-
Notifications
You must be signed in to change notification settings - Fork 18
/
Copy pathprune_distill_step2.py
346 lines (238 loc) · 13 KB
/
prune_distill_step2.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
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
import numpy as np
from torch.utils.data import DataLoader
import torch
from segment_anything_kd import sam_model_registry
from load_sam_json import SamDataset
from segment_anything_kd.utils.transforms import ResizeLongestSide
from prune_funcs import calculate_iou, get_pos_init, del_pos_init, prune_sam_step2_global
import json
from pycocotools import mask as mask_utils
import argparse
seed = 0
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
parser = argparse.ArgumentParser(description='SlimSAM')
parser.add_argument('--traindata_path', type=str,default = '')
parser.add_argument('--valdata_path', type=str,default = '')
parser.add_argument('--trainsize', type=int,default = 10000)
parser.add_argument('--gradsize', type=int,default = 1000)
parser.add_argument('--valsize', type=int,default = 50)
parser.add_argument('--epochs', type=int,default = 20)
parser.add_argument('--norm_type', type=str,default = 'gaussian')
parser.add_argument('--imptype', type=str,default = 'Disturb')
parser.add_argument('--global_way', type=bool,default = True)
parser.add_argument('--prune_ratio', type=float,default = 0.5)
parser.add_argument('--model_path', type=str,default = 'checkpoints/vit_b_slim_step1_.pth')
args, unparsed = parser.parse_known_args()
def train_model():
# torch.backends.cudnn.deterministic = True
device = torch.device("cuda")
print("CUDA visible devices: " + str(torch.cuda.device_count()))
print("CUDA Device Name: " + str(torch.cuda.get_device_name(device)))
train_root_folder = args.traindata_path
val_root_folder = args.valdata_path
TRAIN_SIZE = args.trainsize
VAL_SIZE = args.valsize
GRAD_SIZE = args.gradsize
num_train_epochs = args.epochs
batch_size = 1
model_path = args.model_path
# Creating dataset loaders
grad_dataset = SamDataset(root_folder=train_root_folder, dataset_size=GRAD_SIZE, val=False)
grad_loader = DataLoader(dataset=grad_dataset, batch_size=1, shuffle=False, num_workers=4,
pin_memory=True, drop_last=True)
train_dataset = SamDataset(root_folder=train_root_folder, dataset_size=TRAIN_SIZE, val=False)
train_loader = DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True, num_workers=4,
pin_memory=True, drop_last=True)
val_dataset = SamDataset(root_folder=val_root_folder, dataset_size=VAL_SIZE, val=True)
val_loader = DataLoader(dataset=val_dataset, batch_size=1, shuffle=False, num_workers=2,
pin_memory=True, drop_last=False)
# student model
model = torch.load(model_path)
model.image_encoder = model.image_encoder.module
# rewrite the forward function of image encoder
def forward(self, x):
block_outputs = []
x = self.patch_embed(x)
if self.pos_embed is not None:
x = x + self.pos_embed
block_outputs.append(x)
for blk in self.blocks:
x,qkv_emb,mid_emb,x_emb = blk(x)
block_outputs.append(x_emb)
x = self.neck(x.permute(0, 3, 1, 2))
block_emb = block_outputs[0]
for emb in block_outputs[1:]:
block_emb = torch.cat([block_emb,emb],dim=0)
return x, block_emb
# teacher model
teacher_model_type = 'vit_b'
checkpoint = 'checkpoints/sam_vit_b_qkv.pth'
teacher_model = sam_model_registry[teacher_model_type](checkpoint=checkpoint)
teacher_model.to(device)
teacher_model.eval()
# load the pruned model
pruned_model = torch.load(model_path)
pruned_model.image_encoder = pruned_model.image_encoder.module
pruned_model.to(device)
pruned_model.eval()
# Rewrite forward functions
import types
funcType = types.MethodType
model.image_encoder.forward = funcType(forward, model.image_encoder)
pruned_model.image_encoder.forward = funcType(forward, pruned_model.image_encoder)
teacher_model.image_encoder.forward = funcType(forward, teacher_model.image_encoder)
MSE_loss = torch.nn.MSELoss()
lr = 1e-4
ratio = args.prune_ratio
loss_fn = torch.nn.MSELoss()
transform = ResizeLongestSide(1024)
norm_type = args.norm_type
imptype = args.imptype
global_way = args.global_way
a_weight = 0.5
round_to = model.image_encoder.num_heads
print("===========================Parameter Settings===========================")
print("Pruning Ratio:",ratio)
print("VIT num_heads:",round_to)
print("norm_type:",norm_type)
print("imptype:",imptype)
print("global:",global_way)
print("learning rate:",lr)
print("a_weight:",a_weight)
print("round_to",round_to)
print("TRAIN_SIZE",TRAIN_SIZE,"VAL_SIZE",VAL_SIZE, "GRAD_SIZE",GRAD_SIZE,"Epochs",num_train_epochs)
model_name = teacher_model_type
example_inputs = torch.randn(1, 3, 1024, 1024)
for k in range(1):
############################################get initial grad for importance estimation############################################
best_iou = 0
model.to(device)
model.image_encoder.train()
grad_iter = iter(grad_loader)
for i in range(len(grad_iter)):
batch = next(grad_iter)
input_image = batch["input_image"].to(device)
with torch.no_grad():
teacher_embedding,_ = pruned_model.image_encoder(input_image)
teacher_embedding += torch.normal(mean=0,std=0.01,size=(1, 256, 64, 64)).to(device) #Disturbed image embedding
student_embedding, _= model.image_encoder(input_image)
loss = loss_fn(teacher_embedding, student_embedding)
loss.backward()
#########################################################################################################
print("===========================Pruning Start===========================")
#Bottleneck Pruning
model.cpu().eval()
model = del_pos_init(model)
##Global pruning QKV Attention
model.image_encoder = prune_sam_step2_global(model=model.image_encoder, example_inputs=example_inputs, model_name=model_name, round_to=round_to, ratio=ratio, imptype = imptype, norm_type=norm_type, global_way=global_way, gs=1)
##Global pruning MLP Layer
model.image_encoder = prune_sam_step2_global(model=model.image_encoder, example_inputs=example_inputs, model_name=model_name, round_to=round_to, ratio=ratio, imptype = imptype, norm_type=norm_type, global_way=global_way, gs=2)
model = get_pos_init(model)
model.to(device)
model.image_encoder = torch.nn.DataParallel(model.image_encoder)
model.image_encoder.train()
pruned_model.image_encoder = torch.nn.DataParallel(pruned_model.image_encoder)
pruned_model.image_encoder.eval()
optimizer = torch.optim.Adam(model.image_encoder.parameters(), lr=lr)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'max',factor=0.5,patience=4,verbose=True)
model.zero_grad()
teacher_model.zero_grad()
#Embedding Aligning
for epoch in range(num_train_epochs):
teacher_model.image_encoder = torch.nn.DataParallel(teacher_model.image_encoder)
teacher_model.image_encoder.eval()
torch.cuda.empty_cache()
train_iter = iter(train_loader)
if epoch<11:
a_weight = (11-epoch-1)/11
print("Dynamic weight:",a_weight)
else:
a_weight = 0
print("Dynamic weight:",a_weight)
for i in range(len(train_iter)):
batch = next(train_iter)
input_image = batch["input_image"].to(device)
with torch.no_grad():
teacher_embedding,teacher_block_emb = teacher_model.image_encoder(input_image)
pruned_embedding,pruned_block_emb = pruned_model.image_encoder(input_image)
student_embedding,student_block_emb = model.image_encoder(input_image)
#loss = loss_fn(student_embedding, teacher_embedding)
loss = (1-a_weight)*loss_fn(student_embedding, teacher_embedding)+a_weight*loss_fn(student_block_emb, pruned_block_emb)+a_weight*loss_fn(student_embedding, pruned_embedding)
loss.backward()
#### batchsize×4 ####
if i%4==3:
optimizer.step()
optimizer.zero_grad()
#validation
if i == len(train_iter)-1:
teacher_model.image_encoder = teacher_model.image_encoder.module
iou = 0
model.image_encoder.eval()
with torch.no_grad():
val_iter = iter(val_loader)
for j in range(len(val_iter)):
batch = next(val_iter)
input_image = batch["input_image"].to(device)
input_size = batch["input_size"]
original_image_size = batch["original_image_size"]
original_image_size[0] = original_image_size[0].numpy()[0]
original_image_size[1] = original_image_size[1].numpy()[0]
original_image_size = ([original_image_size[0],original_image_size[1]])
input_size[0] = input_size[0].numpy()[0]
input_size[1] = input_size[1].numpy()[0]
input_size = ([input_size[0],input_size[1]])
id = batch["id"]
annot = batch["annot"][0]
path = id[0]
with open(annot, encoding="utf-8") as f:
dict_data = json.load(f)
dict_data = dict_data["annotations"]
sub_count = 0
sub_iou = 0
for example in dict_data:
sub_count += 1
input_point = np.array(example['point_coords'])
input_label = np.array([1])
mask = mask_utils.decode(example["segmentation"])
point_coords = transform.apply_coords(input_point, original_image_size)
coords_torch = torch.as_tensor(point_coords, dtype=torch.float, device=device)
labels_torch = torch.as_tensor(input_label, dtype=torch.int, device=device)
coords_torch, labels_torch = coords_torch[None, :, :], labels_torch[None, :]
points = (coords_torch, labels_torch)
# Model inference
image_embedding,_ = model.image_encoder(input_image)
sparse_embeddings, dense_embeddings = model.prompt_encoder(
points=points,
boxes=None,
masks=None,
)
low_res_masks, iou_predictions = model.mask_decoder(
image_embeddings=image_embedding,
image_pe=model.prompt_encoder.get_dense_pe(),
sparse_prompt_embeddings=sparse_embeddings,
dense_prompt_embeddings=dense_embeddings,
multimask_output=False,
)
student_masks = teacher_model.postprocess_masks(low_res_masks, input_size, original_image_size)
student_masks = student_masks > teacher_model.mask_threshold
student_masks = student_masks[0].detach().cpu().numpy()[0]
sub_iou += calculate_iou(student_masks, mask)
sub_iou = sub_iou/sub_count
iou += sub_iou
iou = iou/len(val_iter)
model.image_encoder.train()
model.image_encoder.eval()
if iou>=best_iou:
best_iou = iou
filename = 'checkpoints/vit_b_slim_step2_'+'.pth'
torch.save(model, filename)
print("save checkpoint")
model.image_encoder.train()
scheduler.step(iou)
print("epoch:",epoch)
print("IOU: {} Best IOU {}".format(iou,best_iou))
if __name__ == '__main__':
train_model()