def _build(self):
self.global_step = tf.Variable(0, name='global_step')
self.create_models(self.cfg)
self.vocab_list = self.cfg.vocab_list
self.vocab_to_int, self.int_to_vocab = wv.create_look_up_table(
self.vocab_list)
self.decode_fn = wv.decode_with_lookup_table(self.int_to_vocab)
def __init__(self, cfg):
# making session
self.cfg = cfg
self.step = 0
self.epsilon = 1.0
self._use_labeler_as_reward = cfg.use_labeler_as_reward
self._use_oracle_instruction = cfg.use_oracle_instruction
self._use_synonym_for_rollout = cfg.use_synonym_for_rollout
# Vocab loading
vocab_path = get_vocab_path(cfg)
self.vocab_list = wv.load_vocab_list(vocab_path)
self.vocab_list = ['eos', 'sos', 'nothing'] + self.vocab_list[1:]
v2i, i2v = wv.create_look_up_table(self.vocab_list)
self.encode_fn = wv.encode_text_with_lookup_table(
v2i, max_sequence_length=self.cfg.max_sequence_length)
self.decode_fn = wv.decode_with_lookup_table(i2v)
labeler_config = get_labeler_config(cfg, self.vocab_list)
if self._use_labeler_as_reward or not self._use_oracle_instruction:
self.labeler = Labeler(labeler_config=labeler_config)
self.labeler.set_captioning_model(
labeler_config, labeler_config['captioning_weight_path'])
self.labeler.set_answering_model(labeler_config,
labeler_config['answering_weight_path'])
def __init__(self, cfg):
# making session
self.cfg = cfg
self.step = 0
self.epsilon = 1.0
# Vocab loading
vocab_path = get_vocab_path(cfg)
self.vocab_list = wv.load_vocab_list(vocab_path)
v2i, i2v = wv.create_look_up_table(self.vocab_list)
self.encode_fn = wv.encode_text_with_lookup_table(
v2i, max_sequence_length=self.cfg.max_sequence_length)
self.decode_fn = wv.decode_with_lookup_table(i2v)
def main(_):
tf.enable_v2_behavior()
##############################################################################
######################### Data loading and processing ########################
##############################################################################
print('Loading data')
# with gfile.GFile(transition_path, 'r') as f:
# transitions = np.load(f)
with gfile.GFile(transition_state_path, 'r') as f:
states = np.load(f)
states = np.float32(states)
with gfile.GFile(transition_label_path, 'r') as f:
captions = pickle.load(f)
with gfile.GFile(answer_path, 'r') as f:
answers = pickle.load(f)
with gfile.GFile(vocab_path, 'r') as f:
vocab_list = f.readlines()
vocab_list = [w[:-1].decode('utf-8') for w in vocab_list]
vocab_list = ['eos', 'sos', 'nothing'] + vocab_list
vocab_list[-1] = 'to'
v2i, i2v = wv.create_look_up_table(vocab_list)
encode_fn = wv.encode_text_with_lookup_table(v2i)
decode_fn = wv.decode_with_lookup_table(i2v)
caption_decoding_map = {v: k for k, v in captions[0].items()}
decompressed_captions = []
for caption in captions[1:]:
new_caption = []
for c in caption:
new_caption.append(caption_decoding_map[c])
decompressed_captions.append(new_caption)
captions = decompressed_captions
encoded_captions = []
new_answers = []
for i, all_cp in enumerate(captions):
for cp in all_cp:
encoded_captions.append(np.array(encode_fn(cp)))
for a in answers[i]:
new_answers.append(float(a))
all_caption_n = len(encoded_captions)
encoded_captions = np.array(encoded_captions)
encoded_captions = pad_to_max_length(encoded_captions)
answers = np.float32(new_answers)
obs_idx, caption_idx = [], []
curr_caption_idx = 0
for i, _ in enumerate(states):
for cp in captions[i]:
obs_idx.append(i)
caption_idx.append(curr_caption_idx)
curr_caption_idx += 1
assert curr_caption_idx == all_caption_n
obs_idx = np.array(obs_idx)
caption_idx = np.array(caption_idx)
all_idx = np.arange(len(caption_idx))
train_idx = all_idx[:int(len(all_idx) * 0.7)]
test_idx = all_idx[int(len(all_idx) * 0.7):]
print('Number of training examples: {}'.format(len(train_idx)))
print('Number of test examples: {}\n'.format(len(test_idx)))
##############################################################################
############################# Training Setup #################################
##############################################################################
embedding_dim = 32
units = 64
vocab_size = len(vocab_list)
batch_size = 128
max_sequence_length = 21
encoder_config = {'name': 'state', 'embedding_dim': 64}
decoder_config = {
'name': 'state',
'word_embedding_dim': 64,
'hidden_units': 512,
'vocab_size': len(vocab_list),
}
encoder = get_answering_encoder(encoder_config)
decoder = get_answering_decoder(decoder_config)
projection_layer = tf.keras.layers.Dense(1,
activation='sigmoid',
name='answering_projection')
optimizer = tf.keras.optimizers.Adam(1e-4)
bce = tf.keras.losses.BinaryCrossentropy()
@tf.function
def compute_loss(obs, instruction, target):
instruction = tf.expand_dims(instruction, axis=-1)
hidden = decoder.reset_state(batch_size=target.shape[0])
features = encoder(obs)
for i in tf.range(max_sequence_length):
_, hidden, _ = decoder(instruction[:, i], features, hidden)
projection = tf.squeeze(projection_layer(hidden), axis=1)
loss = bce(target, projection)
return loss, projection
@tf.function
def train_step(obs, instruction, target):
with tf.GradientTape() as tape:
loss, _ = compute_loss(obs, instruction, target)
trainable_variables = encoder.trainable_variables + decoder.trainable_variables + projection_layer.trainable_variables
gradients = tape.gradient(loss, trainable_variables)
optimizer.apply_gradients(zip(gradients, trainable_variables))
return loss
##############################################################################
############################# Training Loop ##################################
##############################################################################
print('Start training...\n')
start_epoch = 0
if FLAGS.save_dir:
checkpoint_path = FLAGS.save_dir
ckpt = tf.train.Checkpoint(encoder=encoder,
decoder=decoder,
projection_layer=projection_layer,
optimizer=optimizer)
ckpt_manager = tf.train.CheckpointManager(ckpt,
checkpoint_path,
max_to_keep=5)
if ckpt_manager.latest_checkpoint:
start_epoch = int(ckpt_manager.latest_checkpoint.split('-')[-1])
epochs = 400
step_per_epoch = int(all_caption_n / batch_size)
previous_best, previous_best_accuracy = 100., 0.0
for epoch in range(start_epoch, epochs):
start = time.time()
total_loss = 0
for batch in range(step_per_epoch):
batch_idx = np.random.choice(train_idx, size=batch_size)
input_tensor = tf.convert_to_tensor(states[obs_idx[batch_idx], :])
instruction = tf.convert_to_tensor(
encoded_captions[caption_idx[batch_idx]])
target = tf.convert_to_tensor(answers[caption_idx[batch_idx]])
batch_loss = train_step(input_tensor, instruction, target)
total_loss += batch_loss
if batch % 1000 == 0:
print('Epoch {} Batch {} Loss {:.4f}'.format(
epoch, batch, batch_loss.numpy()))
if epoch % 5 == 0 and FLAGS.save_dir:
test_total_loss = 0
accuracy = 0
for batch in range(10):
batch_idx = np.arange(batch_size) + batch * batch_size
idx = test_idx[batch_idx]
input_tensor = tf.convert_to_tensor(states[obs_idx[idx], :])
instruction = tf.convert_to_tensor(
encoded_captions[caption_idx[idx]])
target = tf.convert_to_tensor(answers[caption_idx[idx]])
t_loss, prediction = compute_loss(input_tensor, instruction,
target)
test_total_loss += t_loss
accuracy += np.mean(
np.float32(np.float32(prediction > 0.5) == target))
test_total_loss /= 10.
accuracy /= 10.
if accuracy > previous_best_accuracy:
previous_best_accuracy, previous_best = accuracy, test_total_loss
ckpt_manager.save(checkpoint_number=epoch)
print('\nEpoch {} | Loss {:.6f} | Val loss {:.6f} | Accuracy {:.3f}'.
format(epoch + 1, total_loss / step_per_epoch, previous_best,
previous_best_accuracy))
print('Time taken for 1 epoch {:.6f} sec\n'.format(time.time() -
start))
if epoch % 10 == 0:
test_total_loss = 0
accuracy = 0
for batch in range(len(test_idx) // batch_size):
batch_idx = np.arange(batch_size) + batch * batch_size
idx = test_idx[batch_idx]
input_tensor = tf.convert_to_tensor(states[obs_idx[idx], :])
instruction = tf.convert_to_tensor(
encoded_captions[caption_idx[idx]])
target = tf.convert_to_tensor(answers[caption_idx[idx]])
t_loss, prediction = compute_loss(input_tensor, instruction,
target)
test_total_loss += t_loss
accuracy += np.mean(
np.float32(np.float32(prediction > 0.5) == target))
test_total_loss /= (len(test_idx) // batch_size)
accuracy /= (len(test_idx) // batch_size)
if accuracy > previous_best_accuracy and FLAGS.save_dir:
previous_best_accuracy, previous_best = accuracy, test_total_loss
ckpt_manager.save(checkpoint_number=epoch)
print('\n====================================================')
print('Test Loss {:.6f} | Test Accuracy {:.3f}'.format(
test_total_loss, accuracy))
print('====================================================\n')
def main(_):
tf.enable_v2_behavior()
##############################################################################
######################### Data loading and processing ########################
##############################################################################
print('Loading data')
with gfile.GFile(transition_path, 'r') as f:
transitions = np.load(f)
if np.max(transitions) > 1.0:
transitions = transitions / 255.0
with gfile.GFile(synthetic_transition_path, 'r') as f:
synthetic_transitions = np.load(f)
if np.max(synthetic_transitions) > 1.0:
synthetic_transitions = synthetic_transitions / 255.0
with gfile.GFile(transition_label_path, 'r') as f:
captions = pickle.load(f)
with gfile.GFile(synthetic_transition_label_path, 'r') as f:
synthetic_captions = pickle.load(f)
with gfile.GFile(vocab_path, 'r') as f:
vocab_list = f.readlines()
vocab_list = [w[:-1].decode('utf-8') for w in vocab_list]
vocab_list = ['eos', 'sos'] + vocab_list
v2i, i2v = wv.create_look_up_table(vocab_list)
encode_fn = wv.encode_text_with_lookup_table(v2i)
decode_fn = wv.decode_with_lookup_table(i2v)
encoded_captions = []
for all_cp in captions:
for cp in all_cp:
cp = 'sos ' + cp + ' eos'
encoded_captions.append(np.array(encode_fn(cp)))
synthetic_encoded_captions = []
for all_cp in synthetic_captions:
for cp in all_cp:
cp = 'sos ' + cp + ' eos'
synthetic_encoded_captions.append(np.array(encode_fn(cp)))
all_caption_n = len(encoded_captions)
all_synthetic_caption_n = len(synthetic_encoded_captions)
encoded_captions = np.array(encoded_captions)
encoded_captions = pad_to_max_length(encoded_captions, max_l=15)
synthetic_encoded_captions = np.array(synthetic_encoded_captions)
synthetic_encoded_captions = pad_to_max_length(synthetic_encoded_captions,
max_l=15)
obs_idx, caption_idx, negative_caption_idx = [], [], []
curr_caption_idx = 0
for i, _ in enumerate(transitions):
for cp in captions[i]:
obs_idx.append(i)
if 'nothing' not in cp:
caption_idx.append(curr_caption_idx)
else:
negative_caption_idx.append(curr_caption_idx)
curr_caption_idx += 1
assert curr_caption_idx == all_caption_n
synthetic_obs_idx, synthetic_caption_idx = [], []
synthetic_negative_caption_idx = []
curr_caption_idx = 0
for i, _ in enumerate(synthetic_transitions):
for cp in synthetic_captions[i]:
synthetic_obs_idx.append(i)
if 'nothing' not in cp:
synthetic_caption_idx.append(curr_caption_idx)
else:
synthetic_negative_caption_idx.append(curr_caption_idx)
curr_caption_idx += 1
assert curr_caption_idx == all_synthetic_caption_n
obs_idx = np.array(obs_idx)
caption_idx = np.array(caption_idx)
negative_caption_idx = np.array(negative_caption_idx)
all_idx = np.arange(len(caption_idx))
train_idx = all_idx[:int(len(all_idx) * 0.8)]
test_idx = all_idx[int(len(all_idx) * 0.8):]
print('Number of training examples: {}'.format(len(train_idx)))
print('Number of test examples: {}\n'.format(len(test_idx)))
synthetic_obs_idx = np.array(synthetic_obs_idx)
synthetic_caption_idx = np.array(synthetic_caption_idx)
synthetic_negative_caption_idx = np.array(synthetic_negative_caption_idx)
synthetic_all_idx = np.arange(len(synthetic_caption_idx))
synthetic_train_idx = synthetic_all_idx[:int(len(synthetic_all_idx) * 0.8)]
synthetic_test_idx = synthetic_all_idx[int(len(synthetic_all_idx) * 0.8):]
print('Number of synthetic training examples: {}'.format(
len(synthetic_train_idx)))
print('Number of synthetic test examples: {}\n'.format(
len(synthetic_test_idx)))
def sample_batch(data_type, batch_size, mode='train'):
is_synthetic = data_type == 'synthetic'
transitions_s = synthetic_transitions if is_synthetic else transitions
encoded_captions_s = synthetic_encoded_captions if is_synthetic else encoded_captions
obs_idx_s = synthetic_obs_idx if is_synthetic else obs_idx
caption_idx_s = synthetic_caption_idx if is_synthetic else caption_idx
all_idx_s = synthetic_all_idx if is_synthetic else all_idx
train_idx_s = synthetic_train_idx if is_synthetic else train_idx
test_idx_s = synthetic_test_idx if is_synthetic else test_idx
if mode == 'train':
batch_idx_s = np.random.choice(train_idx_s, size=batch_size)
else:
batch_idx_s = np.random.choice(test_idx_s, size=batch_size)
input_tensor = tf.convert_to_tensor(
np.concatenate([
transitions_s[obs_idx_s[batch_idx_s], 1, :],
transitions_s[obs_idx_s[batch_idx_s], 1, :]
]))
positive_idx = caption_idx_s[batch_idx_s]
negative_idx = caption_idx_s[np.random.choice(train_idx_s,
size=batch_size)]
caption_tensor = tf.convert_to_tensor(
np.concatenate([
encoded_captions_s[positive_idx],
encoded_captions_s[negative_idx]
],
axis=0))
target_tensor = tf.convert_to_tensor(
np.float32(
np.concatenate([np.ones(batch_size),
np.zeros(batch_size)],
axis=0)))
return input_tensor, caption_tensor, target_tensor
##############################################################################
############################# Training Setup #################################
##############################################################################
embedding_dim = 32
units = 64
vocab_size = len(vocab_list)
batch_size = 64
max_sequence_length = 15
encoder_config = {'name': 'image', 'embedding_dim': 64}
decoder_config = {
'name': 'attention',
'word_embedding_dim': 64,
'hidden_units': 256,
'vocab_size': len(vocab_list),
}
encoder = get_answering_encoder(encoder_config)
decoder = get_answering_decoder(decoder_config)
projection_layer = tf.keras.layers.Dense(1,
activation='sigmoid',
name='answering_projection')
optimizer = tf.keras.optimizers.Adam(1e-4)
bce = tf.keras.losses.BinaryCrossentropy()
@tf.function
def compute_loss(obs, instruction, target, training):
print('Build compute loss...')
instruction = tf.expand_dims(instruction, axis=-1)
hidden = decoder.reset_state(batch_size=target.shape[0])
features = encoder(obs, training=training)
for i in tf.range(max_sequence_length):
_, hidden, _ = decoder(instruction[:, i],
features,
hidden,
training=training)
projection = tf.squeeze(projection_layer(hidden), axis=1)
loss = bce(target, projection)
return loss, projection
@tf.function
def train_step(obs, instruction, target):
print('Build train step...')
with tf.GradientTape() as tape:
loss, _ = compute_loss(obs, instruction, target, True)
trainable_variables = encoder.trainable_variables + decoder.trainable_variables + projection_layer.trainable_variables
print('num trainable: ', len(trainable_variables))
gradients = tape.gradient(loss, trainable_variables)
optimizer.apply_gradients(zip(gradients, trainable_variables))
return loss
##############################################################################
############################# Training Loop ##################################
##############################################################################
print('Start training...\n')
start_epoch = 0
if FLAGS.save_dir:
checkpoint_path = FLAGS.save_dir
ckpt = tf.train.Checkpoint(encoder=encoder,
decoder=decoder,
projection_layer=projection_layer,
optimizer=optimizer)
ckpt_manager = tf.train.CheckpointManager(ckpt,
checkpoint_path,
max_to_keep=5)
if ckpt_manager.latest_checkpoint:
start_epoch = int(ckpt_manager.latest_checkpoint.split('-')[-1])
epochs = 400
step_per_epoch = int(all_caption_n / batch_size)
previous_best, previous_best_accuracy = 100., 0.0
# input_tensor, instruction, target = sample_batch('synthetic', batch_size,
# 'train')
for epoch in range(start_epoch, epochs):
start = time.time()
total_loss = 0
for batch in range(step_per_epoch):
input_tensor, instruction, target = sample_batch(
'synthetic', batch_size, 'train')
batch_loss = train_step(input_tensor, instruction, target)
total_loss += batch_loss
# print(batch, batch_loss)
# print(instruction[0])
# print(encode_fn('nothing'))
# print('====================================')
if batch % 1000 == 0:
print('Epoch {} Batch {} Loss {:.4f}'.format(
epoch, batch, batch_loss.numpy()))
if epoch % 5 == 0 and FLAGS.save_dir:
test_total_loss = 0
accuracy = 0
for batch in range(10):
input_tensor, instruction, target = sample_batch(
'synthetic', batch_size, 'test')
t_loss, prediction = compute_loss(input_tensor, instruction,
target, False)
test_total_loss += t_loss
accuracy += np.mean(
np.float32(np.float32(prediction > 0.5) == target))
test_total_loss /= 10.
accuracy /= 10.
if accuracy > previous_best_accuracy:
previous_best_accuracy, previous_best = accuracy, test_total_loss
ckpt_manager.save(checkpoint_number=epoch)
print('\nEpoch {} | Loss {:.6f} | Val loss {:.6f} | Accuracy {:.3f}'.
format(epoch + 1, total_loss / step_per_epoch, previous_best,
previous_best_accuracy))
print('Time taken for 1 epoch {:.6f} sec\n'.format(time.time() -
start))
if epoch % 10 == 0:
test_total_loss = 0
accuracy = 0
for batch in range(len(test_idx) // batch_size):
input_tensor, instruction, target = sample_batch(
'synthetic', batch_size, 'test')
t_loss, prediction = compute_loss(input_tensor,
instruction,
target,
training=False)
test_total_loss += t_loss
accuracy += np.mean(
np.float32(np.float32(prediction > 0.5) == target))
test_total_loss /= (len(test_idx) // batch_size)
accuracy /= (len(test_idx) // batch_size)
if accuracy > previous_best_accuracy and FLAGS.save_dir:
previous_best_accuracy, previous_best = accuracy, test_total_loss
ckpt_manager.save(checkpoint_number=epoch)
print('\n====================================================')
print('Test Loss {:.6f} | Test Accuracy {:.3f}'.format(
test_total_loss, accuracy))
print('====================================================\n')
def main(_):
tf.enable_v2_behavior()
##############################################################################
######################### Data loading and processing ########################
##############################################################################
print('Loading data')
with gfile.GFile(_TRANSITION_PATH, 'r') as f:
transitions = np.load(f)
if np.max(transitions) > 1.0:
transitions = transitions / 255.0
with gfile.GFile(_SYNTHETIC_TRANSITION_PATH, 'r') as f:
synthetic_tran_sitions = np.load(f)
if np.max(synthetic_transitions) > 1.0:
synthetic_transitions = synthetic_transitions / 255.0
with gfile.GFile(transition_label_path, 'r') as f:
captions = pickle.load(f)
with gfile.GFile(_SYNTHETIC_TRANSITION_LABEL_PATH, 'r') as f:
synthetic_captions = pickle.load(f)
with gfile.GFile(vocab_path, 'r') as f:
vocab_list = f.readlines()
vocab_list = [w[:-1].decode('utf-8') for w in vocab_list]
vocab_list = ['eos', 'sos'] + vocab_list
v2i, i2v = wv.create_look_up_table(vocab_list)
encode_fn = wv.encode_text_with_lookup_table(v2i)
decode_fn = wv.decode_with_lookup_table(i2v)
encoded_captions = []
for all_cp in captions:
for cp in all_cp:
cp = 'sos ' + cp + ' eos'
encoded_captions.append(np.array(encode_fn(cp)))
synthetic_encoded_captions = []
for all_cp in synthetic_captions:
for cp in all_cp:
cp = 'sos ' + cp + ' eos'
synthetic_encoded_captions.append(np.array(encode_fn(cp)))
all_caption_n = len(encoded_captions)
all_synthetic_caption_n = len(synthetic_encoded_captions)
encoded_captions = np.array(encoded_captions)
encoded_captions = pad_to_max_length(encoded_captions, max_l=15)
synthetic_encoded_captions = np.array(synthetic_encoded_captions)
synthetic_encoded_captions = pad_to_max_length(synthetic_encoded_captions,
max_l=15)
obs_idx, caption_idx = [], []
curr_caption_idx = 0
for i, _ in enumerate(transitions):
for cp in captions[i]:
obs_idx.append(i)
caption_idx.append(curr_caption_idx)
curr_caption_idx += 1
assert curr_caption_idx == all_caption_n
synthetic_obs_idx, synthetic_caption_idx = [], []
curr_caption_idx = 0
for i, _ in enumerate(synthetic_transitions):
for cp in synthetic_captions[i]:
synthetic_obs_idx.append(i)
synthetic_caption_idx.append(curr_caption_idx)
curr_caption_idx += 1
assert curr_caption_idx == all_synthetic_caption_n
obs_idx = np.array(obs_idx)
caption_idx = np.array(caption_idx)
all_idx = np.arange(len(caption_idx))
train_idx = all_idx[:int(len(all_idx) * 0.8)]
test_idx = all_idx[int(len(all_idx) * 0.8):]
print('Number of training examples: {}'.format(len(train_idx)))
print('Number of test examples: {}\n'.format(len(test_idx)))
synthetic_obs_idx = np.array(synthetic_obs_idx)
synthetic_caption_idx = np.array(synthetic_caption_idx)
synthetic_all_idx = np.arange(len(synthetic_caption_idx))
synthetic_train_idx = synthetic_all_idx[:int(len(synthetic_all_idx) * 0.8)]
synthetic_test_idx = synthetic_all_idx[int(len(synthetic_all_idx) * 0.8):]
print('Number of synthetic training examples: {}'.format(
len(synthetic_train_idx)))
print('Number of synthetic test examples: {}\n'.format(
len(synthetic_test_idx)))
##############################################################################
############################# Training Setup #################################
##############################################################################
embedding_dim = 32
units = 64
vocab_size = len(vocab_list)
batch_size = 64
max_sequence_length = 15
encoder_config = {'name': 'image', 'embedding_dim': 32}
decoder_config = {
'name': 'attention',
'word_embedding_dim': 64,
'hidden_units': 256,
'vocab_size': len(vocab_list),
}
encoder = get_captioning_encoder(encoder_config)
decoder = get_captioning_decoder(decoder_config)
optimizer = tf.keras.optimizers.Adam()
loss_object = tf.keras.losses.SparseCategoricalCrossentropy(
from_logits=True, reduction='none')
def loss_function(real, pred, sos_symbol=1):
mask = tf.math.logical_not(tf.math.equal(real, sos_symbol))
loss_ = loss_object(real, pred)
mask = tf.cast(mask, dtype=loss_.dtype)
loss_ *= mask
return tf.reduce_mean(loss_)
@tf.function
def train_step(input_tensor, target):
"""Traing on a batch of data."""
loss = 0
# initializing the hidden state for each batch
# because the captions are not related from image to image
hidden = decoder.reset_state(batch_size=target.shape[0])
dec_input = tf.expand_dims([1] * target.shape[0], 1)
with tf.GradientTape() as tape:
features = encoder(input_tensor, training=True)
for i in range(1, target.shape[1]):
# passing the features through the decoder
predictions, hidden, _ = decoder(dec_input,
features,
hidden,
training=True)
loss += loss_function(target[:, i], predictions)
# using teacher forcing
dec_input = tf.expand_dims(target[:, i], 1)
total_loss = (loss / int(target.shape[1]))
trainable_variables = encoder.trainable_variables + decoder.trainable_variables
gradients = tape.gradient(loss, trainable_variables)
optimizer.apply_gradients(zip(gradients, trainable_variables))
return loss, total_loss
@tf.function
def evaluate_batch(input_tensor, target):
"""Evaluate loss on a batch of data."""
loss = 0
# initializing the hidden state for each batch
# because the captions are not related from image to image
hidden = decoder.reset_state(batch_size=target.shape[0])
dec_input = tf.expand_dims([1] * target.shape[0], 1)
features = encoder(input_tensor, training=False)
for i in range(1, target.shape[1]):
# passing the features through the decoder
predictions, hidden, _ = decoder(dec_input,
features,
hidden,
training=False)
loss += loss_function(target[:, i], predictions)
# using teacher forcing
dec_input = tf.expand_dims(target[:, i], 1)
total_loss = (loss / int(target.shape[1]))
return total_loss
##############################################################################
############################# Training Loop ##################################
##############################################################################
print('Start training...\n')
start_epoch = 0
if FLAGS.save_dir:
checkpoint_path = FLAGS.save_dir
ckpt = tf.train.Checkpoint(encoder=encoder,
decoder=decoder,
optimizer=optimizer)
ckpt_manager = tf.train.CheckpointManager(ckpt,
checkpoint_path,
max_to_keep=5)
if ckpt_manager.latest_checkpoint:
start_epoch = int(ckpt_manager.latest_checkpoint.split('-')[-1])
epochs = 400
step_per_epoch = int(len(captions) / batch_size) * 10
previous_best = 100.
mixing_ratio = 0.4
syn_bs = int(batch_size * 2 * mixing_ratio)
true_bs = int(batch_size * 2 * (1 - mixing_ratio))
for epoch in range(start_epoch, epochs):
start = time.time()
total_loss = 0
for batch in range(step_per_epoch):
batch_idx = np.random.choice(train_idx, size=true_bs)
synthetic_batch_idx = np.random.choice(synthetic_train_idx,
size=syn_bs)
input_tensor = transitions[obs_idx[batch_idx], :]
synthetic_input_tensor = synthetic_transitions[
synthetic_obs_idx[synthetic_batch_idx], :]
input_tensor = np.concatenate(
[input_tensor, synthetic_input_tensor], axis=0)
input_tensor = encoder.preprocess(input_tensor)
target = encoded_captions[caption_idx[batch_idx]]
sythetic_target = synthetic_encoded_captions[
synthetic_caption_idx[synthetic_batch_idx]]
target = np.concatenate([target, sythetic_target], axis=0)
batch_loss, t_loss = train_step(input_tensor, target)
total_loss += t_loss
if batch % 100 == 0:
print('Epoch {} Batch {} Loss {:.4f}'.format(
epoch + 1, batch,
batch_loss.numpy() / int(target.shape[1])))
if epoch % 5 == 0 and FLAGS.save_dir:
test_total_loss = 0
for batch in range(3):
batch_idx = np.clip(
np.arange(true_bs) + batch * true_bs, 0, 196)
idx = test_idx[batch_idx]
input_tensor = transitions[obs_idx[idx], :]
target = encoded_captions[caption_idx[idx]]
t_loss = evaluate_batch(input_tensor, target)
test_total_loss += t_loss
batch_idx = np.arange(syn_bs) + batch * syn_bs
idx = synthetic_test_idx[batch_idx]
input_tensor = synthetic_transitions[synthetic_obs_idx[idx], :]
target = synthetic_encoded_captions[synthetic_caption_idx[idx]]
t_loss = evaluate_batch(input_tensor, target)
test_total_loss += t_loss
test_total_loss /= 6.
if test_total_loss < previous_best:
previous_best = test_total_loss
ckpt_manager.save(checkpoint_number=epoch)
print('Epoch {} | Loss {:.6f} | Val loss {:.6f}'.format(
epoch + 1, total_loss / step_per_epoch, previous_best))
print('Time taken for 1 epoch {:.6f} sec\n'.format(time.time() -
start))
if epoch % 20 == 0:
total_loss = 0
for batch in range(len(test_idx) // batch_size):
batch_idx = np.arange(batch_size) + batch * batch_size
idx = test_idx[batch_idx]
input_tensor = transitions[obs_idx[idx], :]
target = encoded_captions[caption_idx[idx]]
# input_tensor = input_tensor[:, 0] - input_tensor[:, 1]
t_loss = evaluate_batch(input_tensor, target)
total_loss += t_loss
print('====================================================')
print('Test Loss {:.6f}'.format(total_loss /
(len(test_idx) // batch_size)))
print('====================================================\n')
