-
Notifications
You must be signed in to change notification settings - Fork 27
/
Copy pathtrain.py
144 lines (117 loc) · 6.89 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
import tensorflow as tf
import argparse
import os
from functools import partial
import train_utils
from vocab import Vocab
from model import LISAModel
import numpy as np
import sys
import util
arg_parser = argparse.ArgumentParser(description='')
arg_parser.add_argument('--train_files', required=True,
help='Comma-separated list of training data files')
arg_parser.add_argument('--dev_files', required=True,
help='Comma-separated list of development data files')
arg_parser.add_argument('--save_dir', required=True,
help='Directory to save models, outputs, etc.')
# todo load this more generically, so that we can have diff stats per task
arg_parser.add_argument('--transition_stats',
help='Transition statistics between labels')
arg_parser.add_argument('--hparams', type=str,
help='Comma separated list of "name=value" hyperparameter settings.')
arg_parser.add_argument('--debug', dest='debug', action='store_true',
help='Whether to run in debug mode: a little faster and smaller')
arg_parser.add_argument('--data_config', required=True,
help='Path to data configuration json')
arg_parser.add_argument('--model_configs', required=True,
help='Comma-separated list of paths to model configuration json.')
arg_parser.add_argument('--task_configs', required=True,
help='Comma-separated list of paths to task configuration json.')
arg_parser.add_argument('--layer_configs', required=True,
help='Comma-separated list of paths to layer configuration json.')
arg_parser.add_argument('--attention_configs',
help='Comma-separated list of paths to attention configuration json.')
arg_parser.add_argument('--num_gpus', type=int,
help='Number of GPUs for distributed training.')
arg_parser.add_argument('--keep_k_best_models', type=int,
help='Number of best models to keep.')
arg_parser.add_argument('--best_eval_key', required=True, type=str,
help='Key corresponding to the evaluation to be used for determining early stopping.')
arg_parser.set_defaults(debug=False, num_gpus=1, keep_k_best_models=1)
args, leftovers = arg_parser.parse_known_args()
util.init_logging(tf.logging.INFO)
# Load all the various configurations
# todo: validate json
data_config = train_utils.load_json_configs(args.data_config)
model_config = train_utils.load_json_configs(args.model_configs)
task_config = train_utils.load_json_configs(args.task_configs, args)
layer_config = train_utils.load_json_configs(args.layer_configs)
attention_config = train_utils.load_json_configs(args.attention_configs)
# attention_config = {}
# if args.attention_configs and args.attention_configs != '':
# attention_config = train_utils.load_json_configs(args.attention_configs)
# Combine layer, task and layer, attention maps
# todo save these maps in save_dir
layer_task_config, layer_attention_config = util.combine_attn_maps(layer_config, attention_config, task_config)
hparams = train_utils.load_hparams(args, model_config)
# Set the random seed. This defaults to int(time.time()) if not otherwise set.
np.random.seed(hparams.random_seed)
tf.set_random_seed(hparams.random_seed)
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
train_filenames = args.train_files.split(',')
dev_filenames = args.dev_files.split(',')
vocab = Vocab(data_config, args.save_dir, train_filenames)
vocab.update(dev_filenames)
embedding_files = [embeddings_map['pretrained_embeddings'] for embeddings_map in model_config['embeddings'].values()
if 'pretrained_embeddings' in embeddings_map]
def train_input_fn():
return train_utils.get_input_fn(vocab, data_config, train_filenames, hparams.batch_size,
num_epochs=hparams.num_train_epochs, shuffle=True, embedding_files=embedding_files,
shuffle_buffer_multiplier=hparams.shuffle_buffer_multiplier)
def dev_input_fn():
return train_utils.get_input_fn(vocab, data_config, dev_filenames, hparams.batch_size, num_epochs=1, shuffle=False,
embedding_files=embedding_files)
# Generate mappings from feature/label names to indices in the model_fn inputs
feature_idx_map, label_idx_map = util.load_feat_label_idx_maps(data_config)
# feature_idx_map = {}
# label_idx_map = {}
# for i, f in enumerate([d for d in data_config.keys() if
# ('feature' in data_config[d] and data_config[d]['feature']) or
# ('label' in data_config[d] and data_config[d]['label'])]):
# if 'feature' in data_config[f] and data_config[f]['feature']:
# feature_idx_map[f] = i
# if 'label' in data_config[f] and data_config[f]['label']:
# if 'type' in data_config[f] and data_config[f]['type'] == 'range':
# idx = data_config[f]['conll_idx']
# j = i + idx[1] if idx[1] != -1 else -1
# label_idx_map[f] = (i, j)
# else:
# label_idx_map[f] = (i, i+1)
# Initialize the model
model = LISAModel(hparams, model_config, layer_task_config, layer_attention_config, feature_idx_map, label_idx_map,
vocab)
if args.debug:
tf.logging.log(tf.logging.INFO, "Created trainable variables: %s" % str([v.name for v in tf.trainable_variables()]))
# Distributed training
distribution = tf.contrib.distribute.MirroredStrategy(num_gpus=args.num_gpus) if args.num_gpus > 1 else None
# Set up the Estimator
checkpointing_config = tf.estimator.RunConfig(save_checkpoints_steps=hparams.eval_every_steps, keep_checkpoint_max=1,
train_distribute=distribution)
estimator = tf.estimator.Estimator(model_fn=model.model_fn, model_dir=args.save_dir, config=checkpointing_config)
# Set up early stopping -- always keep the model with the best F1
export_assets = {"%s.txt" % vocab_name: "%s/assets.extra/%s.txt" % (args.save_dir, vocab_name)
for vocab_name in vocab.vocab_names_sizes.keys()}
tf.logging.log(tf.logging.INFO, "Exporting assets: %s" % str(export_assets))
save_best_exporter = tf.estimator.BestExporter(compare_fn=partial(train_utils.best_model_compare_fn,
key=args.best_eval_key),
serving_input_receiver_fn=train_utils.serving_input_receiver_fn,
assets_extra=export_assets,
exports_to_keep=args.keep_k_best_models)
# Train forever until killed
train_spec = tf.estimator.TrainSpec(input_fn=train_input_fn)
eval_spec = tf.estimator.EvalSpec(input_fn=dev_input_fn, throttle_secs=hparams.eval_throttle_secs,
exporters=[save_best_exporter])
# Run training
tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)