-
Notifications
You must be signed in to change notification settings - Fork 10
/
retrieve_captions.py
177 lines (144 loc) · 6.06 KB
/
retrieve_captions.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
import argparse
from pathlib import Path
import h5py
import numpy as np
import math
import faiss
from tqdm import tqdm
import shutil
import torch
from torch.utils.data import DataLoader
from torchvision import transforms as T
from pytorch_lightning import Trainer, LightningModule, seed_everything
from transformers import CLIPModel, CLIPProcessor
import sys
sys.path.append('.')
from dataset import CocoImageCrops, collate_crops
class CaptionRetriever(LightningModule):
def __init__(self, caption_db, save_dir, k):
super().__init__()
self.save_dir = Path(save_dir)
self.k = k
self.keys, self.features, self.text = self.load_caption_db(caption_db)
self.index = self.build_index(idx_file=self.save_dir/"faiss.index")
self.clip = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
@staticmethod
def load_caption_db(caption_db):
print("Loading caption db")
keys, features, text = [], [], []
with h5py.File(caption_db, "r") as f:
for i in tqdm(range(len(f))):
keys_i = f[f"{i}/keys"][:]
features_i = f[f"{i}/features"][:]
text_i = [str(x, "utf-8") for x in f[f"{i}/captions"][:]]
keys.append(keys_i)
features.append(features_i)
text.extend(text_i)
keys = np.concatenate(keys)
features = np.concatenate(features)
return keys, features, text
def build_index(self, idx_file):
print("Building db index")
n, d = self.keys.shape
K = round(8 * math.sqrt(n))
index = faiss.index_factory(d, f"IVF{K},Flat", faiss.METRIC_INNER_PRODUCT)
assert not index.is_trained
index.train(self.keys)
assert index.is_trained
index.add(self.keys)
index.nprobe = max(1, K//10)
faiss.write_index(index, str(idx_file))
return index
def search(self, images, topk):
features = self.clip.vision_model(pixel_values=images)[1]
query = self.clip.visual_projection(features)
query = query / query.norm(dim=-1, keepdim=True)
D, I = self.index.search(query.detach().cpu().numpy(), topk)
return D, I
def test_step(self, batch, batch_idx):
orig_imgs, five_imgs, nine_imgs, gt_caps, ids = batch
N = len(orig_imgs)
with h5py.File(self.save_dir/"txt_ctx.hdf5", "a") as f:
D_o, I_o = self.search(orig_imgs, topk=self.k) # N x self.k
D_f, I_f = self.search(torch.flatten(five_imgs, end_dim=1), topk=self.k) # N*5 x self.k
D_f, I_f = D_f.reshape(N, 5, self.k), I_f.reshape(N, 5, self.k)
D_n, I_n = self.search(torch.flatten(nine_imgs, end_dim=1), topk=self.k) # N*9 x self.k
D_n, I_n = D_n.reshape(N, 9, self.k), I_n.reshape(N, 9, self.k)
for i in range(N):
g1 = f.create_group(str(int(ids[i])))
texts = [self.text[j] for j in I_o[i]]
features = self.features[I_o[i]]
scores = D_o[i]
g2 = g1.create_group("whole")
g2.create_dataset("features", data=features)
g2.create_dataset("scores", data=scores)
g2.create_dataset("texts", data=texts)
texts = [
[
self.text[I_f[i, j, k]]
for k in range(self.k)
]
for j in range(5)
]
features = self.features[I_f[i].flatten()].reshape((5, self.k, -1))
scores = D_f[i]
g3 = g1.create_group("five")
g3.create_dataset("features", data=features)
g3.create_dataset("scores", data=scores)
g3.create_dataset("texts", data=texts)
texts = [
[
self.text[I_n[i, j, k]]
for k in range(self.k)
]
for j in range(9)
]
features = self.features[I_n[i].flatten()].reshape((9, self.k, -1))
scores = D_n[i]
g4 = g1.create_group("nine")
g4.create_dataset("features", data=features)
g4.create_dataset("scores", data=scores)
g4.create_dataset("texts", data=texts)
def build_ctx_caps(args):
transform = T.Compose([
CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32").feature_extractor,
lambda x: torch.FloatTensor(x["pixel_values"][0]),
])
dset = CocoImageCrops(args.dataset_root/"annotations", args.dataset_root, transform)
dloader = DataLoader(
dataset=dset,
batch_size=args.batch_size,
shuffle=False,
drop_last=False,
num_workers=args.num_workers,
collate_fn=collate_crops
)
cap_retr = CaptionRetriever(
caption_db=args.caption_db,
save_dir=args.save_dir,
k=args.k
)
trainer = Trainer(
gpus=[args.device, ],
deterministic=True,
benchmark=False,
default_root_dir=args.save_dir
)
trainer.test(cap_retr, dloader)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='Retrieve captions')
parser.add_argument('--device', type=int, default=0)
parser.add_argument('--exp_name', type=str, default='retrieved_captions')
parser.add_argument('--dataset_root', type=str, default='datasets/coco_captions')
parser.add_argument('--caption_db', type=str, default='outputs/captions_db/caption_db.hdf5')
parser.add_argument('--k', type=int, default=16)
parser.add_argument('--batch_size', type=int, default=32)
parser.add_argument('--num_workers', type=int, default=12)
args = parser.parse_args()
args.dataset_root = Path(args.dataset_root)
setattr(args, "save_dir", Path("outputs")/args.exp_name)
shutil.rmtree(args.save_dir, ignore_errors=True)
args.save_dir.mkdir(parents=True, exist_ok=True)
print(args)
seed_everything(1, workers=True)
build_ctx_caps(args)