resnet.py

import torch
import torch.nn as nn
import torch.nn.functional as F
from torchsummary import summary
from io import BytesIO
import numpy as np
import os
from pytorch_lightning import LightningModule, Trainer
from torch import nn
from torch.nn import functional as F
from torch.utils.data import DataLoader, random_split
from torchmetrics import Accuracy
from torchvision import transforms
from torchvision.datasets import CIFAR10
from torch_lr_finder import LRFinder
import math
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 PIL import Image
import torch
from torch.utils.data import DataLoader, random_split
import torchvision.transforms as transforms
import torchvision.datasets as datasets
import pytorch_lightning as pl
import matplotlib.pyplot as plt



PATH_DATASETS = os.environ.get("PATH_DATASETS", ".")
BATCH_SIZE = 256


# Model
class custom_ResNet(pl.LightningModule):
    def __init__(self, data_dir=PATH_DATASETS, learning_rate=2e-4):
        super(custom_ResNet, self).__init__()

      # Set our init args as class attributes
      # Hardcode some dataset specific attributes
        self.data_dir = data_dir
        self.learning_rate = learning_rate
        self.classes = ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']
        self.num_classes = 10
        self.train_transform = transforms.Compose([
            transforms.RandomCrop(32, padding=4),
            transforms.RandomHorizontalFlip(),
            transforms.ToTensor(),  # Convert PIL image to tensor
            transforms.Normalize((0.4914, 0.4822, 0.4465), (0.247, 0.243, 0.261))
        ])

        self.test_transform = transforms.Compose([
            transforms.ToTensor(),  # Convert PIL image to tensor
            transforms.Normalize((0.4914, 0.4822, 0.4465), (0.247, 0.243, 0.261))
        ])

        # Define PyTorch model
        # PREPARATION BLOCK
        self.prepblock = nn.Sequential(
            nn.Conv2d(in_channels=3, out_channels=64, kernel_size=(3, 3), padding=1, dilation=1, stride=1, bias=False),
            nn.ReLU(),nn.BatchNorm2d(64))
            # output_size = 32, RF=3


        # CONVOLUTION BLOCK 1
        self.convblock1_l1 = nn.Sequential(
            nn.Conv2d(in_channels=64, out_channels=128, kernel_size=(3, 3), padding=1, dilation=1, stride=1, bias=False),
            # output_size = 32, RF=5
            nn.MaxPool2d(2, 2),nn.ReLU(),nn.BatchNorm2d(128))
            # output_size = 16, RF=6

        self.convblock1_r1 = nn.Sequential(
            nn.Conv2d(in_channels=128, out_channels=128, kernel_size=(3, 3), padding=1, dilation=1, stride=1, bias=False),
            nn.ReLU(),nn.BatchNorm2d(128),
            # output_size = 16, RF=10
            nn.Conv2d(in_channels=128, out_channels=128, kernel_size=(3, 3), padding=1, dilation=1, stride=1, bias=False),
            nn.ReLU(),nn.BatchNorm2d(128))
            # output_size = 16, RF=14


        # CONVOLUTION BLOCK 2
        self.convblock2_l1 = nn.Sequential(
            nn.Conv2d(in_channels=128, out_channels=256, kernel_size=(3, 3), padding=1, dilation=1, stride=1, bias=False),
            # output_size = 16, RF=18
            nn.MaxPool2d(2, 2),nn.ReLU(),nn.BatchNorm2d(256))
            # output_size = 8, RF=20


        # CONVOLUTION BLOCK 3
        self.convblock3_l1 = nn.Sequential(
            nn.Conv2d(in_channels=256, out_channels=512, kernel_size=(3, 3), padding=1, dilation=1, stride=1, bias=False),
            # output_size = 8, RF=28
            nn.MaxPool2d(2, 2),
            nn.ReLU(),nn.BatchNorm2d(512))
            # output_size = 4, RF=32


        self.convblock3_r2 = nn.Sequential(
            nn.Conv2d(in_channels=512, out_channels=512, kernel_size=(3, 3), padding=1, dilation=1, stride=1, bias=False),
            nn.ReLU(),nn.BatchNorm2d(512),
             # output_size = 4, RF=48
            nn.Conv2d(in_channels=512, out_channels=512, kernel_size=(3, 3), padding=1, dilation=1, stride=1, bias=False),
            nn.ReLU(),nn.BatchNorm2d(512))
            # output_size = 4, RF=64


        # CONVOLUTION BLOCK 4
        self.convblock4_mp = nn.Sequential(nn.MaxPool2d(4))
        # output_size = 1, RF = 88


        # OUTPUT BLOCK - Fully Connected layer
        self.output_block = nn.Sequential(nn.Linear(in_features=512, out_features=10, bias=False))
        # output_size = 1, RF = 88


    def forward(self, x):

        # Preparation Block
        x1 = self.prepblock(x)

        # Convolution Block 1
        x2 = self.convblock1_l1(x1)
        x3 = self.convblock1_r1(x2)
        x4 = x2 + x3

        # Convolution Block 2
        x5 = self.convblock2_l1(x4)

        # Convolution Block 3
        x6 = self.convblock3_l1(x5)
        x7 = self.convblock3_r2(x6)
        x8 = x7 + x6

        # Convolution Block 4
        x9 = self.convblock4_mp(x8)

        # Output Block
        x9 = x9.view(x9.size(0), -1)
        x10 = self.output_block(x9)
        return F.log_softmax(x10, dim=1)

    def training_step(self, batch, batch_idx):
        x, y = batch
        y_hat = self.forward(x)
        loss = F.cross_entropy(y_hat, y)
        pred = y_hat.argmax(dim=1, keepdim=True)
        acc = pred.eq(y.view_as(pred)).float().mean()
        self.log('train_loss', loss, on_step=True, on_epoch=True, prog_bar=True)
        self.log('train_acc', acc, on_step=True, on_epoch=True, prog_bar=True)
        return loss

    def validation_step(self, batch, batch_idx):
        x, y = batch
        y_hat = self.forward(x)
        loss = F.cross_entropy(y_hat, y)
        pred = y_hat.argmax(dim=1, keepdim=True)
        acc = pred.eq(y.view_as(pred)).float().mean()
        self.log('val_loss', loss, prog_bar=True)
        self.log('val_acc', acc, prog_bar=True)
        return loss

    def test_step(self, batch, batch_idx):
        x, y = batch
        y_hat = self.forward(x)
        loss = F.cross_entropy(y_hat, y)
        pred = y_hat.argmax(dim=1, keepdim=True)
        acc = pred.eq(y.view_as(pred)).float().mean()
        self.log('test_loss', loss, prog_bar=True)
        self.log('test_acc', acc, prog_bar=True)
        return pred  # Return predictions instead of loss

    def configure_optimizers(self):
        optimizer = torch.optim.Adam(self.parameters(), lr=0.001)
        return optimizer


    ####################
    # DATA RELATED HOOKS
    ####################

    def prepare_data(self):
        # download
        CIFAR10(self.data_dir, train=True, download=True)
        CIFAR10(self.data_dir, train=False, download=True)

    def setup(self, stage=None):

        # Assign train/val datasets for use in dataloaders
        if stage == "fit" or stage is None:
            cifar_full = CIFAR10(self.data_dir, train=True, transform=self.train_transform)
            self.cifar_train, self.cifar_val = random_split(cifar_full, [45000, 5000])

        # Assign test dataset for use in dataloader(s)
        if stage == "test" or stage is None:
            self.cifar_test = CIFAR10(self.data_dir, train=False, download=True, transform=self.test_transform)

    def train_dataloader(self):
        return DataLoader(self.cifar_train, batch_size=BATCH_SIZE, num_workers=os.cpu_count())

    def val_dataloader(self):
        return DataLoader(self.cifar_val, batch_size=BATCH_SIZE, num_workers=os.cpu_count())

    def test_dataloader(self):
        return DataLoader(self.cifar_test, batch_size=BATCH_SIZE, num_workers=os.cpu_count())

    def collect_misclassified_images(self, num_images):
        misclassified_images = []
        misclassified_true_labels = []
        misclassified_predicted_labels = []
        num_collected = 0

        for batch in self.test_dataloader():
            x, y = batch
            y_hat = self.forward(x)
            pred = y_hat.argmax(dim=1, keepdim=True)
            misclassified_mask = pred.eq(y.view_as(pred)).squeeze()
            misclassified_images.extend(x[~misclassified_mask].detach())  # Detach here to avoid CPU transfer
            misclassified_true_labels.extend(y[~misclassified_mask].detach())  # Detach here to avoid CPU transfer
            misclassified_predicted_labels.extend(pred[~misclassified_mask].detach())  # Detach here to avoid CPU transfer

            num_collected += sum(~misclassified_mask)

            if num_collected >= num_images:
                break

        return misclassified_images[:num_images], misclassified_true_labels[:num_images], misclassified_predicted_labels[:num_images], len(misclassified_images)


    def normalize_image(self, img_tensor):
        min_val = img_tensor.min()
        max_val = img_tensor.max()
        return (img_tensor - min_val) / (max_val - min_val)

    def get_gradcam_images(self, target_layer=-1, transparency=0.5, num_images=10):
        misclassified_images, true_labels, predicted_labels, num_misclassified = self.collect_misclassified_images(num_images)
        count = 0
        k = 0
        misclassified_images_converted = list()
        gradcam_images = list()

        if target_layer == -2:
          target_layer = self.convblock2_l1.cpu()
        else:
          target_layer = self.convblock3_l1.cpu()

        dataset_mean, dataset_std = np.array([0.49139968, 0.48215841, 0.44653091]), np.array([0.24703223, 0.24348513, 0.26158784])
        grad_cam = GradCAM(model=self.cpu(), target_layers=target_layer, use_cuda=False)  # Move model to CPU

        for i in range(0, num_images):
            img_converted = misclassified_images[i].cpu().numpy().transpose(1, 2, 0)  # Convert tensor to numpy and transpose to (H, W, C)
            img_converted = dataset_std * img_converted + dataset_mean
            img_converted = np.clip(img_converted, 0, 1)
            misclassified_images_converted.append(img_converted)
            targets = [ClassifierOutputTarget(true_labels[i])]
            grayscale_cam = grad_cam(input_tensor=misclassified_images[i].unsqueeze(0).cpu(), targets=targets)  # Move input to CPU
            grayscale_cam = grayscale_cam[0, :]
            output = show_cam_on_image(img_converted, grayscale_cam, use_rgb=True, image_weight=transparency)
            gradcam_images.append(output)

        return gradcam_images

    # Add a 'use_gradcam' parameter to the show_misclassified_images function
    def show_misclassified_images(self, num_images=10, use_gradcam=False, gradcam_layer=-1, transparency=0.5):
        misclassified_images, true_labels, predicted_labels, num_misclassified = self.collect_misclassified_images(num_images)
          
        # Create subplots based on the number of columns required
        num_rows = num_images
        num_cols = 2 if use_gradcam else 1  # Show GradCAM images side by side with misclassified images if 'use_gradcam' is True

        fig, axs = plt.subplots(num_rows, num_cols, figsize=(8, 5 * num_rows))

        if use_gradcam:
            grad_cam_images = self.get_gradcam_images(target_layer=gradcam_layer, transparency=transparency, num_images=num_images)

        for i in range(num_images):
            img = misclassified_images[i].numpy().transpose((1, 2, 0))  # Convert tensor to numpy and transpose to (H, W, C)
            img = self.normalize_image(img)  # Normalize the image

            if num_cols > 1:  # Use multiple columns for subplots
                axs[i, 0].imshow(img)
                axs[i, 0].set_title(f"True label: {self.classes[true_labels[i]]}\nPredicted: {self.classes[predicted_labels[i]]}")
                axs[i, 0].axis("off")

                if use_gradcam:
                    # gradcam_img = grad_cam_images[i].numpy().transpose((1, 2, 0))  # Convert tensor to numpy and transpose to (H, W, C)
                    gradcam_img = self.normalize_image(grad_cam_images[i])  # Normalize the image
                    axs[i, 1].imshow(gradcam_img)
                    axs[i, 1].set_title("GradCAM")
                    axs[i, 1].axis("off")
            else:  # Use a single column for subplots
                axs[i].imshow(img)
                axs[i].set_title(f"True label: {self.classes[true_labels[i]]}\nPredicted: {self.classes[predicted_labels[i]]}")
                axs[i].axis("off")

        fig.tight_layout()
        return fig