Spaces:
Sleeping
Sleeping
Commit
·
80d1ab2
1
Parent(s):
902ac2f
Initial commit
Browse files- .gitignore +7 -0
- app.py +228 -0
- requirements.txt +4 -0
.gitignore
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
gradio_cached_examples/
|
| 2 |
+
*.png
|
| 3 |
+
*.jpg
|
| 4 |
+
flagged/
|
| 5 |
+
*.pt
|
| 6 |
+
*.json
|
| 7 |
+
*.npy
|
app.py
ADDED
|
@@ -0,0 +1,228 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import cv2
|
| 3 |
+
import numpy as np
|
| 4 |
+
from PIL import Image
|
| 5 |
+
import json
|
| 6 |
+
import gradio as gr
|
| 7 |
+
import matplotlib.pyplot as plt
|
| 8 |
+
|
| 9 |
+
import torch
|
| 10 |
+
from torch import nn
|
| 11 |
+
import torch.nn.functional as F
|
| 12 |
+
import timm
|
| 13 |
+
from transformers import DistilBertModel, DistilBertConfig, DistilBertTokenizer
|
| 14 |
+
|
| 15 |
+
class CFG:
|
| 16 |
+
image_path = './images'
|
| 17 |
+
captions_path = './captions'
|
| 18 |
+
batch_size = 64
|
| 19 |
+
num_workers = 4
|
| 20 |
+
head_lr = 1e-3
|
| 21 |
+
image_encoder_lr = 1e-4
|
| 22 |
+
text_encoder_lr = 1e-5
|
| 23 |
+
weight_decay = 1e-3
|
| 24 |
+
patience = 1
|
| 25 |
+
factor = 0.8
|
| 26 |
+
epochs = 2
|
| 27 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 28 |
+
|
| 29 |
+
model_name = 'resnet50'
|
| 30 |
+
image_embedding = 2048
|
| 31 |
+
text_encoder_model = "distilbert-base-uncased"
|
| 32 |
+
text_embedding = 768
|
| 33 |
+
text_tokenizer = "distilbert-base-uncased"
|
| 34 |
+
max_length = 200
|
| 35 |
+
|
| 36 |
+
pretrained = True # for both image encoder and text encoder
|
| 37 |
+
trainable = True # for both image encoder and text encoder
|
| 38 |
+
temperature = 1.0
|
| 39 |
+
|
| 40 |
+
# image size
|
| 41 |
+
size = 224
|
| 42 |
+
|
| 43 |
+
# for projection head; used for both image and text encoders
|
| 44 |
+
num_projection_layers = 1
|
| 45 |
+
projection_dim = 256
|
| 46 |
+
dropout = 0.1
|
| 47 |
+
|
| 48 |
+
class ImageEncoder(nn.Module):
|
| 49 |
+
"""
|
| 50 |
+
Encode images to a fixed size vector
|
| 51 |
+
"""
|
| 52 |
+
|
| 53 |
+
def __init__(
|
| 54 |
+
self, model_name=CFG.model_name, pretrained=CFG.pretrained, trainable=CFG.trainable
|
| 55 |
+
):
|
| 56 |
+
super().__init__()
|
| 57 |
+
self.model = timm.create_model(
|
| 58 |
+
model_name, pretrained, num_classes=0, global_pool="avg"
|
| 59 |
+
)
|
| 60 |
+
for p in self.model.parameters():
|
| 61 |
+
p.requires_grad = trainable
|
| 62 |
+
|
| 63 |
+
def forward(self, x):
|
| 64 |
+
return self.model(x)
|
| 65 |
+
|
| 66 |
+
class TextEncoder(nn.Module):
|
| 67 |
+
def __init__(self, model_name=CFG.text_encoder_model, pretrained=CFG.pretrained, trainable=CFG.trainable):
|
| 68 |
+
super().__init__()
|
| 69 |
+
if pretrained:
|
| 70 |
+
self.model = DistilBertModel.from_pretrained(model_name)
|
| 71 |
+
else:
|
| 72 |
+
self.model = DistilBertModel(config=DistilBertConfig())
|
| 73 |
+
|
| 74 |
+
for p in self.model.parameters():
|
| 75 |
+
p.requires_grad = trainable
|
| 76 |
+
|
| 77 |
+
# we are using the CLS token hidden representation as the sentence's embedding
|
| 78 |
+
self.target_token_idx = 0
|
| 79 |
+
|
| 80 |
+
def forward(self, input_ids, attention_mask):
|
| 81 |
+
output = self.model(input_ids=input_ids, attention_mask=attention_mask)
|
| 82 |
+
last_hidden_state = output.last_hidden_state
|
| 83 |
+
return last_hidden_state[:, self.target_token_idx, :]
|
| 84 |
+
|
| 85 |
+
class ProjectionHead(nn.Module):
|
| 86 |
+
def __init__(
|
| 87 |
+
self,
|
| 88 |
+
embedding_dim,
|
| 89 |
+
projection_dim=CFG.projection_dim,
|
| 90 |
+
dropout=CFG.dropout
|
| 91 |
+
):
|
| 92 |
+
super().__init__()
|
| 93 |
+
self.projection = nn.Linear(embedding_dim, projection_dim)
|
| 94 |
+
self.gelu = nn.GELU()
|
| 95 |
+
self.fc = nn.Linear(projection_dim, projection_dim)
|
| 96 |
+
self.dropout = nn.Dropout(dropout)
|
| 97 |
+
self.layer_norm = nn.LayerNorm(projection_dim)
|
| 98 |
+
|
| 99 |
+
def forward(self, x):
|
| 100 |
+
projected = self.projection(x)
|
| 101 |
+
x = self.gelu(projected)
|
| 102 |
+
x = self.fc(x)
|
| 103 |
+
x = self.dropout(x)
|
| 104 |
+
x = x + projected
|
| 105 |
+
x = self.layer_norm(x)
|
| 106 |
+
return x
|
| 107 |
+
|
| 108 |
+
class CLIPModel(nn.Module):
|
| 109 |
+
def __init__(
|
| 110 |
+
self,
|
| 111 |
+
temperature=CFG.temperature,
|
| 112 |
+
image_embedding=CFG.image_embedding,
|
| 113 |
+
text_embedding=CFG.text_embedding,
|
| 114 |
+
):
|
| 115 |
+
super().__init__()
|
| 116 |
+
self.image_encoder = ImageEncoder()
|
| 117 |
+
self.text_encoder = TextEncoder()
|
| 118 |
+
self.image_projection = ProjectionHead(embedding_dim=image_embedding)
|
| 119 |
+
self.text_projection = ProjectionHead(embedding_dim=text_embedding)
|
| 120 |
+
self.temperature = temperature
|
| 121 |
+
|
| 122 |
+
def forward(self, batch):
|
| 123 |
+
# Getting Image and Text Features
|
| 124 |
+
image_features = self.image_encoder(batch["image"])
|
| 125 |
+
text_features = self.text_encoder(
|
| 126 |
+
input_ids=batch["input_ids"], attention_mask=batch["attention_mask"]
|
| 127 |
+
)
|
| 128 |
+
# Getting Image and Text Embeddings (with same dimension)
|
| 129 |
+
image_embeddings = self.image_projection(image_features)
|
| 130 |
+
text_embeddings = self.text_projection(text_features)
|
| 131 |
+
|
| 132 |
+
# Calculating the Loss
|
| 133 |
+
images_similarity = image_embeddings @ text_embeddings.T / self.temperature
|
| 134 |
+
texts_similarity = images_similarity.T
|
| 135 |
+
labels = torch.arange(batch["image"].shape[0]).long().to(CFG.device)
|
| 136 |
+
|
| 137 |
+
total_loss = (
|
| 138 |
+
F.cross_entropy(images_similarity, labels) +
|
| 139 |
+
F.cross_entropy(texts_similarity, labels)
|
| 140 |
+
) / 2
|
| 141 |
+
|
| 142 |
+
return total_loss
|
| 143 |
+
|
| 144 |
+
def find_matches_cpu(model, image_embeddings, query, image_filenames, n=4):
|
| 145 |
+
tokenizer = DistilBertTokenizer.from_pretrained(CFG.text_tokenizer)
|
| 146 |
+
encoded_query = tokenizer([query])
|
| 147 |
+
batch = {
|
| 148 |
+
key: torch.tensor(values).to('cpu')
|
| 149 |
+
for key, values in encoded_query.items()
|
| 150 |
+
}
|
| 151 |
+
with torch.no_grad():
|
| 152 |
+
text_features = model.text_encoder(
|
| 153 |
+
input_ids=batch["input_ids"], attention_mask=batch["attention_mask"]
|
| 154 |
+
)
|
| 155 |
+
text_embeddings = model.text_projection(text_features)
|
| 156 |
+
|
| 157 |
+
image_embeddings_n = F.normalize(image_embeddings, p=2, dim=-1)
|
| 158 |
+
text_embeddings_n = F.normalize(text_embeddings, p=2, dim=-1)
|
| 159 |
+
dot_similarity = text_embeddings_n @ image_embeddings_n.T
|
| 160 |
+
|
| 161 |
+
values, indices = torch.topk(dot_similarity.squeeze(0), n * 5)
|
| 162 |
+
matches = [image_filenames[idx] for idx in indices[::5]]
|
| 163 |
+
return matches
|
| 164 |
+
|
| 165 |
+
def rle_decode(img_rle_array, img_name, img_size):
|
| 166 |
+
encoded_image = img_rle_array
|
| 167 |
+
# Initialize variables for decoding
|
| 168 |
+
decoded_image = []
|
| 169 |
+
for i in range(0, len(encoded_image), 2):
|
| 170 |
+
pixel_value = encoded_image[i]
|
| 171 |
+
run_length = encoded_image[i + 1]
|
| 172 |
+
decoded_image.extend([pixel_value] * run_length)
|
| 173 |
+
|
| 174 |
+
# Convert the decoded image back to a NumPy array
|
| 175 |
+
decoded_array = np.array(decoded_image, dtype=np.uint8)
|
| 176 |
+
|
| 177 |
+
# Reshape the decoded array to the original image shape (224, 224)
|
| 178 |
+
decoded_image = decoded_array.reshape(img_size) # Use original shape
|
| 179 |
+
|
| 180 |
+
# Create a PIL Image from the decoded array
|
| 181 |
+
decoded_image = Image.fromarray(decoded_image)
|
| 182 |
+
|
| 183 |
+
decoded_image_save_path = './' + str(img_name)
|
| 184 |
+
# Save or display the decoded image
|
| 185 |
+
decoded_image.save(decoded_image_save_path) # Save the decoded image to a file
|
| 186 |
+
return decoded_image_save_path
|
| 187 |
+
|
| 188 |
+
def get_matched_image(matches, val_file_dict_loaded):
|
| 189 |
+
img_size = (112, 112)
|
| 190 |
+
match_img_list = []
|
| 191 |
+
for img_name in matches:
|
| 192 |
+
img_rle_array = val_file_dict_loaded[img_name]
|
| 193 |
+
decoded_image_path = rle_decode(img_rle_array, img_name, img_size)
|
| 194 |
+
match_img_list.append(decoded_image_path)
|
| 195 |
+
return match_img_list
|
| 196 |
+
|
| 197 |
+
def get_grayscale_image(text_query):
|
| 198 |
+
model_inf = CLIPModel().to('cpu')
|
| 199 |
+
model_inf.load_state_dict(torch.load('./best_clip_model_cpu.pt', map_location='cpu'))
|
| 200 |
+
|
| 201 |
+
clip_image_embeddings_np_inf = np.load('./clip_image_embeddings.npy')
|
| 202 |
+
image_embeddings_inf = torch.tensor(clip_image_embeddings_np_inf)
|
| 203 |
+
|
| 204 |
+
img_file_names = np.load('./val_img_file_names.npy',allow_pickle=True)
|
| 205 |
+
|
| 206 |
+
with open("./val_imgs_rle_encode.json", "r") as json_file:
|
| 207 |
+
val_file_dict_loaded = json.load(json_file)
|
| 208 |
+
|
| 209 |
+
matches = find_matches_cpu(model_inf,
|
| 210 |
+
image_embeddings_inf,
|
| 211 |
+
query=text_query,
|
| 212 |
+
image_filenames=img_file_names,
|
| 213 |
+
n=1)
|
| 214 |
+
|
| 215 |
+
matched_images = get_matched_image(matches, val_file_dict_loaded)
|
| 216 |
+
return matched_images
|
| 217 |
+
|
| 218 |
+
def gradio_fn(text):
|
| 219 |
+
text_query = str(text)
|
| 220 |
+
match_img_list = get_grayscale_image(text_query)
|
| 221 |
+
pil_img = Image.open(match_img_list[0])
|
| 222 |
+
pil_img = pil_img.resize((224, 224))
|
| 223 |
+
np_img_array = np.array(pil_img)
|
| 224 |
+
return np_img_array
|
| 225 |
+
|
| 226 |
+
demo = gr.Interface(fn=gradio_fn, inputs="text", outputs="image", title="CLIP Image Search")
|
| 227 |
+
|
| 228 |
+
demo.launch(share=True)
|
requirements.txt
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
transformers
|
| 2 |
+
gradio
|
| 3 |
+
timm
|
| 4 |
+
torch
|