LambdaResNet reproduction

By Xuan Li, Xiao Feng, Changran Huang, Xiangying Wei Transfer from https://github.com/eziaowonder/Lambda-

link to paper, YouTube This project implements Lambda-Resnet and analyzes its performance on small datasets namely ImageNette, the subset of ImageNet. With the same training configurations, we compare the differences of both training and validation loss on this dataset, with different levels of noisy labels and model architecture.

Fullgrad of lambdaResNet50(pre-trained in ImageNet)(left) and ResNet50(pre-trained)(right)

Repo Structure

• data The imageNette data class and image augmentation transform.
• figure Some grad cam figures and their original image(from imageNette and gradCAM repo).
• grad_cam The grad CAM code(edit from grad CAM repo).
• layers Lambda layer and SE attention block.
• model The ResNet with lambda layer.
• models Our trained models.
• reports Our analysis paper.
• KaggleTraining.ipynb The original code we use for training our model.
• train.py
• test.ipynb Some tests during implementation.

How to run

• kaggle or colab(recommand)
  Using KaggleTraining.ipynb in kaggle or colab, search imagenette to obtain the dataset, then run all the sections.
• python train.py
  You can modified the Configs inside this file.

Results

available in wandb

Dataset structure

You can simply search imagenette in Kaggle if you use the KaggleTraining.ipynb. Alternatively, you can refer to the original repo

    ../input/imagenette/imagenette
    ├── train
    │   ├── n01440764
    │   ├── n02102040
    │   ├── n02979186
    │   ├── n03000684
    │   ├── n03028079
    │   ├── n03394916
    │   ├── n03417042
    │   ├── n03425413
    │   ├── n03445777
    │   └── n03888257
    ├── train_noisy_imagenette.csv
    ├── val
    │   ├── n01440764
    │   ├── n02102040
    │   ├── n02979186
    │   ├── n03000684
    │   ├── n03028079
    │   ├── n03394916
    │   ├── n03417042
    │   ├── n03425413
    │   ├── n03445777
    │   └── n03888257
    └── val_noisy_imagenette.csv

ref

• some training tricks
• lamdbda convolution
• ResNet code
• training framework ref
• pre-trained models
• grad CAM