed_size).to(device)
dictionary = pd.read_csv(args.dictionary,
header=0,
encoding='unicode_escape',
error_bad_lines=False)
dictionary = list(dictionary['keys'])
decoder = DecoderRNN(len(dictionary), args.hidden_size, len(vocab),
args.num_layers).to(device)
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters(
)) # + list(encoder.linear.parameters()) + list(encoder.bn.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, (array, captions, lengths) in enumerate(data_loader):
# Set mini-batch dataset
array = array.to(device)
captions = captions.to(device)
targets = pack_padded_sequence(captions, lengths,
batch_first=True)[0]
# Forward, backward and optimize
#features = encoder(images)
outputs = decoder(array, captions, lengths)
loss = criterion(outputs, targets)
decoder.zero_grad()
#encoder.zero_grad()
loss.backward()
optimizer.step()
# Print log info
if i % args.log_step == 0:
print(
'Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, Perplexity: {:5.4f}'
.format(epoch, args.num_epochs, i, total_step, loss.item(),
np.exp(loss.item())))
# Save the model checkpoints
if (i + 1) % args.save_step == 0:
torch.save(
decoder.state_dict(),
os.path.join(args.model_path,
'decoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
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.RandomCrop(args.crop_size),
# transforms.RandomHorizontalFlip(),
# transforms.ToTensor(),
# transforms.Normalize((0.485, 0.456, 0.406),
# (0.229, 0.224, 0.225))])
transform = transforms.Compose([
transforms.RandomHorizontalFlip(),
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)
# data_loader = get_loader(args.image_dir, args.caption_path, vocab,
# transform, args.batch_size,
# shuffle=True, num_workers=args.num_workers)
sasr_data_loader = SASR_Data_Loader(vocab, transform)
sasr_data_loader.load_data(args.data_file, args.init_flag)
frogger_data_loader = sasr_data_loader.data_loader(
args.batch_size, transform, shuffle=True, num_workers=args.num_workers)
# Build the models
encoder = EncoderCNN(args.embed_size)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers)
if torch.cuda.is_available():
encoder.cuda()
decoder.cuda()
# Loss and Optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(
encoder.linear.parameters()) + list(encoder.bn.parameters())
optimizer = torch.optim.Adam(params, lr=args.learning_rate)
total_step = len(frogger_data_loader)
for epoch in range(args.num_epochs):
for i, (images, captions, lengths) in enumerate(frogger_data_loader):
images = to_var(images, volatile=True)
if (list(images.size())[0] != 1):
captions = to_var(captions)
# print(list(images.size())[0])
# print(captions)
# exit(0)
targets = pack_padded_sequence(captions,
lengths,
batch_first=True)[0]
decoder.zero_grad()
encoder.zero_grad()
features = encoder(images)
outputs = decoder(features, captions, lengths)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
# 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])))
# 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(
encoder.state_dict(),
os.path.join(args.model_path,
'encoder-%d-%d.pkl' % (epoch + 1, i + 1)))
def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Load vocabulary wrapper.
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
# Image preprocessing
# For normalization, see https://github.com/pytorch/vision#models
transform = transforms.Compose([
transforms.RandomCrop(args.crop_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
val_loader = get_loader('./data/val_resized2014/',
'./data/annotations/captions_val2014.json', vocab,
transform, 1, False, 1)
start_epoch = 0
encoder_state = args.encoder
decoder_state = args.decoder
# Build the models
encoder = EncoderCNN(args.embed_size)
if not args.train_encoder:
encoder.eval()
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers)
if args.restart:
encoder_state, decoder_state = 'new', 'new'
if encoder_state == '': encoder_state = 'new'
if decoder_state == '': decoder_state = 'new'
if decoder_state != 'new':
start_epoch = int(decoder_state.split('-')[1])
print("Using encoder: {}".format(encoder_state))
print("Using decoder: {}".format(decoder_state))
# Build data loader
data_loader = get_loader(args.image_dir,
args.caption_path,
vocab,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
""" Make logfile and log output """
with open(args.model_path + args.logfile, 'a+') as f:
f.write("Training on vanilla loss (using new model). Started {} .\n".
format(str(datetime.now())))
f.write("Using encoder: new\nUsing decoder: new\n\n")
if torch.cuda.is_available():
encoder.cuda()
decoder.cuda()
# Loss and Optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(
encoder.linear.parameters()) + list(encoder.bn.parameters())
optimizer = torch.optim.Adam(params, lr=args.learning_rate)
batch_loss = []
batch_acc = []
# Train the Models
total_step = len(data_loader)
for epoch in range(start_epoch, args.num_epochs):
for i, (images, captions, lengths, _, _) in enumerate(data_loader):
# Set mini-batch dataset
images = to_var(images, volatile=True)
captions = to_var(captions)
targets = pack_padded_sequence(captions, lengths,
batch_first=True)[0]
# Forward, Backward and Optimize
decoder.zero_grad()
encoder.zero_grad()
features = encoder(images)
out = decoder(features, captions, lengths)
loss = criterion(out, targets)
batch_loss.append(loss.data[0])
loss.backward()
optimizer.step()
# # Print log info
# if i % args.log_step == 0:
# print('Epoch [%d/%d], Step [%d/%d], Loss: %.4f, Perplexity: %5.4f, Val: %.5f, %.5f'
# %(epoch, args.num_epochs, i, total_step,
# loss.data[0], np.exp(loss.data[0]), acc, gt_acc))
# with open(args.model_path + args.logfile, 'a') as f:
# f.write('Epoch [%d/%d], Step [%d/%d], Loss: %.4f, Perplexity: %5.4f, Val: %.5f, %.5f\n'
# %(epoch, args.num_epochs, i, total_step,
# loss.data[0], np.exp(loss.data[0]), acc, gt_acc))
# 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(
encoder.state_dict(),
os.path.join(args.model_path,
'encoder-%d-%d.pkl' % (epoch + 1, i + 1)))
with open(args.model_path + 'training_loss.pkl', 'w+') as f:
pickle.dump(batch_loss, f)
with open(args.model_path + 'training_val.pkl', 'w+') as f:
pickle.dump(batch_acc, f)
with open(args.model_path + args.logfile, 'a') as f:
f.write("Training finished at {} .\n\n".format(str(datetime.now())))
def main(args):
#setup tensorboard
if args.tensorboard:
cc = CrayonClient(hostname="localhost")
print(cc.get_experiment_names())
#if args.name in cc.get_experiment_names():
try:
cc.remove_experiment(args.name)
except:
print("experiment didnt exist")
cc_server = cc.create_experiment(args.name)
# Create model directory
full_model_path = args.model_path + "/" + args.name
if not os.path.exists(full_model_path):
os.makedirs(full_model_path)
with open(full_model_path + "/parameters.json", 'w') as f:
f.write((json.dumps(vars(args))))
# Image preprocessing
transform = transforms.Compose([
transforms.Scale(args.crop_size),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
mini_transform = transforms.Compose(
[transforms.ToPILImage(),
transforms.Scale(20),
transforms.ToTensor()])
# Load vocabulary wrapper.
if args.vocab_path is not None:
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
else:
print("building new vocab")
vocab = build_vocab(args.image_dir, 1, None)
with open((full_model_path + "/vocab.pkl"), 'wb') as f:
pickle.dump(vocab, f)
# Build data loader
data_loader = get_loader(args.image_dir,
vocab,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
code_data_set = ProcessingDataset(root=args.image_dir,
vocab=vocab,
transform=transform)
train_ds, val_ds = validation_split(code_data_set)
train_loader = torch.utils.data.DataLoader(train_ds, collate_fn=collate_fn)
test_loader = torch.utils.data.DataLoader(val_ds, collate_fn=collate_fn)
train_size = len(train_loader)
test_size = len(test_loader)
# Build the models
encoder = EncoderCNN(args.embed_size, args.train_cnn)
print(encoder)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers)
print(decoder)
if torch.cuda.is_available():
encoder.cuda()
decoder.cuda()
# Loss and Optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(
encoder.linear.parameters()) + list(encoder.bn.parameters())
#params = list(decoder.parameters()) #+ list(encoder.linear.parameters()) + list(encoder.bn.parameters())
optimizer = torch.optim.Adam(params, lr=args.learning_rate)
start_time = time.time()
add_log_entry(args.name, start_time, vars(args))
# Train the Models
total_step = len(data_loader)
for epoch in range(args.num_epochs):
for i, (images, captions, lengths) in enumerate(data_loader):
decoder.train()
encoder.train()
# Set mini-batch dataset
image_ts = to_var(images, volatile=True)
captions = to_var(captions)
targets = pack_padded_sequence(captions, lengths,
batch_first=True)[0]
count = images.size()[0]
# Forward, Backward and Optimize
decoder.zero_grad()
encoder.zero_grad()
features = encoder(image_ts)
outputs = decoder(features, captions, lengths)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
total = targets.size(0)
max_index = outputs.max(dim=1)[1]
#correct = (max_index == targets).sum()
_, predicted = torch.max(outputs.data, 1)
correct = predicted.eq(targets.data).cpu().sum()
accuracy = 100. * correct / total
if args.tensorboard:
cc_server.add_scalar_value("train_loss", loss.data[0])
cc_server.add_scalar_value("perplexity", np.exp(loss.data[0]))
cc_server.add_scalar_value("accuracy", accuracy)
# Print log info
if i % args.log_step == 0:
print(
'Epoch [%d/%d], Step [%d/%d], Loss: %.4f, accuracy: %2.2f Perplexity: %5.4f'
% (epoch, args.num_epochs, i, total_step, loss.data[0],
accuracy, np.exp(loss.data[0])))
# Save the models
if (i + 1) % args.save_step == 0:
torch.save(
decoder.state_dict(),
os.path.join(full_model_path,
'decoder-%d-%d.pkl' % (epoch + 1, i + 1)))
torch.save(
encoder.state_dict(),
os.path.join(full_model_path,
'encoder-%d-%d.pkl' % (epoch + 1, i + 1)))
if 1 == 2 and i % int(train_size / 10) == 0:
encoder.eval()
#decoder.eval()
correct = 0
for ti, (timages, tcaptions,
tlengths) in enumerate(test_loader):
timage_ts = to_var(timages, volatile=True)
tcaptions = to_var(tcaptions)
ttargets = pack_padded_sequence(tcaptions,
tlengths,
batch_first=True)[0]
tfeatures = encoder(timage_ts)
toutputs = decoder(tfeatures, tcaptions, tlengths)
print(ttargets)
print(toutputs)
print(ttargets.size())
print(toutputs.size())
#correct = (ttargets.eq(toutputs[0].long())).sum()
accuracy = 100 * correct / test_size
print('accuracy: %.4f' % (accuracy))
if args.tensorboard:
cc_server.add_scalar_value("accuracy", accuracy)
torch.save(
decoder.state_dict(),
os.path.join(full_model_path,
'decoder-%d-%d.pkl' % (epoch + 1, i + 1)))
torch.save(
encoder.state_dict(),
os.path.join(full_model_path,
'encoder-%d-%d.pkl' % (epoch + 1, i + 1)))
end_time = time.time()
print("finished training, runtime: %d", [(end_time - start_time)])
def main():
# Configuration for hyper-parameters
config = Config()
# Image preprocessing
transform = config.train_transform
# Load vocabulary wrapper
with open(os.path.join(config.vocab_path, 'vocab.pkl'), 'rb') as f:
vocab = pickle.load(f)
# Build data loader
image_path = os.path.join(config.image_path, 'train2014')
json_path = os.path.join(config.caption_path, 'captions_train2014.json')
train_loader = get_data_loader(image_path, json_path, vocab,
transform, config.batch_size,
shuffle=True, num_workers=config.num_threads)
total_step = len(train_loader)
# Build Models
teachercnn = EncoderCNN(config.embed_size)
teachercnn.eval()
studentcnn = StudentCNN_Model1(config.embed_size)
#Load the best teacher model
teachercnn.load_state_dict(torch.load(os.path.join('../TrainedModels/TeacherCNN', config.trained_encoder)))
studentlstm = DecoderRNN(config.embed_size, config.hidden_size/2,
len(vocab), config.num_layers/2)
if torch.cuda.is_available():
teachercnn.cuda()
studentcnn.cuda()
studentlstm.cuda()
# Loss and Optimizer
criterion_lstm = nn.CrossEntropyLoss()
criterion_cnn = nn.MSELoss()
params = list(studentlstm.parameters()) + list(studentcnn.parameters())
optimizer_lstm = torch.optim.Adam(params, lr=config.learning_rate)
optimizer_cnn = torch.optim.Adam(studentcnn.parameters(), lr=config.cnn_learningrate)
print('entering in to training loop')
# Train the Models
for epoch in range(config.num_epochs):
for i, (images, captions, lengths, img_ids) in enumerate(train_loader):
images = Variable(images)
captions = Variable(captions)
if torch.cuda.is_available():
images = images.cuda()
captions = captions.cuda()
targets = pack_padded_sequence(captions, lengths, batch_first=True)[0]
# Forward, Backward and Optimize
optimizer_lstm.zero_grad()
optimizer_cnn.zero_grad()
features_tr = teachercnn(images)
features_st = studentcnn(images)
outputs = studentlstm(features_st, captions, lengths)
loss = criterion(features_st, features_tr.detach()) + criterion_lstm(outputs, targets)
loss.backward()
optimizer_cnn.step()
optimizer_lstm.step()
# Print log info
if i % config.log_step == 0:
print('Epoch [%d/%d], Step [%d/%d], Loss: %.4f, Perplexity: %5.4f'
%(epoch, config.num_epochs, i, total_step,
loss.data[0], np.exp(loss.data[0])))
# Save the Model
if (i+1) % config.save_step == 0:
torch.save(studentlstm.state_dict(),
os.path.join(config.student_lstm_path,
'decoder-%d-%d.pkl' %(epoch+1, i+1)))
torch.save(studentcnn.state_dict(),
os.path.join(config.student_cnn_path,
'encoder-%d-%d.pkl' %(epoch+1, i+1)))
# Print training statistics to file.
f.write(stats + '\n')
f.flush()
# Print training statistics (on different line).
if i_step % print_every == 0:
print('\r' + stats)
epoch_loss += loss.item()
epoch_loss /= total_step
# Save the weights.
if save_every == -1:
# Only save the best one so far!
if epoch_loss <= smallest_loss:
torch.save(
decoder.state_dict(),
os.path.join(
'./models',
"{:02d}-decoder-{:.4f}.pkl".format(epoch, epoch_loss)))
torch.save(
encoder.state_dict(),
os.path.join(
'./models',
"{:02d}-encoder-{:.4f}.pkl".format(epoch, epoch_loss)))
smallest_loss = epoch_loss
elif epoch % save_every == 0:
torch.save(
decoder.state_dict(),
os.path.join('./models',
"{:02d}-decoder-{:.4f}.pkl".format(epoch,
epoch_loss)))
# Calculate the batch loss.
loss = criterion(outputs.view(-1, vocab_size), captions.view(-1))
# Backward pass.
loss.backward()
# Update the parameters in the optimizer.
optimizer.step()
# Get training statistics.
stats = 'Epoch [%d/%d], Step [%d/%d], Loss: %.4f, Perplexity: %5.4f' % (epoch, args.num_epochs, i_step, total_step, loss.item(), np.exp(loss.item()))
# Log the loss to Azure ML
run.log('loss', loss.item())
run.log('perplexity', np.exp(loss.item()))
run.log('stats', stats)
# Print training statistics (on same line).
print('\r' + stats, end="")
sys.stdout.flush()
# Print training statistics (on different line).
if i_step % args.print_every == 0:
print('\r' + stats)
# Save the weights.
if epoch % args.save_every == 0:
torch.save(decoder.state_dict(), os.path.join('./models', 'decoder-%d.pkl' % epoch))
torch.save(encoder.state_dict(), os.path.join('./models', 'encoder-%d.pkl' % epoch))
def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Image preprocessing
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.033, 0.032, 0.033), (0.027, 0.027, 0.027))
])
# Build vocab
vocab = build_vocab(args.root_path, threshold=0)
vocab_path = args.vocab_path
with open(vocab_path, 'wb') as f:
pickle.dump(vocab, f)
len_vocab = vocab.idx
print(vocab.idx2word)
# Build data loader
data_loader = get_loader(args.root_path,
vocab,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
# Build the models
encoder = ResNet(ResidualBlock, [3, 3, 3], args.embed_size)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers)
#Build atten models
if torch.cuda.is_available():
encoder.cuda(1)
decoder.cuda(1)
# Loss and Optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(encoder.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, lengths) in enumerate(data_loader):
# make one hot
# cap_ = torch.unsqueeze(captions,2)
# one_hot_ = torch.FloatTensor(captions.size(0),captions.size(1),len_vocab).zero_()
# one_hot_caption = one_hot_.scatter_(2, cap_, 1)
# Set mini-batch dataset
images = to_var(images)
captions = to_var(captions)
#captions_ = to_var(one_hot_caption)
targets = pack_padded_sequence(captions, lengths,
batch_first=True)[0]
# Forward, Backward and Optimize
optimizer.zero_grad()
features = encoder(images)
outputs = decoder(features, captions, lengths)
captions = captions.view(-1)
outputs = outputs.view(-1, len_vocab)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
#print(targets)
#print(outputs)
# 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])))
#test set accuracy
#print(outputs.max(1)[1])
outputs_np = outputs.max(1)[1].cpu().data.numpy()
targets_np = targets.cpu().data.numpy()
print(outputs_np)
print(targets_np)
location_match = 0
size_match = 0
shape_match = 0
exact_match = 0
for i in range(len(targets_np)):
if outputs_np[i] == targets_np[i]:
exact_match += 1
if i >= args.batch_size and i < args.batch_size * 2 and outputs_np[
i] == targets_np[i]:
shape_match += 1
elif i >= args.batch_size * 2 and i < args.batch_size * 3 and outputs_np[
i] == targets_np[i]:
location_match += 1
elif i >= args.batch_size * 3 and i < args.batch_size * 4 and outputs_np[
i] == targets_np[i]:
size_match += 1
print(
'location match : %.4f, shape match : %.4f, exact_match: %.4f'
% (location_match / (args.batch_size), shape_match /
args.batch_size, exact_match / len(targets_np)))
# 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(
encoder.state_dict(),
os.path.join(args.model_path,
'encoder-%d-%d.pkl' % (epoch + 1, i + 1)))
def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Image preprocessing, normalization for the pretrained resnet
transform = transforms.Compose([
transforms.RandomCrop(args.crop_size),
transforms.RandomHorizontalFlip(),
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)
# Build data loader
data_loader = get_loader(args.image_dir,
args.caption_path,
vocab,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
# Build the models
encoder = EncoderCNN(args.embed_size).to(device)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers).to(device)
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(
encoder.linear.parameters()) + list(encoder.bn.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, lengths) in enumerate(data_loader):
# Set mini-batch dataset
images = images.to(device)
captions = captions.to(device)
targets = pack_padded_sequence(captions, lengths,
batch_first=True)[0]
# Forward, backward and optimize
features = encoder(images)
outputs = decoder(features, captions, lengths)
loss = criterion(outputs, targets)
decoder.zero_grad()
encoder.zero_grad()
loss.backward()
optimizer.step()
# Print log info
if i % args.log_step == 0:
print(
'Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, Perplexity: {:5.4f}'
.format(epoch, args.num_epochs, i, total_step, loss.item(),
np.exp(loss.item())))
# Save the model checkpoints
if (i + 1) % args.save_step == 0:
torch.save(
decoder.state_dict(),
os.path.join(args.model_path,
'decoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
torch.save(
encoder.state_dict(),
os.path.join(args.model_path,
'encoder-{}-{}.ckpt'.format(epoch + 1,
i + 1)))
def main():
# Configuration for hyper-parameters
torch.cuda.set_device(0)
config = Config()
# Image preprocessing
transform = config.train_transform
# Load vocabulary wrapper
with open(os.path.join(config.vocab_path, 'vocab.pkl'), 'rb') as f:
vocab = pickle.load(f)
# Build data loader
train_image_path = os.path.join(config.image_path, 'train2017')
json_path = os.path.join(config.caption_path, 'captions_train2017.json')
train_loader = get_data_loader(train_image_path,
json_path,
vocab,
transform,
config.batch_size,
shuffle=False,
num_workers=config.num_threads)
val_image_path = os.path.join(config.image_path, 'val2017')
json_path = os.path.join(config.caption_path, 'captions_val2017.json')
val_loader = get_data_loader(val_image_path,
json_path,
vocab,
transform,
config.batch_size,
shuffle=False,
num_workers=config.num_threads)
total_step = len(train_loader)
# Build Models
encoder = EncoderCNN(config.embed_size)
encoder.eval()
decoder = DecoderRNN(config.embed_size, config.hidden_size, len(vocab),
config.num_layers)
if torch.cuda.is_available():
encoder.cuda()
decoder.cuda()
# Loss and Optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(encoder.resnet.fc.parameters())
optimizer = torch.optim.Adam(params, lr=config.learning_rate)
print('entering in to training loop')
# Train the Models
with open('train1_log.txt', 'w') as logfile:
logfile.write('Validation Error,Training Error')
for epoch in range(0, 25):
for i, (images, captions, lengths,
img_ids) in enumerate(train_loader):
images = Variable(images)
captions = Variable(captions)
if torch.cuda.is_available():
images = images.cuda()
captions = captions.cuda()
targets = pack_padded_sequence(captions,
lengths,
batch_first=True)[0]
# Forward, Backward and Optimize
optimizer.zero_grad()
features = encoder(images)
outputs = decoder(features, captions, lengths)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
# Print log info
if i % config.log_step == 0:
print(
'Epoch [%d/%d], Step [%d/%d], Loss: %.4f, Perplexity: %5.4f'
% (epoch, config.num_epochs, i, total_step,
loss.data[0], np.exp(loss.data[0])))
# Save the Model
if (i + 1) % config.save_step == 0:
torch.save(
encoder.state_dict(),
os.path.join(config.teacher_cnn_path,
'encoder-%d-%d.pkl' % (epoch + 1, i + 1)))
torch.save(
decoder.state_dict(),
os.path.join(config.teacher_lstm_path,
'decoder-%d-%d.pkl' % (epoch + 1, i + 1)))
print('Just Completed an Epoch, Initite Validation Error Test')
avgvalloss = 0
for j, (images, captions, lengths,
img_ids) in enumerate(val_loader):
images = Variable(images)
captions = Variable(captions)
if torch.cuda.is_available():
images = images.cuda()
captions = captions.cuda()
targets = pack_padded_sequence(captions,
lengths,
batch_first=True)[0]
optimizer.zero_grad()
features = encoder(images)
outputs = decoder(features, captions, lengths)
valloss = criterion(outputs, targets)
if j == 0:
avgvalloss = valloss.data[0]
avgvalloss = (avgvalloss + valloss.data[0]) / 2
if ((j + 1) % 1000 == 0):
print('Average Validation Loss: %.4f' % (avgvalloss))
logfile.write(
str(avgvalloss) + ',' + str(loss.data[0]) + str('\n'))
break
# Print training statistics (on same line).
print('\r' + stats, end="")
sys.stdout.flush()
# Print training statistics to file.
f.write(stats + '\n')
f.flush()
# Print training statistics (on different line).
if i_step % print_every == 0:
print('\r' + stats)
# Save the weights.
if epoch % save_every == 0:
torch.save(decoder.state_dict(),
os.path.join('./models', 'decoder-%d.pkl' % epoch))
torch.save(encoder.state_dict(),
os.path.join('./models', 'encoder-%d.pkl' % epoch))
# Close the training log file.
f.close()
# <a id='step3'></a>
# ## Step 3: (Optional) Validate your Model
#
# To assess potential overfitting, one approach is to assess performance on a validation set. If you decide to do this **optional** task, you are required to first complete all of the steps in the next notebook in the sequence (**3_Inference.ipynb**); as part of that notebook, you will write and test code (specifically, the `sample` method in the `DecoderRNN` class) that uses your RNN decoder to generate captions. That code will prove incredibly useful here.
#
# If you decide to validate your model, please do not edit the data loader in **data_loader.py**. Instead, create a new file named **data_loader_val.py** containing the code for obtaining the data loader for the validation data. You can access:
# - the validation images at filepath `'/opt/cocoapi/images/train2014/'`, and
# - the validation image caption annotation file at filepath `'/opt/cocoapi/annotations/captions_val2014.json'`.
def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Image preprocessing, normalization for the pretrained resnet
transform = transforms.Compose([
transforms.RandomCrop(args.crop_size),
transforms.RandomHorizontalFlip(),
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)
# Build data loader
data_loader = get_loader(args.image_dir, args.caption_path, vocab,
transform, args.batch_size,
shuffle=True, num_workers=args.num_workers)
# Build the models
encoder = EncoderCNN(args.embed_size).to(device)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab), args.num_layers).to(device)
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(encoder.linear.parameters()) + list(encoder.bn.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, lengths) in enumerate(data_loader):
# Set mini-batch dataset
images = images.to(device)
captions = captions.to(device)
targets = pack_padded_sequence(captions, lengths, batch_first=True)[0]
# Forward, backward and optimize
features = encoder(images)
outputs = decoder(features, captions, lengths)
loss = criterion(outputs, targets)
decoder.zero_grad()
encoder.zero_grad()
loss.backward()
optimizer.step()
# Print log info
if i % args.log_step == 0:
print('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, Perplexity: {:5.4f}'
.format(epoch, args.num_epochs, i, total_step, loss.item(), np.exp(loss.item())))
# Save the model checkpoints
if (i+1) % args.save_step == 0:
torch.save(decoder.state_dict(), os.path.join(
args.model_path, 'decoder-{}-{}.ckpt'.format(epoch+1, i+1)))
torch.save(encoder.state_dict(), os.path.join(
args.model_path, 'encoder-{}-{}.ckpt'.format(epoch+1, i+1)))
