-
Notifications
You must be signed in to change notification settings - Fork 68
/
train_directly_runner.py
242 lines (210 loc) · 8.27 KB
/
train_directly_runner.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
from utils.run import get_full_experiment_name, make_log_and_ckpt_dirs
import os
import subprocess
import time
import numpy as np
root = '/atlas/u/chrisyeh/africa_poverty/'
default_params = {
# 'ckpt_dir': os.path.join(root, 'ckpts/DHSIncountry/'),
# 'log_dir': os.path.join(root, 'logs/DHSIncountry/'),
# 'ooc': False,
'ckpt_dir': os.path.join(root, 'ckpts/DHS_OOC/'),
'log_dir': os.path.join(root, 'logs/DHS_OOC/'),
'ooc': True,
'batch_size': 64,
}
# for OOC training
# (ls_bands, nl_band, dataset, hs_weight_init) # optionally + (reg, lr), for keep_frac training
HPARAMS = [
# ('ms', 'None', '2009-17A', 'samescaled', 1e-2, 1e-4),
# ('ms', 'None', '2009-17B', 'samescaled', 1e-3, 1e-4),
# ('ms', 'None', '2009-17C', 'samescaled', 1e-3, 1e-3),
# ('ms', 'None', '2009-17D', 'samescaled', 1e-3, 1e-2),
# ('ms', 'None', '2009-17E', 'samescaled', 1e-2, 1e-3),
# ('ms', 'split', '2009-17A', 'samescaled'),
# ('ms', 'split', '2009-17B', 'samescaled'),
# ('ms', 'split', '2009-17C', 'samescaled'),
# ('ms', 'split', '2009-17D', 'samescaled'),
# ('ms', 'split', '2009-17E', 'samescaled'),
('None', 'split', '2009-17A', 'random', 1.0, 1e-4),
('None', 'split', '2009-17B', 'random', 1.0, 1e-4),
('None', 'split', '2009-17C', 'random', 1.0, 1e-4),
('None', 'split', '2009-17D', 'random', 1.0, 1e-2),
('None', 'split', '2009-17E', 'random', 1.0, 1e-4),
# ('rgb', 'None', '2009-17A', 'same'),
# ('rgb', 'None', '2009-17B', 'same'),
# ('rgb', 'None', '2009-17C', 'same'),
# ('rgb', 'None', '2009-17D', 'same'),
# ('rgb', 'None', '2009-17E', 'same'),
# ('rgb', 'split', '2009-17A', 'samescaled'),
# ('rgb', 'split', '2009-17B', 'samescaled'),
# ('rgb', 'split', '2009-17C', 'samescaled'),
# ('rgb', 'split', '2009-17D', 'samescaled'),
# ('rgb', 'split', '2009-17E', 'samescaled'),
]
# for incountry training
# (ls_bands, nl_band, dataset, hs_weight_init)
# HPARAMS = [
# ('ms', 'None', 'incountryA', 'samescaled'),
# ('ms', 'None', 'incountryB', 'samescaled'),
# ('ms', 'None', 'incountryC', 'samescaled'),
# ('ms', 'None', 'incountryD', 'samescaled'),
# ('ms', 'None', 'incountryE', 'samescaled'),
# ('ms', 'split', 'incountryA', 'samescaled'),
# ('ms', 'split', 'incountryB', 'samescaled'),
# ('ms', 'split', 'incountryC', 'samescaled'),
# ('ms', 'split', 'incountryD', 'samescaled'),
# ('ms', 'split', 'incountryE', 'samescaled'),
# ('None', 'split', 'incountryA', 'random'),
# ('None', 'split', 'incountryB', 'random'),
# ('None', 'split', 'incountryC', 'random'),
# ('None', 'split', 'incountryD', 'random'),
# ('None', 'split', 'incountryE', 'random'),
# ('rgb', 'None', 'incountryA', 'same'),
# ('rgb', 'None', 'incountryB', 'same'),
# ('rgb', 'None', 'incountryC', 'same'),
# ('rgb', 'None', 'incountryD', 'same'),
# ('rgb', 'None', 'incountryE', 'same'),
# ('rgb', 'split', 'incountryA', 'samescaled'),
# ('rgb', 'split', 'incountryB', 'samescaled'),
# ('rgb', 'split', 'incountryC', 'samescaled'),
# ('rgb', 'split', 'incountryD', 'samescaled'),
# ('rgb', 'split', 'incountryE', 'samescaled'),
# ]
PYTHON_COMMAND_TEMPLATE = 'python train_directly.py \
--label_name wealthpooled \
--batcher base \
--model_name resnet --num_layers 18 \
--lr_decay 0.96 \
--batch_size {batch_size} \
--augment True \
--max_epochs 200 --eval_every 1 --print_every 40 \
--ooc {ooc} \
--keep_frac "{keep_frac}" \
--gpu 0 --num_threads 5 \
--ckpt_dir {ckpt_dir} \
--log_dir {log_dir} \
--cache train,train_eval,val \
--seed "{seed}" \
\
--experiment_name "{experiment_name}" \
--dataset "{dataset}" \
--ls_bands "{ls_bands}" \
--nl_band "{nl_band}" \
--lr "{lr}" \
--fc_reg "{reg}" --conv_reg "{reg}" \
--imagenet_weights_path "{imagenet_weights_path}" \
--hs_weight_init "{hs_weight_init}"'
def hparams_to_command(hparams_tup):
ls_bands, nl_band, dataset, hs_weight_init, reg, lr, keep_frac, seed = hparams_tup
if keep_frac < 1:
experiment_name = '{dataset}_18preact_{bands}_{init}_keep{keep_frac:g}_seed{seed}'.format(
dataset=dataset,
bands=get_bandname(ls_bands, nl_band),
init=hs_weight_init,
keep_frac=keep_frac,
seed=seed)
else:
experiment_name = '{dataset}_18preact_{bands}_{init}'.format(
dataset=dataset,
bands=get_bandname(ls_bands, nl_band),
init=hs_weight_init)
imagenet_weights_path = 'None'
if hs_weight_init in ['same', 'samescaled']:
imagenet_weights_path = '/atlas/group/model_weights/imagenet_resnet18_tensorpack.npz'
command = PYTHON_COMMAND_TEMPLATE.format(
batch_size=default_params['batch_size'],
ckpt_dir=default_params['ckpt_dir'],
log_dir=default_params['log_dir'],
ooc=default_params['ooc'],
keep_frac=keep_frac,
seed=seed,
experiment_name=experiment_name,
dataset=dataset,
ls_bands=ls_bands,
nl_band=nl_band,
lr=lr,
reg=reg,
imagenet_weights_path=imagenet_weights_path,
hs_weight_init=hs_weight_init)
full_experiment_name = get_full_experiment_name(
experiment_name, default_params['batch_size'], reg, reg, lr)
log_dir, _ = make_log_and_ckpt_dirs(
default_params['log_dir'], default_params['ckpt_dir'], full_experiment_name)
output_path = os.path.join(log_dir, 'output.log')
command = f'{command} >& "{output_path}" &'
return command
# Empirical memory requirements:
# - DHS incountryD
# - MS: 44 GB RAM, 5439MiB GPU
# - NL: 13 GB RAM, 4415MiB GPU
def get_mem(bands: str) -> int:
return {
'rgb': 20,
'rgbnl': 32,
'ms': 44,
'msnl': 56,
'nl': 15,
}[bands]
def get_bandname(ls_bands: str, nl_band: str) -> str:
return {
('ms', 'None'): 'ms',
('ms', 'split'): 'msnl',
('None', 'split'): 'nl',
('rgb', 'None'): 'rgb',
('rgb', 'split'): 'rgbnl',
}[(ls_bands, nl_band)]
def get_mem_for_hparams(hparams):
bandname = get_bandname(hparams[0], hparams[1])
keep_frac = hparams[-2]
min_mem = 10
buffer = 2
mem = int(get_mem(bandname) * keep_frac + buffer)
mem = max(min_mem, mem)
return mem
len_hparams = len(HPARAMS[0])
assert all(np.array(list(map(len, HPARAMS))) == len_hparams)
# order of params: ls_bands, nl_band, dataset, hs_weight_init, reg, lr, keep_frac, seed
all_hparams = []
if len_hparams == 4: # (ls_bands, nl_band, dataset, hs_weight_init)
keep = 1.0
seed = 123
for hparams_tup in HPARAMS:
for reg in [1e-0, 1e-1, 1e-2, 1e-3]:
for lr in [1e-2, 1e-3, 1e-4]:
new_tup = tuple(list(hparams_tup) + [reg, lr, keep, seed])
all_hparams.append(new_tup)
elif len_hparams == 6: # (ls_bands, nl_band, dataset, hs_weight_init, reg, lr)
for hparams_tup in HPARAMS:
for keep in [0.05, 0.10, 0.25, 0.5]:
for seed in [123, 456, 789]:
new_tup = tuple(list(hparams_tup) + [keep, seed])
all_hparams.append(new_tup)
# sort hparams by ls_bands
all_hparams = sorted(all_hparams, key=lambda x: x[0])
outputs_dir = os.path.join(root, 'outputs', str(int(time.time())))
os.makedirs(outputs_dir, exist_ok=True)
for i in range(0, len(all_hparams), 2):
print(i)
hparams1 = all_hparams[i]
command1 = hparams_to_command(hparams1)
mem1 = get_mem_for_hparams(hparams1)
command2 = ''
mem2 = 0
if i + 1 < len(all_hparams):
hparams2 = all_hparams[i + 1]
command2 = hparams_to_command(hparams2)
mem2 = get_mem_for_hparams(hparams2)
command = command1 + '\n\n' + command2
mem = mem1 + mem2
with open(os.path.join(root, 'scripts', 'train_model_slurm.sh'), 'r') as template_file:
template = template_file.read()
full_slurm_script = template.format(
SLURM_MEM=f'{mem}G',
SLURM_JOB_NAME=f'slurm_{i}',
SLURM_OUTPUT_LOG=os.path.join(outputs_dir, f'slurm_{i}.log'),
content=command)
slurm_sh_path = os.path.join(outputs_dir, f'slurm_{i}.sh')
with open(slurm_sh_path, 'w') as f:
f.write(full_slurm_script)
subprocess.run(['sbatch', slurm_sh_path])