def train(model_folder, num_epochs, learning_rate, batch_size, stroke_length,
steps_per_char, save_every):
char_seq_length = stroke_length // steps_per_char
train_data, val_data, test_data, metadata = get_preprocessed_data_splits()
train_data_gen = conditional_batch_generator(train_data,
stroke_length=stroke_length,
char_length=char_seq_length,
batch_size=batch_size)
epoch_size = len(train_data[0]) // batch_size
model = ConditionalStrokeModel(model_folder,
learning_rate=learning_rate,
batch_size=batch_size,
rnn_steps=stroke_length,
is_train=True,
char_dict=metadata['char_dict'],
char_seq_len=char_seq_length)
for epoch in range(num_epochs):
for i in range(epoch_size):
input_batch, target_batch, char_batch = next(train_data_gen)
loss, _ = model.train(input_batch, target_batch, char_batch)
if not epoch % save_every and epoch != 0:
model.save()
if not epoch % save_every and epoch != 0:
model = ConditionalStrokeModel.load(model_folder,
batch_size=1,
rnn_steps=1,
is_train=False,
char_seq_len=30)
strokes = decode(model)
plotting.plot_stroke(
strokes, os.path.join(model_folder, 'test{}'.format(epoch)))
model = ConditionalStrokeModel.load(model_folder,
learning_rate=learning_rate,
batch_size=batch_size,
rnn_steps=stroke_length,
is_train=True,
char_seq_len=char_seq_length)
_print_loss(model, train_data, epoch, stroke_length, char_seq_length,
batch_size, 'Train')
_print_loss(model, val_data, epoch, stroke_length, char_seq_length,
batch_size, ' Val')
model.save()
def train(model_folder, learning_rate, num_epochs, num_layers, num_steps,
hidden_size, lr_decay, batch_size):
train_data, _, test_data, metadata = get_preprocessed_data_splits(
sentences_to_int=True)
train_data_gen = character_batch_generator(train_data,
batch_size=batch_size)
vocab_size = len(metadata['char_dict'])
epoch_size = len(train_data[0]) // batch_size
model = CharacterGenModel(model_folder,
batch_size,
num_steps,
vocab_size,
num_layers,
hidden_size,
char_dict=metadata['char_dict'])
for i in range(num_epochs):
learning_rate *= 1 / (1. + lr_decay * i)
model.sess.run(model.lr_update, {model.new_lr: learning_rate})
print('Epoch {}, lr {}'.format(i, model.sess.run(model.lr)))
total_cost = 0
for _ in range(epoch_size):
input_batch, target_batch = next(train_data_gen)
cost = model.train(input_batch, target_batch)[0]
total_cost += cost
model.save()
train_perplexity = np.exp(total_cost / (epoch_size))
print('Epoch: {} Train perplexity: {}'.format(i, train_perplexity))
total_cost = 0
test_epochs = len(test_data[0]) // batch_size
for _ in range(test_epochs):
test_input, test_target = next(character_batch_generator(test_data))
cost = model.sess.run(model.cost, {
model.inputs: test_input,
model.targets: test_target
})
total_cost += cost
test_perplexity = np.exp(total_cost / test_epochs)
print("Test perplexity: {}".format(test_perplexity))
print('BPC: {}'.format(total_cost / test_epochs))
def train(model_folder, num_epochs, learning_rate, batch_size, stroke_length,
save_every):
train_data, val_data, test_data, metadata = get_preprocessed_data_splits()
train_data_gen = stroke_batch_generator(train_data,
length=stroke_length,
batch_size=batch_size)
model = UnconditionalStrokeModel(model_folder,
learning_rate=learning_rate,
batch_size=batch_size,
rnn_steps=stroke_length,
is_train=True)
epoch_size = len(train_data[0]) // batch_size
for epoch in range(num_epochs):
for i in range(epoch_size):
input_batch, target_batch = next(train_data_gen)
model.train(input_batch, target_batch)
if not epoch % save_every and epoch != 0:
model.save()
if not epoch % save_every and epoch != 0:
# make some stroke plots
model = UnconditionalStrokeModel.load(model_folder,
batch_size=1,
rnn_steps=1,
is_train=False)
strokes = decode(model, 650, epoch)
plotting.plot_stroke(
strokes, os.path.join(model_folder, 'test{}'.format(epoch)))
model = UnconditionalStrokeModel.load(model_folder,
learning_rate=learning_rate,
batch_size=batch_size,
rnn_steps=stroke_length)
_print_loss(model, train_data, epoch, stroke_length, batch_size,
'Train')
_print_loss(model, val_data, epoch, stroke_length, batch_size,
' Val')
model.save()