-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrun_sparsify.py
217 lines (205 loc) · 8.54 KB
/
run_sparsify.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
import os
from train_model import (
process_args,
prepare_model_train,
prepare_dataset,
prepare_model,
model_indx_is_conv,
)
from dataset import MIPBatchLoader
from sparsify import SparsifyModel, SparsifySequential, SparsifyBackward, SparsifyDgl
import numpy as np
from training.utils import get_storage_dir, log_config, Mode
import time
"""script used to sparsify a trained model by computing the importance score of each neuron and pruning non-critical neurons
"""
def prepare_arguments():
parser = process_args()
parser.add_argument("--num-samples", "-n", default=10, type=int)
parser.add_argument("--sparsification-weight", "-sw", default=5, type=int)
parser.add_argument("--threshold", "-tt", default=1e-1, type=float)
parser.add_argument("--mean-threshold", "-mth", action="store_true")
parser.add_argument("--force", "-f", action="store_true")
parser.add_argument("--heat-map", "-hm", action="store_true")
parser.add_argument("--fine-tune", "-ft", action="store_true")
parser.add_argument("--sequential-run", "-seq", action="store_true")
parser.add_argument("--relaxed", "-rl", action="store_true")
parser.add_argument("--tuned-epochs", "-te", default=1, type=int)
parser.add_argument("--prune-from", "-pf", type=int)
parser.add_argument("--layer-by-layer", "-bll", action="store_true")
return parser
def prepare_config():
parser = prepare_arguments()
config = parser.parse_args()
return config
def prepare_images_mip_input(mip_batch_loader):
X, y, initial_bounds = next(mip_batch_loader)
input_size, n_channels = mip_batch_loader.get_input_n_channels()
n_output_classes = mip_batch_loader.get_n_output_classes()
return X, y, initial_bounds, input_size, n_output_classes, n_channels
if __name__ == "__main__":
config = prepare_config()
data_loaders = prepare_dataset(config)
train_loader = data_loaders["train"]
val_loader = data_loaders["val"]
test_loader = data_loaders["test"]
mip_data_loader = MIPBatchLoader(
config, val_loader, epsilon=1e-5, is_conv_model=model_indx_is_conv[config.model]
)
(
X,
y,
initial_bounds,
input_size,
n_output_classes,
n_channels,
) = prepare_images_mip_input(mip_data_loader)
exp_indx = 0
all_exp_results = {}
finetuning_prefix = "Fine Tuned"
pruning_percentage_prefix = "Pruning Percentage"
tuning_epochs = config.tuned_epochs
sparsification_time_list = []
model_train = None
while True:
model = prepare_model(
config,
input_size=input_size,
n_channels=n_channels,
n_output_classes=n_output_classes,
)
storage_parent_dir = get_storage_dir(config, model.name, exp_indx)
del model
if not (os.path.isdir(storage_parent_dir)):
break
model_train = prepare_model_train(
config,
input_size,
n_channels=n_channels,
n_output_classes=n_output_classes,
exp_indx=exp_indx,
)
log_config(model_train._logger, config)
mip_data_loader.set_model(
model_train.model
) # evaluator is accuracy computation the default one
X, y, initial_bounds = next(mip_data_loader)
if config.sequential_run:
# running independently on each class then taking the average
sparsify = SparsifySequential(
model_train,
mip_data_loader,
threshold=config.threshold,
sparsification_weight=config.sparsification_weight,
relaxed_constraints=config.relaxed,
n_output_classes=n_output_classes,
mean_threshold=config.mean_threshold,
)
else:
sparsify = SparsifyModel(
model_train,
threshold=config.threshold,
sparsification_weight=config.sparsification_weight,
relaxed_constraints=config.relaxed,
mean_threshold=config.mean_threshold,
)
if not (config.decoupled_train):
sparsify.create_bounds(initial_bounds)
if config.layer_by_layer:
sparsify = SparsifyBackward(
sparsify, mip_data_loader, n_output_classes=n_output_classes,
)
elif config.decoupled_train:
sparsify = SparsifyDgl(
sparsify, mip_data_loader, n_output_classes=n_output_classes,
)
parameters_removed_percentage = 0
for mode in [Mode.MASK, Mode.Random, Mode.CRITICAL]:
start_sparsify_time = time.time()
removal_percentage = sparsify.sparsify_model(
X,
y,
mode=mode,
use_cached=not (config.force),
start_pruning_from=config.prune_from,
)
sparsification_time = time.time() - start_sparsify_time
if mode == Mode.MASK:
parameters_removed_percentage = removal_percentage
sparsification_time_list.append(sparsification_time)
model_results = model_train.print_results(
train_loader,
val_loader,
test_loader,
save_heat_map=config.heat_map,
mode_name=mode.name,
)
model_results[-1][pruning_percentage_prefix] = parameters_removed_percentage
for model_indx, mode_name in enumerate(["original", mode.name]):
if (pruning_percentage_prefix) not in model_results[model_indx]:
model_results[model_indx][pruning_percentage_prefix] = 0
for metric_name in model_results[model_indx]:
key_results_name = mode_name + metric_name
if key_results_name not in all_exp_results:
all_exp_results[key_results_name] = []
all_exp_results[key_results_name].append(
model_results[model_indx][metric_name]
)
# Fine tune all modes to compare results
if config.fine_tune and tuning_epochs > 0:
model_train.train(
train_loader,
val_loader=None,
num_epochs=tuning_epochs,
finetune_masked=True,
)
finetuned_model_results = model_train.print_results(
train_loader,
val_loader,
test_loader,
mode_name=mode.name,
test_original_model=False,
)
finetuned_model_results[0][
pruning_percentage_prefix
] = parameters_removed_percentage
for metric_name in finetuned_model_results[0]:
key_results_name = finetuning_prefix + mode.name + metric_name
if key_results_name not in all_exp_results:
all_exp_results[key_results_name] = []
all_exp_results[key_results_name].append(
finetuned_model_results[0][metric_name]
)
sparsify.reset()
exp_indx += 1
# Now logging mean and variance of sparsification results
if exp_indx > 0:
list_modes = ["original", Mode.MASK.name, Mode.Random.name, Mode.CRITICAL.name]
if config.fine_tune:
list_modes.insert(2, finetuning_prefix + Mode.MASK.name)
list_modes.insert(3, finetuning_prefix + Mode.Random.name)
list_modes.insert(4, finetuning_prefix + Mode.CRITICAL.name)
for mode in list_modes:
for metric_name in [
"loss_train",
"acc_train",
"loss_test",
"acc_test",
pruning_percentage_prefix,
]:
results_list = all_exp_results[mode + metric_name]
metric_name_clean = " ".join(metric_name.split("_")).capitalize()
model_train._logger.info(
"{} {} mean {} +- {}".format(
mode,
metric_name_clean,
np.mean(results_list),
np.std(results_list),
)
)
if model_train is not None:
model_train._logger.info(
"Sparsification time including swapping pytorch layers mean {} +- {}".format(
np.mean(sparsification_time_list), np.std(sparsification_time_list)
)
)