Spaces:
Sleeping
Sleeping
Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,160 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import zipfile
|
3 |
+
import torch
|
4 |
+
from torch import nn, optim
|
5 |
+
from torch.utils.data import DataLoader, Dataset
|
6 |
+
from torchvision import transforms
|
7 |
+
from PIL import Image
|
8 |
+
from transformers import CLIPModel, CLIPProcessor
|
9 |
+
import gradio as gr
|
10 |
+
|
11 |
+
# Step 1: Unzip the dataset
|
12 |
+
if not os.path.exists("data"):
|
13 |
+
os.makedirs("data")
|
14 |
+
|
15 |
+
print("Extracting Data.zip...")
|
16 |
+
with zipfile.ZipFile("Data.zip", 'r') as zip_ref:
|
17 |
+
zip_ref.extractall("data")
|
18 |
+
print("Extraction complete.")
|
19 |
+
|
20 |
+
# Step 2: Dynamically find the 'safe' and 'unsafe' folders
|
21 |
+
def find_dataset_path(root_dir):
|
22 |
+
for root, dirs, files in os.walk(root_dir):
|
23 |
+
if 'safe' in dirs and 'unsafe' in dirs:
|
24 |
+
return root
|
25 |
+
return None
|
26 |
+
|
27 |
+
# Look for 'safe' and 'unsafe' inside 'data/Data'
|
28 |
+
dataset_path = find_dataset_path("data/Data")
|
29 |
+
if dataset_path is None:
|
30 |
+
print("Debugging extracted structure:")
|
31 |
+
for root, dirs, files in os.walk("data"):
|
32 |
+
print(f"Root: {root}")
|
33 |
+
print(f"Directories: {dirs}")
|
34 |
+
print(f"Files: {files}")
|
35 |
+
raise FileNotFoundError("Expected 'safe' and 'unsafe' folders not found inside 'data/Data'. Please check the Data.zip structure.")
|
36 |
+
print(f"Dataset path found: {dataset_path}")
|
37 |
+
|
38 |
+
# Step 3: Define Custom Dataset Class
|
39 |
+
class CustomImageDataset(Dataset):
|
40 |
+
def __init__(self, root_dir, transform=None):
|
41 |
+
self.root_dir = root_dir
|
42 |
+
self.transform = transform
|
43 |
+
self.image_paths = []
|
44 |
+
self.labels = []
|
45 |
+
|
46 |
+
for label, folder in enumerate(["safe", "unsafe"]): # 0 = safe, 1 = unsafe
|
47 |
+
folder_path = os.path.join(root_dir, folder)
|
48 |
+
if not os.path.exists(folder_path):
|
49 |
+
raise FileNotFoundError(f"Folder '{folder}' not found in '{root_dir}'")
|
50 |
+
for filename in os.listdir(folder_path):
|
51 |
+
if filename.endswith((".jpg", ".jpeg", ".png")): # Only load image files
|
52 |
+
self.image_paths.append(os.path.join(folder_path, filename))
|
53 |
+
self.labels.append(label)
|
54 |
+
|
55 |
+
def __len__(self):
|
56 |
+
return len(self.image_paths)
|
57 |
+
|
58 |
+
def __getitem__(self, idx):
|
59 |
+
image_path = self.image_paths[idx]
|
60 |
+
image = Image.open(image_path).convert("RGB")
|
61 |
+
label = self.labels[idx]
|
62 |
+
if self.transform:
|
63 |
+
image = self.transform(image)
|
64 |
+
return image, label
|
65 |
+
|
66 |
+
# Step 4: Data Transformations
|
67 |
+
transform = transforms.Compose([
|
68 |
+
transforms.Resize((224, 224)), # Resize to 224x224 pixels
|
69 |
+
transforms.ToTensor(), # Convert to tensor
|
70 |
+
transforms.Normalize((0.5,), (0.5,)), # Normalize image values
|
71 |
+
])
|
72 |
+
|
73 |
+
# Step 5: Load the Dataset
|
74 |
+
train_dataset = CustomImageDataset(dataset_path, transform=transform)
|
75 |
+
train_loader = DataLoader(train_dataset, batch_size=8, shuffle=True)
|
76 |
+
|
77 |
+
# Debugging: Check the dataset
|
78 |
+
print(f"Number of samples in the dataset: {len(train_dataset)}")
|
79 |
+
if len(train_dataset) == 0:
|
80 |
+
raise ValueError("The dataset is empty. Please check if 'Data.zip' is correctly unzipped and contains 'safe' and 'unsafe' folders.")
|
81 |
+
|
82 |
+
# Step 6: Load Pretrained CLIP Model
|
83 |
+
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
|
84 |
+
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
|
85 |
+
|
86 |
+
# Add a Classification Layer
|
87 |
+
model.classifier = nn.Linear(model.visual_projection.out_features, 2) # 2 classes: safe, unsafe
|
88 |
+
|
89 |
+
# Define Optimizer and Loss Function
|
90 |
+
optimizer = optim.Adam(model.classifier.parameters(), lr=1e-4)
|
91 |
+
criterion = nn.CrossEntropyLoss()
|
92 |
+
|
93 |
+
# Step 7: Fine-Tune the Model
|
94 |
+
model.train()
|
95 |
+
for epoch in range(3): # Number of epochs
|
96 |
+
total_loss = 0
|
97 |
+
for images, labels in train_loader:
|
98 |
+
optimizer.zero_grad()
|
99 |
+
images = torch.stack([img.to(torch.float32) for img in images]) # Batch of images
|
100 |
+
outputs = model.get_image_features(pixel_values=images)
|
101 |
+
logits = model.classifier(outputs)
|
102 |
+
loss = criterion(logits, labels)
|
103 |
+
loss.backward()
|
104 |
+
optimizer.step()
|
105 |
+
total_loss += loss.item()
|
106 |
+
print(f"Epoch {epoch+1}, Loss: {total_loss / len(train_loader)}")
|
107 |
+
|
108 |
+
# Save the Fine-Tuned Model
|
109 |
+
model.save_pretrained("fine-tuned-model")
|
110 |
+
processor.save_pretrained("fine-tuned-model")
|
111 |
+
print("Model fine-tuned and saved successfully.")
|
112 |
+
|
113 |
+
# Step 8: Define Gradio Inference Function
|
114 |
+
def classify_image(image, class_names):
|
115 |
+
# Load Fine-Tuned Model
|
116 |
+
model = CLIPModel.from_pretrained("fine-tuned-model")
|
117 |
+
processor = CLIPProcessor.from_pretrained("fine-tuned-model")
|
118 |
+
|
119 |
+
# Split class names from comma-separated input
|
120 |
+
labels = [label.strip() for label in class_names.split(",") if label.strip()]
|
121 |
+
if not labels:
|
122 |
+
return {"Error": "Please enter at least one valid class name."}
|
123 |
+
|
124 |
+
# Process the image and labels
|
125 |
+
inputs = processor(text=labels, images=image, return_tensors="pt", padding=True)
|
126 |
+
outputs = model(**inputs)
|
127 |
+
logits_per_image = outputs.logits_per_image
|
128 |
+
probs = logits_per_image.softmax(dim=1) # Convert logits to probabilities
|
129 |
+
|
130 |
+
# Extract labels with their corresponding probabilities
|
131 |
+
result = {label: probs[0][i].item() for i, label in enumerate(labels)}
|
132 |
+
return dict(sorted(result.items(), key=lambda item: item[1], reverse=True))
|
133 |
+
|
134 |
+
# Step 9: Set Up Gradio Interface
|
135 |
+
iface = gr.Interface(
|
136 |
+
fn=classify_image,
|
137 |
+
inputs=[
|
138 |
+
gr.Image(type="pil"),
|
139 |
+
gr.Textbox(label="Possible class names (comma-separated)", placeholder="e.g., safe, unsafe")
|
140 |
+
],
|
141 |
+
outputs=gr.Label(num_top_classes=2),
|
142 |
+
title="Content Safety Classification",
|
143 |
+
description="Classify images as 'safe' or 'unsafe' using a fine-tuned CLIP model.",
|
144 |
+
)
|
145 |
+
|
146 |
+
# Launch Gradio Interface
|
147 |
+
if __name__ == "__main__":
|
148 |
+
iface.launch()
|
149 |
+
|
150 |
+
|
151 |
+
|
152 |
+
|
153 |
+
|
154 |
+
|
155 |
+
|
156 |
+
|
157 |
+
|
158 |
+
|
159 |
+
|
160 |
+
|