File size: 5,391 Bytes
b9288df
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import numpy as np
import time
import random
import torch
import torchvision.transforms as transforms
#import requests
import gradio as gr
import matplotlib.pyplot as plt

from models import get_model
from dotmap import DotMap
from PIL import Image


# args
args = DotMap()
args.deploy = 'vanilla'
args.arch = 'dino_small_patch16'
args.device = 'cuda:7'
#args.resume = '/fast_scratch/hushell/fluidstack/FS125_few-shot-transformer/outputs/dinosmall_1e-4/best_converted.pth'
args.resume = 'https://huggingface.co/hushell/pmf_dinosmall_lr1e-4/resolve/main/best_converted.pth'
args.api_key = 'AIzaSyAFkOGnXhy-2ZB0imDvNNqf2rHb98vR_qY'
args.cx = '06d75168141bc47f1'


# model
device = torch.device(args.device)
model = get_model(args)
model.to(device)
#checkpoint = torch.load(args.resume, map_location='cpu')
checkpoint = torch.hub.load_state_dict_from_url(args.resume, map_location='cpu')
model.load_state_dict(checkpoint['model'], strict=True)


# image transforms
def test_transform():
    def _convert_image_to_rgb(im):
        return im.convert('RGB')

    return transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        _convert_image_to_rgb,
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406],
                             std=[0.229, 0.224, 0.225]),
        ])

preprocess = test_transform()

@torch.no_grad()
def denormalize(x, mean, std):
    # 3, H, W
    t = x.clone()
    t.mul_(std).add_(mean)
    return torch.clamp(t, 0, 1)


# Google image search
from google_images_search import GoogleImagesSearch

# define search params
# option for commonly used search param are shown below for easy reference.
# For param marked with '##':
#   - Multiselect is currently not feasible. Choose ONE option only
#   - This param can also be omitted from _search_params if you do not wish to define any value
_search_params = {
    'q': '...',
    'num': 10,
    'fileType': 'png', #'jpg|gif|png',
    'rights': 'cc_publicdomain', #'cc_publicdomain|cc_attribute|cc_sharealike|cc_noncommercial|cc_nonderived',
    #'safe': 'active|high|medium|off|safeUndefined', ##
    'imgType': 'photo', #'clipart|face|lineart|stock|photo|animated|imgTypeUndefined', ##
    #'imgSize': 'huge|icon|large|medium|small|xlarge|xxlarge|imgSizeUndefined', ##
    #'imgDominantColor': 'black|blue|brown|gray|green|orange|pink|purple|red|teal|white|yellow|imgDominantColorUndefined', ##
    'imgColorType': 'color', #'color|gray|mono|trans|imgColorTypeUndefined' ##
}


# Gradio UI
def inference(query, labels, n_supp=10):
    '''
    query: PIL image
    labels: list of class names
    '''
    labels = labels.split(',')
    n_supp = int(n_supp)

    #print(f'#rows={len(labels)}, #cols={n_supp}')
    fig, axs = plt.subplots(len(labels), n_supp, figsize=(n_supp*4, len(labels)*4))

    with torch.no_grad():
        # query image
        query = preprocess(query).unsqueeze(0).unsqueeze(0).to(device) # (1, 1, 3, H, W)

        supp_x = []
        supp_y = []

        # search support images
        for idx, y in enumerate(labels):
            with GoogleImagesSearch(args.api_key, args.cx) as gis:
                _search_params['q'] = y
                _search_params['num'] = n_supp
                gis.search(search_params=_search_params, custom_image_name='my_image')
                gis._custom_image_name = 'my_image'

                for j, x in enumerate(gis.results()):
                    #url = x.url
                    #x_im = Image.open(requests.get(url, stream=True).raw)
                    x.download('./')
                    x_im = Image.open(x.path)

                    # vis
                    axs[idx, j].imshow(x_im)
                    axs[idx, j].set_title(f'{y}{j}')
                    axs[idx, j].axis('off')

                    x_im = preprocess(x_im) # (3, H, W)
                    supp_x.append(x_im)
                    supp_y.append(idx)

        print('Searching for support images is done.')

        supp_x = torch.stack(supp_x, dim=0).unsqueeze(0).to(device) # (1, n_supp*n_labels, 3, H, W)
        supp_y = torch.tensor(supp_y).long().unsqueeze(0).to(device) # (1, n_supp*n_labels)

        with torch.cuda.amp.autocast(True):
            output = model(supp_x, supp_y, query) # (1, 1, n_labels)

        probs = output.softmax(dim=-1).detach().cpu().numpy()

        return {k: float(v) for k, v in zip(labels, probs[0, 0])}, fig


# DEBUG
#query = Image.open('../labrador-puppy.jpg')
##labels = 'dog, cat'
#labels = 'girl, boy'
#output = inference(query, labels, n_supp=2)
#print(output)


gr.Interface(fn=inference,
             inputs=[
                 gr.inputs.Image(label="Image to classify", type="pil"),
                 gr.inputs.Textbox(lines=1, label="Class hypotheses:", placeholder="Enter class names separated by ','",),
                 #gr.inputs.Number(default=1, label="Number of support examples from Google")
                 gr.inputs.Slider(minimum=2, maximum=10, step=1, label="Number of support examples from Google")
             ],
             theme="grass",
             outputs=[
                 gr.outputs.Label(label="Predicted class probabilities"),
                 gr.outputs.Image(type='plot', label="Support examples from Google image search"),
             ],
             description="PMF few-shot learning with Google image search").launch()