Grad-CAM from Scratch with PyTorch Hooks

automobile stops abruptly. Worryingly, there is no such thing as a cease check in sight. The engineers can solely make guesses as to why the automobile’s neural community grew to become confused. It might be a tumbleweed rolling throughout the road, a automobile coming down the opposite lane or the crimson billboard within the background. To seek out the true motive, they flip to Grad-CAM [1].

Grad-CAM is an explainable AI (XAI) approach that helps reveal why a convolutional neural community (CNN) made a specific determination. The tactic produces a heatmap that highlights the areas in a picture which might be an important for a prediction. For our self-driving automobile instance, this might present if the pixels from the weed, automobile or billboard precipitated the automobile to cease.

Now, Grad-CAM is one in every of many XAI strategies for Pc Imaginative and prescient. Attributable to its pace, flexibility and reliability, it has shortly turn into one of the in style. It has additionally impressed many associated strategies. So, if you’re excited by XAI, it’s value understanding precisely how this methodology works. To try this, we will likely be implementing Grad-CAM from scratch utilizing Python.

Particularly, we will likely be counting on PyTorch Hooks. As you will notice, these enable us to dynamically extract gradients and activations from a community throughout ahead and backwards passes. These are sensible expertise that won’t solely can help you implement Grad-CAM but in addition any gradient-based XAI methodology. See the total challenge on GitHub.

The speculation behind Grad-CAM

Earlier than we get to the code, it’s value bearing on the idea behind Grad-CAM. If you would like a deep dive, then take a look at the video beneath. If you wish to find out about different strategies, then see this free XAI for Pc Imaginative and prescient course.

To summarise, when creating Grad-CAM heatmaps, we begin with a educated CNN. We then do a ahead go by this community with a single pattern picture. This may activate all convolutional layers within the community. We name these function maps ($A^ok$). They are going to be a set of 2D matrices that comprise totally different options detected within the pattern picture.

With Grad-CAM, we’re sometimes within the maps from the final convolutional layer of the community. After we apply the strategy to VGG16, you will notice that its closing layer has 512 function maps. We use these as they comprise options with essentially the most detailed semantic data whereas nonetheless retaining spatial data. In different phrases, they inform us what was used for a prediction and the place within the picture it was taken from.

The issue is that these maps additionally comprise options which might be necessary for different courses. To mitigate this, we observe the method proven in Determine 1. As soon as we’ve the function maps ($A^ok$), we weight them by how necessary they’re to the category of curiosity ($y_c$). We do that utilizing $a_k^c$ — the typical gradient of the rating for $y_c$ w.r.t. to the weather within the function map. We then do element-wise summation. For VGG16, you will notice we go from 512 maps of 14×14 pixels to a single 14×14 map.

The gradients for a person factor ($frac{partial y^c}{partial A_{ij}^ok}$) inform us how a lot the rating will change with a small change within the factor. Which means that giant common gradients point out that your complete function map was necessary and will contribute extra to the ultimate heatmap. So, after we weight and sum the maps, those that comprise options for different courses will seemingly contribute much less.

The ultimate steps are to use the ReLU activation operate to make sure all unfavourable parts can have a price of zero. Then we upsample with interpolation so the heatmap has the identical dimensions because the pattern picture. The ultimate map is summarised by the method beneath. You may recognise it from the Grad-CAM paper [1].

$$ L_{Grad-CAM}^c = ReLUleft( sum_{ok} a_k^c A^ok proper) $$

Grad-CAM from Scratch

Don’t fear if the idea just isn’t utterly clear. We’ll stroll by it step-by-step as we apply the strategy from scratch. You could find the total challenge on GitHub. To begin, we’ve our imports beneath. These are all widespread imports for pc imaginative and prescient issues.

import matplotlib.pyplot as plt
import numpy as np

import cv2
from PIL import Picture

import torch
import torch.nn.purposeful as F
from torchvision import fashions, transforms

import urllib.request

Load pretrained mannequin from PyTorch

We’ll be making use of Grad-CAM to VGG16 pretrained on ImageNet. To assist, we’ve the 2 features beneath. The primary will format a picture within the appropriate means for enter into the mannequin. The normalisation values used are the imply and commonplace deviation of the pictures in ImageNet. The 224×224 measurement can be commonplace for ImageNet fashions.

def preprocess_image(img_path):

    """Load and preprocess photos for PyTorch fashions."""

    img = Picture.open(img_path).convert("RGB")

    #Transforms utilized by imagenet fashions
    remodel = transforms.Compose([
        transforms.Resize((224, 224)),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])

    return remodel(img).unsqueeze(0)

ImageNet has many courses. The second operate will format the output of the mannequin so we show the courses with the best predicted possibilities.

def display_output(output,n=5):

    """Show the highest n classes predicted by the mannequin."""
    
    # Obtain the classes
    url = "https://uncooked.githubusercontent.com/pytorch/hub/grasp/imagenet_classes.txt"
    urllib.request.urlretrieve(url, "imagenet_classes.txt")

    with open("imagenet_classes.txt", "r") as f:
        classes = [s.strip() for s in f.readlines()]

    # Present high classes per picture
    possibilities = torch.nn.purposeful.softmax(output[0], dim=0)
    top_prob, top_catid = torch.topk(possibilities, n)

    for i in vary(top_prob.measurement(0)):
        print(classes[top_catid[i]], top_prob[i].merchandise())

    return top_catid[0]

We now load the pretrained VGG16 mannequin (line 2), transfer it to a GPU (strains 5-8) and set it to analysis mode (line 11). You may see a snippet of the mannequin output in Determine 2. VGG16 is manufactured from 16 weighted layers. Right here, you possibly can see the final 2 of 13 convolutional layers and the three absolutely linked layers.

# Load the pre-trained mannequin (e.g., VGG16)
mannequin = fashions.vgg16(pretrained=True)

# Set the mannequin to gpu
gadget = torch.gadget('mps' if torch.backends.mps.is_built() 
                      else 'cuda' if torch.cuda.is_available() 
                      else 'cpu')
mannequin.to(gadget)

# Set the mannequin to analysis mode
mannequin.eval()

The names you see in Determine 2 are necessary. Later, we’ll use them to reference a particular layer within the community to entry its activations and gradients. Particularly, we’ll use mannequin.options[28]. That is the ultimate convolutional layer within the community. As you possibly can see within the snapshot, this layer accommodates 512 function maps.

Ahead go with pattern picture

We will likely be explaining a prediction from this mannequin. To do that, we’d like a pattern picture that will likely be fed into the mannequin. We downloaded one from Wikipedia Commons (strains 2-3). We then load it (strains 5-6), crop it to have equal top and width (line 7) and show it (strains 9-10). In Determine 3, you possibly can see we’re utilizing a picture of a whale shark in an aquarium.

# Load a pattern picture from the net
img_url = "https://add.wikimedia.org/wikipedia/commons/thumb/a/a1/Male_whale_shark_at_Georgia_Aquarium.jpg/960px-Male_whale_shark_at_Georgia_Aquarium.jpg"
urllib.request.urlretrieve(img_url, "sample_image.jpg")[0]

img_path = "sample_image.jpg"
img = Picture.open(img_path).convert("RGB")
img = img.crop((320, 0, 960, 640))  # Crop to 640x640

plt.imshow(img)
plt.axis("off")

ImageNet has no devoted class for whale sharks, so it will likely be fascinating to see what the mannequin predicts. To do that, we begin by processing our picture (line 2) and transferring it to the GPU (line 3). We then do a ahead go to get a prediction (line 6) and show the highest 5 possibilities (line 7). You may see these in Determine 4.

# Preprocess the picture
img_tensor = preprocess_image(img_path)
img_tensor = img_tensor.to(gadget)

# Ahead go
predictions = mannequin(img_tensor)
display_output(predictions,n=5)

Given the out there courses, these appear affordable. They’re all marine life and the highest two are sharks. Now, let’s see how we will clarify this prediction. We wish to perceive what areas of the picture contribute essentially the most to the best predicted class — hammerhead.

PyTorch hooks naming conventions

Grad-CAM heatmaps are created utilizing each activations from a ahead go and gradients from a backwards go. To entry these, we’ll use PyTorch hooks. These are features that can help you save the inputs and outputs of a layer. We gained’t do it right here, however they even can help you alter these features. For instance, Guided Backpropagation might be utilized by guaranteeing solely constructive gradients are propagated utilizing a backwards hook.

You may see some examples of those features beneath. A forwards_hook will likely be known as throughout a ahead go. It will likely be registered on a given module (i.e. layer). By default, the operate receives three arguments — the module, its enter and its output. Equally, a backwards_hook is triggered throughout a backwards go with the module and gradients of the enter and output.

# Instance of a forwards hook operate
def fowards_hook(module, enter, output):
    """Parameters:
            module (nn.Module): The module the place the hook is utilized.
            enter (tuple of Tensors): Enter to the module.
            output (Tensor): Output of the module."""
    ...

# Instance of a backwards hook operate 
def backwards_hook(module, grad_in, grad_out):
    """Parameters:
            module (nn.Module): The module the place the hook is utilized.
            grad_in (tuple of Tensors): Gradients w.r.t. the enter of the module.
            grad_out (tuple of Tensors): Gradients w.r.t. the output of the module."""
    ...

To keep away from confusion, let’s make clear the parameter names utilized by these features. Check out the overview of the usual backpropagation process for a convolutional layer in Determine 5. This layer consists of a set of kernels, $Ok$, and biases, $b$. The opposite elements are the:

• enter – a set of function maps or a picture
• output – set of function maps
• grad_in is the gradient of the loss w.r.t. the layer’s enter.
• grad_out is the gradient of the loss w.r.t. the layer’s output.

We have now labelled these utilizing the identical names of the arguments used to name the hook features that we apply later.

Consider, we gained’t use the gradients in the identical means as backpropagation. Normally, we use the gradients of a batch of photos to replace $Ok$ and $b$. Now, we’re solely excited by grad_out of a single pattern picture. This may give us the gradients of the weather within the layer’s function maps. In different phrases, the gradients we use to weight the function maps.

Activations with PyTorch ahead hook

Our VGG16 community has been created utilizing ReLU with inplace=True. These modify tensors in reminiscence, so the unique values are misplaced. That’s, tensors used as enter are overwritten by the ReLU operate. This may result in issues when making use of hooks, as we might have the unique enter. So we use the code beneath to exchange all ReLU features with inplace=False ones. This won’t affect the output of the mannequin, however it’s going to improve its reminiscence utilization.

# Change all in-place ReLU activations with out-of-place ones
def replace_relu(mannequin):

    for title, youngster in mannequin.named_children():
        if isinstance(youngster, torch.nn.ReLU):
            setattr(mannequin, title, torch.nn.ReLU(inplace=False))
            print(f"Changing ReLU activation in layer: {title}")
        else:
            replace_relu(youngster)  # Recursively apply to submodules

# Apply the modification to the VGG16 mannequin
replace_relu(mannequin)

Under we’ve our first hook operate — save_activations. This may append the output from a module (line 6) to a listing of activations (line 2). In our case, we’ll solely register the hook onto one module (i.e. the final convolutional layer), so this listing will solely comprise one factor. Discover how we format the output (line 6). We detach it from the computational graph so the community just isn’t affected. We additionally format them as a numpy array and squeeze the batch dimension.

# Record to retailer activations
activations = []

# Perform to save lots of activations
def save_activations(module, enter, output):
    activations.append(output.detach().cpu().numpy().squeeze())

To make use of the hook operate, we register it on the final convolutional layer — mannequin.options[28]. That is executed utilizing the register_forward_hook operate.

# Register the hook to the final convolutional layer
hook = mannequin.options[28].register_forward_hook(save_activations)

Now, after we do a ahead go (line 2), the save_activations hook operate will likely be known as for this layer. In different phrases, its output will likely be saved to the activations listing.

# Ahead go by the mannequin to get activations
prediction = mannequin(img_tensor)

Lastly, it’s good observe to take away the hook operate when it’s now not wanted (line 2). This implies the ahead hook operate won’t be triggered if we do one other ahead go.

# Take away the hook after use
hook.take away()  

The form of those activations is (512, 14, 14). In different phrases, we’ve 512 function maps and every map is 14×14 pixels. You may see some examples of those in Determine 6. A few of these maps could comprise options necessary for different courses or those who lower the likelihood of the anticipated class. So let’s see how we will discover gradients to assist establish an important maps.

act_shape = np.form(activations[0])
print(f"Form of activations: {act_shape}") # (512, 14, 14)

Gradients with PyTorch backwards hooks

To get gradients, we observe an identical course of to earlier than. The important thing distinction is that we now use the register_full_backward_hook to register the save_gradients operate (line 7). This may be certain that it’s known as throughout a backwards go. Importantly, we do the backwards go (line 16) from the output for the category with the best rating (line 13). This successfully units the rating for this class to 1 and all different scores to 0. In different phrases, we get the gradients of the hammerhead class w.r.t. to the weather of the function maps.

gradients = []

def save_gradient(module, grad_in, grad_out):
    gradients.append(grad_out[0].cpu().numpy().squeeze())

# Register the backward hook on a convolutional layer
hook = mannequin.options[28].register_full_backward_hook(save_gradient)

# Ahead go
output = mannequin(img_tensor)

# Decide the category with highest rating
rating = output[0].max()

# Backward go from the rating
rating.backward()

# Take away the hook after use
hook.take away()

We can have a gradient for each factor of the function maps. So, once more, the form is (512, 14, 14). Determine 7 visualises a few of these. You may see some are inclined to have larger values. Nevertheless, we’re not so involved with the person gradients. After we create a Grad-CAM heatmap, we’ll use the typical gradient of every function map.

grad_shape = np.form(gradients[0])
print(f"Form of gradients: {grad_shape}") # (512, 14, 14)

Lastly, earlier than we transfer on, it’s good observe to reset the mannequin’s gradients (line 2). That is notably necessary in case you plan to run the code for a number of photos, as gradients might be amassed with every backwards go.

# Reset gradients
mannequin.zero_grad() 

Creating Grad-CAM heatmaps

First, we discover the imply gradients for every function map. There will likely be 512 of those common gradients. Plotting a histogram of them, you possibly can see most are typically round 0. In different phrases, these don’t have a lot affect on the anticipated rating. There are a number of that are inclined to have a unfavourable affect and a constructive affect. It’s these function maps we wish to give extra weight to.

# Step 1: mixture the gradients
gradients_aggregated = np.imply(gradients[0], axis=(1, 2))

We mix all of the activations by doing element-wise summation (strains 2-4). After we do that, we weight every function map by its common gradient (line 3). Ultimately, we can have one 14×14 array.

# Step 2: weight the activations by the aggregated gradients and sum them up
weighted_activations = np.sum(activations[0] * 
                              gradients_aggregated[:, np.newaxis, np.newaxis], 
                              axis=0)

These weighted activations will comprise each constructive and unfavourable pixels. We are able to contemplate the unfavourable pixels to be suppressing the anticipated rating. In different phrases, a rise within the worth of those areas tends to lower the rating. Since we’re solely within the constructive contributions—areas that help the category prediction—we apply a ReLU activation to the ultimate heatmap (line 2). You may see the distinction within the heatmaps in Determine 9.

# Step 3: ReLU summed activations
relu_weighted_activations = np.most(weighted_activations, 0)

You may see the heatmap in Determine 9 is kind of coarse. It will be extra helpful if it had the size of the unique picture. For this reason the final step for creating Grad-CAM heatmaps is to upsample to the dimension of the enter picture (strains 2-4). On this case, we’ve a 224×224 picture.

#Step 4: Upsample the heatmap to the unique picture measurement
upsampled_heatmap = cv2.resize(relu_weighted_activations, 
                               (img_tensor.measurement(3), img_tensor.measurement(2)), 
                               interpolation=cv2.INTER_LINEAR)

print(np.form(upsampled_heatmap))  # Needs to be (224, 224)

Determine 10 provides us our closing visualisation. We show the pattern picture (strains 5-7) subsequent to the heatmap (strains 10-15). For the latter, we create a transparent visualisation with the assistance of Canny Edge detection (line 10). This offers us an edge map (i.e. define) of the pattern picture. We are able to then overlay the heatmap on high of this (line 14).

# Step 5: visualise the heatmap
fig, ax = plt.subplots(1, 2, figsize=(8, 8))

# Enter picture
resized_img = img.resize((224, 224))
ax[0].imshow(resized_img)
ax[0].axis("off")

# Edge map for the enter picture
edge_img = cv2.Canny(np.array(resized_img), 100, 200)
ax[1].imshow(255-edge_img, alpha=0.5, cmap='grey')

# Overlay the heatmap 
ax[1].imshow(upsampled_heatmap, alpha=0.5, cmap='coolwarm')
ax[1].axis("off")

our Grad-CAM heatmap, there may be some noise. Nevertheless, it seems the mannequin is counting on the tail fin and, to a lesser extent, the pectoral fin to make its predictions. It’s beginning to make sense why the mannequin labeled this shark as a hammerhead. Maybe each animals share these traits.

For some additional investigation, we apply the identical course of however now utilizing an precise picture of a hammerhead. On this case, the mannequin seems to be counting on the identical options. This can be a bit regarding. Would we not count on the mannequin to make use of one of many shark’s defining options— the hammerhead? Finally, this will lead VGG16 to confuse various kinds of sharks.

With this instance, we see how Grad-CAM can spotlight potential flaws in our mannequin. We can’t solely get their predictions but in addition perceive how they made them. We are able to perceive if the options used will result in unexpected predictions down the road. This may doubtlessly save us lots of time, cash and within the case of extra consequential purposes, lives!

If you wish to be taught extra about XAI for CV take a look at one in every of these articles. Or see this Free XAI for CV course.