Adaptive-and-Iteratively-Improving-Recurrent-Lateral-Connections

An official Pytorch implementation of "Adaptive and Iteratively Improving Recurrent Lateral Connections" https://arxiv.org/abs/1910.11105

Prerequisites

• ubuntu18.04
• python 3.6
• torch==1.2
• torchvision==0.4
• numpy==1.17.4
• cv2
• tensorboard

Results

We expiremented our method on two tasks, five different datasets and six models, in this repo we will show how to reproduce our on Imagenet using Resnet50,Resnet110,Resnet20 (which will appear in the updated version of the paper) and MultiFiberNet2D.

ResNet20 and ResNet110 on Cifar10

This section is built on top of Yerlan Idelbayev's wonderful repo: https://github.com/akamaster/pytorch_resnet_cifar10.

ResNet20

Evaluate backbone:

Let's first evaluate the backbone using this cmdline

python trainer20.py --save_path <save_best_weight> --tb_filename <save_tensorboard_path> --pretrained_path pretrained_models/resnet20-12fca82f.th   --epochs 1 --pretrained --arch resnet20 --model resnet20 --evaluate --dataset_path <path2dataset> &  

Expected Top1: 91.73

Integrate block and FT:

Let's integrate the block and finetune the network

python trainer20.py --save_path <path> --tb_filename <tb_path> --pretrained_path pretrained_models/resnet20-12fca82f.th --epochs 200 --pretrained --num_loops 2 --arch resnet20 --model resnet20_feedback --lr 0.1 --alpha 0.0001 --original_weights pretrained_models/resnet20-12fca82f.th --dataset_path <path2dataset> 

Evaluate model

python trainer20.py --pretrained_path <weights2evaluate>  --epochs 200 --pretrained --num_loops 2 --arch resnet20 --model resnet20_feedback --alpha 0.0001 --original_weights pretrained_models/resnet20-12fca82f.th --dataset_path <path2dataset> --evaluate

A pretrained model can be evaluated:

python trainer20.py --pretrained_path pretrained_models/2loops_ep138_acc93.16.pth  --epochs 200 --pretrained --num_loops 2 --arch resnet20 --model resnet20_feedback --alpha 0.0001 --original_weights pretrained_models/resnet20-12fca82f.th --dataset_path <path2dataset> --evaluate

Expected Top1: 93.16

ResNet50 on ImageNet (Coming soon)

We will build our model on top of PyTorch Resnet50. First download the backbone weights from: Resnet50 Weights

Evaluate backbone:

Let's first evaluate PyTorch's Resnet50, expected results:76.13

python3 main_resnet.py --lr 0.0005 --num_loops 2  --batch-size 240 --pretrained_path resnet50-19c8e357.pth --model resnet --pretrained True --evaluate --pretrained True

Train Feedback:

python3 main_resnet.py --lr 0.0005 --num_loops 2  --batch-size 240 --pretrained_path res
net50-19c8e357.pth   --model resnet_feedback --pretrained True

Evaluate Feedback:

python3 main_resnet.py --lr 0.0005 --num_loops 2  --batch-size 240 --pretrained_path <path2weights> --model resnet_feedback --pretrained True --evaluate zx