[Re] Ablation-CAM

Reproduction of Ablation-CAM [paper], a technique to visualize and interpret features learnt by Convolutional Neural Networks (CNNs). Also check out my [critique report]

Quickstart

The main.py file facilitates running of all the functionality offered by the code. It accepts several command line arguments, which are explained below.

Argument	Possible values	Description
task	train, viz	Pass train to finetune a model on CIFAR-10 dataset or viz to visualize the localization maps of few images from a specified dataset.
model	vgg, resnet18, resnet50	Architecture to be used. Current options are vgg (VGG-16), resnet18 (ResNet-18) and resnet50 (ResNet-50).
data-root	data/cifar10, data/imagenet/train	Path to directory containing images to be used. Currently the code supports use of ImageNet and CIFAR-10 datasets, and they need to be organized inside main as follows: For CIFAR-10: Specify the value for this argument as data/cifar10. The torchvision version of the dataset will be downloaded inside main/data/cifar10 and will be ready to use. For ImageNet: I've used a smaller version of the dataset, called ImageNette which contains 10 handpicked classes out of 1000 with few images in each. Download the dataset from HERE and extract it inside main/data/imagenet such that train/ and val/ are available directly inside main/data/imagenet. Then pass data/imagenet/train to this argument.
model-ckpt	-	Path to model checkpoint to be loaded. This is used when visualization is being done on CIFAR-10.
batch-size	Default: 16	Batch size for training or visualization.
method	channel, spatial	Pass channel (default) to use the paper's method or spatial to perform location-wise ablation instead of channel-wise.
final-activation	relu, softmax	Activation function for the localization map after it is generated from the weighted sum. Pass relu (default) or softmax.
clf-activation	softmax, sigmoid	Activation function to convert classifier logits into probabilities. Pass softmax (default) or sigmoid.
num-batches	Default: 1	Number of batches to process for visualization task. Ignored for train task.

For example, to visualize the localization maps for one batch of ImageNette using a pretrained ResNet-18, run the following inside main/:

python3 main.py --task viz                          \
                --model resnet18                    \
                --data-root data/imagenet/train     \
                --batch-size 16                     \
                --method channel                    \
                --final-activation relu             \
                --clf-activation softmax            \
                --num-batches 1

As another example, to visualize the localization maps for one batch of CIFAR-10 using fine-tuned VGG-16 saved at main/outputs/models/vgg.ckpt, run the following inside main/:

python3 main.py --task viz                                    \
                --model vgg                                   \
                --data-root data/cifar10                      \  
                --batch-size 16                               \
                --method channel                              \
                --final-activation relu                       \
                --clf-activation softmax                      \
                --model-ckpt outputs/models/vgg.ckpt          \
                --num-batches 1

CIFAR-10 fine-tuned models

I have fine-tuned some models on CIFAR-10 to perform visualization on simpler images compared to ImageNet (classifier modified to have 10 outputs and trained for 10 epochs, images resized to 128 because Ablation-CAM fails on small images sizes). Each model achieved close to 93% validation accuracy. The trained weights can be found below.

• VGG-16 : Download
• ResNet-18 : Download
• ResNet-50 : Download