# train loop:
hps = hps_model
start = time.time()
for local_step in range(hps.num_steps):
step = rnn.sess.run(rnn.global_step)
curr_learning_rate = (hps.learning_rate - hps.min_learning_rate) * (
hps.decay_rate)**step + hps.min_learning_rate
raw_z, raw_a = random_batch()
inputs = np.concatenate((raw_z[:, :-1, :], raw_a[:, :-1, :]), axis=2)
outputs = raw_z[:, 1:, :] # teacher forcing (shift by one predictions)
feed = {
rnn.input_x: inputs,
rnn.output_x: outputs,
rnn.lr: curr_learning_rate
}
(train_cost, state, train_step, _) = rnn.sess.run(
[rnn.cost, rnn.final_state, rnn.global_step, rnn.train_op], feed)
if (step % 20 == 0 and step > 0):
end = time.time()
time_taken = end - start
start = time.time()
output_log = "step: %d, lr: %.6f, cost: %.4f, train_time_taken: %.4f" % (
step, curr_learning_rate, train_cost, time_taken)
print(output_log)
rnn.save_json(os.path.join(model_save_path, args.name + "_rnn.json"))
# save the model (don't bother with tf checkpoints json all the way ...)
rnn.save_json(os.path.join(model_save_path, args.name + "_rnn.json"))
# train loop:
hps = hps_model
start = time.time()
for local_step in range(hps.num_steps):
step = rnn.sess.run(rnn.global_step)
curr_learning_rate = (hps.learning_rate - hps.min_learning_rate) * (
hps.decay_rate)**step + hps.min_learning_rate
raw_z, raw_a = random_batch()
inputs = np.concatenate((raw_z[:, :-1, :], raw_a[:, :-1, :]), axis=2)
outputs = raw_z[:, 1:, :] # teacher forcing (shift by one predictions)
feed = {
rnn.input_x: inputs,
rnn.output_x: outputs,
rnn.lr: curr_learning_rate
}
(train_cost, state, train_step, _) = rnn.sess.run(
[rnn.cost, rnn.final_state, rnn.global_step, rnn.train_op], feed)
if (step % 20 == 0 and step > 0):
end = time.time()
time_taken = end - start
start = time.time()
output_log = "step: %d, lr: %.6f, cost: %.4f, train_time_taken: %.4f" % (
step, curr_learning_rate, train_cost, time_taken)
print(output_log)
# save the model (don't bother with tf checkpoints json all the way ...)
rnn.save_json(os.path.join(model_save_path, "rnn.json"))
with open(os.path.join("tf_initial_z", "initial_z.json"), 'wt') as outfile:
json.dump([initial_mu, initial_logvar], outfile, sort_keys=True, indent=0, separators=(',', ': '))
reset_graph()
rnn = MDNRNN(hps_model)
# train loop:w
hps = hps_model
start = time.time()
for local_step in range(hps.num_steps):
step = rnn.sess.run(rnn.global_step)
curr_learning_rate = (hps.learning_rate-hps.min_learning_rate) * (hps.decay_rate) ** step + hps.min_learning_rate
raw_z, raw_a = random_batch()
print(raw_z.shape, raw_a.shape)
inputs = np.concatenate((raw_z[:, :-1, :], raw_a[:, :-1, :]), axis=2)
outputs = raw_z[:, 1:, :] # teacher forcing (shift by one predictions)
feed = {rnn.input_x: inputs, rnn.output_x: outputs, rnn.lr: curr_learning_rate}
(train_cost, state, train_step, _) = rnn.sess.run([rnn.cost, rnn.final_state, rnn.global_step, rnn.train_op], feed)
if (step%20==0 and step > 0):
end = time.time()
time_taken = end-start
start = time.time()
output_log = "step: %d, lr: %.6f, cost: %.4f, train_time_taken: %.4f" % (step, curr_learning_rate, train_cost, time_taken)
print(output_log)
rnn.save_json(os.path.join(arglist.model_save_path, "rnn.json"))
# save the model (don't bother with tf checkpoints json all the way ...)
rnn.save_json(os.path.join(arglist.model_save_path, "rnn.json"))