zomux /
deepy
| @@ 19-42 (lines=24) @@ | ||
| 16 | ||
| 17 | default_model = os.path.join(os.path.dirname(__file__), "models", "class_based_rnnlm.gz") |
|
| 18 | ||
| 19 | if __name__ == '__main__': |
|
| 20 | ap = ArgumentParser() |
|
| 21 | ap.add_argument("--model", default="") |
|
| 22 | ap.add_argument("--small", action="store_true") |
|
| 23 | args = ap.parse_args() |
|
| 24 | ||
| 25 | vocab, lmdata = load_data(small=args.small, history_len=5, batch_size=64) |
|
| 26 | import pdb; pdb.set_trace() |
|
| 27 | model = NeuralLM(vocab.size) |
|
| 28 | model.stack(RNN(hidden_size=100, output_type="sequence", hidden_activation='sigmoid', |
|
| 29 | persistent_state=True, batch_size=lmdata.size, |
|
| 30 | reset_state_for_input=0), |
|
| 31 | ClassOutputLayer(output_size=100, class_size=100)) |
|
| 32 | ||
| 33 | if os.path.exists(args.model): |
|
| 34 | model.load_params(args.model) |
|
| 35 | ||
| 36 | trainer = SGDTrainer(model, {"learning_rate": LearningRateAnnealer.learning_rate(1.2), |
|
| 37 | "weight_l2": 1e-7}) |
|
| 38 | annealer = LearningRateAnnealer() |
|
| 39 | ||
| 40 | trainer.run(lmdata, epoch_controllers=[annealer]) |
|
| 41 | ||
| 42 | model.save_params(default_model) |
|
| 43 | ||
| @@ 19-41 (lines=23) @@ | ||
| 16 | ||
| 17 | default_model = os.path.join(os.path.dirname(__file__), "models", "lstm_rnnlm.gz") |
|
| 18 | ||
| 19 | if __name__ == '__main__': |
|
| 20 | ap = ArgumentParser() |
|
| 21 | ap.add_argument("--model", default="") |
|
| 22 | ap.add_argument("--small", action="store_true") |
|
| 23 | args = ap.parse_args() |
|
| 24 | ||
| 25 | vocab, lmdata = load_data(small=args.small, history_len=5, batch_size=64) |
|
| 26 | model = NeuralLM(vocab.size) |
|
| 27 | model.stack(LSTM(hidden_size=100, output_type="sequence", |
|
| 28 | persistent_state=True, batch_size=lmdata.size, |
|
| 29 | reset_state_for_input=0), |
|
| 30 | FullOutputLayer(vocab.size)) |
|
| 31 | ||
| 32 | if os.path.exists(args.model): |
|
| 33 | model.load_params(args.model) |
|
| 34 | ||
| 35 | trainer = SGDTrainer(model, {"learning_rate": LearningRateAnnealer.learning_rate(1.2), |
|
| 36 | "weight_l2": 1e-7}) |
|
| 37 | annealer = LearningRateAnnealer() |
|
| 38 | ||
| 39 | trainer.run(lmdata, epoch_controllers=[annealer]) |
|
| 40 | ||
| 41 | model.save_params(default_model) |
|
| 42 | ||
| @@ 19-41 (lines=23) @@ | ||
| 16 | ||
| 17 | default_model = os.path.join(os.path.dirname(__file__), "models", "baseline_rnnlm.gz") |
|
| 18 | ||
| 19 | if __name__ == '__main__': |
|
| 20 | ap = ArgumentParser() |
|
| 21 | ap.add_argument("--model", default="") |
|
| 22 | ap.add_argument("--small", action="store_true") |
|
| 23 | args = ap.parse_args() |
|
| 24 | ||
| 25 | vocab, lmdata = load_data(small=args.small, history_len=5, batch_size=64) |
|
| 26 | model = NeuralLM(vocab.size) |
|
| 27 | model.stack(RNN(hidden_size=100, output_type="sequence", hidden_activation="sigmoid", |
|
| 28 | persistent_state=True, batch_size=lmdata.size, |
|
| 29 | reset_state_for_input=0), |
|
| 30 | FullOutputLayer(vocab.size)) |
|
| 31 | ||
| 32 | if os.path.exists(args.model): |
|
| 33 | model.load_params(args.model) |
|
| 34 | ||
| 35 | trainer = SGDTrainer(model, {"learning_rate": LearningRateAnnealer.learning_rate(1.2), |
|
| 36 | "weight_l2": 1e-7}) |
|
| 37 | annealer = LearningRateAnnealer() |
|
| 38 | ||
| 39 | trainer.run(lmdata, epoch_controllers=[annealer]) |
|
| 40 | ||
| 41 | model.save_params(default_model) |
|
| 42 | ||
