def main(start_epoch):
(model, graph, saver_seq2seq) = build_model()
config = gpu_config()
with tf.Session(graph=graph, config=config) as sess:
sess.run(tf.global_variables_initializer())
print("Initialized.")
restore_ckpt = None
if start_epoch > -1:
if force_restore_point != "":
restore_ckpt = force_restore_point
else:
restore_ckpt = f"{ckpt_path}/seq2seq_RL{restore_extra_str}_{start_epoch}"
if restore_ckpt is not None:
saver_seq2seq.restore(sess, restore_ckpt)
print("Restored from", restore_ckpt)
for i in xrange(num_epochs):
if not infer_only: # for getting perplexity of test data, use train branch
mode = "train"
start_epoch += 1
run_vhred(model, sess, mode, start_epoch)
if not no_validation and not glimpse_training and start_epoch % 5 == 0 and start_epoch >= 10:
mode = "valid"
zipped_responses = run_vhred(model, sess, mode,
start_epoch)
else:
continue
else:
print("Inferring on test set...")
mode = "test"
zipped_responses = run_vhred(model, sess, mode, start_epoch)
# Make sure sent is not empty and always ends with an eou
flattened = [
decode2string(index2token,
sent,
end_token=end_token_str,
remove_END_TOKEN=True) for tp in zipped_responses
for sent in tp
]
flattened = [maybe_add(sent, eou_str) for sent in flattened]
# now we mark sentences that are generated by our model
marked_G = [("G: " + sent) if k % 3 == 1 else sent
for (k, sent) in enumerate(flattened)]
marked_M = [("M: " + sent) if k % 3 == 2 else sent
for (k, sent) in enumerate(marked_G)]
filename = f"{output_path}/{extra_str[1:]}_{mode}_result_{start_epoch}.txt"
write_lines(filename, marked_M)
# only need 1 epoch for inferring or getting PPL
if infer_only:
break
assert len(polite_responses) == len(neutral_responses) == len(rude_responses)
num_responses = len(polite_responses)
zipped = zip_lsts(
[source_test[:num_responses],
target_test[:num_responses],
polite_responses,
neutral_responses,
rude_responses])
flattened = [decode2string(index2token, sent, remove_END_TOKEN=True)
for tp in zipped for sent in tp]
# now we mark sentences that are generated by our model
marked_G = [("G: " + sent)
if k % 5 == 1 else sent
for (k, sent) in enumerate(flattened)]
marked_P = [("P: " + sent)
if k % 5 == 2 else sent
for (k, sent) in enumerate(marked_G)]
marked_N = [("N: " + sent)
if k % 5 == 3 else sent
for (k, sent) in enumerate(marked_P)]
marked_R = [("R: " + sent)
if k % 5 == 4 else sent
for (k, sent) in enumerate(marked_N)]
write_lines(filename, marked_R)
dump_pickle(
data_path + "/fusion_%.1f_%d_infer.pkl" %
(fusion_rate, seq2seq_epoch), responses)
responses_lst.append(responses)
num_responses = len(responses_lst[0])
print("Generated %d responses for each fusion rate." % num_responses)
# add in source sents and ground truths
zipped_responses = zip_lsts([source_test[:num_responses]] +
[target_test[:num_responses]] + responses_lst)
# Write results to file
filename = data_path + "fusion_responses_%.1f.txt" % fusion_rate_candidates[
0]
text_zipped_responses = [[
label + decode2string(index2token, response, remove_END_TOKEN=True)
for (label, response) in
zip(["", "G: "] +
convert_list_to_str_list(fusion_rate_candidates), responses)
] for responses in zipped_responses]
flattened_text_responses = [
response for responses in text_zipped_responses
for response in responses
]
write_lines(filename, flattened_text_responses)