Image-Classification-on-small-datasets-in-Pytorch

In this repo, we will learn how to finetune and feature extract the torchvision models, all of which have been pretrained on the 1000-class Imagenet dataset. This repo will give an indepth look at how to work with modern CNN architectures, and will build an intuition for finetuning any PyTorch model. Since each model architecture is different, there is no boilerplate finetuning code that will work in all scenarios. Rather, the researcher must look at the existing architecture and make custom adjustments for each model.

Original	Grad-CAM-Video

CNN and Vision Transformer models

Models	Paper Links
VGG	https://arxiv.org/abs/1409.1556
AlexNet	https://arxiv.org/abs/1803.01164
ResNet	https://arxiv.org/abs/1512.03385
Squeezenet	https://arxiv.org/abs/1602.07360v4
Densenet	https://arxiv.org/abs/1608.06993
Googlenet	https://arxiv.org/abs/1409.4842
Shufflenet	https://arxiv.org/abs/1707.01083
MNASNet	https://arxiv.org/abs/1807.11626
RegNet	https://arxiv.org/abs/2003.13678
ResNeXt	https://arxiv.org/abs/1611.05431
EfficientNet	https://arxiv.org/abs/1905.11946
EfficientNet_v2	https://arxiv.org/abs/2104.00298
ViT	https://arxiv.org/abs/2010.11929
Swin	https://arxiv.org/abs/2103.14030

Class names

rhino	buffalo

elephant	zebra

Quantitative results

Squeezenet:

Training Accuracy	Training Loss	Confusion Matrix

class names	precision	recall	f1-score	support
buffalo	0.93	0.97	0.95	113
elephant	0.91	0.93	0.92	113
rhino	0.94	0.91	0.93	113
zebra	0.98	0.96	0.97	113

accuracy			0.94	452
macro avg	0.94	0.94	0.94	452
weighted avg	0.94	0.94	0.94	452

Qualitative Results

Grad-CAM images

Models	rhino	buffalo	elephant	zebra
VGG
AlexNet
ResNet
Squeezenet
Densenet
Shufflenet
RegNet
ResNeXt
EfficientNet_v2

Using codes

split dataset

python split_dataset.py

  ├── Dataset folder 
    ├── train
        ├── class1
              ├── 1111.png
              ├── 2222.png
        ├── class2
              ├── 1111.png
              ├── 2222.png
    ├── val
        ├── class1
              ├── 1111.png
              ├── 2222.png
        ├── class2
              ├── 1111.png
              ├── 2222.png

Train model

python train.py

# uncomment one of the CNN models in train.py and run 

def train():
    data_loader = Dataloader(data_dir='dataset', image_size=IMAGE_SIZE, batch_size=32)

    '''---Choose model---'''

    model, params_to_update = cnn_models.initialize_model(model_name='vgg', num_classes=4, feature_extract=True)
    # model, params_to_update = cnn_models.initialize_model(model_name='densenet', num_classes=4, feature_extract=True)
    # model, params_to_update = cnn_models.initialize_model(model_name='resnet', num_classes=4, feature_extract=True)
    # model, params_to_update = cnn_models.initialize_model(model_name='squeezenet', num_classes=4, feature_extract=True)
    # model, params_to_update = cnn_models.initialize_model(model_name='alexnet', num_classes=4, feature_extract=True)
    # model, params_to_update = cnn_models.initialize_model(model_name='resnext', num_classes=4, feature_extract=True)
    # model, params_to_update = cnn_models.initialize_model(model_name='googlenet', num_classes=4, feature_extract=True)
    # model, params_to_update = cnn_models.initialize_model(model_name='mnasnet', num_classes=4, feature_extract=True)
    # model, params_to_update = cnn_models.initialize_model(model_name='regnet', num_classes=4, feature_extract=True)
    # model, params_to_update = cnn_models.initialize_model(model_name='shufflenet', num_classes=4, feature_extract=True)
    # model, params_to_update = cnn_models.initialize_model(model_name='efficientnet_b7', num_classes=4, feature_extract=True)
    # model, params_to_update = cnn_models.initialize_model(model_name='efficientnet_v2', num_classes=4, feature_extract=True)
    # model, params_to_update = cnn_models.initialize_model(model_name='ViT', num_classes=4, feature_extract=True)
    # model, params_to_update = cnn_models.initialize_model(model_name='Swin', num_classes=4, feature_extract=True)

Generate GRAD-CAM images

python grad_cam_results.py results/wild-animal-dataset.pth samples1/zebra_sample_image.jpg

# uncomment one of the CNN models in grad_cam_results.py and run 

model = models.vgg19(weights=None, num_classes=4)
# model = models.densenet161(weights=None, num_classes=4)
# model = models.resnet50(weights=None, num_classes=4)
# model = models.squeezenet1_0(weights=None, num_classes=4)
# model = models.alexnet(weights=None, num_classes=4)
# model = models.resnext101_64x4d(weights=None, num_classes=4)
# model = models.googlenet(weights=None, num_classes=4)
# model = models.mnasnet1_0(weights=None, num_classes=4)
# model = models.regnet_x_16gf(weights=None, num_classes=4)
# model = models.shufflenet_v2_x2_0(weights=None, num_classes=4)
# model = models.efficientnet_b7(weights=None, num_classes=4)
# model = models.efficientnet_v2_s(weights=None, num_classes=4)
# model = models.vit_l_16(weights=None, num_classes=4)
# model = models.swin_b(weights=None, num_classes=4)

P.S

Please give ⭐ to this repo if you find it useful. Stay tuned!

Harry-KIT/Image-Classification-on-small-datasets-in-Pytorch