Deep Residual Learning for Image Recognition

Overview

Deep Residual Learning for Image Recognition

This is a Torch implementation of "Deep Residual Learning for Image Recognition",Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun the winners of the 2015 ILSVRC and COCO challenges.

What's working: CIFAR converges, as per the paper.

What's not working yet: Imagenet. I also have only implemented Option (A) for the residual network bottleneck strategy.

Table of contents

Changes

  • 2016-02-01: Added others' preliminary results on ImageNet for the architecture. (I haven't found time to train ImageNet yet)
  • 2016-01-21: Completed the 'alternate solver' experiments on deep networks. These ones take quite a long time.
  • 2016-01-19:
    • New results: Re-ran the 'alternate building block' results on deeper networks. They have more of an effect.
    • Added a table of contents to avoid getting lost.
    • Added experimental artifacts (log of training loss and test error, the saved model, the any patches used on the source code, etc) for two of the more interesting experiments, for curious folks who want to reproduce our results. (These artifacts are hereby released under the zlib license.)
  • 2016-01-15:
    • New CIFAR results: I re-ran all the CIFAR experiments and updated the results. There were a few bugs: we were only testing on the first 2,000 images in the training set, and they were sampled with replacement. These new results are much more stable over time.
  • 2016-01-12: Release results of CIFAR experiments.

How to use

  • You need at least CUDA 7.0 and CuDNN v4.
  • Install Torch.
  • Install the Torch CUDNN V4 library: git clone https://github.com/soumith/cudnn.torch; cd cudnn; git co R4; luarocks make This will give you cudnn.SpatialBatchNormalization, which helps save quite a lot of memory.
  • Install nninit: luarocks install nninit.
  • Download CIFAR 10. Use --dataRoot to specify the location of the extracted CIFAR 10 folder.
  • Run train-cifar.lua.

CIFAR: Effect of model size

For this test, our goal is to reproduce Figure 6 from the original paper:

figure 6 from original paper

We train our model for 200 epochs (this is about 7.8e4 of their iterations on the above graph). Like their paper, we start at a learning rate of 0.1 and reduce it to 0.01 at 80 epochs and then to 0.01 at 160 epochs.

Training loss

Training loss curve

Testing error

Test error curve

Model My Test Error Reference Test Error from Tab. 6 Artifacts
Nsize=3, 20 layers 0.0829 0.0875 Model, Loss and Error logs, Source commit + patch
Nsize=5, 32 layers 0.0763 0.0751 Model, Loss and Error logs, Source commit + patch
Nsize=7, 44 layers 0.0714 0.0717 Model, Loss and Error logs, Source commit + patch
Nsize=9, 56 layers 0.0694 0.0697 Model, Loss and Error logs, Source commit + patch
Nsize=18, 110 layers, fancy policy¹ 0.0673 0.0661² Model, Loss and Error logs, Source commit + patch

We can reproduce the results from the paper to typically within 0.5%. In all cases except for the 32-layer network, we achieve very slightly improved performance, though this may just be noise.

¹: For this run, we started from a learning rate of 0.001 until the first 400 iterations. We then raised the learning rate to 0.1 and trained as usual. This is consistent with the actual paper's results.

²: Note that the paper reports the best run from five runs, as well as the mean. I consider the mean to be a valid test protocol, but I don't like reporting the 'best' score because this is effectively training on the test set. (This method of reporting effectively introduces an extra parameter into the model--which model to use from the ensemble--and this parameter is fitted to the test set)

CIFAR: Effect of model architecture

This experiment explores the effect of different NN architectures that alter the "Building Block" model inside the residual network.

The original paper used a "Building Block" similar to the "Reference" model on the left part of the figure below, with the standard convolution layer, batch normalization, and ReLU, followed by another convolution layer and batch normalization. The only interesting piece of this architecture is that they move the ReLU after the addition.

We investigated two alternate strategies.

Three different alternate CIFAR architectures

  • Alternate 1: Move batch normalization after the addition. (Middle) The reasoning behind this choice is to test whether normalizing the first term of the addition is desirable. It grew out of the mistaken belief that batch normalization always normalizes to have zero mean and unit variance. If this were true, building an identity building block would be impossible because the input to the addition always has unit variance. However, this is not true. BN layers have additional learnable scale and bias parameters, so the input to the batch normalization layer is not forced to have unit variance.

  • Alternate 2: Remove the second ReLU. The idea behind this was noticing that in the reference architecture, the input cannot proceed to the output without being modified by a ReLU. This makes identity connections technically impossible because negative numbers would always be clipped as they passed through the skip layers of the network. To avoid this, we could either move the ReLU before the addition or remove it completely. However, it is not correct to move the ReLU before the addition: such an architecture would ensure that the output would never decrease because the first addition term could never be negative. The other option is to simply remove the ReLU completely, sacrificing the nonlinear property of this layer. It is unclear which approach is better.

To test these strategies, we repeat the above protocol using the smallest (20-layer) residual network model.

(Note: The other experiments all use the leftmost "Reference" model.)

Training loss

Testing error

Architecture Test error
ReLU, BN before add (ORIG PAPER reimplementation) 0.0829
No ReLU, BN before add 0.0862
ReLU, BN after add 0.0834
No ReLU, BN after add 0.0823

All methods achieve accuracies within about 0.5% of each other. Removing ReLU and moving the batch normalization after the addition seems to make a small improvement on CIFAR, but there is too much noise in the test error curve to reliably tell a difference.

CIFAR: Effect of model architecture on deep networks

The above experiments on the 20-layer networks do not reveal any interesting differences. However, these differences become more pronounced when evaluated on very deep networks. We retry the above experiments on 110-layer (Nsize=19) networks.

Training loss

Testing error

Results:

  • For deep networks, it's best to put the batch normalization before the addition part of each building block layer. This effectively removes most of the batch normalization operations from the input skip paths. If a batch normalization comes after each building block, then there exists a path from the input straight to the output that passes through several batch normalizations in a row. This could be problematic because each BN is not idempotent (the effects of several BN layers accumulate).

  • Removing the ReLU layer at the end of each building block appears to give a small improvement (~0.6%)

Architecture Test error Artifacts
ReLU, BN before add (ORIG PAPER reimplementation) 0.0697 Model, Loss and Error logs, Source commit + patch
No ReLU, BN before add 0.0632 Model, Loss and Error logs, Source commit + patch
ReLU, BN after add 0.1356 Model, Loss and Error logs, Source commit + patch
No ReLU, BN after add 0.1230 Model, Loss and Error logs, Source commit + patch

ImageNet: Effect of model architecture (preliminary)

@ducha-aiki is performing preliminary experiments on imagenet. For ordinary CaffeNet networks, @ducha-aiki found that putting batch normalization after the ReLU layer may provide a small benefit compared to putting it before.

Second, results on CIFAR-10 often contradicts results on ImageNet. I.e., leaky ReLU > ReLU on CIFAR, but worse on ImageNet.

@ducha-aiki's more detailed results here: https://github.com/ducha-aiki/caffenet-benchmark/blob/master/batchnorm.md

CIFAR: Alternate training strategies (RMSPROP, Adagrad, Adadelta)

Can we improve on the basic SGD update rule with Nesterov momentum? This experiment aims to find out. Common wisdom suggests that alternate update rules may converge faster, at least initially, but they do not outperform well-tuned SGD in the long run.

Training loss curve

Testing error curve

In our experiments, vanilla SGD with Nesterov momentum and a learning rate of 0.1 eventually reaches the lowest test error. Interestingly, RMSPROP with learning rate 1e-2 achieves a lower training loss, but overfits.

Strategy Test error
Original paper: SGD + Nesterov momentum, 1e-1 0.0829
RMSprop, learrning rate = 1e-4 0.1677
RMSprop, 1e-3 0.1055
RMSprop, 1e-2 0.0945
Adadelta¹, rho = 0.3 0.1093
Adagrad, 1e-3 0.3536
Adagrad, 1e-2 0.1603
Adagrad, 1e-1 0.1255

¹: Adadelta does not use a learning rate, so we did not use the same learning rate policy as in the paper. We just let it run until convergence.

See Andrej Karpathy's CS231N notes for more details on each of these learning strategies.

CIFAR: Alternate training strategies on deep networks

Deeper networks are more prone to overfitting. Unlike the earlier experiments, all of these models (except Adagrad with a learning rate of 1e-3) achieve a loss under 0.1, but test error varies quite wildly. Once again, using vanilla SGD with Nesterov momentum achieves the lowest error.

Training loss

Testing error

Solver Testing error
Nsize=18, Original paper: Nesterov, 1e-1 0.0697
Nsize=18, RMSprop, 1e-4 0.1482
Nsize=18, RMSprop, 1e-3 0.0821
Nsize=18, RMSprop, 1e-2 0.0768
Nsize=18, RMSprop, 1e-1 0.1098
Nsize=18, Adadelta 0.0888
Nsize=18, Adagrad, 1e-3 0.3022
Nsize=18, Adagrad, 1e-2 0.1321
Nsize=18, Adagrad, 1e-1 0.1145

Effect of batch norm momentum

For our experiments, we use batch normalization using an exponential running mean and standard deviation with a momentum of 0.1, meaning that the running mean and std changes by 10% of its value at each batch. A value of 1.0 would cause the batch normalization layer to calculate the mean and standard deviation across only the current batch, and a value of 0 would cause the batch normalization layer to stop accumulating changes in the running mean and standard deviation.

The strictest interpretation of the original batch normalization paper is to calculate the mean and standard deviation across the entire training set at every update. This takes too long in practice, so the exponential average is usually used instead.

We attempt to see whether batch normalization momentum affects anything. We try different values away from the default, along with a "dynamic" update strategy that sets the momentum to 1 / (1+n), where n is the number of batches seen so far (N resets to 0 at every epoch). At the end of training for a certain epoch, this means the batch normalization's running mean and standard deviation is effectively calculated over the entire training set.

None of these effects appear to make a significant difference.

Test error curve

Strategy Test Error
BN, momentum = 1 just for fun 0.0863
BN, momentum = 0.01 0.0835
Original paper: BN momentum = 0.1 0.0829
Dynamic, reset every epoch. 0.0822

TODO: Imagenet

Owner
Kimmy
Kimmy
Implementation for "Domain-Specific Bias Filtering for Single Labeled Domain Generalization"

DSBF Introduction This repository contains the implementation code for paper: Domain-Specific Bias Filtering for Single Labeled Domain Generalization

ScottYuan 7 Jan 05, 2023
🔥 TensorFlow Code for technical report: "YOLOv3: An Incremental Improvement"

🆕 Are you looking for a new YOLOv3 implemented by TF2.0 ? If you hate the fucking tensorflow1.x very much, no worries! I have implemented a new YOLOv

3.6k Dec 26, 2022
DeRF: Decomposed Radiance Fields

DeRF: Decomposed Radiance Fields Daniel Rebain, Wei Jiang, Soroosh Yazdani, Ke Li, Kwang Moo Yi, Andrea Tagliasacchi Links Paper Project Page Abstract

UBC Computer Vision Group 24 Dec 02, 2022
Framework for evaluating ANNS algorithms on billion scale datasets.

Billion-Scale ANN http://big-ann-benchmarks.com/ Install The only prerequisite is Python (tested with 3.6) and Docker. Works with newer versions of Py

Harsha Vardhan Simhadri 132 Dec 24, 2022
Ludwig is a toolbox that allows to train and evaluate deep learning models without the need to write code.

Translated in 🇰🇷 Korean/ Ludwig is a toolbox that allows users to train and test deep learning models without the need to write code. It is built on

Ludwig 8.7k Jan 05, 2023
Official PyTorch implementation of paper: Standardized Max Logits: A Simple yet Effective Approach for Identifying Unexpected Road Obstacles in Urban-Scene Segmentation (ICCV 2021 Oral Presentation)

SML (ICCV 2021, Oral) : Official Pytorch Implementation This repository provides the official PyTorch implementation of the following paper: Standardi

SangHun 61 Dec 27, 2022
This script runs neural style transfer against the provided content image.

Neural Style Transfer Content Style Output Description: This script runs neural style transfer against the provided content image. The content image m

Martynas Subonis 0 Nov 25, 2021
Training deep models using anime, illustration images.

animeface deep models for anime images. Datasets anime-face-dataset Anime faces collected from Getchu.com. Based on Mckinsey666's dataset. 63.6K image

Tomoya Sawada 61 Dec 25, 2022
3D AffordanceNet is a 3D point cloud benchmark consisting of 23k shapes from 23 semantic object categories, annotated with 56k affordance annotations and covering 18 visual affordance categories.

3D AffordanceNet This repository is the official experiment implementation of 3D AffordanceNet benchmark. 3D AffordanceNet is a 3D point cloud benchma

49 Dec 01, 2022
PConv-Keras - Unofficial implementation of "Image Inpainting for Irregular Holes Using Partial Convolutions". Try at: www.fixmyphoto.ai

Partial Convolutions for Image Inpainting using Keras Keras implementation of "Image Inpainting for Irregular Holes Using Partial Convolutions", https

Mathias Gruber 871 Jan 05, 2023
Source codes of CenterTrack++ in 2021 ICME Workshop on Big Surveillance Data Processing and Analysis

MOT Tracked object bounding box association (CenterTrack++) New association method based on CenterTrack. Two new branches (Tracked Size and IOU) are a

36 Oct 04, 2022
[CVPR 2022] Pytorch implementation of "Templates for 3D Object Pose Estimation Revisited: Generalization to New objects and Robustness to Occlusions" paper

template-pose Pytorch implementation of "Templates for 3D Object Pose Estimation Revisited: Generalization to New objects and Robustness to Occlusions

Van Nguyen Nguyen 92 Dec 28, 2022
Source code of SIGIR2021 Paper 'One Chatbot Per Person: Creating Personalized Chatbots based on Implicit Profiles'

DHAP Source code of SIGIR2021 Long Paper: One Chatbot Per Person: Creating Personalized Chatbots based on Implicit User Profiles . Preinstallation Fir

ZYMa 32 Dec 06, 2022
This repository stores the code to reproduce the results published in "TiWS-iForest: Isolation Forest in Weakly Supervised and Tiny ML scenarios"

TinyWeaklyIsolationForest This repository stores the code to reproduce the results published in "TiWS-iForest: Isolation Forest in Weakly Supervised a

2 Mar 21, 2022
AAAI 2022: Stationary diffusion state neural estimation

Stationary Diffusion State Neural Estimation Although many graph-based clustering methods attempt to model the stationary diffusion state in their obj

绽琨 33 Nov 24, 2022
Detectron2 for Document Layout Analysis

Detectron2 trained on PubLayNet dataset This repo contains the training configurations, code and trained models trained on PubLayNet dataset using Det

Himanshu 163 Nov 21, 2022
Open source implementation of "A Self-Supervised Descriptor for Image Copy Detection" (SSCD).

A Self-Supervised Descriptor for Image Copy Detection (SSCD) This is the open-source codebase for "A Self-Supervised Descriptor for Image Copy Detecti

Meta Research 68 Jan 04, 2023
One-Shot Neural Ensemble Architecture Search by Diversity-Guided Search Space Shrinking

One-Shot Neural Ensemble Architecture Search by Diversity-Guided Search Space Shrinking This is an official implementation for NEAS presented in CVPR

Multimedia Research 19 Sep 08, 2022
This is an official PyTorch implementation of Task-Adaptive Neural Network Search with Meta-Contrastive Learning (NeurIPS 2021, Spotlight).

NeurIPS 2021 (Spotlight): Task-Adaptive Neural Network Search with Meta-Contrastive Learning This is an official PyTorch implementation of Task-Adapti

Wonyong Jeong 15 Nov 21, 2022
Lightweight stereo matching network based on MobileNetV1 and MobileNetV2

MobileStereoNet: Towards Lightweight Deep Networks for Stereo Matching

Cognitive Systems Research Group 139 Nov 30, 2022