Create clip_base.py
Browse files- clip_base.py +145 -0
clip_base.py
ADDED
@@ -0,0 +1,145 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import torch
|
3 |
+
import requests
|
4 |
+
import zipfile
|
5 |
+
from PIL import Image
|
6 |
+
from io import BytesIO
|
7 |
+
import tensorflow as tf
|
8 |
+
import pandas as pd
|
9 |
+
from transformers import CLIPProcessor, TFCLIPModel
|
10 |
+
|
11 |
+
|
12 |
+
class TfliteConverter:
|
13 |
+
|
14 |
+
def __init__(self, model, image_data, title_list, unique_title_list):
|
15 |
+
self.model = model
|
16 |
+
self.image_data = image_data
|
17 |
+
self.title_list = title_list
|
18 |
+
self.unique_title_list = unique_title_list
|
19 |
+
|
20 |
+
def build_model(self):
|
21 |
+
concrete_func = self.serving_fn.get_concrete_function()
|
22 |
+
|
23 |
+
converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
|
24 |
+
|
25 |
+
converter.target_spec.supported_ops = [
|
26 |
+
tf.lite.OpsSet.TFLITE_BUILTINS,
|
27 |
+
tf.lite.OpsSet.SELECT_TF_OPS
|
28 |
+
]
|
29 |
+
|
30 |
+
tflite_model = converter.convert()
|
31 |
+
|
32 |
+
model_path = os.path.join(os.getcwd(), 'clip_tflite_model.tflite')
|
33 |
+
json_path = os.path.join(os.getcwd(), 'categories.json')
|
34 |
+
zip_path = os.path.join(os.getcwd(), 'model_package.zip')
|
35 |
+
|
36 |
+
with open(model_path, 'wb') as f:
|
37 |
+
f.write(tflite_model)
|
38 |
+
|
39 |
+
categories_df = pd.DataFrame({
|
40 |
+
'id': range(len(self.unique_title_list)),
|
41 |
+
'title': self.unique_title_list
|
42 |
+
})
|
43 |
+
categories_df.to_json(json_path, orient='records', indent=2)
|
44 |
+
|
45 |
+
# Create ZIP file containing both files
|
46 |
+
with zipfile.ZipFile(zip_path, 'w', zipfile.ZIP_DEFLATED) as zipf:
|
47 |
+
zipf.write(model_path, os.path.basename(model_path))
|
48 |
+
zipf.write(json_path, os.path.basename(json_path))
|
49 |
+
|
50 |
+
# # Clean up temporary files
|
51 |
+
# os.remove(model_path)
|
52 |
+
# os.remove(json_path)
|
53 |
+
|
54 |
+
return zip_path
|
55 |
+
|
56 |
+
@tf.function(input_signature=[
|
57 |
+
tf.TensorSpec(shape=[1, 224, 224, 3], dtype=tf.float32, name='input_images')
|
58 |
+
])
|
59 |
+
def serving_fn(self, input_images):
|
60 |
+
x = tf.transpose(input_images, (0, 3, 1, 2))
|
61 |
+
|
62 |
+
with torch.no_grad():
|
63 |
+
img_embeddings = self.model.get_image_features(pixel_values=x)
|
64 |
+
|
65 |
+
labels_embeddings = [item['embeddings'] for item in self.image_data]
|
66 |
+
|
67 |
+
labels_embeddings = tf.stack(labels_embeddings, axis=0)
|
68 |
+
|
69 |
+
similarities = tf.reduce_sum(tf.multiply(img_embeddings, labels_embeddings), axis=-1)
|
70 |
+
norm_img = tf.norm(img_embeddings, axis=-1)
|
71 |
+
norm_labels = tf.norm(labels_embeddings, axis=-1)
|
72 |
+
|
73 |
+
cosine_similarity = similarities / (norm_img * norm_labels)
|
74 |
+
|
75 |
+
name_to_score = {}
|
76 |
+
|
77 |
+
names = self.title_list
|
78 |
+
static_names = self.unique_title_list
|
79 |
+
|
80 |
+
for i in range(len(names)):
|
81 |
+
name = names[i]
|
82 |
+
score = cosine_similarity[i]
|
83 |
+
|
84 |
+
if name not in name_to_score or score > name_to_score[name]:
|
85 |
+
name_to_score[name] = score
|
86 |
+
|
87 |
+
result = [name_to_score.get(name, 0) for name in static_names]
|
88 |
+
return tf.convert_to_tensor(result, dtype=tf.float32)
|
89 |
+
|
90 |
+
|
91 |
+
class OpenAiClipModel:
|
92 |
+
|
93 |
+
def __init__(self, payload):
|
94 |
+
self.payload = payload
|
95 |
+
self.processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
|
96 |
+
self.model = TFCLIPModel.from_pretrained("openai/clip-vit-base-patch32")
|
97 |
+
self.df = pd.DataFrame([(key, value) for key, values in self.payload.items() for value in values], columns=['image_category', 'image_path'])
|
98 |
+
self.labels_embeddings = []
|
99 |
+
self.image_data = []
|
100 |
+
self.title_list = self.df["image_category"].tolist()
|
101 |
+
self.unique_title_list = list(set(self.title_list))
|
102 |
+
|
103 |
+
|
104 |
+
def generate_text_embedding(self, text):
|
105 |
+
inputs = self.processor(text=[text], return_tensors="pt", padding=True)
|
106 |
+
with torch.no_grad():
|
107 |
+
text_embedding = self.model.get_text_features(**inputs)
|
108 |
+
return text_embedding
|
109 |
+
|
110 |
+
def generate_image_embeddings(self):
|
111 |
+
for index, row in self.df.iterrows():
|
112 |
+
image_info = row.to_dict()
|
113 |
+
image_path = image_info.get("image_path")
|
114 |
+
|
115 |
+
if os.path.exists(image_path):
|
116 |
+
try:
|
117 |
+
image = Image.open(image_path)
|
118 |
+
inputs = self.processor(images=image, return_tensors="tf")
|
119 |
+
outputs = self.model.get_image_features(**inputs)
|
120 |
+
image_embedding = outputs.numpy().flatten()
|
121 |
+
image_info['embeddings'] = image_embedding.tolist()
|
122 |
+
self.labels_embeddings.append(outputs)
|
123 |
+
except Exception as e:
|
124 |
+
image_info['embeddings'] = None
|
125 |
+
else:
|
126 |
+
image_info['embeddings'] = None
|
127 |
+
|
128 |
+
self.image_data.append(image_info)
|
129 |
+
|
130 |
+
|
131 |
+
def build_model(self):
|
132 |
+
self.generate_image_embeddings()
|
133 |
+
self.title_list = self.df["image_category"].tolist()
|
134 |
+
|
135 |
+
tflite_client = TfliteConverter(
|
136 |
+
model=self.model,
|
137 |
+
image_data=self.image_data,
|
138 |
+
title_list=self.title_list,
|
139 |
+
unique_title_list=self.unique_title_list
|
140 |
+
)
|
141 |
+
model_file = tflite_client.build_model()
|
142 |
+
return model_file
|
143 |
+
|
144 |
+
|
145 |
+
|