Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,133 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import json
|
2 |
+
import gradio as gr
|
3 |
+
import torch
|
4 |
+
import PIL
|
5 |
+
import os
|
6 |
+
from models import ViTClassifier
|
7 |
+
from datasets import load_dataset
|
8 |
+
from transformers import TrainingArguments, ViTConfig, ViTForImageClassification
|
9 |
+
from torchvision import transforms
|
10 |
+
import pandas as pd
|
11 |
+
|
12 |
+
def load_config(config_path):
|
13 |
+
with open(config_path, 'r') as f:
|
14 |
+
config = json.load(f)
|
15 |
+
print("Config Loaded:", config) # Debugging
|
16 |
+
return config
|
17 |
+
|
18 |
+
def load_model(config, device='cuda'):
|
19 |
+
device = torch.device(device if torch.cuda.is_available() else 'cpu')
|
20 |
+
ckpt = torch.load(config['checkpoint_path'], map_location=device)
|
21 |
+
print("Checkpoint Loaded:", ckpt.keys()) # Debugging
|
22 |
+
model = ViTClassifier(config, device=device, dtype=torch.float32)
|
23 |
+
print("Model Loaded:", model) # Debugging
|
24 |
+
model.load_state_dict(ckpt['model'])
|
25 |
+
return model.to(device).eval()
|
26 |
+
|
27 |
+
def prepare_model_for_push(model, config):
|
28 |
+
# Create a VisionTransformerConfig
|
29 |
+
vit_config = ViTConfig(
|
30 |
+
image_size=config['model']['input_size'],
|
31 |
+
patch_size=config['model']['patch_size'],
|
32 |
+
hidden_size=config['model']['hidden_size'],
|
33 |
+
num_heads=config['model']['num_attention_heads'],
|
34 |
+
num_layers=config['model']['num_hidden_layers'],
|
35 |
+
mlp_ratio=4, # Common default for ViT
|
36 |
+
hidden_dropout_prob=config['model']['hidden_dropout_prob'],
|
37 |
+
attention_probs_dropout_prob=config['model']['attention_probs_dropout_prob'],
|
38 |
+
layer_norm_eps=config['model']['layer_norm_eps'],
|
39 |
+
num_classes=config['model']['num_classes']
|
40 |
+
)
|
41 |
+
# Create a VisionTransformer model
|
42 |
+
vit_model = ViTForImageClassification(vit_config)
|
43 |
+
# Copy the weights from your custom model to the VisionTransformer model
|
44 |
+
state_dict = vit_model.state_dict()
|
45 |
+
for key in state_dict.keys():
|
46 |
+
if key in model.state_dict():
|
47 |
+
state_dict[key] = model.state_dict()[key]
|
48 |
+
vit_model.load_state_dict(state_dict)
|
49 |
+
return vit_model, vit_config
|
50 |
+
|
51 |
+
def run_inference(input_image, model):
|
52 |
+
print("Input Image Type:", type(input_image)) # Debugging
|
53 |
+
# Directly use the PIL Image object
|
54 |
+
fake_prob = model.forward(input_image).item()
|
55 |
+
result_description = get_result_description(fake_prob)
|
56 |
+
return {
|
57 |
+
"Fake Probability": fake_prob,
|
58 |
+
"Result Description": result_description
|
59 |
+
}
|
60 |
+
|
61 |
+
def get_result_description(fake_prob):
|
62 |
+
if fake_prob > 0.5:
|
63 |
+
return "The image is likely a fake."
|
64 |
+
else:
|
65 |
+
return "The image is likely real."
|
66 |
+
|
67 |
+
def run_evaluation(dataset_name, model, config, device):
|
68 |
+
dataset = load_dataset(dataset_name)
|
69 |
+
eval_df, accuracy = evaluate_model(model, dataset, config, device)
|
70 |
+
return accuracy, eval_df.to_csv(index=False)
|
71 |
+
|
72 |
+
def evaluate_model(model, dataset, config, device):
|
73 |
+
device = torch.device(device if torch.cuda.is_available() else 'cpu')
|
74 |
+
model.to(device).eval()
|
75 |
+
norm_mean = config['preprocessing']['norm_mean']
|
76 |
+
norm_std = config['preprocessing']['norm_std']
|
77 |
+
resize_size = config['preprocessing']['resize_size']
|
78 |
+
crop_size = config['preprocessing']['crop_size']
|
79 |
+
augment_list = [
|
80 |
+
transforms.Resize(resize_size),
|
81 |
+
transforms.CenterCrop(crop_size),
|
82 |
+
transforms.ToTensor(),
|
83 |
+
transforms.Normalize(mean=norm_mean, std=norm_std),
|
84 |
+
transforms.ConvertImageDtype(torch.float32),
|
85 |
+
]
|
86 |
+
preprocess = transforms.Compose(augment_list)
|
87 |
+
true_labels = []
|
88 |
+
predicted_probs = []
|
89 |
+
predicted_labels = []
|
90 |
+
with torch.no_grad():
|
91 |
+
for sample in dataset:
|
92 |
+
image = sample['image']
|
93 |
+
label = sample['label']
|
94 |
+
image = preprocess(image).unsqueeze(0).to(device)
|
95 |
+
output = model.forward(image)
|
96 |
+
prob = output.item()
|
97 |
+
true_labels.append(label)
|
98 |
+
predicted_probs.append(prob)
|
99 |
+
predicted_labels.append(1 if prob > 0.5 else 0)
|
100 |
+
eval_df = pd.DataFrame({
|
101 |
+
'True Label': true_labels,
|
102 |
+
'Predicted Probability': predicted_probs,
|
103 |
+
'Predicted Label': predicted_labels
|
104 |
+
})
|
105 |
+
accuracy = (eval_df['True Label'] == eval_df['Predicted Label']).mean()
|
106 |
+
return eval_df, accuracy
|
107 |
+
|
108 |
+
def main():
|
109 |
+
# Load configuration
|
110 |
+
config_path = "config.json"
|
111 |
+
config = load_config(config_path)
|
112 |
+
# Load model
|
113 |
+
device = config['device']
|
114 |
+
model = load_model(config, device=device)
|
115 |
+
# Define Gradio interface for inference
|
116 |
+
def gradio_interface(input_image):
|
117 |
+
return run_inference(input_image, model)
|
118 |
+
# Create Gradio Tabs
|
119 |
+
with gr.Blocks() as demo:
|
120 |
+
gr.Markdown("# Deepfake Detection")
|
121 |
+
with gr.Tab("Image Inference"):
|
122 |
+
with gr.Row():
|
123 |
+
with gr.Column():
|
124 |
+
gr.Markdown("## Upload Image for Evaluation")
|
125 |
+
input_image = gr.Image(type="pil", label="Upload Image")
|
126 |
+
with gr.Column():
|
127 |
+
output = gr.JSON(label="Classification Result")
|
128 |
+
input_image.change(fn=gradio_interface, inputs=input_image, outputs=output)
|
129 |
+
# Launch the Gradio app
|
130 |
+
demo.launch()
|
131 |
+
|
132 |
+
if __name__ == "__main__":
|
133 |
+
main()
|