-
Notifications
You must be signed in to change notification settings - Fork 8
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
replace partition_num with blank_num, add gan_hyperparams
- Loading branch information
Showing
7 changed files
with
218 additions
and
15 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,208 @@ | ||
| # -*- coding: utf-8 -*- | ||
| """ | ||
| configurate the hyperparameters, based on command line arguments. | ||
| """ | ||
| import argparse | ||
| import os | ||
|
|
||
| from texar.data import SpecialTokens | ||
|
|
||
|
|
||
| class Hyperparams: | ||
| """ | ||
| config dictionrary, initialized as an empty object. | ||
| The specific values are passed on with the ArgumentParser | ||
| """ | ||
| def __init__(self): | ||
| self.help = "the hyperparams dictionary to use" | ||
|
|
||
|
|
||
| def load_hyperparams(): | ||
| """ | ||
| main function to define hyperparams | ||
| """ | ||
| # pylint: disable=too-many-statements | ||
| args = Hyperparams() | ||
| argparser = argparse.ArgumentParser() | ||
| argparser.add_argument('--mask_rate', type=float, default=0.5) | ||
| argparser.add_argument('--blank_num', type=int, default=1) | ||
| argparser.add_argument('--batch_size', type=int, default=400) # 4096 | ||
| argparser.add_argument('--test_batch_size', type=int, default=10) | ||
| argparser.add_argument('--max_seq_length', type=int, default=16) # 256 | ||
| argparser.add_argument('--hidden_dim', type=int, default=512) | ||
| argparser.add_argument('--running_mode', type=str, | ||
| default='train_and_evaluate', | ||
| help='can also be test mode') | ||
| argparser.add_argument('--max_training_steps', type=int, default=2500000) | ||
| argparser.add_argument('--warmup_steps', type=int, default=10000) | ||
| argparser.add_argument('--max_train_epoch', type=int, default=150) | ||
| argparser.add_argument('--bleu_interval', type=int, default=5) | ||
| argparser.add_argument('--log_disk_dir', type=str, default='./') | ||
| argparser.add_argument('--filename_prefix', type=str, default='yahoo.') | ||
| argparser.add_argument('--data_dir', type=str, | ||
| default='./yahoo_data/') | ||
| argparser.add_argument('--save_eval_output', default=1, | ||
| help='save the eval output to file') | ||
| argparser.add_argument('--lr_constant', type=float, default=1) | ||
| argparser.add_argument('--learning_rate_strategy', type=str, default='dynamic') # 'static' | ||
| argparser.add_argument('--zero_pad', type=int, default=0) | ||
| argparser.add_argument('--bos_pad', type=int, default=0, | ||
| help='use all-zero embedding for bos') | ||
| argparser.add_argument('--random_seed', type=int, default=1234) | ||
| argparser.add_argument('--beam_width', type=int, default=2) | ||
| argparser.parse_args(namespace=args) | ||
|
|
||
| args.present_rate = 1 - args.mask_rate | ||
| args.pretrain_epoch = args.max_train_epoch * 0.8 | ||
| args.max_decode_len = args.max_seq_length | ||
| args.data_dir = os.path.abspath(args.data_dir) | ||
| args.filename_suffix = '.txt' | ||
| args.train_file = os.path.join(args.data_dir, | ||
| '{}train{}'.format(args.filename_prefix, args.filename_suffix)) | ||
| args.valid_file = os.path.join(args.data_dir, | ||
| '{}valid{}'.format(args.filename_prefix, args.filename_suffix)) | ||
| args.test_file = os.path.join(args.data_dir, | ||
| '{}test{}'.format(args.filename_prefix, args.filename_suffix)) | ||
| args.vocab_file = os.path.join(args.data_dir, 'vocab.txt') | ||
| log_params_dir = 'log_dir/{}bsize{}.epoch{}.seqlen{}.{}_lr.present{}.partition{}.hidden{}.seq2seq/'.format( | ||
| args.filename_prefix, args.batch_size, args.max_train_epoch, args.max_seq_length, | ||
| args.learning_rate_strategy, args.present_rate, args.blank_num, args.hidden_dim) | ||
| args.log_dir = os.path.join(args.log_disk_dir, log_params_dir) | ||
| print('train_file:{}'.format(args.train_file)) | ||
| print('valid_file:{}'.format(args.valid_file)) | ||
| train_dataset_hparams = { | ||
| "num_epochs": 1, | ||
| "seed": args.random_seed, | ||
| "shuffle": True, | ||
| "dataset": { | ||
| "files": args.train_file, | ||
| "vocab_file": args.vocab_file, | ||
| "max_seq_length": args.max_seq_length, | ||
| "bos_token": SpecialTokens.BOS, | ||
| "eos_token": SpecialTokens.EOS, | ||
| "length_filter_mode": "truncate", | ||
| }, | ||
| 'batch_size': args.batch_size, | ||
| 'allow_smaller_final_batch': True, | ||
| } | ||
| eval_dataset_hparams = { | ||
| "num_epochs": 1, | ||
| 'seed': args.random_seed, | ||
| 'shuffle': False, | ||
| 'dataset': { | ||
| 'files': args.valid_file, | ||
| 'vocab_file': args.vocab_file, | ||
| "max_seq_length": args.max_seq_length, | ||
| "bos_token": SpecialTokens.BOS, | ||
| "eos_token": SpecialTokens.EOS, | ||
| "length_filter_mode": "truncate", | ||
| }, | ||
| 'batch_size': args.test_batch_size, | ||
| 'allow_smaller_final_batch': True, | ||
| } | ||
| test_dataset_hparams = { | ||
| "num_epochs": 1, | ||
| "seed": args.random_seed, | ||
| "shuffle": False, | ||
| "dataset": { | ||
| "files": args.test_file, | ||
| "vocab_file": args.vocab_file, | ||
| "max_seq_length": args.max_seq_length, | ||
| "bos_token": SpecialTokens.BOS, | ||
| "eos_token": SpecialTokens.EOS, | ||
| "length_filter_mode": "truncate", | ||
| }, | ||
| 'batch_size': args.test_batch_size, | ||
| 'allow_smaller_final_batch': True, | ||
| } | ||
| args.word_embedding_hparams = { | ||
| 'name': 'lookup_table', | ||
| 'dim': args.hidden_dim, | ||
| 'initializer': { | ||
| 'type': 'random_normal_initializer', | ||
| 'kwargs': { | ||
| 'mean': 0.0, | ||
| 'stddev': args.hidden_dim**-0.5, | ||
| }, | ||
| } | ||
| } | ||
| cell = { | ||
| "type": "LSTMBlockCell", | ||
| "kwargs": { | ||
| "num_units": args.hidden_dim*4, | ||
| "forget_bias": 0. | ||
| }, | ||
| "dropout": {"output_keep_prob": 1-0.1}, | ||
| "num_layers": 1 | ||
| } | ||
| output_layer = { | ||
| "num_layers": 0, | ||
| "layer_size": args.hidden_dim*4, | ||
| "activation": "identity", | ||
| "final_layer_activation": None, | ||
| "other_dense_kwargs": None, | ||
| "dropout_layer_ids": [], | ||
| "dropout_rate": 0.1, | ||
| "variational_dropout": False, | ||
| "@no_typecheck": ["activation", "final_layer_activation", | ||
| "layer_size", "dropout_layer_ids"] | ||
| } | ||
| encoder_hparams = { | ||
| "rnn_cell": cell, | ||
| "output_layer": output_layer, | ||
| "name": "unidirectional_rnn_encoder" | ||
| } | ||
| decoder_hparams = { | ||
| "rnn_cell": cell, | ||
| "max_decoding_length_train": args.max_seq_length+2, | ||
| "max_decoding_length_infer": args.max_seq_length+2, | ||
| "name": "basic_rnn_decoder" | ||
| } | ||
| classifier_hparams = { | ||
| 'kernel_size': [3, 4, 5], | ||
| 'filters': 128, | ||
| 'other_conv_kwargs': {'padding': 'same'}, | ||
| 'dropout_conv': [1], | ||
| 'dropout_rate': 0.5, | ||
| 'num_dense_layers': 0, | ||
| 'num_classes': 1 | ||
| } | ||
|
|
||
| loss_hparams = { | ||
| 'label_confidence': 0.9, | ||
| } | ||
|
|
||
| opt_hparams = { | ||
| 'learning_rate_schedule': args.learning_rate_strategy, | ||
| 'lr_constant': args.lr_constant, | ||
| 'warmup_steps': args.warmup_steps, | ||
| 'max_training_steps': args.max_training_steps, | ||
| 'Adam_beta1': 0.9, | ||
| 'Adam_beta2': 0.997, | ||
| 'Adam_epsilon': 1e-9, | ||
| } | ||
| d_opt = { | ||
| 'optimizer': { | ||
| 'type': 'AdamOptimizer', | ||
| 'kwargs': { | ||
| 'learning_rate': 5e-4, | ||
| }, | ||
| }, | ||
| } | ||
| print('logdir:{}'.format(args.log_dir)) | ||
| if not os.path.exists(args.log_dir): | ||
| os.makedirs(args.log_dir) | ||
| if not os.path.exists(args.log_dir + 'img/'): | ||
| os.makedirs(args.log_dir + 'img/') | ||
| return { | ||
| 'train_dataset_hparams': train_dataset_hparams, | ||
| 'eval_dataset_hparams': eval_dataset_hparams, | ||
| 'test_dataset_hparams': test_dataset_hparams, | ||
| 'encoder_hparams': encoder_hparams, | ||
| 'decoder_hparams': decoder_hparams, | ||
| 'classifier_hparams': classifier_hparams, | ||
| 'loss_hparams': loss_hparams, | ||
| 'opt_hparams': opt_hparams, | ||
| 'd_opt': d_opt, | ||
| 'args': args, | ||
| } |