-
Notifications
You must be signed in to change notification settings - Fork 0
/
clip_embeddings.py
275 lines (241 loc) · 9.27 KB
/
clip_embeddings.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
import argparse
import json
import os
import pickle
import sys
from collections import defaultdict
from copy import deepcopy
from multiprocessing.dummy import Pool
from pathlib import Path
import clip
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
from PIL import Image
from tqdm import tqdm
batch_size = 64
def embed_text(sentences, model, device):
tokenized = clip.tokenize(sentences, truncate=True).to(device, non_blocking=True)
with torch.no_grad():
text_embeddings = model.encode_text(tokenized)
text_embeddings /= text_embeddings.norm(dim=-1, keepdim=True)
return text_embeddings
def get_text_embeddings(anns, model, device, gpu_id):
texts, ids = zip(*anns)
texts, ids = list(texts), list(ids)
text_embeddings = {}
for i in tqdm(
range(0, len(texts), batch_size),
total=len(range(0, len(texts), batch_size)),
desc=f"get_text_emb_{gpu_id}",
position=gpu_id,
):
batch_ids = ids[i : i + batch_size]
embed = embed_text(texts[i : i + batch_size], model, device).detach().cpu().numpy()
for j in range(len(batch_ids)):
text_embeddings[batch_ids[j]] = embed[j]
return text_embeddings
def embed_image(images, model):
with torch.no_grad():
image_embedddings = model.encode_image(images)
image_embedddings /= image_embedddings.norm(dim=-1, keepdim=True)
return image_embedddings
def get_valid_images(ids, image_paths, preprocess):
invalid_ids = []
images = []
for i, image_path in enumerate(image_paths):
try:
image = Image.open(image_path)
image = preprocess(image).unsqueeze(0)
images.append(image)
except Exception as e:
print(e, ids[i])
invalid_ids.append(ids[i])
return images, set(invalid_ids)
def get_image_embeddings(anns, model, preprocess, device, gpu_id) -> dict[str, np.ndarray]:
image_paths, ids = zip(*anns)
image_paths, ids = list(image_paths), list(ids)
image_embeddings = {}
for i in tqdm(
range(0, len(image_paths), batch_size),
total=len(list(range(0, len(image_paths), batch_size))),
position=gpu_id,
desc=f"get_img_emb_{gpu_id}",
):
try:
batch_ids = ids[i : i + batch_size]
batch_images, invalid_batch_ids = get_valid_images(
batch_ids, image_paths[i : i + batch_size], preprocess
)
images_tensor = torch.vstack(batch_images).to(device, non_blocking=True)
embed = embed_image(images_tensor, model).detach().cpu().numpy()
for j in range(embed.shape[0]):
if batch_ids[j] not in invalid_batch_ids:
image_embeddings[batch_ids[j]] = embed[j]
except Exception as e:
print(e)
return image_embeddings
def load_pretrained(args, device: torch.device, model: nn.Module, checkpoint_path: Path):
if args.gpu is None:
checkpoint = torch.load(checkpoint_path)
else:
# Map model to be loaded to specified single gpu.
checkpoint = torch.load(checkpoint_path, map_location=device)
sd = checkpoint["state_dict"]
if next(iter(sd.items()))[0].startswith("module"):
sd = {k[len("module.") :]: v for k, v in sd.items()}
model.load_state_dict(sd)
print(f"Loaded checkpoint '{checkpoint_path}' (epoch {checkpoint['epoch']}), device: {device}")
def get_text_anns(args):
"""
Return a list of 2-tuples. First tuple item is sentence text, second is sentence text (or a
unique id for the sentence if one exists). Return value looks like::
[("cars are driving on the left", "cars are driving on the left), ...]
"""
if args.dataset == "geoframes":
args.data_file_name = f"geoframes_clip_samples_fixed_window_{args.split}"
data = pd.read_csv(
f"/shared/gbiamby/geo/captions/clip_samples_fixed_window_{args.split}_openclip.csv"
)
texts = [p for p in set(data.sentence.values)]
ids = texts
if args.dataset == "gptj_clues":
args.data_file_name = f"gptj_clues"
data = json.load(
open("/shared/g-luo/geoguessr/data/data/guidebook/kb/v3/cleaned_clues.json")
)
texts = list(set([t["text"] for t in data["clues"]]))
ids = texts
else:
raise NotImplementedError()
# De-duplicate by "id", i.e., sentence text:
id_to_path = {}
for path, _id in zip(texts, ids):
id_to_path[_id] = path
print(f"Loaded dataset {args.dataset}-{args.data_type} with {len(ids)} items")
return list(zip(list(id_to_path.values()), list(id_to_path.keys())))
def get_image_anns(args):
"""
Return a list of 2-tuples. First tuple item is full image path, second is a unique identifier
for the image (which can also just be the image path)
Return value looks like::
[("/shared/you/geo/img_01.jpg", "/shared/you/geo/img_01.jpg"), ...]
"""
if args.dataset in ["flickr", "yfcc", "yfcc100m"]:
args.data_file_name = "placing2014_reference"
img_base_path = Path("/shared/group/yfcc100m/placing2014/data/images")
dataset = json.load(
open(f"/shared/g-luo/geoguessr/models/clip_zs/{args.data_file_name}.json")
)
image_paths = [img_base_path / ann["IMG_ID"] for ann in dataset]
ids = [ann["hash"] for ann in dataset]
elif args.dataset == "geoframes":
args.data_file_name = f"geoframes_clip_samples_fixed_window_{args.split}"
data = pd.read_csv(
f"/shared/gbiamby/geo/captions/clip_samples_fixed_window_{args.split}_openclip.csv"
)
image_paths = [p for p in set(data.file_path.values)]
ids = image_paths
elif args.dataset == "streetview":
raise NotImplementedError()
else:
raise NotImplementedError()
# De-duplicate by "id":
id_to_path = {}
for path, _id in zip(image_paths, ids):
id_to_path[_id] = path
print(f"Loaded dataset {args.dataset} with {len(ids)} items")
return list(zip(list(id_to_path.values()), list(id_to_path.keys())))
def get_embeddings_worker(args, gpu_id: int, anns):
args.gpu = gpu_id
print(f"gpu_id: {args.gpu}, num_items: {len(anns)}")
device = torch.device(f"cuda:{gpu_id}")
model, preprocess = clip.load(args.model_name, device=device)
if "checkpoint" in args and args.checkpoint is not None and args.checkpoint:
load_pretrained(args, device, model, args.checkpoint)
model = model.to(device)
if args.data_type in ["img", "image"]:
embeddings = get_image_embeddings(anns, model, preprocess, device, gpu_id)
else:
embeddings = get_text_embeddings(anns, model, device, gpu_id)
return embeddings
def get_embeddings(args):
if args.data_type in ["img", "image"]:
anns = get_image_anns(args)
else:
anns = get_text_anns(args)
all_embeddings = {}
# Divide the work into roughly equal chunks across num_gpus:
gpu_to_anns = defaultdict(lambda: [])
for i, ann in enumerate(anns):
gpu_id = i % args.num_gpu
gpu_to_anns[gpu_id].append(ann)
worker_args = ((deepcopy(args), gpu_id, _anns) for gpu_id, _anns in gpu_to_anns.items())
# Compute
with Pool(processes=args.num_gpu) as pool:
embeddings = pool.starmap(get_embeddings_worker, worker_args)
# Combine results from each thread into a single dict:
for emb in embeddings:
all_embeddings.update(emb)
print("Total embeddings: ", len(all_embeddings))
save_file = (
Path("/shared/gbiamby/geo/models/clip_ft")
/ (args.model_name.lower().replace("/", "") + "_zeroshot")
/ f"{args.data_file_name}_{args.data_type}.pkl"
)
print("Saving results to: ", save_file)
save_file.parent.mkdir(exist_ok=True, parents=True)
pickle.dump(all_embeddings, open(save_file, "wb"))
def main(args):
get_embeddings(args)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"--model_name",
type=str,
default="ViT-B/32",
help="CLIP model name",
)
parser.add_argument(
"--dataset",
type=str,
default="flickr",
help="Name of dataset to generate embeddings for.",
)
parser.add_argument(
"--split",
type=str,
default="val",
help="(Optional) Split (train/val/test) within the dataset, if applicable, to generate emb's for.",
)
parser.add_argument(
"--data_type",
type=str,
default="image",
help="'image' or 'text'",
)
parser.add_argument(
"--gpu",
type=int,
default=None,
help="Specify a single GPU to run the code on for debugging."
"Leave at None to use all available GPUs.",
)
parser.add_argument(
"--num_gpu",
type=int,
default=1,
help="Num gpu threads to use",
)
parser.add_argument(
"--checkpoint",
type=Path,
# default=Path(
# "/home/gbiamby/proj/im2gps_kb/lib/open_clip/logs/lr=1e-06_wd=0.1_agg=True_model=ViT-B/32_batchsize=96_workers=4_date=2022-03-15-04-33-55/checkpoints/epoch_2.pt"
# ),
default=None,
help="Path to model checkpoint. If None then loads off the shelf pretrained CLIP",
)
args = parser.parse_args()
main(args)