### Validation
if conf.do_validate and batch_count % conf.valid_freq == 0:
logging.info("VALIDATION: Feed-forward pass.")
# Do the predictions
tv = time.time()
y_pred_dict = model.predict({'input': X_val})
logging.info("VALIDATION: Feed-forward pass took %.2f seconds." % (time.time() - tv))
# TODO: Numpy should have a solution for this
for i in range(X_val.shape[1]):
y_pred_val[:, i, :] = y_pred_dict['output_%d' % i]
# Take argmax of output machines and report BLEU
one_best = get_1best_translations(y_pred_dict, c.trg_idx_to_sent)
one_best_bleu = scorer.score_sentences(y_val_strings, one_best)
logging.info("VALIDATION: Direct 1-best (argmax) %s" % one_best_bleu)
# Beam search + 1-best + BLEU
logging.info("VALIDATION: Starting beam search...")
d = decoder.process(y_pred_val)
# Take 1-best of beam-search and report BLEU
hyps = [c.trg_idx_to_sent(sent[0][0]) for sent in d]
bleu = scorer.score_sentences(y_val_strings, hyps)
logging.info("VALIDATION: 1-best of %d-best %s" % (conf.beam_size, bleu))
# If BLEU is better, save the weights
if bleu > best_val_bleu:
# Reset patience counter
no_val_improvement = 0
# Load vocabulary files and invert them
print "Loading vocabulary files..."
src_vocab = load_object(args.source_vocab)
trg_vocab = load_object(args.target_vocab)
corp.set_src_vocab(src_vocab)
corp.set_trg_vocab(trg_vocab)
print "Completed."
# Map source sentences
mapped_sents = corp.map_sentences(src_sents, corp.src_vocab, conf.max_seq_len, skip_unk=args.remove_unk)
print "%d source sentences mapped." % len(mapped_sents)
# Create decoder instance
decoder = Decoder(trg_vocab["</s>"], trg_vocab["<unk>"], args.beamsize, args.generate_unk)
# Validation prediction placeholder
y_pred = np.ndarray((mapped_sents.shape[0], mapped_sents.shape[1], len(trg_vocab))).astype(np.float32)
# Feed-forward pass
y_pred_dict = model.predict({'input': mapped_sents})
for i in range(y_pred.shape[1]):
y_pred[:, i, :] = y_pred_dict['output_%d' % i]
scorer = MultiBleuScorer()
d = decoder.process(y_pred)
# Generate 1-best sentences
hyps = [corp.trg_idx_to_sent(hyp[0]) for hyp in d[0]]
bleu = scorer.score_sentences(ref_trans_file, hyps)
print bleu
