def run(arguments) -> None:
print("Loading data ...")
model = LanguageModel.restore(arguments["TRAINED_MODEL"])
print(f" Loaded trained model from {arguments['TRAINED_MODEL']}.")
test_data = load_data_from_dir(
model.vocab,
length=model.hyperparameters["max_seq_length"],
data_dir=arguments["TEST_DATA_DIR"],
max_num_files=arguments.get("--max-num-files"),
)
print(
f" Loaded {test_data.shape[0]} test samples from {arguments['TEST_DATA_DIR']}."
)
test_loss, test_acc = model.run_one_epoch(
get_minibatch_iterator(
test_data,
model.hyperparameters["batch_size"],
is_training=False,
drop_remainder=False,
),
training=False,
)
print(f"Test: Loss {test_loss:.4f}, Acc {test_acc:.3f}")
def run(arguments) -> None:
model = LanguageModel.restore(arguments["TRAINED_MODEL"])
def compute_next_token(token_seq: List[str], num_cands: int = 3) -> str:
tensorised_seq = tensorise_token_sequence(model.vocab, len(token_seq) + 1, token_seq)
next_tok_probs = model.predict_next_token(tensorised_seq)
top_idxs = (-next_tok_probs).argsort()[:num_cands]
return [(model.vocab.get_name_for_id(top_idx),
next_tok_probs[top_idx])
for top_idx in top_idxs]
tokens = arguments['TOKENS']
for idx in range(int(arguments['--num-steps'])):
cands = compute_next_token(tokens)
print("Prediction at step %i (tokens %s):" % (idx, tokens))
for (token, prob) in cands:
print(" Prob %.3f: %s" % (prob, token))
next_tok = cands[0][0]
if next_tok == END_SYMBOL:
print('Reached end of sequence. Stopping.')
break
print("Continuing with token %s" % next_tok)
tokens.append(next_tok)