if summarize:
loss = np.mean(last_save_losses)
llprint("\n\tAvg. Logistic Loss: %.4f\n" % (loss))
if np.isnan(loss):
raise Exception('nan Loss')
if summarize and rnn.debug:
loss = np.mean(last_save_losses)
last_save_losses = []
if take_checkpoint:
llprint("\nSaving Checkpoint ... "),
curr_state = {
"epoch": epoch,
"rnn_state": rnn.state_dict(),
"opti_state": optimizer.state_dict()
}
# check_ptr = os.path.join(ckpts_dir, 'step_{}.pth'.format(epoch))
# T.save(curr_state, check_ptr)
check_ptr = os.path.join(ckpts_dir, 'best.pth')
T.save(curr_state, check_ptr)
llprint("Done!\n")
for i in range(args.test_iterations):
llprint("\rIteration %d/%d" % (i, args.test_iterations))
input_data, target_output = dataset.generate_data(
1, args.bits, args.cuda)
if args.debug:
output, (chx, mhx, rv), v = rnn(input_data, (None, mhx, None),
last_100_losses.append(loss_value)
if summarize:
llprint("\rIteration %d/%d" % (epoch, iterations))
llprint("\nAvg. Logistic Loss: %.4f\n" %
(np.mean(last_100_losses)))
output = output.data.cpu().numpy()
print("Real value: ", ' = ' + str(int(target_output[0])))
print("Predicted: ",
' = ' + str(int(output // 1)) + " [" + str(output) + "]")
last_100_losses = []
if take_checkpoint:
llprint("\nSaving Checkpoint ... "),
check_ptr = os.path.join(ckpts_dir, 'step_{}.pth'.format(epoch))
cur_weights = rnn.state_dict()
T.save(cur_weights, check_ptr)
llprint("Done!\n")
llprint("\nTesting generalization...\n")
rnn.eval()
for i in range(int((iterations + 1) / 10)):
llprint("\nIteration %d/%d" % (i, iterations))
# We test now the learned generalization using sequence_max_length examples
random_length = np.random.randint(2, int(sequence_max_length) * 10 + 1)
input_data, target_output, sums_text = generate_data(
random_length, input_size)
if rnn.debug:
