-
Notifications
You must be signed in to change notification settings - Fork 3
/
finetune_seg.py
227 lines (180 loc) · 8.51 KB
/
finetune_seg.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
from tqdm.auto import tqdm, trange
from torch import nn
from transformers import AutoProcessor, CLIPSegForImageSegmentation
from transformers import CLIPProcessor, CLIPModel
from collections import namedtuple
from torchvision.transforms import Resize
from argparse import ArgumentParser
import os
import pandas as pd
from seg_utils.hash_data import HashDS
from seg_utils.crop_data import search_crops, CropDS
from seg_utils.plurals import to_plur_aug
import torch
from torch.utils.data import DataLoader
import json
def get_opts():
parser = ArgumentParser()
parser.add_argument('--clip_ft', '-c', type=str, default="data/clip_ckpt/0_CLIPModel", help="directory of fine-tuned CLIP checkpoint")
parser.add_argument('--data_dir', '-d', type=str, default="data/wikiscenes", help="directory WikiScenes data is stored in")
parser.add_argument('--crop_metadata', '-cm', type=str, default="data/seg_crop_metadata.csv", help="crop metadata filename")
parser.add_argument('--hash_metadata', '-hm', type=str, default="data/seg_hash_metadata.csv", help="hash metadata filename")
parser.add_argument('--hash_data', '-hd', type=str, default="data/hashdata", help="hash data (correspondence-based data) directory")
parser.add_argument('--pseudolabels', '-p', type=str, default="data/pseudolabels.csv", help="filename of pseudolabel csv file")
parser.add_argument('--crop_batch_size', '-cb', type=int, default=4, help="crop data batch size")
parser.add_argument('--epochs', '-e', type=int, default=10, help="epochs")
parser.add_argument('--lr', '-l', type=float, default=1e-4, help="learning rate")
parser.add_argument('--grad_acc_steps', '-g', type=int, default=1, help="gradient accumulation steps")
parser.add_argument('--n_crops', '-nc', type=int, default=2, help="number of crops to try per step")
parser.add_argument('--L0', '-L0', type=float, default=1.0, help="loss coefficient 0")
parser.add_argument('--L1', '-L1', type=float, default=1.0, help="loss coefficient 1")
parser.add_argument('--L2', '-L2', type=float, default=1.0, help="loss coefficient 2")
parser.add_argument('--L3', '-L3', type=float, default=1.0, help="loss coefficient 3")
parser.add_argument('--output', '-o', type=str, default="data/seg_ckpt", help="output checkpoint directory")
parser.add_argument('--output_losses', '-ol', type=str, default="data/seg_losses.json", help="output losses log")
return parser.parse_args()
def main():
args = get_opts()
assert os.path.exists(args.crop_metadata), f'Missing crop metadata file: {args.crop_metadata}'
assert os.path.exists(args.hash_metadata), f'Missing hash metadata file: {args.hash_metadata}'
assert os.path.exists(args.hash_data), f'Missing hash data directory: {args.hash_data}'
assert os.path.exists(args.clip_ft), f'Missing finetuned CLIP directory: {args.clip_ft}'
print("Loading models...")
processor = AutoProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
model = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined")
model.to('cuda')
model_orig = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined")
model_orig.to('cuda')
model_orig.eval()
for param in model_orig.parameters():
param.requires_grad = False
clip_proc = CLIPProcessor.from_pretrained(args.clip_ft)
clip = CLIPModel.from_pretrained(args.clip_ft)
clip.to('cuda')
clip.eval()
print("Models loaded")
print("Loading data...")
name2spl = pd.read_csv(args.pseudolabels,
usecols=['name', 'spl']).set_index('name').spl.to_dict()
name2bt = pd.read_csv(args.pseudolabels,
usecols=['name', 'building_type']).set_index('name').building_type.to_dict()
hash_ds = HashDS(args.hash_data, args.hash_metadata, name2spl, name2bt, neg_only=False)
# hash_ds_test = HashDS(args.hash_data, args.hash_metadata, name2spl, name2bt, neg_only=True, spl='test')
crop_ds = CropDS(args.data_dir, args.crop_metadata)
print("Setting up data loaders...")
CropMinibatch = namedtuple("Minibatch", "imgs labels coords coords_ gts")
def crop_collate(B):
inp = processor(
images=[r.img.crop(r.coords) for r in B],
text = [r.label for r in B],
padding=True,
return_tensors="pt").to('cuda')
with torch.no_grad():
gt = model_orig(**inp).decoder_output.logits.sigmoid()
return CropMinibatch(
[r.img for r in B],
[to_plur_aug(r.label) for r in B],
[r.coords for r in B],
[r.coords_ for r in B],
gt
)
crop_dl = DataLoader(
crop_ds,
batch_size=args.crop_batch_size,
num_workers=0,
collate_fn=crop_collate
)
hash_dl = DataLoader(
hash_ds,
shuffle=True,
num_workers=0,
collate_fn=lambda x: x[0]
)
print("Starting training...")
h, w = 352, 352
rsz = Resize((h, w), antialias=True)
loss_fn = nn.BCEWithLogitsLoss()
epochs = args.epochs
lr = args.lr
grad_acc_steps = args.grad_acc_steps
n_crops = args.n_crops
L0, L1, L2, L3 = args.L0, args.L1, args.L2, args.L3
optimizer = torch.optim.AdamW(model.parameters(), lr=lr)
losses = []
losses0 = []
losses1 = []
losses2 = []
losses3 = []
for i in trange(epochs):
for j, x in enumerate(tqdm(hash_dl)):
if j % grad_acc_steps == 0:
optimizer.step()
optimizer.zero_grad()
label_orig = x.label
label = to_plur_aug(x.label)
inp = processor(
images=x.img_,
text = label,
padding=True,
return_tensors="pt").to('cuda')
out = model(**inp).decoder_output.logits
gt = torch.tensor(x.gt_).to('cuda')
mask = torch.tensor(x.mask_).to('cuda')
mask_flat = mask.ravel() > 0
loss = loss_fn(out.ravel()[mask_flat], gt.ravel()[mask_flat])
loss = loss * L0
losses0.append(loss.item())
ent = loss_fn(out, out.sigmoid())
loss_e = ent.mean()
loss += loss_e * L1
losses1.append(loss_e.item())
c = search_crops(x.img_, label_orig, clip_proc, clip, n_crops=n_crops)
out_crop = out[c.y0:c.y1, c.x0:c.x1]
out_crop = rsz(out_crop[None])[0]
inp_crop = processor(
images=c.C,
text = label,
padding=True,
return_tensors="pt").to('cuda')
# out_crop = model(**inp_crop).decoder_output.logits
with torch.no_grad():
gt_crop = model_orig(**inp_crop).decoder_output.logits.sigmoid()
loss_c = loss_fn(out_crop.ravel(), gt_crop.ravel())
loss += loss_c * L2
losses2.append(loss_c.item())
R = next(iter(crop_dl))
new_gt = R.gts
inp_new = processor(
images=R.imgs,
text = R.labels,
padding=True,
return_tensors="pt").to('cuda')
out_new = model(**inp_new).decoder_output.logits
out_new_crops = []
for ON, co in zip(out_new, R.coords_):
x0, y0, x1, y1 = co
ON_c = ON[y0:y1, x0:x1]
ON_c = rsz(ON_c[None])[0]
out_new_crops.append(ON_c)
out_new_crops = torch.stack(out_new_crops) # (b, 352, 352)
loss_new = loss_fn(out_new_crops.ravel(), new_gt.ravel())
loss += loss_new * L3
losses3.append(loss_new.item())
loss.backward()
losses.append(loss.item())
optimizer.step()
optimizer.zero_grad()
print("Training finished")
print("Saving losses to:", args.output_losses)
if os.path.exists(args.output_losses):
print(f"Warning: {args.output_losses} exists; overwriting")
losses_obj = { 'losses': losses, 'losses0': losses0, 'losses1': losses1, 'losses2': losses2, 'losses3': losses3 }
with open(args.output_losses, 'w') as f:
json.dump(losses_obj, f, indent=4)
print("Saving trained model to:", args.output)
if os.path.exists(args.output):
print(f"Warning: {args.output} exists; overwriting")
model.save_pretrained(args.output)
print("done")
if __name__ == "__main__":
main()