def evaluate(input_sentences, output_sentences, input_vocab, output_vocab,
input_reverse, output_reverse, hy, writer):
dataset = NMTDataset(input_sentences, output_sentences, input_vocab,
output_vocab, input_reverse, output_reverse)
loader = DataLoader(dataset,
batch_size=hy.batch_size,
shuffle=True,
drop_last=True)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
input_vocab_size = len(input_vocab.keys())
output_vocab_size = len(output_vocab.keys())
encoder = EncoderRNN(input_vocab_size, hy.embedding_size, hy.hidden_size,
hy.rnn_layers, hy.bidirectional, device)
decoder = DecoderRNN(output_vocab_size, hy.embedding_size, hy.hidden_size,
hy.rnn_layers, hy.bidirectional, device)
accuracies = []
for epoch in range(1, hy.num_epochs + 1):
encoder.load_state_dict(
torch.load("saved_runs/encoder_{}_weights.pt".format(epoch)))
decoder.load_state_dict(
torch.load("saved_runs/decoder_{}_weights.pt".format(epoch)))
accuracy = compute_model_accuracy(encoder, decoder, loader, device,
epoch, writer)
accuracies.append(accuracy)
print("=" * 80)
print("Final Accuracy = {:.1f}".format(100. * np.max(accuracies)))
print("=" * 80)
return accuracies
def main(args):
"""Output learned representation of RRMs from CNN_LSTM autoencoder."""
# Load pickled vocab
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
# Load the already preprocessed data
df_aligned = preprocess(preprocessed=True,
RRM_path=args.processed_RRM_path,
output_path=args.processed_RRM_path,
vocab=vocab)
# Data loader
loader = RRM_Sequence(df_aligned, vocab)
loader = DataLoader(loader, 16, shuffle=True, collate_fn=collate_fn)
encoderCNN = ResNetEncoder(84, 26, 64) # TODO don't hardcode?
decoderRNN = DecoderRNN(64, 128, 26, 1) # TODO don't hardcode?
# Use CUDA if available
if torch.cuda.is_available():
encoderCNN.cuda()
decoderRNN.cuda()
# Load pickled models
with open(args.encoder_path, 'rb') as encoder:
encoderCNN.load_state_dict(torch.load(encoder))
with open(args.decoder_path, 'rb') as decoder:
decoderRNN.load_state_dict(torch.load(decoder))
# Loop over data
for batch_idx, (names, rrms_aligned, rrms_unaligned,
lengths) in enumerate(loader):
rrms_aligned = to_var(rrms_aligned)
rrms_unaligned = to_var(rrms_unaligned)
features = encoderCNN(rrms_aligned)
if args.which_representation == 'encoder':
hiddens = features
else:
hiddens = forward(decoderRNN, features, rrms_unaligned, lengths)
hiddens = hiddens.data.cpu().numpy()
# hiddens have shape (16, 64) for encoder, (16, 128) for decoder
if batch_idx == 0:
df = pd.DataFrame(hiddens)
df['name'] = names
else:
df1 = pd.DataFrame(hiddens)
df1['name'] = names
df = pd.concat([df, df1])
# Write to file
df.to_csv(args.hidden_path)
def __init__(self, **kwargs):
dp = DataPreprocessor()
file_name_formatted = dp.write_to_file()
dc = DataCleaner(file_name_formatted)
dc.clean_data_pipeline().trim_rare_words()
self.data_loader = DataLoader(dc.vocabulary, dc.pairs)
self.dp = dp
self.dc = dc
load_embedding = kwargs.get('pretrained_embedding', False)
embedding_file = kwargs.get('pretrained_embedding_file', None)
load_enc_dec = kwargs.get('pretrained_enc_dec', False)
load_enc_file = kwargs.get('pretrained_enc_file', None)
load_dec_file = kwargs.get('pretrained_dec_file', None)
self.model_name = kwargs.get('model_name', 'cb_model')
attn_model = kwargs.get('attention_type', 'dot')
self.hidden_size = kwargs.get('hidden_size', 500)
self.encoder_nr_layers = kwargs.get('enc_nr_layers', 2)
self.decoder_nr_layers = kwargs.get('dec_nr_layers', 2)
dropout = kwargs.get('dropout', 0.1)
self.batch_size = kwargs.get('batch_size', 64)
self.clip = kwargs.get('clip', 50.0)
self.teacher_forcing_ratio = kwargs.get('teacher_forcing_ratio', 1.0)
self.learning_rate = kwargs.get('lr', 0.0001)
self.decoder_learning_ratio = kwargs.get('decoder_learning_ratio', 5.0)
self.nr_iteration = kwargs.get('nr_iterations', 4000)
self.print_every = kwargs.get('print_every', 1)
self.save_every = 500
self.embedding = nn.Embedding(self.dc.vocabulary.num_words, self.hidden_size)
if load_embedding:
self.embedding.load_state_dict(embedding_file)
# Initialize encoder & decoder models
encoder = EncoderRNN(self.hidden_size, self.embedding, self.encoder_nr_layers, dropout)
decoder = DecoderRNN(
attn_model,
self.embedding,
self.hidden_size,
self.dc.vocabulary.num_words,
self.decoder_nr_layers,
dropout
)
if load_enc_dec:
encoder.load_state_dict(load_enc_file)
decoder.load_state_dict(load_dec_file)
# Use appropriate device
encoder = encoder.to(device)
decoder = decoder.to(device)
self.encoder = encoder
self.decoder = decoder
self.encoder_optimizer = optim.Adam(encoder.parameters(), lr=self.learning_rate)
self.decoder_optimizer = optim.Adam(decoder.parameters(), lr=self.learning_rate * self.decoder_learning_ratio)
return
47: 'dribble',
48: 'ride_bike',
49: 'situp'
}
abnormal_length = 0
device = torch.device(
"cuda:1" if torch.cuda.is_available() else "cpu") # use CPU or GPU
three_d_encoder = Res3d().to(device).eval()
embed_encoder = CNN_fc_EmbedEncoder().to(device).eval()
rnn_decoder = DecoderRNN().to(device).eval()
embed_encoder.load_state_dict(
torch.load('./ckpt/cnn_encoder_epoch201.pth'))
rnn_decoder.load_state_dict(torch.load('./ckpt/rnn_decoder_epoch201.pth'))
print('---state dict loaded----')
video_dir = '../dataset/val_data'
# get videos
with torch.no_grad():
lib_vidos = os.listdir(video_dir)
lib_vidos.sort(key=lambda x: int(x[:-4]))
for video_name in lib_vidos:
assert '.avi' in video_name
video_file_path = os.path.join(video_dir, video_name)
video_cap = cv2.VideoCapture(video_file_path)
total_frams = video_cap.get(7)
class CorrelationNetwork(Sequence2SequenceNetwork):
def __init__(self, config):
self.init_writer()
self.load_configuration(config)
self.load_vocabulary()
self.prepare_data()
self.build_model()
self.load_pretrained_model()
self.train_model()
self.save_model(self.n_epochs)
self.evaluate_all()
self.close_writer()
def build_model(self):
if self.use_embeddings:
self.embedding = nn.Embedding.from_pretrained(self.embedding_wts)
else:
self.embedding = nn.Embedding(self.vocab.n_words,
self.embedding_dim)
self.encoders = []
self.encoder_optimizers = []
# Note: No embeddings used in the encoders
for m in ['v', 's']:
encoder = EncoderRNN(self.enc_input_dim[m], self.hidden_size,
self.enc_n_layers, self.dropout, self.unit,
m).to(self.device)
encoder_optimizer = optim.Adam(encoder.parameters(), lr=self.lr)
if self.modality == 'ss-vv':
checkpoint = torch.load(self.pretrained_modality[m],
map_location=self.device)
encoder.load_state_dict(checkpoint['en'])
encoder_optimizer.load_state_dict(checkpoint['en_op'])
self.encoders.append(encoder)
self.encoder_optimizers.append(encoder_optimizer)
self.decoder = DecoderRNN(self.attn_model, self.embedding_dim,
self.hidden_size, self.vocab.n_words,
self.unit, self.dec_n_layers, self.dropout,
self.embedding).to(self.device)
text_checkpoint = torch.load(self.pretrained_modality['t'],
map_location=self.device)
self.decoder.load_state_dict(text_checkpoint['de'])
self.project_factor = self.encoders[0].project_factor
self.latent2hidden = nn.Linear(self.latent_dim, self.hidden_size *
self.project_factor).to(self.device)
self.epoch = 0
def train_model(self):
best_score = 1e-200
plot_losses = []
print_loss_total = 0 # Reset every epoch
saving_skipped = 0
for epoch in range(self.epoch, self.n_epochs):
random.shuffle(self.pairs)
for iter in range(0, self.n_iters, self.batch_size):
training_batch = batch2TrainData(
self.vocab, self.pairs[iter:iter + self.batch_size],
self.modality)
# Extract fields from batch
vid_vec, lengths, speech_vec, tar_lengths = training_batch
# Run a training iteration with the current batch
loss = self.train(vid_vec, lengths, speech_vec, tar_lengths,
iter)
self.writer.add_scalar('{}loss'.format(self.data_dir), loss,
iter)
print_loss_total += loss
print_loss_avg = print_loss_total * self.batch_size / self.n_iters
print_loss_total = 0
print('Epoch: [{}/{}] Loss: {:.4f}'.format(epoch, self.n_epochs,
print_loss_avg))
# evaluate and save the model
curr_score = self.evaluate_all()
self.writer.add_scalar('{}bleu_score'.format(self.data_dir),
curr_score)
if curr_score > best_score:
saving_skipped = 0
best_score = curr_score
self.save_model(epoch)
saving_skipped += 1
if self.use_scheduler and saving_skipped > 3:
saving_skipped = 0
new_lr = self.lr * 0.5
print('Entered the dungeon...')
if new_lr > self.lr_lower_bound: # lower bound on lr
self.lr = new_lr
print('lr decreased to => {}'.format(self.lr))
def train(self, input_variable, lengths, target_variable, tar_lengths,
iter):
for i, _ in enumerate(self.encoders):
self.encoders[i].train()
self.encoders[i].zero_grad()
input_variable = input_variable.to(self.device)
lengths = lengths.to(self.device)
target_variable = target_variable.to(self.device)
tar_lengths = tar_lengths.to(self.device)
# Initialize variables
loss = 0
print_losses = []
n_totals = 0
# Forward pass through encoder
enc_out_1, enc_hidden_1 = self.encoders[0](input_variable, lengths)
enc_out_2, enc_hidden_2 = self.encoders[1](target_variable,
tar_lengths)
if self.unit == 'gru':
latent_1 = enc_hidden_1
latent_2 = enc_hidden_2
else: # lstm
(latent_1, cs_1) = enc_hidden_1
(latent_2, cs_2) = enc_hidden_2
loss = self.mean_square_error(latent_1, latent_2)
loss.backward()
# Clip gradients: gradients are modified in place
for i, _ in enumerate(self.encoders):
torch.nn.utils.clip_grad_norm_(self.encoders[i].parameters(),
self.clip)
self.encoder_optimizers[i].step()
return loss.item()
def mean_square_error(self, inp, target):
criterion = nn.MSELoss()
return criterion(inp, target)
def save_model(self, epoch):
directory = '{}'.format(self.save_dir)
if not os.path.exists(directory):
os.makedirs(directory)
torch.save(
{
'epoch': epoch,
'en_1': self.encoders[0].state_dict(),
'en_2': self.encoders[1].state_dict(),
'en_op1': self.encoder_optimizers[0].state_dict(),
'en_op2': self.encoder_optimizers[1].state_dict(),
'de': self.decoder.state_dict()
}, '{}{}-{}-{}.pth'.format(directory, self.modality, self.langs,
epoch))
def evaluate_all(self):
for i, _ in enumerate(self.encoders):
self.encoders[i].eval()
self.decoder.eval()
searcher = GreedySearchDecoder(self.encoders[0], self.decoder,
self.latent2hidden, self.device,
self.SOS_TOKEN)
refs = []
hyp = []
for pair in self.test_pairs:
output_words = self.evaluate(searcher, self.vocab, pair[0])
if output_words:
final_output = []
for x in output_words:
if x == '<EOS>':
break
final_output.append(x)
refs.append([pair[2].split()])
hyp.append(final_output)
bleu_scores = calculateBleuScores(refs, hyp)
print('Bleu score: {bleu_1} | {bleu_2} | {bleu_3} | {bleu_4}'.format(
**bleu_scores))
eg_idx = random.choice(range(len(hyp)))
print(hyp[eg_idx], refs[eg_idx])
return bleu_scores['bleu_4']
def evaluate(self,
searcher,
vocab,
sentence_or_vector,
max_length=conf['MAX_LENGTH']):
with torch.no_grad():
input_batch, lengths = inputVarVec([sentence_or_vector],
self.modality)
# Use appropriate device
input_batch = input_batch.to(self.device)
lengths = lengths.to(self.device)
# Decode sentence with searcher
tokens, scores = searcher(input_batch, lengths, max_length)
# indexes -> words
decoded_words = [
vocab.index2word[token.item()] for token in tokens
]
return decoded_words
class Sequence2SequenceNetwork(object):
def __init__(self, config):
self.init_writer()
self.load_configuration(config)
self.load_vocabulary()
self.prepare_data()
self.build_model()
self.load_pretrained_model()
self.train_model()
self.save_model(self.n_epochs)
self.evaluate_all()
self.close_writer()
def init_writer(self):
self.writer = SummaryWriter()
def load_configuration(self, config):
# Load configuration
self.iter_num = 0
self.lr = config['lr']
self.gpu = config['gpu']
self.unit = config['unit']
self.clip = config['clip']
self.beta1 = config['beta1']
self.beta2 = config['beta2']
self.langs = config['langs']
self.fusion = config['fusion']
self.log_tb = config['log_tb']
self.epsilon = config['epsilon']
self.attn_model = config['attn']
self.dropout = config['dropout']
self.emb_mode = config['emb_mode']
self.save_dir = config['save_dir']
self.data_dir = config['data_dir']
self.n_epochs = config['n_epochs']
self.SOS_TOKEN = config['SOS_TOKEN']
self.EOS_TOKEN = config['EOS_TOKEN']
self.MAX_LENGTH = config['MAX_LENGTH']
self.latent_dim = config['latent_dim']
self.batch_size = config['batch_size']
self.model_code = config['model_code']
self.vocab_path = config['vocab_path']
self.hidden_size = config['hidden_size']
self.use_cuda = torch.cuda.is_available()
self.log_tb_every = config['log_tb_every']
self.enc_n_layers = config['enc_n_layers']
self.dec_n_layers = config['dec_n_layers']
self.dec_learning_ratio = config['dec_lr']
self.bidirectional = config['bidirectional']
self.enc_input_dim = config['enc_input_dim']
self.embedding_dim = config['embedding_dim']
self.use_scheduler = config['use_scheduler']
self.use_embeddings = config['use_embeddings']
self.lr_lower_bound = config['lr_lower_bound']
self.teacher_forcing_ratio = config['tf_ratio']
self.load_model_name = config['load_model_name']
self.modality = config[
'modalities'] # no splitting as it's not multimodal case
if self.modality in ['ss-vv', 'v-s']:
self.pretrained_modality = config['pretrained_modality']
self.generate_word_embeddings = config['generate_word_embeddings']
self.device = torch.device(
'cuda:{}'.format(self.gpu) if self.use_cuda else 'cpu')
def load_vocabulary(self):
try:
with open(self.vocab_path, 'rb') as f:
self.vocab = pickle.load(f)
except FileNotFoundError as e: # build vocab if it doesn't exist
self.vocab = buildVocab()
def prepare_data(self):
# Note: The below workaround is used a lot and doing so is okay
# because this script would only be run for unimodal cases
self.pairs = prepareData(self.langs, [self.modality])[self.modality]
num_pairs = len(self.pairs)
self.pairs = self.pairs[:self.batch_size *
(num_pairs // self.batch_size)]
random.shuffle(self.pairs)
self.n_iters = len(self.pairs)
print('\nLoading test data pairs')
self.test_pairs = prepareData(self.langs, [self.modality],
train=False)[self.modality]
random.shuffle(self.test_pairs)
print(random.choice(self.pairs))
if self.use_embeddings:
if self.generate_word_embeddings:
self.embedding_wts = generateWordEmbeddings(
self.vocab, self.emb_mode)
else:
self.embedding_wts = loadWordEmbeddings(self.emb_mode)
def build_model(self):
if self.use_embeddings:
self.embedding = nn.Embedding.from_pretrained(self.embedding_wts)
else:
self.embedding = nn.Embedding(self.vocab.n_words,
self.embedding_dim)
if self.modality == 't': # Need embedding only for t2t mode
self.encoder = EncoderRNN(self.embedding_dim,
self.hidden_size,
self.enc_n_layers,
self.dropout,
self.unit,
self.modality,
self.embedding,
fusion_or_unimodal=True).to(self.device)
else:
# Note: no embedding used here
self.encoder = EncoderRNN(self.enc_input_dim,
self.hidden_size,
self.enc_n_layers,
self.dropout,
self.unit,
self.modality,
fusion_or_unimodal=True).to(self.device)
self.decoder = DecoderRNN(self.attn_model, self.embedding_dim,
self.hidden_size, self.vocab.n_words,
self.unit, self.dec_n_layers, self.dropout,
self.embedding).to(self.device)
self.encoder_optimizer = optim.Adam(self.encoder.parameters(),
lr=self.lr)
self.decoder_optimizer = optim.Adam(self.decoder.parameters(),
lr=self.lr *
self.dec_learning_ratio)
self.epoch = 0 # define here to add resume training feature
self.project_factor = self.encoder.project_factor
self.latent2hidden = nn.Linear(self.latent_dim, self.hidden_size *
self.project_factor).to(self.device)
def load_pretrained_model(self):
if self.load_model_name:
checkpoint = torch.load(self.load_model_name,
map_location=self.device)
print('Loaded {}'.format(self.load_model_name))
self.epoch = checkpoint['epoch']
self.encoder.load_state_dict(checkpoint['en'])
self.decoder.load_state_dict(checkpoint['de'])
self.encoder_optimizer.load_state_dict(checkpoint['en_op'])
self.decoder_optimizer.load_state_dict(checkpoint['de_op'])
self.embedding.load_state_dict(checkpoint['embedding'])
def train_model(self):
best_score = 1e-200
print_loss_total = 0 # Reset every epoch
saving_skipped = 0
for epoch in range(self.epoch, self.n_epochs):
incomplete = False
for iter in range(0, self.n_iters, self.batch_size):
pairs = self.pairs[iter:iter + self.batch_size]
# Skip incomplete batch
if len(pairs) < self.batch_size:
incomplete = True
continue
training_batch = batch2TrainData(self.vocab, pairs,
self.modality)
# Extract fields from batch
input_variable, lengths, target_variable, \
mask, max_target_len, _ = training_batch
if incomplete:
break
# Run a training iteration with the current batch
loss = self.train(input_variable, lengths, target_variable,
mask, max_target_len, iter)
self.writer.add_scalar('{}loss'.format(self.data_dir), loss,
iter)
print_loss_total += loss
print_loss_avg = print_loss_total * self.batch_size / self.n_iters
print_loss_total = 0
print('Epoch: [{}/{}] Loss: {:.4f}'.format(epoch, self.n_epochs,
print_loss_avg))
# evaluate and save the model
curr_score = self.evaluate_all()
self.writer.add_scalar('{}bleu_score'.format(self.data_dir),
curr_score)
if curr_score > best_score:
saving_skipped = 0
best_score = curr_score
self.save_model(epoch)
saving_skipped += 1
if self.use_scheduler and saving_skipped > 3:
saving_skipped = 0
new_lr = self.lr * 0.5
print('Entered the dungeon...')
if new_lr > self.lr_lower_bound: # lower bound on lr
self.lr = new_lr
print('lr decreased to => {}'.format(self.lr))
def train(self, input_variable, lengths, target_variable, mask,
max_target_len, iter):
self.encoder.train()
self.decoder.train()
self.encoder_optimizer.zero_grad()
self.decoder_optimizer.zero_grad()
input_variable = input_variable.to(self.device)
lengths = lengths.to(self.device)
target_variable = target_variable.to(self.device)
mask = mask.to(self.device)
# Initialize variables
loss = 0
print_losses = []
n_totals = 0
# Forward pass through encoder
encoder_outputs, encoder_hidden = self.encoder(input_variable, lengths)
# Create initial decoder input (start with SOS tokens for each sentence)
decoder_input = torch.LongTensor([[self.SOS_TOKEN] * self.batch_size])
decoder_input = decoder_input.to(self.device)
# Set initial decoder hidden state to the encoder's final hidden state
if self.unit == 'gru':
decoder_hidden = encoder_hidden[:self.decoder.n_layers]
else:
decoder_hidden = (encoder_hidden[0][:self.decoder.n_layers],
encoder_hidden[1][:self.decoder.n_layers])
if iter % conf['log_tb_every'] == 0:
# Visualize latent space
if self.unit == 'gru':
vis_hidden = decoder_hidden[-1, :, :]
else:
vis_hidden = decoder_hidden[0][-1, :, :]
self.writer.add_embedding(vis_hidden,
tag='decoder_hidden_{}'.format(iter))
use_teacher_forcing = True if random.random(
) < self.teacher_forcing_ratio else False
if use_teacher_forcing:
for t in range(max_target_len):
decoder_output, decoder_hidden = self.decoder(
decoder_input, decoder_hidden, encoder_outputs)
# Teacher forcing: next input is current target
decoder_input = target_variable[t].view(1, -1)
# Calculate and accumulate loss
mask_loss, nTotal = self.mask_nll_loss(decoder_output,
target_variable[t],
mask[t])
loss += mask_loss
print_losses.append(mask_loss.item() * nTotal)
n_totals += nTotal
else:
for t in range(max_target_len):
decoder_output, decoder_hidden = self.decoder(
decoder_input, decoder_hidden, encoder_outputs)
# No teacher forcing: next input is decoder's own current output
_, topi = decoder_output.topk(1)
decoder_input = torch.LongTensor(
[[topi[i][0] for i in range(self.batch_size)]])
decoder_input = decoder_input.to(self.device)
# Calculate and accumulate loss
mask_loss, nTotal = self.mask_nll_loss(decoder_output,
target_variable[t],
mask[t])
loss += mask_loss
print_losses.append(mask_loss.item() * nTotal)
n_totals += nTotal
loss.backward()
# Clip gradients: gradients are modified in place
torch.nn.utils.clip_grad_norm_(self.encoder.parameters(), self.clip)
torch.nn.utils.clip_grad_norm_(self.decoder.parameters(), self.clip)
self.encoder_optimizer.step()
self.decoder_optimizer.step()
return sum(print_losses) / n_totals
def mask_nll_loss(self, inp, target, mask):
n_total = mask.sum()
cross_entropy = -torch.log(
torch.gather(inp, 1, target.view(-1, 1)).squeeze(1))
loss = cross_entropy.masked_select(mask).sum()
loss = loss.to(self.device)
return loss, n_total.item()
def save_model(self, epoch):
directory = self.save_dir
if not os.path.exists(directory):
os.makedirs(directory)
torch.save(
{
'epoch': epoch,
'en': self.encoder.state_dict(),
'de': self.decoder.state_dict(),
'en_op': self.encoder_optimizer.state_dict(),
'de_op': self.decoder_optimizer.state_dict(),
'embedding': self.embedding.state_dict()
}, '{}{}-{}-{}-{}.pth'.format(directory, self.model_code,
self.modality, self.langs, epoch))
def evaluate_all(self):
self.encoder.eval()
self.decoder.eval()
searcher = GreedySearchDecoder(self.encoder, self.decoder, None,
self.device, self.SOS_TOKEN)
refs = []
hyp = []
for pair in self.test_pairs:
output_words = self.evaluate(self.encoder, self.decoder, searcher,
self.vocab, pair[0])
if output_words:
final_output = []
for x in output_words:
if x == '<EOS>':
break
final_output.append(x)
refs.append([pair[1].split()])
hyp.append(final_output)
bleu_scores = calculateBleuScores(refs, hyp)
print('Bleu score: {bleu_1} | {bleu_2} | {bleu_3} | {bleu_4}'.format(
**bleu_scores))
eg_idx = random.choice(range(len(hyp)))
print(hyp[eg_idx], refs[eg_idx])
return bleu_scores['bleu_4']
def evaluate(self,
encoder,
decoder,
searcher,
vocab,
sentence_or_vector,
max_length=conf['MAX_LENGTH']):
with torch.no_grad():
if self.modality == 't': # `sentence_or_vector` ~> sentence
# Format input sentence as a batch
# words => indexes
indexes_batch = [
indexesFromSentence(vocab, sentence_or_vector)
]
if None in indexes_batch:
return None
for idx, indexes in enumerate(indexes_batch):
indexes_batch[idx] = indexes_batch[idx] + [self.EOS_TOKEN]
# Create lengths tensor
lengths = torch.tensor(
[len(indexes) for indexes in indexes_batch])
# Transpose dimensions of batch to match models' expectations
input_batch = torch.LongTensor(indexes_batch).transpose(0, 1)
else: # `sentence_or_vector` ~> vector
input_batch, lengths = inputVarVec([sentence_or_vector],
self.modality)
# Use appropriate device
input_batch = input_batch.to(self.device)
lengths = lengths.to(self.device)
# Decode sentence with searcher
tokens, scores = searcher(input_batch, lengths, max_length)
# indexes -> words
decoded_words = [
vocab.index2word[token.item()] for token in tokens
]
return decoded_words
def close_writer(self):
self.writer.close()
class FusionNetwork(Sequence2SequenceNetwork):
def __init__(self, config):
self.init_writer()
self.load_configuration(config)
self.load_vocabulary()
self.prepare_data()
self.build_model()
self.load_pretrained_model()
self.train_model()
self.save_model(self.n_epochs)
self.evaluate_all()
self.close_writer()
def prepare_data(self):
self.modality = self.modality.split('-')
self.pairs = prepareData(self.langs, self.modality) # dict: m => pairs
num_pairs = len(random.choice(list(self.pairs.values())))
rand_indices = random.sample(list(range(num_pairs)), num_pairs)
self.n_iters = num_pairs
for m in self.modality:
self.pairs[m] = self.pairs[m][: self.batch_size * (
num_pairs // self.batch_size)]
# Shuffle all modalities the same way
self.pairs[m] = [p for p, _ in sorted(zip(self.pairs[m], rand_indices))]
print(random.choice(self.pairs[m]))
print('\nLoading test data pairs')
self.test_pairs = prepareData(self.langs, self.modality, train=False)
self.num_test_pairs = len(random.choice(list(self.test_pairs.values())))
rand_indices = random.sample(list(range(self.num_test_pairs)),
self.num_test_pairs)
if self.use_embeddings:
if self.generate_word_embeddings:
self.embedding_wts = generateWordEmbeddings(self.vocab,
self.emb_mode)
else:
self.embedding_wts = loadWordEmbeddings(self.emb_mode)
def build_model(self):
if self.use_embeddings:
self.embedding = nn.Embedding.from_pretrained(self.embedding_wts)
else:
self.embedding = nn.Embedding(self.vocab.n_words,
self.embedding_dim)
self.encoder = MultimodalEncoderRNN(self.fusion, self.hidden_size,
self.enc_n_layers, self.dropout,
self.unit, self.modality,
self.embedding,
self.device).to(self.device)
if self.fusion == 'early' or self.fusion is None:
parameter_list = self.encoder.parameters()
else:
parameter_list = []
for m in self.encoder.modalities:
parameter_list += list(self.encoder.encoder[m].parameters())
# Need to expand hidden layer according to # modalities for early fusion
self.decoder = DecoderRNN(self.attn_model, self.embedding_dim,
self.hidden_size, self.vocab.n_words,
self.unit, self.dec_n_layers, self.dropout,
self.embedding).to(self.device)
self.encoder_optimizer = optim.Adam(parameter_list,
lr=self.lr)
self.decoder_optimizer = optim.Adam(self.decoder.parameters(),
lr=self.lr*self.dec_learning_ratio)
self.epoch = 0 # define here to add resume training feature
def load_pretrained_model(self):
if self.load_model_name:
checkpoint = torch.load(self.load_model_name,
map_location=self.device)
print('Loaded {}'.format(self.load_model_name))
self.epoch = checkpoint['epoch']
self.encoder.load_state_dict(checkpoint['en'])
self.decoder.load_state_dict(checkpoint['de'])
self.encoder_optimizer.load_state_dict(checkpoint['en_op'])
self.decoder_optimizer.load_state_dict(checkpoint['de_op'])
self.embedding.load_state_dict(checkpoint['embedding'])
self.vocab.__dict__ = checkpoint['vocab_dict']
self.evaluate_all()
def train_model(self):
best_score = 1e-200
print_loss_total = 0 # Reset every epoch
num_pairs = {}
for m in self.modality:
num_pairs[m] = len(self.pairs[m])
saving_skipped = 0
for epoch in range(self.epoch, self.n_epochs):
incomplete = False
for iter in range(0, self.n_iters, self.batch_size):
training_batch = {}
input_variable = {}
lengths = {}
for m in self.modality:
pairs = self.pairs[m][iter: iter + self.batch_size]
# Skip incomplete batch
if len(pairs) < self.batch_size:
incomplete = True
continue
training_batch[m] = batch2TrainData(
self.vocab, pairs, m)
# Extract fields from batch
input_variable[m], lengths[m], target_variable, \
mask, max_target_len, _ = training_batch[m]
if incomplete:
break
# Run a training iteration with the current batch
loss = self.train(input_variable, lengths, target_variable,
mask, max_target_len, epoch, iter)
self.writer.add_scalar('{}loss'.format(self.data_dir), loss, iter)
print_loss_total += loss
print_loss_avg = print_loss_total * self.batch_size / self.n_iters
print_loss_total = 0
print('Epoch: [{}/{}] Loss: {:.4f}'.format(
epoch, self.n_epochs, print_loss_avg))
# evaluate and save the model
curr_score = self.evaluate_all()
self.writer.add_scalar('{}bleu_score'.format(self.data_dir), curr_score, iter)
if curr_score > best_score:
saving_skipped = 0
best_score = curr_score
self.save_model(epoch)
saving_skipped += 1
if self.use_scheduler and saving_skipped > 3:
saving_skipped = 0
new_lr = self.lr * 0.5
print('Entered the dungeon...')
if new_lr > self.lr_lower_bound: # lower bound on lr
self.lr = new_lr
print('lr decreased to => {}'.format(self.lr))
def train(self, input_variable, lengths, target_variable, mask,
max_target_len, epoch, iter):
self.encoder.train()
self.decoder.train()
self.encoder_optimizer.zero_grad()
self.decoder_optimizer.zero_grad()
for m in self.modality:
input_variable[m] = input_variable[m].to(self.device)
lengths[m] = lengths[m].to(self.device)
target_variable = target_variable.to(self.device)
mask = mask.to(self.device)
# Initialize variables
loss = 0
print_losses = []
n_totals = 0
# Forward pass through encoder
encoder_outputs, encoder_hidden = self.encoder(input_variable, lengths)
# Create initial decoder input (start with SOS tokens for each sentence)
decoder_input = torch.LongTensor([[self.SOS_TOKEN] * self.batch_size])
decoder_input = decoder_input.to(self.device)
# Set initial decoder hidden state to the encoder's final hidden state
if self.unit == 'gru':
decoder_hidden = encoder_hidden[:self.decoder.n_layers]
else:
decoder_hidden = (encoder_hidden[0][:self.decoder.n_layers],
encoder_hidden[1][:self.decoder.n_layers])
if iter % conf['log_tb_every'] == 1:
# Visualize latent space
if self.unit == 'gru':
vis_hidden = decoder_hidden[-1, :, :]
else:
vis_hidden = decoder_hidden[0][-1, :, :]
self.writer.add_embedding(vis_hidden,
tag='decoder_hidden_{}_{}'.format(
epoch, iter))
use_teacher_forcing = True if random.random() < self.teacher_forcing_ratio else False
if use_teacher_forcing:
for t in range(max_target_len):
decoder_output, decoder_hidden = self.decoder(
decoder_input, decoder_hidden, encoder_outputs)
# Teacher forcing: next input is current target
decoder_input = target_variable[t].view(1, -1)
# Calculate and accumulate loss
mask_loss, nTotal = self.mask_nll_loss(decoder_output,
target_variable[t],
mask[t])
loss += mask_loss
print_losses.append(mask_loss.item() * nTotal)
n_totals += nTotal
else:
for t in range(max_target_len):
decoder_output, decoder_hidden = self.decoder(
decoder_input, decoder_hidden, encoder_outputs
)
# No teacher forcing: next input is decoder's own current output
_, topi = decoder_output.topk(1)
decoder_input = torch.LongTensor(
[[topi[i][0] for i in range(self.batch_size)]])
decoder_input = decoder_input.to(self.device)
# Calculate and accumulate loss
mask_loss, nTotal = self.mask_nll_loss(
decoder_output, target_variable[t], mask[t])
loss += mask_loss
print_losses.append(mask_loss.item() * nTotal)
n_totals += nTotal
loss.backward()
# Clip gradients: gradients are modified in place
torch.nn.utils.clip_grad_norm_(self.encoder.parameters(), self.clip)
torch.nn.utils.clip_grad_norm_(self.decoder.parameters(), self.clip)
self.encoder_optimizer.step()
self.decoder_optimizer.step()
return sum(print_losses) / n_totals
def mask_nll_loss(self, inp, target, mask):
n_total = mask.sum()
cross_entropy = -torch.log(torch.gather(inp, 1,
target.view(-1, 1)).squeeze(1))
loss = cross_entropy.masked_select(mask).sum()
loss = loss.to(self.device)
return loss, n_total.item()
def save_model(self, epoch):
directory = self.save_dir
if not os.path.exists(directory):
os.makedirs(directory)
torch.save({
'epoch': epoch,
'en': self.encoder.state_dict(),
'de': self.decoder.state_dict(),
'en_op': self.encoder_optimizer.state_dict(),
'de_op': self.decoder_optimizer.state_dict(),
'vocab_dict': self.vocab.__dict__,
'embedding': self.embedding.state_dict()},
'{}{}-{}-{}.pth'.format(directory, self.model_code, epoch, iter))
def evaluate_all(self):
self.encoder.eval()
self.decoder.eval()
searcher = GreedySearchDecoder(
self.encoder, self.decoder, None, self.device, self.SOS_TOKEN)
refs = []
hyp = []
for id in range(self.num_test_pairs):
# Sample test pairs of each modality
output_words, reference = self.evaluate(
searcher, self.vocab, self.test_pairs, id)
if output_words:
final_output = []
for x in output_words:
if x == '<EOS>':
break
final_output.append(x)
refs.append(reference.split())
hyp.append(final_output)
bleu_scores = calculateBleuScores(refs, hyp)
print('Bleu score: {bleu_1} | {bleu_2} | {bleu_3} | {bleu_4}'.format(
**bleu_scores))
eg_idx = random.choice(range(len(hyp)))
print(hyp[eg_idx], refs[eg_idx])
return bleu_scores['bleu_4']
def evaluate(self, searcher, vocab, test_pairs, id,
max_length=conf['MAX_LENGTH']):
lengths = {}
input_batch = {}
with torch.no_grad():
reference = random.choice(list(test_pairs.values()))[id][1]
for m in self.modality:
sentence_or_vector = test_pairs[m][id][0]
if m == 't': # `sentence_or_vector` ~> sentence
# Format input sentence as a batch
# words => indexes
indexes_batch = [indexesFromSentence(vocab, sentence_or_vector)]
if None in indexes_batch:
return None
for idx, indexes in enumerate(indexes_batch):
indexes_batch[idx] = indexes_batch[idx] + [self.EOS_TOKEN]
# Create lengths tensor
lengths[m] = torch.tensor(
[len(indexes) for indexes in indexes_batch])
# Transpose dimensions of batch to match models' expectations
input_batch[m] = torch.LongTensor(
indexes_batch).transpose(0, 1)
else: # `sentence_or_vector` ~> vector
input_batch[m], lengths[m] = \
inputVarVec([sentence_or_vector], m)
# Use appropriate device
input_batch[m] = input_batch[m].to(self.device)
lengths[m] = lengths[m].to(self.device)
# Decode sentence with searcher
tokens, scores = searcher(input_batch, lengths, max_length)
# indexes -> words
decoded_words = [vocab.index2word[token.item()] for token in tokens]
return decoded_words, reference
def close_writer(self):
self.writer.close()