-
Notifications
You must be signed in to change notification settings - Fork 5
/
show_wti.py
201 lines (163 loc) · 8.06 KB
/
show_wti.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
from os.path import join,dirname,abspath
import matplotlib
import matplotlib.pyplot as plt
from matplotlib.offsetbox import OffsetImage, AnnotationBbox
import torch
from tvr.models.modeling import DRL
from tvr.models.tokenization_clip import SimpleTokenizer
import numpy as np
from PIL import Image
from decord import VideoReader, cpu
from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize, InterpolationMode
def show_single_pair(t_feat, v_feat, t_tokens, v_frames, save_fig=None, dense_text_fc_weight=None,
dense_vision_fc_weight=None, i_min=None, i_max=None):
with torch.no_grad():
if dense_text_fc_weight:
t_weight = torch.softmax(dense_text_fc_weight(t_feat).squeeze(1), dim=0)
else:
t_weight = torch.ones(t_feat.shape[0]) / t_feat.shape[0]
if dense_vision_fc_weight:
v_weight = torch.softmax(dense_vision_fc_weight(v_feat).squeeze(1), dim=0)
else:
v_weight = torch.ones(v_feat.shape[0]) / v_feat.shape[0]
t_feat = t_feat / t_feat.norm(dim=-1, keepdim=True)
v_feat = v_feat / v_feat.norm(dim=-1, keepdim=True)
corr = torch.einsum('tc,vc->tv', [t_feat, v_feat]) # txv
t_max = corr.max(dim=1)[0]
v_max = corr.max(dim=0)[0]
t2v = (t_weight.T * t_max).sum()
v2t = (v_weight * v_max).sum()
n_t = len(t_feat)
n_v = len(v_feat)
fig = plt.figure(figsize=(20, 20))
ax = fig.add_subplot(1, 1, 1)
cmap = plt.cm.rainbow
if (i_min is not None) and (i_max is not None):
norm = matplotlib.colors.Normalize(vmin=i_min, vmax=i_max)
else:
norm = matplotlib.colors.Normalize(vmin=corr.min(), vmax=corr.max())
t_w_norm = matplotlib.colors.Normalize(vmin=t_weight[1:-1].min(), vmax=t_weight[1:-1].max())
v_w_norm = matplotlib.colors.Normalize(vmin=v_weight.min(), vmax=v_weight.max())
t_loc = [i / n_t for i in range(n_t)]
# text_len = len(' '.join(t_tokens))
# cum_len = np.cumsum([len(_t) for _t in t_tokens]).tolist()
# t_loc = [(i+cum_len[i])/text_len for i in range(n_t)]
for i in range(n_t):
plt.text(t_loc[i], 1, t_tokens[i], fontsize=30, horizontalalignment='center', verticalalignment='bottom')
plt.text(t_loc[i], 1 + 0.05, '%.3f' % (t_max[i].item()), fontsize=25, horizontalalignment='center',
verticalalignment='bottom')
plt.text(t_loc[i], 1 + 0.08, 'x', fontsize=35, horizontalalignment='center',
verticalalignment='bottom')
ax.scatter(t_loc[i], 1.04, s=1000, c=[cmap(norm(t_max[i].item()))])
plt.text(t_loc[i], 1 + 0.11, '%.3f' % (t_weight[i].item()), fontsize=25, horizontalalignment='center',
verticalalignment='bottom')
ax.scatter(t_loc[i], 1.15, s=1000, c=[cmap(t_w_norm(t_weight[i].item()))])
plt.text(0.5, 1 + 0.17, '(%.3f + %.3f)/2 = %.3f' % (t2v.item(), v2t.item(), (t2v.item() + v2t.item()) / 2.0),
fontsize=35, horizontalalignment='center',
verticalalignment='bottom')
for j in range(n_v):
ax.add_artist( # ax can be added image as artist.
AnnotationBbox(
OffsetImage(v_frames[j])
, (j / n_v, 0.5)
, frameon=False
, box_alignment=(0.5, 1)
)
)
plt.text(j / n_v, 0.5 - 0.12, '%.3f' % (v_max[j].item()), fontsize=25, horizontalalignment='center',
verticalalignment='top')
plt.text(j / n_v, 0.5 - 0.15, "x", fontsize=35, horizontalalignment='center',
verticalalignment='top')
ax.scatter(j / n_v, 0.4, s=1000, c=[cmap(norm(v_max[j].item()))])
plt.text(j / n_v, 0.5 - 0.18, '%.3f' % (v_weight[j].item()), fontsize=25, horizontalalignment='center',
verticalalignment='top')
ax.scatter(j / n_v, 0.5 - 0.22, s=1000, c=[cmap(v_w_norm(v_weight[j].item()))])
for i in range(n_t):
for j in range(n_v):
_norm = norm(corr[i, j])
plt.plot([t_loc[i], j / n_v], [1, 0.5], color=cmap(_norm), linewidth=_norm * 4, alpha=_norm)
plt.axis('off')
plt.axis('equal')
plt.tight_layout()
if save_fig:
fig.savefig(save_fig, dpi=100)
plt.show()
plt.close('all')
def preprocess_video(video_path, image_resolution=224, max_frames=12):
transform = Compose([
Resize(image_resolution, interpolation=InterpolationMode.BICUBIC),
CenterCrop(image_resolution),
lambda image: image.convert("RGB"),
ToTensor(),
Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
# Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
video_mask = torch.zeros(max_frames, dtype=torch.long)
max_video_length = 0
# T x 3 x H x W
video = torch.zeros((max_frames, 3, image_resolution, image_resolution), dtype=torch.float32)
vreader = VideoReader(video_path, ctx=cpu(0))
fps = int(vreader.get_avg_fps())
f_start = 0
f_end = int(len(vreader) - 1)
num_frames = f_end - f_start + 1
if num_frames > 0:
# T x 3 x H x W
all_pos = list(range(f_start, f_end + 1, fps))
if len(all_pos) > max_frames:
sample_pos = [all_pos[_] for _ in np.linspace(0, len(all_pos) - 1, num=max_frames, dtype=int)]
else:
sample_pos = all_pos
video_raw = [Image.fromarray(f) for f in vreader.get_batch(sample_pos).asnumpy()]
patch_images = torch.stack([transform(img) for img in video_raw])
slice_len = patch_images.shape[0]
max_video_length = slice_len
if slice_len < 1:
pass
else:
video[:slice_len, ...] = patch_images
video_mask[:max_video_length] = torch.tensor([1] * max_video_length)
return video_raw, video, video_mask
def preprocess_text(text_input, max_words=32):
tokenizer = SimpleTokenizer()
words = tokenizer.tokenize(text_input)
SPECIAL_TOKEN = {"CLS_TOKEN": "<|startoftext|>", "SEP_TOKEN": "<|endoftext|>",
"MASK_TOKEN": "[MASK]", "UNK_TOKEN": "[UNK]", "PAD_TOKEN": "[PAD]"}
words = [SPECIAL_TOKEN["CLS_TOKEN"]] + words
total_length_with_CLS = max_words - 1
if len(words) > total_length_with_CLS:
words = words[:total_length_with_CLS]
words = words + [SPECIAL_TOKEN["SEP_TOKEN"]]
input_ids = tokenizer.convert_tokens_to_ids(words)
t_tokens = [tokenizer.decode([_t_id]) for _t_id in input_ids]
t_tokens = [t.replace('<|startoftext|>', '[CLS]').replace('<|endoftext|>', '[SEP]').strip() for t in t_tokens]
return t_tokens, torch.tensor(input_ids)
def main():
# init model
device = "cuda" if torch.cuda.is_available() else "cpu"
from collections import namedtuple
Config = namedtuple('Config', ["base_encoder", "agg_module", "interaction", "wti_arch"])
model = DRL(Config("ViT-B/32", "seqTransf", "wti", 2)).to(device).eval()
# please finetune a wti model first
finetune_path = join(dirname(abspath(__file__)), '../ckpts/ckpt_msrvtt_wti/pytorch_model.bin.4')
model.load_state_dict(torch.load(finetune_path, map_location="cpu"))
video_path = "video7500.mp4"
text_input = "a soccer team walking out on the field"
# preprocess text and video
t_tokens, text = preprocess_text(text_input, 32)
text = text.to(device)
text_mask = (text > -1).type(torch.long).to(device)
raw_video_data, video, video_mask = preprocess_video(video_path)
video = video.unsqueeze(0).to(device)
video_mask = video_mask.unsqueeze(0).to(device)
with torch.no_grad():
text_feat = model.get_text_feat(text, text_mask)
video_feat = model.get_video_feat(video, video_mask, False)
t2v, v2t, _ = model.get_similarity_logits(text_feat, video_feat, text_mask, video_mask)
print(f"t2v similarity: {t2v.item()} and v2t similarity: {v2t.item()}")
v_frames = [img.resize([112, 112]) for img in raw_video_data]
show_single_pair(text_feat.squeeze(0), video_feat.squeeze(0), t_tokens, v_frames, 'wti.jpg',
model.text_weight_fc, model.video_weight_fc)
return True
if __name__ == "__main__":
main()