vocabulary_name=hparams.vocabulary_name,
corpus_name=hparams.corpus_name,
cached_data_dir=args.cached_dataset_dir,
hparams=hparams)
data.prepare_data()
else:
setup_tf_logging(args.model_dir)
log_hparams(args.model_dir)
warm_start = None
if args.params_from_cudnn_rnn is not None:
platform_independent_to_cudnn_variables = {
"affine_projection_pseudo_cell/w": "affine_projection_layer/w",
"affine_projection_pseudo_cell/b": "affine_projection_layer/b",
"rnn/predict_next/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/kernel":
"cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/kernel",
"rnn/predict_next/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/bias":
"cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/bias",
"rnn/predict_next/multi_rnn_cell/cell_1/cudnn_compatible_lstm_cell/kernel":
"cudnn_lstm/rnn/multi_rnn_cell/cell_1/cudnn_compatible_lstm_cell/kernel",
"rnn/predict_next/multi_rnn_cell/cell_1/cudnn_compatible_lstm_cell/bias":
"cudnn_lstm/rnn/multi_rnn_cell/cell_1/cudnn_compatible_lstm_cell/bias",
}
warm_start = tf.estimator.WarmStartSettings(
ckpt_to_initialize_from=args.params_from_cudnn_rnn,
var_name_to_prev_var_name=platform_independent_to_cudnn_variables,
)
logger.info("Running with parameters: {}".format(hparams.to_json()))
train_and_eval(args.cached_dataset_dir, args.model_dir, hparams, warm_start)
best_acc = acc
torch.save({
"epoch": current_epoch,
"losses": losses,
"model_state_dict": model.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
}, os.path.join(exp_dir, 'best.pth'))
print("| Epoch: {:3d}, Eval loss: {:0.4f}, current acc: {:2.3f}%, the best: {:2.3f}%".format(current_epoch, eval_loss, acc*100, best_acc*100))
if __name__=="__main__":
args = docopt(__doc__)
db_path = args['<db_path>']
feat_path = args['<feat_path>']
exp_dir = args['<exp/task>']
preset = args["--preset"]
if preset is not None:
with open(preset) as f:
hparams.parse_json(f.read())
# make a dir, like exp/2019task1b
os.makedirs(exp_dir, exist_ok=True)
fp = open(os.path.join(exp_dir, "config.json"),'w')
fp.write(hparams.to_json(indent=' '*4))
fp.close()
print(hparams_debug_string())
train(db_path, feat_path, exp_dir)
parser.add_argument("model_dir")
parser.add_argument("--cached_dataset_dir")
parser.add_argument('--hparams',
type=str,
help='Comma separated list of "name=value" pairs.')
args = parser.parse_args()
if args.mode == "train":
if args.hparams:
hparams.parse(args.hparams)
if args.cached_dataset_dir is None:
train_lm_on_simple_examples_with_glove(args.model_dir)
else:
with open(Path(args.model_dir) / "hparams.json",
"wt") as params_file:
print(hparams.to_json(), file=params_file)
train_lm_on_cached_simple_examples_with_glove(
args.cached_dataset_dir, args.model_dir, hparams)
elif args.mode == "prepare":
prepare_training_dataset(args.model_dir)
elif args.mode == "predict":
if args.hparams:
hparams.parse(args.hparams)
eval_lm_on_cached_simple_examples_with_glove(args.cached_dataset_dir,
args.model_dir, "train",
hparams)
elif args.mode == "predict_with_lm_training_process":
from lm_training_process import eval_lm_on_cached_simple_examples_with_glove_check
if args.hparams:
hparams.parse(args.hparams)
eval_lm_on_cached_simple_examples_with_glove_check(