forked from dandelin/ViLT
-
Notifications
You must be signed in to change notification settings - Fork 0
/
demo_vqa.py
269 lines (206 loc) · 9.34 KB
/
demo_vqa.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
# import gradio as gr
import torch
import copy
import time
import requests
import io
import numpy as np
import re
import json
import urllib.request
from scipy.sparse.linalg import eigsh
from scipy.sparse import diags
import torch.nn.functional as F
import matplotlib.pyplot as plt
from pymatting.util.util import row_sum
# import ipdb
from PIL import Image
from vilt.config import ex
from vilt.modules import ViLTransformerSS
from vilt.transforms import pixelbert_transform
from vilt.datamodules.datamodule_base import get_pretrained_tokenizer
import sys
import cv2
IMAGE_SIZE = 224
@ex.automain
def main(_config):
_config = copy.deepcopy(_config)
loss_names = {
"itm": 0,
"mlm": 0,
"mpp": 0,
"vqa": 1,
"imgcls": 0,
"nlvr2": 0,
"irtr": 0,
"arc": 0,
}
tokenizer = get_pretrained_tokenizer(_config["tokenizer"])
with urllib.request.urlopen(
"https://github.com/dandelin/ViLT/releases/download/200k/vqa_dict.json"
) as url:
id2ans = json.loads(url.read().decode())
_config.update(
{
"loss_names": loss_names,
}
)
model = ViLTransformerSS(_config)
model.setup("test")
model.eval()
device = "cuda:0" if _config["num_gpus"] > 0 else "cpu"
model.to(device)
def infer(url, text):
try:
if "http" in url:
res = requests.get(url)
image = Image.open(io.BytesIO(res.content)).convert("RGB")
else:
image = Image.open(url)
# orig_shape = np.array(image).shape
img = pixelbert_transform(size=IMAGE_SIZE)(image)
# print("pixelberted image shape: {}".format(img.shape))
img = img.unsqueeze(0).to(device)
except:
return False
batch = {"text": [text], "image": [img]}
# with torch.no_grad():
encoded = tokenizer(batch["text"])
text_tokens = tokenizer.tokenize(batch["text"][0])
print(text_tokens)
batch["text_ids"] = torch.tensor(encoded["input_ids"]).to(device)
batch["text_labels"] = torch.tensor(encoded["input_ids"]).to(device)
batch["text_masks"] = torch.tensor(encoded["attention_mask"]).to(device)
ret = model.infer(batch)
vqa_logits = model.vqa_classifier(ret["cls_feats"])
answer = id2ans[str(vqa_logits.argmax().item())]
# print("cls features shape: {}".format(ret['image_feats'][0].shape))
# print(np.array(image), answer)
# return [np.array(image), answer], ret['image_feats'][0], orig_shape
# return [np.array(image), answer], ret['image_feats'][0], img
return answer, ret['image_feats'][0], ret['text_feats'][0], img, text_tokens
# question = "What is the colour of her pants?"
# question = "Does he have earphones?"
# question = "Is he wearing glasses?"
# question = "Is there an owl?"
# question = "Is the man swimming?"
# question = "What animals are shown?"
# question = "What animal hat did she wear?"
# question = "What is the color of the flowers?"
# question = "How many windows are there?"
# question = "Is there a laptop?"
question = "What is the girl in white dress doing?"
# result, image_feats, text_feats, image, text_tokens = infer('easy_test.jpg', 'What is the colour of the ball?')
# result, image_feats, text_feats, image, text_tokens = infer('images/weird_dj.jpg', question)
# result, image_feats, text_feats, image, text_tokens = infer('images/clock_owl.jpg', question)
# result, image_feats, text_feats, image, text_tokens = infer('images/swim.jpg', question)
# result, image_feats, text_feats, image, text_tokens = infer('images/nee-sama.jpeg', question)
# result, image_feats, text_feats, image, text_tokens = infer('../../nii_depressed.jpg', question)
# result, image_feats, text_feats, image, text_tokens = infer('images/skii.jpg', question)
# result, image_feats, text_feats, image, text_tokens = infer('images/cows.jpg', question)
# result, image_feats, text_feats, image, text_tokens = infer('images/shiv.jpg', question)
result, image_feats, text_feats, image, text_tokens = infer('images/demon.jpg', question)
# result, image_feats, text_feats, image, text_tokens = infer("https://s3.geograph.org.uk/geophotos/06/21/24/6212487_1cca7f3f_1024x1024.jpg", question)
PATCH_SIZE = 32
print(f"QUESTION: {question}")
print("Answer: {}".format(result))
pp_img_shape = (image.shape[2], image.shape[3])
# print("Feature shape: {} | orig_img shape: {}".format(feats.shape, pp_img_shape))
# sys.exit()
def get_eigen (feats):
feats = F.normalize(feats, p = 2, dim = -1)
W_feat = (feats @ feats.T)
W_feat = (W_feat * (W_feat > 0))
W_feat = W_feat / W_feat.max()
W_feat = W_feat.cpu().detach().numpy()
def get_diagonal (W):
D = row_sum(W)
D[D < 1e-12] = 1.0 # Prevent division by zero.
D = diags(D)
return D
D = np.array(get_diagonal(W_feat).todense())
L = D - W_feat
# print(L)
# print("here")
# L[ np.isnan(L) ] = 0
# L[ L == np.inf ] = 0
try:
eigenvalues, eigenvectors = eigsh(L, k = 5, which = 'LM', sigma = -0.5, M = D)
except:
try:
eigenvalues, eigenvectors = eigsh(L, k = 5, which = 'SM', sigma = -0.5, M = D)
except:
eigenvalues, eigenvectors = eigsh(L, k = 5, which = 'LM', M = D)
eigenvalues, eigenvectors = torch.from_numpy(eigenvalues), torch.from_numpy(eigenvectors.T).float()
n_tuple = torch.kthvalue(eigenvalues.real, 2)
fev_idx = n_tuple.indices
return eigenvectors[fev_idx]
# return eigenvectors[1]
text_relevance = torch.abs(get_eigen(text_feats[1:-1]))
# text_relevance = get_eigen(text_feats[1:-1])
# dim = int(image_relevance.numel() ** 0.5)
image_relevance = torch.abs(get_eigen(image_feats[1:, :]))
image_relevance = image_relevance.reshape(1, 1, pp_img_shape[0]//PATCH_SIZE, pp_img_shape[1]//PATCH_SIZE)
image_relevance = torch.nn.functional.interpolate(image_relevance, size=pp_img_shape, mode='bilinear')
image_relevance = image_relevance.reshape(pp_img_shape[0], pp_img_shape[1]).cpu().numpy()
# image_relevance = image_relevance.reshape(pp_img_shape[0]//PATCH_SIZE, pp_img_shape[1]//PATCH_SIZE).cpu().numpy()
image_relevance = (image_relevance - image_relevance.min()) / (image_relevance.max() - image_relevance.min())
def show_cam_on_image(img, mask):
heatmap = cv2.applyColorMap(np.uint8(255 * mask), cv2.COLORMAP_JET)
heatmap = np.float32(heatmap) / 255
cam = heatmap + np.float32(img)
cam = cam / np.max(cam)
return cam
image = image[0].permute(1, 2, 0).cpu().numpy()
image = (image - image.min()) / (image.max() - image.min())
vis = show_cam_on_image(image, image_relevance)
vis = np.uint8(255 * vis)
vis = cv2.cvtColor(np.array(vis), cv2.COLOR_RGB2BGR)
fig, axs = plt.subplots(ncols=2, figsize=(20, 5))
axs[0].imshow(vis)
axs[0].axis('off')
axs[0].set_title('Spectral Approach Image Relevance')
ti = axs[1].imshow(text_relevance.unsqueeze(dim = 0).numpy())
axs[1].set_title("Spectral Approach Word Impotance")
plt.sca(axs[1])
plt.xticks(np.arange(len(text_tokens)), text_tokens)
# plt.sca(axs[1])
plt.colorbar(ti, orientation = "horizontal", ax = axs[1])
# plt.imshow(vis)
plt.show()
# feats = feats.unsqueeze(dim = 0)
# print("Answer: {}".format(result[1]))
# print("Feature shape: {} | orig_shape: {}".format(feats.shape, orig_shape))
# feats = F.normalize(feats, p = 2, dim = -1)
# # H_patch, W_patch = orig_shape[0]//PATCH_SIZE, orig_shape[1]//PATCH_SIZE
# # print("H_patch, W_patch: {}, {}".format(H_patch, W_patch))
# # H_pad_lr, W_pad_lr = H_patch, W_patch
# H_patch, W_patch = orig_shape[0] // PATCH_SIZE, orig_shape[1] // PATCH_SIZE
# H_pad, W_pad = H_patch * PATCH_SIZE, W_patch * PATCH_SIZE
# H_pad_lr, W_pad_lr = H_pad // PATCH_SIZE, W_pad // PATCH_SIZE
# if (H_patch, W_patch) != (H_pad_lr, W_pad_lr):
# feats = F.interpolate(
# feats.T.reshape(1, -1, H_patch, W_patch),
# size=(H_pad_lr, W_pad_lr), mode='bilinear', align_corners=False
# ).reshape(-1, H_pad_lr * W_pad_lr).T
# # feats = F.interpolate(
# # feats.T.reshape(1, -1, H_patch, W_patch),
# # size=(H_pad_lr, W_pad_lr), mode='bilinear', align_corners=False
# # ).reshape(-1, H_pad_lr * W_pad_lr).T
# W_feat = (feats @ feats.T)
# W_feat = (W_feat * (W_feat > 0))
# W_feat = W_feat / W_feat.max() # NOTE: If features are normalized, this naturally does nothing
# W_feat = W_feat.cpu().numpy()
# def get_diagonal (W):
# D = row_sum(W)
# D[D < 1e-12] = 1.0 # Prevent division by zero.
# D = diags(D)
# return D
# D = np.array(get_diagonal(W_feat).todense())
# L = D - W_feat
# eigenvalues, eigenvectors = eigsh(L, k = 5, which = 'LM', sigma = 0)
# eigenvalues, eigenvectors = torch.from_numpy(eigenvalues), torch.from_numpy(eigenvectors.T).float()
# # print(eigenvectors[1].shape)
# fiedel_ev = eigenvectors[1].numpy().reshape(H_patch, W_patch)
# plt.imshow(fiedel_ev)
# plt.show()