File size: 6,534 Bytes
ce649db
 
 
 
 
6668c84
6101644
 
ce649db
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c5ad46a
 
 
 
 
ce649db
c5ad46a
 
 
ce649db
 
 
 
 
 
 
 
 
 
 
 
c5ad46a
 
 
 
ce649db
c5ad46a
 
 
ce649db
 
 
 
 
 
 
 
 
 
 
 
 
c5ad46a
 
 
 
71b659e
c5ad46a
 
 
 
 
71b659e
c5ad46a
 
 
 
ce649db
 
a01e989
 
 
 
 
 
 
 
 
 
 
 
 
 
6101644
 
 
 
 
 
 
bca94f8
ce649db
a9e905c
bca94f8
 
 
a9e905c
 
 
 
 
 
b0cc17a
d92e39c
a9e905c
01c6c72
a9e905c
bca94f8
 
ce649db
01c6c72
 
 
 
ab4fce6
 
 
01c6c72
 
0789e97
fa00cfe
0789e97
01c6c72
 
0789e97
01c6c72
 
0789e97
01c6c72
 
0789e97
01c6c72
 
ab4fce6
 
 
 
 
 
0789e97
c5ad46a
 
a9e905c
c5ad46a
6101644
0789e97
6668c84
01c6c72
ce649db
c5ad46a
 
 
 
ce649db
c5ad46a
ce649db
 
 
c5ad46a
 
a01e989
 
c5ad46a
 
 
 
 
ce649db
6101644
 
f920e44
6101644
ce649db
 
6101644
 
29302cd
 
 
 
 
624b826
6101644
 
 
a9e905c
6668c84
 
 
 
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
import io
import os
import requests
import zipfile
import natsort
import gc
from PIL import Image
from PIL import UnidentifiedImageError

os.environ["TOKENIZERS_PARALLELISM"] = "false"
from stqdm import stqdm
import streamlit as st
from jax import numpy as jnp
import transformers
from transformers import AutoTokenizer
from torchvision.transforms import Compose, CenterCrop, Normalize, Resize, ToTensor
from torchvision.transforms.functional import InterpolationMode
from modeling_hybrid_clip import FlaxHybridCLIP

import utils


@st.cache(hash_funcs={FlaxHybridCLIP: lambda _: None})
def get_model():
    return FlaxHybridCLIP.from_pretrained("clip-italian/clip-italian")


@st.cache(
    hash_funcs={
        transformers.models.bert.tokenization_bert_fast.BertTokenizerFast: lambda _: None
    }
)
def get_tokenizer():
    return AutoTokenizer.from_pretrained(
        "dbmdz/bert-base-italian-xxl-uncased", cache_dir="./", use_fast=True
    )


@st.cache(suppress_st_warning=True)
def download_images():
    # from sentence_transformers import SentenceTransformer, util
    img_folder = "photos/"
    if not os.path.exists(img_folder) or len(os.listdir(img_folder)) == 0:
        os.makedirs(img_folder, exist_ok=True)

        photo_filename = "unsplash-25k-photos.zip"
        if not os.path.exists(photo_filename):  # Download dataset if does not exist
            print(f"Downloading {photo_filename}...")
            response = requests.get(
                f"http://sbert.net/datasets/{photo_filename}", stream=True
            )
            total_size_in_bytes = int(response.headers.get("content-length", 0))
            block_size = 1024  # 1 Kb
            progress_bar = stqdm(
                total=total_size_in_bytes
            )  # , unit='iB', unit_scale=True
            content = io.BytesIO()
            for data in response.iter_content(block_size):
                progress_bar.update(len(data))
                content.write(data)
            progress_bar.close()
            z = zipfile.ZipFile(content)
            # content.close()
            print("Extracting the dataset...")
            z.extractall(path=img_folder)
    print("Done.")


@st.cache()
def get_image_features(dataset_name):
    if dataset_name == "Unsplash":
        return jnp.load("static/features/features.npy")
    else:
        return jnp.load("static/features/CC_embeddings.npy")


@st.cache()
def load_urls(dataset_name):
    if dataset_name == "CC":
        with open("static/CC_urls.txt") as fp:
            urls = [l.strip() for l in fp.readlines()]
        return urls
    else:
        ValueError(f"{dataset_name} not supported here")


def get_image_transform(image_size):
    return Compose(
        [
            Resize([image_size], interpolation=InterpolationMode.BICUBIC),
            CenterCrop(image_size),
            ToTensor(),
            Normalize(
                (0.48145466, 0.4578275, 0.40821073),
                (0.26862954, 0.26130258, 0.27577711),
            ),
        ]
    )


headers = {
    #'User-Agent': 'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/71.0.3578.98 Safari/537.36',
    'User-Agent': 'Googlebot-Image/1.0',  # Pretend to be googlebot
    'X-Forwarded-For': '64.18.15.200'
}


def app():

    st.title("From Text to Image")
    st.markdown(
        """
    
        ### 👋 Ciao!

        Here you can search for ~150.000 images in the Conceptual Captions dataset (CC) or in the Unsplash 25k Photos dataset.
        Even though we did not train on any of these images you will see most queries make sense. When you see errors, there might be two possibilities: 
        the model is answering in a wrong way or the image you are looking for is not in the dataset and the model is giving you the best answer it can get.
        
        
        
        🤌 Italian mode on! 🤌

        You can choose from one of the following examples:
        """
    )

    suggestions = [
        "Un gatto",
        "Due gatti",
        "Un fiore giallo",
        "Un fiore blu",
        "Una coppia in montagna",
        "Una coppia al tramonto"
    ]
    sugg_idx = -1

    col1, col2, col3, col4, col5, col6 = st.beta_columns([1, 1, 1.2, 1.2, 1.4, 1.4])
    with col1:
        if st.button(suggestions[0]):
            sugg_idx = 0
    with col2:
        if st.button(suggestions[1]):
            sugg_idx = 1
    with col3:
        if st.button(suggestions[2]):
            sugg_idx = 2
    with col4:
        if st.button(suggestions[3]):
            sugg_idx = 3
    with col5:
        if st.button(suggestions[4]):
            sugg_idx = 4
    with col6:
        if st.button(suggestions[5]):
            sugg_idx = 5

    col1, col2 = st.beta_columns([3, 1])
    with col1:
        query = st.text_input("... or insert an Italian query text")
    with col2:
        dataset_name = st.selectbox("IR dataset", ["CC", "Unsplash"])

    query = suggestions[sugg_idx] if sugg_idx > -1 else query if query else ""

    if query:
        with st.spinner("Computing..."):

            if dataset_name == "Unsplash":
                download_images()

            image_features = get_image_features(dataset_name)
            model = get_model()
            tokenizer = get_tokenizer()

            if dataset_name == "Unsplash":
                image_size = model.config.vision_config.image_size
                dataset = utils.CustomDataSet(
                    "photos/", transform=get_image_transform(image_size)
                )
            elif dataset_name == "CC":
                dataset = load_urls(dataset_name)
            else:
                raise ValueError()

            N = 3

            image_paths = utils.find_image(
                query, model, dataset, tokenizer, image_features, N, dataset_name
            )

        for i, image_url in enumerate(image_paths):
            try:
                if dataset_name == "Unsplash":
                    st.image(image_url)
                elif dataset_name == "CC":
                    image_raw = requests.get(image_url, stream=True, allow_redirects=True, headers=headers).raw
                    image = Image.open(image_raw).convert("RGB")
                    st.image(image, use_column_width=True)
                break
            except (UnidentifiedImageError) as e:
                if i == N - 1:
                    st.text(f'Tried to show {N} different image URLS but none of them were reachabele.\nMaybe try a different query?')

        gc.collect()

        sugg_idx = -1