app.py

#!/usr/bin/env python

import torch
import gradio as gr
from torchvision import transforms


# load imagenet labels
with open("imagenet_classes.txt", "r") as f:
    categories = [s.strip() for s in f.readlines()]


# load a resnet18 model pretrained on ImageNet
# and turn off autograd on model's parameters
model = torch.hub.load('pytorch/vision:v0.10.0', 'resnet18', pretrained=True).eval()
for param in model.parameters():
    param.requires_grad = False


# preprocess data
pretrained_std = torch.Tensor([0.229, 0.224, 0.225])
pretrained_mean = torch.Tensor([0.485, 0.456, 0.406])

preprocess = transforms.Compose([
    transforms.Resize(224),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=pretrained_mean, std=pretrained_std),
])


def undo_preprocess(processed_img):
    img = processed_img.mul(pretrained_std.view(-1,1,1)).add(pretrained_mean.view(-1,1,1))
    img = torch.clamp(img, 0, 1).squeeze()
    img = img.numpy().transpose(1,2,0)
    return img


def set_example_image(img):
    return gr.Image.update(value=img[0])


def generate_adversial_image(img, epsilon):
    
    # convert a PIL image to tensor and set requires_grad to True
    processed_img = preprocess(img).unsqueeze(0)
    processed_img.requires_grad = True
    
    # forward pass
    output = model(processed_img)
    
    # get predictions
    with torch.no_grad():
        probs = torch.nn.functional.softmax(output[0], dim=0)
    top5_prob, top5_idx = torch.topk(probs, 5)
    preds = {categories[idx]: prob.item() for idx, prob in zip(top5_idx, top5_prob)}
    
    # compute gradient
    label = torch.Tensor([top5_idx[0]]).type(torch.LongTensor)
    loss = torch.nn.functional.cross_entropy(output, label)
    loss.backward()
    
    # generate adversarial image
    with torch.no_grad():
        adv_img = processed_img + epsilon*processed_img.grad.data.sign()
        adv_output = model(adv_img)
    adv_probs = torch.nn.functional.softmax(adv_output[0], dim=0)
    adv_top5_prob, adv_top5_idx = torch.topk(adv_probs, 5)
    adv_preds = {categories[idx]: prob.item() for idx, prob in zip(adv_top5_idx, adv_top5_prob)}
    
    return undo_preprocess(processed_img.detach()), undo_preprocess(adv_img), preds, adv_preds


def main():

    with gr.Blocks() as demo:
        
        gr.Markdown('''## Generate Adversarial Image with Fast Gradient Sign Method
                    
                    Fast Gradient Sign Method was introduced in the paper [Explaining and Harnessing Adversarial Examples](https://arxiv.org/abs/1412.6572).

                    Given an input image and a neural network, an adversarial image can be generated by applying a small perturbation in the direction of the gradient that maximizes the loss to the input image.
                    
                    <code>adv_img = input_img + epsilon * sign(input_img.grad)</code>.

                    The parameter `epsilon` controls the maximum bound of the perturbation.
                    
                    ''')
                                                 
        with gr.Box():
            input_image = gr.Image(type="pil", label="Input Image")
            example_images = gr.Dataset(components=[input_image],
                                        samples=[['coral.jpg'], ['goldfish.jpg'], ['otter.jpg'], ['panda.jpg']])
            
            with gr.Row():
                epsilon = gr.Slider(minimum=0, maximum=0.5, value=0.02, step=0.01, label="epsilon")
                btn = gr.Button("Generate Adversarial Image")
        
        gr.Markdown('''### Original Image''')
        with gr.Box():
            with gr.Row():
                img_before = gr.Image(label="Original Image")
                label_before = gr.Label(label="Original Prediction")
        
        gr.Markdown('''### Adversarial Image''')
        with gr.Box():
            with gr.Row():
                img_after = gr.Image(label="Adversarial Image")
                label_after = gr.Label(label="New Prediction")
            
        gr.Markdown('''The prediction is done by ResNet18. Example images are from [Unsplash](https://unsplash.com).''')
        
        
        # events
        btn.click(fn=generate_adversial_image, 
                  inputs=[input_image, epsilon], 
                  outputs=[img_before, img_after, label_before, label_after])
        
        example_images.click(fn=set_example_image,
                                 inputs=example_images,
                                 outputs=example_images.components)

    demo.launch()



if __name__ == '__main__':
    main()