def run_epoch(session, m, data, eval_op, config, verbose=False):
"""Run the model on the given data."""
epoch_size = ((len(data) // m.batch_size) - 1) // m.num_steps
start_time = time.time()
costs = 0.0
iters = 0
state = [x.eval() for x in m.initial_state]
for step, (x,
y) in enumerate(data_iterator(data, m.batch_size, m.num_steps)):
noise_x, noise_i, noise_h, noise_o = get_noise(x, m, config.drop_x,
config.drop_i,
config.drop_h,
config.drop_o)
feed_dict = {
m.input_data: x,
m.targets: y,
m.noise_x: noise_x,
m.noise_i: noise_i,
m.noise_h: noise_h,
m.noise_o: noise_o
}
feed_dict.update(
{m.initial_state[i]: state[i]
for i in range(m.num_layers)})
cost, state, _ = session.run([m.cost, m.final_state, eval_op],
feed_dict)
costs += cost
iters += m.num_steps
if verbose and step % (epoch_size // 10) == 10:
print("%.3f perplexity: %.3f speed: %.0f wps" %
(step * 1.0 / epoch_size, np.exp(costs / iters),
iters * m.batch_size / (time.time() - start_time)))
return np.exp(costs / iters)
def run_mc_epoch(seed, session, m, data, eval_op, config, mc_steps, verbose=False):
"""Run the model with noise on the given data multiple times for MC evaluation."""
n_steps = len(data)
all_probs = np.array([0.0]*n_steps)
sum_probs = np.array([0.0]*n_steps)
mc_i = 1
print("Total MC steps to do:", mc_steps)
if not os.path.isdir('./probs'):
print('Creating probs directory')
os.mkdir('./probs')
while mc_i <= mc_steps:
print("MC sample number:", mc_i)
epoch_size = ((len(data) // m.batch_size) - 1) // m.num_steps
start_time = time.time()
costs = 0.0
iters = 0
state = [x.eval() for x in m.initial_state]
for step, (x, y) in enumerate(data_iterator(data, m.batch_size, m.num_steps)):
if step == 0:
noise_x, noise_i, noise_h, noise_o = get_noise(x, m, config.drop_x, config.drop_i, config.drop_h, config.drop_o)
feed_dict = {m.input_data: x, m.targets: y,
m.noise_x: noise_x, m.noise_i: noise_i, m.noise_h: noise_h, m.noise_o: noise_o}
feed_dict.update({m.initial_state[i]: state[i] for i in range(m.num_layers)})
cost, state, _ = session.run([m.cost, m.final_state, eval_op], feed_dict)
costs += cost
iters += m.num_steps
all_probs[step] = np.exp(-cost)
if verbose and step % (epoch_size // 10) == 10:
print("%.3f perplexity: %.3f speed: %.0f wps" % (step * 1.0 / epoch_size, np.exp(costs / iters),
iters * m.batch_size / (time.time() - start_time)))
perplexity = np.exp(costs / iters)
print("Perplexity:", perplexity)
if perplexity < 500:
savefile = 'probs/' + str(seed) + '_' + str(mc_i)
print("Accepted. Saving to:", savefile)
np.save(savefile, all_probs)
sum_probs += all_probs
mc_i += 1
return np.exp(np.mean(-np.log(np.clip(sum_probs/mc_steps, 1e-10, 1-1e-10))))
def run_epoch(m, data, config, is_train, verbose=False, log=None):
"""Run the model on the given data."""
epoch_size = ((len(data) // config.batch_size) - 1) // config.num_steps
start_time = time.time()
costs = 0.0
iters = 0
m.reset_hidden_state()
for step, (x, y) in enumerate(
data_iterator(data, config.batch_size, config.num_steps)):
if is_train:
noise_x = get_noise_x(x, config.drop_x)
cost = m.train(x, y, noise_x)
else:
cost = m.evaluate(x, y)
costs += cost
iters += config.num_steps
if verbose and step % (epoch_size // 10) == 10:
log.info("%.3f perplexity: %.3f speed: %.0f wps" %
(step * 1.0 / epoch_size, np.exp(costs / iters),
iters * config.batch_size / (time.time() - start_time)))
return np.exp(costs / iters)
def run_epoch(data, config, is_train, verbose=False):
"""Run the model on the given data."""
epoch_size = ((len(data) // config.batch_size) - 1) // config.num_steps
start_time = time.time()
costs = 0.0
iters = 0
for hidden_state in hidden_states:
hidden_state.set_value(np.zeros_like(hidden_state.get_value()))
for step, (x, y) in enumerate(
data_iterator(data, config.batch_size, config.num_steps)):
if is_train:
noise_x = get_noise_x(x, config.drop_x)
cost = train(x, y, noise_x)
else:
cost = evaluate(x, y)
costs += cost
iters += config.num_steps
if verbose and step % (epoch_size // 10) == 10:
print("%.3f perplexity: %.3f speed: %.0f wps" %
(step * 1.0 / epoch_size, np.exp(costs / iters),
iters * config.batch_size / (time.time() - start_time)))
return np.exp(costs / iters)
