/
streamlit_demo.py
162 lines (125 loc) · 5.61 KB
/
streamlit_demo.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
import os
import torch
import numpy as np
import streamlit as st
from PIL import Image
import models
import networks
torch.set_grad_enabled(False)
st.set_page_config(layout="wide")
st.session_state.old_buffer = None
@st.cache(hash_funcs={torch.nn.parameter.Parameter: lambda _: None, torch.Tensor: lambda _: None}, allow_output_mutation=True)
def load_model(model_type, mapper_ckpt):
model = models.create_model(model_type).eval().to(device)
real_encoder, synth_encoder = networks.create_encoders(model, 'linear')
real_mapper = real_encoder.mapper
synth_mapper = synth_encoder.mapper
real_mapper.eval().to(device)
synth_mapper.eval().to(device)
all_dicts = torch.load(mapper_ckpt, map_location=device)
real_mapper.load_state_dict(all_dicts['real_mapper'])
synth_mapper.load_state_dict(all_dicts['synth_mapper'])
return model, real_mapper, synth_mapper
@st.cache(hash_funcs={torch.nn.parameter.Parameter: lambda _: None, torch.Tensor: lambda _: None})
def load_laion_subset(model_type, real_mapper, device):
laion_dict = torch.load(
f'feats/{model_type}/laion_subset.pth', map_location=device)
laion_paths = [n.replace('png', 'jpg') for n in laion_dict['paths']]
data_feat = laion_dict['feats']
data_feat_lin = real_mapper(data_feat)
data_feat_lin = normalize_feat(data_feat_lin)
return data_feat_lin, laion_paths
def normalize_feat(feat):
return torch.nn.functional.normalize(feat, dim=-1, p=2)
def load_img(buffer):
image = np.array(Image.open(buffer).convert("RGB"))
return image
def resize_img(img, scale=None, shortside=None, resample=Image.BICUBIC):
# if scale is specified, use it
# otherwise, resize to short side
img_pil = Image.fromarray(img)
H, W = img.shape[0], img.shape[1]
assert(scale is not None or shortside is not None)
if scale is None: # use size
minHW = min(H, W)
scale = 1. * shortside / minHW
Hnew, Wnew = int(H * scale), int(W * scale)
img_pil = img_pil.resize((Wnew, Hnew), resample=resample)
return np.array(img_pil)
# Settings
with st.sidebar:
st.title('Attribution demo')
device = 'cuda'
laion_folder = 'dataset/laion_subset_jpeg'
st.header('Model')
method_types = ['clip (finetuned)', 'dino (finetuned)']
method_type = st.select_slider(
'Attribution features', options=method_types, value='clip (finetuned)')
model_type = method_type.split(' ')[0]
img_file_buffer = st.file_uploader(
"Upload an image", type=["png", "jpg", "jpeg", 'webp'])
if img_file_buffer is not None and img_file_buffer != st.session_state.old_buffer:
st.session_state.old_buffer = img_file_buffer
img = load_img(img_file_buffer)
st.session_state.img = img
else:
img = load_img('images/test.png')
st.header('Display')
N = st.select_slider('Number retrieved', options=(
10**np.linspace(1, 3, 101)).astype('int'), value=100)
num_columns = st.slider('Width', min_value=1, max_value=20, value=5)
res = st.select_slider(
'Resolution', options=np.arange(64, 1024+32, 32), value=256)
pad = st.slider('Padding', min_value=0, max_value=64, value=8)
display_per = 0.007
mapper_ckpt = f'weights/mapper/{model_type}_tuned.pth'
calib_ckpt = f'weights/calibrator/{model_type}_tuned.pth'
calibrator = networks.Calibrator()
calibrator.load_state_dict(torch.load(calib_ckpt))
st.write(
model_type, f'inv temp: {calibrator.inv_tau.item():.4f}; bias: {calibrator.get_bias().item():.4f}')
# base model
model, real_mapper, synth_mapper = load_model(model_type, mapper_ckpt)
# load training image features and paths
data_feat_lin, laion_paths = load_laion_subset(model_type, real_mapper, device)
img_disp = resize_img(img, shortside=res)
st.image(resize_img(img, shortside=res), width=img_disp.shape[1])
img_query = model.preprocess(Image.fromarray(img)).unsqueeze(0).to(device)
query = model(img_query).float()
query_feat_lin = synth_mapper(query)
query_feat_lin = normalize_feat(query_feat_lin)
# Features --> Scores
# N x F (N = number of training images, F = features, 512 for CLIP)
score = data_feat_lin @ query_feat_lin.squeeze(0) # N, similarity scores
# sort & softmax
predictions = torch.argsort(-score).cpu() # sort by score
sorted_score = score[predictions] # highest scores
calibrated_softmax = calibrator(sorted_score.unsqueeze(0))[0]
new_row = True
prev_ind = 0
for i in range(N):
filename = laion_paths[predictions[i]]
img = load_img(os.path.join(laion_folder, filename))
shortside = 1 + \
int((res - 1) * torch.sqrt(calibrated_softmax[i] / display_per))
img_rs = resize_img(img, shortside=shortside)
if new_row:
img_accum = img_rs
new_row = False
else:
H_accum, W_accum = img_accum.shape[0], img_accum.shape[1]
H_new, W_new = img_rs.shape[0], img_rs.shape[1]
img_accum = np.concatenate(
[img_accum, 255 + np.zeros((H_accum, W_new + pad, 3), dtype=img_accum.dtype)], axis=1)
img_accum[int((H_accum - H_new) / 2):int((H_accum - H_new) / 2) +
H_new, W_accum + pad:, :] = img_rs
if img_accum.shape[1] > num_columns * res:
new_row = True
next_ind = i + 1
aggregated_score = torch.sum(calibrated_softmax[prev_ind : next_ind]).item()
st.write(
f'**[{aggregated_score * 100:.2f}%] estimated influence**, **{next_ind - prev_ind} images**')
prev_ind = next_ind
st.image(img_accum, width=img_accum.shape[1])
remaining = 1 - torch.sum(calibrated_softmax[:N]).item()
st.write('Remaining influence:', f'{remaining * 100:.2f}%')