-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathrun_cv.py
329 lines (263 loc) · 11 KB
/
run_cv.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
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
import copy
import itertools
import json
import os
import random
from pprint import pprint
try:
import drexo
except ImportError:
drexo_url = 'https://gitlab.com/mdarcy220/drexo/raw/51c2d3db67f0e7b51df2aeb5c4c9e907053c5aae/drexo/drexo.py?inline=false'
print('Could not import drexo v0.1. Downloading standalone from {}'.format(drexo_url))
import urllib.request
urllib.request.urlretrieve(drexo_url, 'drexo.py')
import drexo
import numpy as np
CV_NUM_FOLDS = 5
NUM_TRIALS = 3
runman = None
def make_cmdargs_from_params(params, output_dir, seed):
core_args = list(itertools.chain.from_iterable(params['core_args']))
grid_args = []
for argname in params['grid_params']:
if argname == '--train_batch_size':
batch_size = params['grid_params'][argname]
grid_args.extend(['--gradient_accumulation_steps', str(batch_size//2)])
if argname == '--num_train_epochs':
num_epochs = params['grid_params'][argname]
grid_args.extend([argname, str(params['grid_params'][argname])])
final_args = ['python3'] + [params['run_filename']] + core_args + grid_args
final_args.extend(['--output_dir', output_dir + os.path.sep])
final_args.extend(['--data_dir', params['data_dir']])
final_args.extend(['--seed', seed])
return final_args
def on_job_finished(full_runconfig):
# Remove saved model after (to save space)
model_save_filename = full_runconfig['params']['model_save_filename']
if '{}' in model_save_filename:
model_save_filename = model_save_filename.format(full_runconfig['params']['grid_params']['--num_train_epochs']-1)
model_save_filepath = os.path.join(full_runconfig['job_output_dir'], model_save_filename)
if os.path.exists(model_save_filepath):
os.remove(model_save_filepath)
def compare_params(cur_params, params_from_cache):
if 'keys_to_compare' not in cur_params:
return False
#if 'keys_to_compare' not in params_from_cache or set(params_from_cache['keys_to_compare']) != set(cur_params['keys_to_compare']):
if 'keys_to_compare' not in params_from_cache or not set(params_from_cache['keys_to_compare']).issubset(set(cur_params['keys_to_compare'])):
return False
cur_params['keys_to_compare']
for key in cur_params['keys_to_compare']:
if key not in cur_params or key not in params_from_cache or cur_params[key] != params_from_cache[key]:
return False
return True
def get_acc_from_dir(output_dir):
label_info = None
with open(os.path.join(output_dir, 'output', 'model_labels.json'), 'r') as f:
label_info = json.load(f)
num_total = 0
num_correct = 0
for ex in label_info:
num_total += 1
if ex['model_label'] == ex['true_label']:
num_correct += 1
return num_correct / num_total
def try_until_good(params, data_dir, callback):
global runman
failure_cutoff_acc = 0.30
max_trials = 5
new_params = copy.deepcopy(params)
new_params['data_dir'] = data_dir
results = []
def end_hook(output_dir, seed):
nonlocal results, failure_cutoff_acc, max_trials, new_params
acc = get_acc_from_dir(output_dir)
results.append(acc)
if acc < failure_cutoff_acc and len(results) < max_trials:
runman.add_to_run_queue(new_params, end_hook, front=True)
return
if acc < failure_cutoff_acc:
acc = sum(results)/len(results)
callback(new_params, output_dir, data_dir, acc)
runman.add_to_run_queue(new_params, end_hook)
def try_params_on_folds(params, data_dirs, callback):
fold_results = dict()
def sub_callback(_1, _2, data_dir, acc):
nonlocal fold_results
fold_results[data_dir] = acc
for d in data_dirs:
if d not in fold_results:
return False
avg_acc = sum([x for x in fold_results.values()]) / len(fold_results)
callback(params, data_dirs, avg_acc)
for data_dir in data_dirs:
try_until_good(params, data_dir, sub_callback)
def get_all_gridsearch_params():
all_params = []
for batch_size in [16, 32]:
for num_epochs in [3, 4, 6]:
for learning_rate in [1e-5, 2e-5, 3e-5]:
all_params.append({'--train_batch_size': batch_size, '--num_train_epochs': num_epochs, '--learning_rate': learning_rate})
return all_params
def combine_params(base_params, grid_params):
params = copy.copy(base_params)
params['grid_params'] = copy.copy(grid_params)
return params
def gridsearch_params_on_subfolds(base_params, main_fold, callback):
global CV_NUM_FOLDS
results = []
num_tested_params = 0
def sub_callback(params, _, avg_acc):
nonlocal results, num_tested_params
results.append((params, avg_acc))
num_tested_params += 1
if num_tested_params != len(get_all_gridsearch_params()):
return False
best_params = None
best_params_value = -1
for p, v in results:
if best_params_value < v:
best_params = p
best_params_value = v
callback(base_params, main_fold, best_params)
data_dirs = []
for fold_num in range(CV_NUM_FOLDS):
data_dirs.append(os.path.join(main_fold, 'fold{}'.format(fold_num)))
for grid_params in get_all_gridsearch_params():
params = combine_params(base_params, grid_params)
try_params_on_folds(params, data_dirs, sub_callback)
def run_cv(model_info, data_path, modelname, config, do_nested_gridsearch=False):
global CV_NUM_FOLDS
arg_pieces = []
model_common_args = [['--do_eval']]
for fold_num in range(CV_NUM_FOLDS):
model_specific_args = []
if modelname == 'bert':
model_specific_args.append(['--do_lower_case'])
fold_path = data_path + str(fold_num)
config_args = []
if config == 'ft_swagonly':
config_args = [copy.copy(model_info[modelname]['swag_load_args'])]
elif config == 'ft_swag_and_codah':
config_args = [['--do_train']] + [copy.copy(model_info[modelname]['swag_load_args'])]
elif config == 'ft_codah_only':
config_args = [['--do_train']]
elif config == 'ft_answeronly':
config_args = [['--do_train'], ['--answer_only']]
else:
print('Unknown config {}'.format(config))
continue
if modelname == 'bert' and config not in {'ft_swagonly', 'ft_swag_and_codah'}:
config_args.append(['--bert_model', 'bert-large-uncased'])
config_args.append(['--model_labels_save_filename', 'model_labels.json'])
core_args = [model_info[modelname]['base_args']] + model_common_args + config_args
core_args = sorted(core_args, key=lambda x: str(x))
base_params = {
'keys_to_compare': ['run_filename', 'core_args', 'model', 'config', 'fold_num', 'foldtype', 'data_dir', 'grid_params'],
'run_filename': model_info[modelname]['run_filename'],
'core_args': core_args,
'grid_params': model_info[modelname]['default_grid_params'].copy(),
'model': modelname,
'config': config,
'data_dir': fold_path,
'fold_num': fold_num,
'foldtype': 'mainfold',
'model_save_filename': model_info[modelname]['model_save_filename'],
}
_run_fold(base_params, fold_path, do_nested_gridsearch)
global_final_results = []
def _run_fold(base_params, fold_path, do_nested_gridsearch):
global runman, NUM_TRIALS, global_final_results
base_params = copy.deepcopy(base_params)
if do_nested_gridsearch:
def callback(_1, _2, best_params):
nonlocal fold_path
def sub_callback(_3, output_dir, _4, acc):
global_final_results.append([best_params, acc, os.path.abspath(output_dir)])
best_params = copy.deepcopy(best_params)
best_params['foldtype'] = 'mainfold'
for i in range(NUM_TRIALS):
try_until_good(best_params, fold_path, sub_callback)
# Only run answer-only with best_params found in codah-only gridsearch
if best_params['config'] != 'ft_codah_only':
return
best_params2 = copy.deepcopy(best_params)
def sub_callback2(_3, output_dir, _4, acc):
global_final_results.append([best_params2, acc, os.path.abspath(output_dir)])
best_params2['core_args'].append(['--answer_only'])
best_params2['config'] = 'ft_answeronly'
best_params2['core_args'] = sorted(best_params2['core_args'], key=lambda x: str(x))
for i in range(NUM_TRIALS):
try_until_good(best_params2, fold_path, sub_callback2)
base_params['foldtype'] = 'subfold'
gridsearch_params_on_subfolds(base_params, fold_path, callback)
else:
def sub_callback(output_dir, seed):
acc = get_acc_from_dir(output_dir)
global_final_results.append([base_params, acc, output_dir])
for i in range(NUM_TRIALS):
runman.add_to_run_queue(base_params, sub_callback)
def calc_result_stats(raw_results):
results_byfold = dict()
for tmp in raw_results:
params = tmp[0]
dataname = os.path.basename(os.path.dirname(params['data_dir']))
modelname = os.path.basename(os.path.dirname(params['model']))
foldname = os.path.basename(params['data_dir'])
key = (modelname, params['config'], dataname, foldname)
if key not in results_byfold:
results_byfold[key] = []
results_byfold[key].append(tmp[1])
results = dict()
for key in results_byfold:
newkey = (key[0], key[1], key[2])
if newkey not in results:
results[newkey] = []
results[newkey].append(results_byfold[key])
for key in results:
# Avg across folds
tmp = np.mean(results[key], axis=0).tolist()
# Avg across trials and convert to percent
results[key] = 'mean={:0.3f}, std={:0.3f}'.format(100*np.mean(tmp), 100*np.std(tmp, ddof=1))
return results
if __name__ == '__main__':
runman = drexo.RunManager('./gitignore/outputs/all_experiments/', make_cmdargs_from_params, params_compare_func=compare_params, job_finished_func=on_job_finished)
model_info = {
'bert': {'run_filename': 'run_classifier.py',
'base_args': ['--save_final_only', '--task_name', 'codah'],
'swag_save_file': './gitignore/saved_models/swag_bert_for_cv/',
'swag_load_args': ['--bert_model', './gitignore/saved_models/swag_bert_for_cv/'],
'model_save_filename': 'pytorch_model_epoch{}.bin',
'default_grid_params': {'--train_batch_size': 16, '--learning_rate': 2e-5, '--num_train_epochs': 3},
},
'gpt1': {'run_filename': 'run_swag_gpt1.py',
'base_args': ['--train_filename', 'train.tsv', '--eval_filename', 'test.tsv', '--data_format', 'codah'],
'swag_save_file': './gitignore/saved_models/swag_gpt1_for_cv/',
'model_save_filename': 'pytorch_model.bin',
'default_grid_params': {'--train_batch_size': 32, '--learning_rate': 6.25e-5, '--num_train_epochs': 3},
},
}
for modelname in model_info:
if 'swag_load_args' not in model_info[modelname]:
model_info[modelname]['swag_load_args'] = ['--load_model_from', os.path.join(model_info[modelname]['swag_save_file'], model_info[modelname]['model_save_filename'])]
configs = ['ft_codah_only', 'ft_swag_and_codah', 'ft_answeronly']
for data_dirname in ['codah_' + x for x in ['20', '40', '60', '80']]:
data_path = os.path.join('./gitignore/data/altsizes/', data_dirname, 'fold')
for config in configs:
for model in model_info:
do_gridsearch = (model == 'bert')
if do_gridsearch and config == 'ft_answeronly':
# Answeronly gets tested automatically
# as part of the ft_codah_only gridsearch
continue
run_cv(model_info, data_path, model, config, do_nested_gridsearch=do_gridsearch)
runman.runner.run()
print()
print(global_final_results)
with open('./gitignore/outputs/global_final_results.json', 'w') as f:
json.dump(global_final_results, f)
results = calc_result_stats(global_final_results)
pprint(results)
if len(runman.get_errors()) > 0:
for tmp in runman.get_errors():
print(tmp)
print('Run had errors!!!')