app.py

## Import the required Modules/Packages
import pandas as pd
import torch
from torch.nn import functional as F
import torchvision
from torchvision import transforms
import numpy as np
import gradio as gr
from PIL import Image
from collections import OrderedDict
from pytorch_grad_cam import GradCAM
from pytorch_grad_cam.utils.image import show_cam_on_image
from pytorch_grad_cam.utils.model_targets import ClassifierOutputTarget

from model import custResNet

classes = {0: 'airplane',
           1: 'automobile',
           2: 'bird',
           3: 'cat',
           4: 'deer',
           5: 'dog',
           6: 'frog',
           7: 'horse',
           8: 'ship',
           9: 'truck'}

mis_classified_df = pd.read_csv('misclassified_images.csv')
mis_classified_df['ground_truths'] = mis_classified_df['ground_truths'].map(classes)
mis_classified_df['predicted_vals'] = mis_classified_df['predicted_vals'].map(classes)
mis_classified_df = mis_classified_df.sample(frac=1)

device = torch.device("cuda")

model1 = custResNet()
model1.load_state_dict(torch.load('cust_resnet_model.pth', map_location=torch.device('cpu')), strict=False)
model1.eval()

transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.49139968, 0.48215827, 0.44653124], std=[0.24703233, 0.24348505, 0.26158768])
])

inv_transform = transforms.Normalize(mean=[-(0.49139968/0.24703233), -(0.48215827/0.24348505), -(0.44653124/0.26158768)], std=[(0.24703233), (1/0.24348505), (1/0.26158768)])

def get_target_layer(target_layer):

    if (target_layer==4):
      result = [model1.block3]
    elif (target_layer==3):
      result = [model1.block2]
    elif (target_layer==2):
      result = [model1.block1]
    elif (target_layer==1):
      result = [model1.prep_block]
    else:
      result = [model1.block3]

    return result

grad_cam_call_list = [GradCAM(model=model1, target_layers=get_target_layer(i), use_cuda=(device == 'cuda')) for i in range(4)]

def classify_image(input_image, top_classes=3, grad_cam=True, target_layers=[2, 3], transparency=0.7):
    input_ = transform(input_image).unsqueeze(0)

    output = model1(input_)
    output = F.softmax(output.flatten(), dim=-1)

    confidences = [(classes[i], float(output[i])) for i in range(10)]
    confidences.sort(key=lambda x: x[1], reverse=True)
    confidences = OrderedDict(confidences[:top_classes])

    label = torch.argmax(output).item()

    results = []
    if grad_cam:
      for layer in target_layers:

        grad_cam = grad_cam_call_list[layer]
        targets = [ClassifierOutputTarget(label)]
        grayscale_cam = grad_cam(input_tensor=input_, targets=targets)
        grayscale_cam = grayscale_cam[0, :]
        output_image = show_cam_on_image(input_image / 255, grayscale_cam, use_rgb=True, image_weight=transparency)
        results.append((output_image, f"Layer {layer - 4}"))
    else:
      results.append((input_image, "Input"))

    return results, confidences

demo1 = gr.Interface(
    fn=classify_image,
    inputs=[
        gr.Image(shape=(32, 32), label="Input Image", value='test_images/cat.jpg'),
        gr.Slider(1, 10, value=3, step=1, label="Number of Top Classes"),
        gr.Checkbox(label="Show GradCAM?", value=True),
        #gr.Slider(-4, -1, value=-2, step=1, label="Which Layer?"),
        gr.CheckboxGroup(["-4", "-3", "-2", "-1"], value=["-2", "-1"], label="Which Network Layer(s)?", type='index'),
        gr.Slider(0, 1, value=0.7, label="Transparency", step=0.1)
    ],
    outputs=[gr.Gallery(label="Output Images", columns=2, rows=2),
             gr.Label(label='Top Classes')],
    examples=[[f'test_images/{k}.jpg'] for k in classes.values()]
)

def show_mis_classifications(num_examples=20, grad_cam=True, target_layer=-2, transparency=0.5):
    result = list()
    for index, row in mis_classified_df.iterrows():
        image = np.asarray(Image.open(f'misclassified_examples/{index}.jpg'))
        output_image, confidence = classify_image(image, top_classes=1, grad_cam=grad_cam, target_layers=[4+target_layer],
                                                    transparency=transparency)
        truth = row['ground_truths']
        predicted = list(confidence)[0]
        if truth != predicted:
            result.append((output_image[0][0], f"{row['ground_truths']} / {predicted}"))
        if len(result) >= num_examples:
            break
    return result

demo2 = gr.Interface(
    fn=show_mis_classifications,
    inputs=[
        gr.Number(value=20, minimum=1, maximum=len(mis_classified_df), label="No. of missclassified Examples", precision=0),
        gr.Checkbox(label="Show GradCAM?", value=True),
        gr.Slider(-4, -1, value=-2, step=1, label="Which Layer?"),
        gr.Slider(0, 1, value=0.7, label="Transparency", step=0.1),
    ],
    outputs=[gr.Gallery(label="Missclassified Images (Truth / Predicted)", columns=4)]
)

demo = gr.TabbedInterface([demo1, demo2], ["Examples", "Misclassified Examples"])
demo.launch(debug=True)