def predict_train(config, sess):
"""Train an NTM for the copy task given a TensorFlow session, which is a
connection to the C++ backend"""
if not os.path.isdir(config.checkpoint_dir):
raise Exception(" [!] Directory %s not found" % config.checkpoint_dir)
# delimiter flag-like vector inputs indicating the start and end
# you can see these in the figure examples in the README
# this is kind of defined redundantly
start_symbol = np.zeros([config.input_dim], dtype=np.float32)
start_symbol[0] = 1
end_symbol = np.zeros([config.input_dim], dtype=np.float32)
end_symbol[1] = 1
# initialise the neural turing machine and the neural-net controller thing
cell = NTMCell(input_dim=config.input_dim,
output_dim=config.output_dim,
controller_layer_size=config.controller_layer_size,
write_head_size=config.write_head_size,
read_head_size=config.read_head_size)
ntm = NTM(cell, sess, config.min_length, config.max_length*3)
print(" [*] Initialize all variables")
tf.initialize_all_variables().run()
print(" [*] Initialization finished")
start_time = time.time()
for idx in xrange(config.epoch):
# generate a sequence of random length
seq_length = randint(config.min_length, config.max_length) * 4
inc_seq, comp_seq = generate_predict_sequence(seq_length, config.input_dim - 2)
# this somehow associates the desired inputs and outputs with the NTM
feed_dict = {input_:vec for vec, input_ in zip(inc_seq, ntm.inputs)}
feed_dict.update(
{true_output:vec for vec, true_output in zip(comp_seq, ntm.true_outputs)}
)
feed_dict.update({
ntm.start_symbol: start_symbol,
ntm.end_symbol: end_symbol
})
# this runs the session and returns the current training loss and step
# I'm kind of surprised it returns the step, but whatevs
_, cost, step = sess.run([ntm.optims[seq_length],
ntm.get_loss(seq_length),
ntm.global_step], feed_dict=feed_dict)
# how does one use these checkpoints?
if idx % 100 == 0:
ntm.save(config.checkpoint_dir, 'copy', step)
if idx % print_interval == 0:
print("[%5d] %2d: %.2f (%.1fs)" \
% (idx, seq_length, cost, time.time() - start_time))
print("Training Copy task finished")
return cell, ntm
def copy_train(config):
sess = config.sess
if not os.path.isdir(config.checkpoint_dir):
raise Exception(" [!] Directory %s not found" % config.checkpoint_dir)
# delimiter flag for start and end
start_symbol = np.zeros([config.input_dim], dtype=np.float32)
start_symbol[0] = 1
end_symbol = np.zeros([config.input_dim], dtype=np.float32)
end_symbol[1] = 1
cell = NTMCell(input_dim=config.input_dim,
output_dim=config.output_dim,
controller_layer_size=config.controller_layer_size,
write_head_size=config.write_head_size,
read_head_size=config.read_head_size)
ntm = NTM(cell, sess, config.min_length, config.max_length)
print(" [*] Initialize all variables")
tf.initialize_all_variables().run()
print(" [*] Initialization finished")
start_time = time.time()
for idx in xrange(config.epoch):
seq_length = randint(config.min_length, config.max_length)
seq = generate_copy_sequence(seq_length, config.input_dim - 2)
feed_dict = {input_: vec for vec, input_ in zip(seq, ntm.inputs)}
feed_dict.update({
true_output: vec
for vec, true_output in zip(seq, ntm.true_outputs)
})
feed_dict.update({
ntm.start_symbol: start_symbol,
ntm.end_symbol: end_symbol
})
_, cost, step = sess.run([
ntm.optims[seq_length],
ntm.get_loss(seq_length), ntm.global_step
],
feed_dict=feed_dict)
if idx % 100 == 0:
ntm.save(config.checkpoint_dir, 'copy', step)
if idx % print_interval == 0:
print("[%5d] %2d: %.2f (%.1fs)" \
% (idx, seq_length, cost, time.time() - start_time))
print("Training Copy task finished")
return cell, ntm
def copy_train(config):
sess = config.sess
if not os.path.isdir(config.checkpoint_dir):
raise Exception(" [!] Directory %s not found" % config.checkpoint_dir)
# delimiter flag for start and end
start_symbol = np.zeros([config.input_dim], dtype=np.float32)
start_symbol[0] = 1
end_symbol = np.zeros([config.input_dim], dtype=np.float32)
end_symbol[1] = 1
cell = NTMCell(input_dim=config.input_dim,
output_dim=config.output_dim,
controller_layer_size=config.controller_layer_size,
write_head_size=config.write_head_size,
read_head_size=config.read_head_size)
ntm = NTM(cell, sess, config.min_length, config.max_length)
print(" [*] Initialize all variables")
tf.initialize_all_variables().run()
print(" [*] Initialization finished")
start_time = time.time()
for idx in xrange(config.epoch):
seq_length = randint(config.min_length, config.max_length)
seq = generate_copy_sequence(seq_length, config.input_dim - 2)
feed_dict = {input_:vec for vec, input_ in zip(seq, ntm.inputs)}
feed_dict.update(
{true_output:vec for vec, true_output in zip(seq, ntm.true_outputs)}
)
feed_dict.update({
ntm.start_symbol: start_symbol,
ntm.end_symbol: end_symbol
})
_, cost, step = sess.run([ntm.optims[seq_length],
ntm.get_loss(seq_length),
ntm.global_step], feed_dict=feed_dict)
if idx % 100 == 0:
ntm.save(config.checkpoint_dir, 'copy', step)
if idx % print_interval == 0:
print("[%5d] %2d: %.2f (%.1fs)" \
% (idx, seq_length, cost, time.time() - start_time))
print("Training Copy task finished")
return cell, ntm
turing_machine.sequence_length:
2 * random_length + 2
})
# TODO: This works for batch size = 1
seq_out = np.round(np.reshape(temp_output,
(1, -1))).tolist()[0]
seq_target = np.reshape(target_output, (1, -1)).tolist()[0]
dist = hamming_distance(seq_out, seq_target)
val = tf.compat.v1.Summary.Value(tag="Hamming_%",
simple_value=dist)
summary2 = tf.compat.v1.Summary(value=[val])
summarizer.add_summary(summary, i)
summarizer.add_summary(summary2, i)
cycle_time = time.perf_counter() - last_time
print(f"Iteration: {i}/{iterations}")
print("Avg. Logistic Loss: %.4f" %
(np.mean(last_100_losses) / batch_size))
print(
"Time needed: %.4f s, sample/sec: %d" %
(cycle_time, log_frequency * batch_size / cycle_time))
last_100_losses = []
last_time = time.perf_counter()
if take_checkpoint:
llprint("\nSaving Checkpoint ... "),
turing_machine.save(session, ckpts_dir, 'step-%d' % (i))
llprint("Done!\n")
print(" [*] Initialization finished")
else:
ntm.load(config['checkpoint_dir'], 'copy')
start_time = time.time()
print('')
for idx in range(config['epoch']):
seq_length = np.random.randint(2, config['length'] + 1)
X, Y, masks = build_seq_batch(seq_length, config['length'],
config['input_dim'] - 2)
feed_dict = {ntm.inputs: X, ntm.true_outputs: Y, ntm.masks: masks}
if idx % print_interval != 0:
_, cost, step = sess.run(
[ntm.optims, ntm.losses, ntm.global_step],
feed_dict=feed_dict)
else:
_, cost, step, Y_pre = sess.run(
[ntm.optims, ntm.losses, ntm.global_step, ntm.outputs],
feed_dict=feed_dict)
print("[%5d] %2d: %.4f (%.1fs)" \
% (idx, seq_length, cost, time.time() - start_time))
Y_pre = np.array(Y_pre)
mask_id = masks.reshape(-1).astype(bool)
print(np.argmax(Y, axis=1)[mask_id])
print(np.argmax(Y_pre, axis=1)[mask_id])
print("Training Copy task finished")
ntm.save(config['checkpoint_dir'], 'n_copy', idx)
