app.py

import torch
import torchvision
from torchvision import transforms
import numpy as np
import gradio as gr
from PIL import Image
from pytorch_grad_cam import GradCAM
from pytorch_grad_cam.utils.image import show_cam_on_image
from resnet_lightning import ResNet18Model
import gradio as gr

model = ResNet18Model.load_from_checkpoint("epoch=19-step=3920.ckpt")

inv_normalize = transforms.Normalize(
    mean = [-0.50/0.23, -0.50/0.23, -0.50/0.23],
    std= [1/0.23, 1/0.23,1/0.23]
)

classes = ('plane', 'car', 'bird', 'cat', 'deer',
           'dog', 'frog', 'horse', 'ship', 'truck')

model_layer_names = ["1", "2", "3"]

def get_layer(layer_name):
    print("layer name:", layer_name)
    if layer_name == 1:
        return [model.layer1[-1]]
    elif layer_name == 2:
        return [model.layer2[-1]]
    elif layer_name == 3:
        return [model.layer3[-1]]
    else:
        return None

def resize_image_pil(image, new_width, new_height):
    img =  Image.fromarray(np.array(image))
    width, height = img.size

    width_scale = new_width/width
    height_scale = new_height/height
    scale = min(width_scale, height_scale)
    resized = img.resize((int(width*scale), int(height*scale)), Image.NEAREST)
    resized = resized.crop((0,0,new_width, new_height))

    return resized


def inference(input_img, show_gradcam, layer_name, num_classes, transparancy = 0.5):
    print(show_gradcam, layer_name, num_classes, transparancy)
    input_img = resize_image_pil(input_img,32,32)
    input_img = np.array(input_img)
    org_img = input_img
    
    input_img= input_img.reshape((32,32,3))
    transform = transforms.ToTensor()
    input_img = transform(input_img)
    input_img = input_img.unsqueeze(0)
    outputs = model(input_img)
    # print(outputs)
    softmax =  torch.nn.Softmax(dim=0)
    o = softmax(outputs.flatten())
    
    output_numpy = np.squeeze(np.asarray(outputs.detach().numpy()))
    index_sort = np.argsort(output_numpy)[::-1]
    
    confidences = {}
    for i in range(int(num_classes)):
        confidences[classes[index_sort[i]]] = float(o[index_sort[i]])

    
    prediction= torch.max(outputs, 1)

    if show_gradcam:
        target_layers = get_layer(layer_name)
        print("target layer",target_layers)
        cam = GradCAM(model=model, target_layers=target_layers)
        grayscale_cam = cam(input_tensor= input_img)
        grayscale_cam = grayscale_cam[0, :]
        visualization = show_cam_on_image(org_img/255,grayscale_cam,use_rgb=True,
                                     image_weight=transparancy)
    else:
        visualization = org_img

    
    return classes[int(prediction[0].item())], visualization, confidences
    
demo = gr.Interface(
    inference,
    inputs = [
        gr.Image(width=256,height=256,label="Input image"),
        gr.Number(value=3, maximum=10, minimum=1,step=1.0, precision=0,label="Number of classes to display"),
        gr.Checkbox(True, label="Show GradCAM Image"),
        gr.Dropdown(model_layer_names, value=3, label="Which layer for Gradcam"),
        gr.Slider(0, 1, value=0.5,label="Overall opacity of the overlay"),
    ],
    outputs = [
        gr.Label(label="Class", container=True, show_label= True),
        gr.Image(width= 256, height=256,label="Output Image"),
        gr.Label(label="Confidences", container=True, show_label= True),
    ],
    title = "CIFAR 10 trained on ResNet model in pytorch lightning with Gradcam",
    description = " A simple gradio inference to infer on resnet18 model",
    examples = [["cat.jpg",1, True, 10, 0.4]]
)

if __name__ == "__main__":
    demo.launch()