🔪 Elimination based Lightweight Neural Net with Pretrained Weights

Last update: Jul 12, 2022

Overview

ELimNet

ELimNet: Eliminating Layers in a Neural Network Pretrained with Large Dataset for Downstream Task

Removed top layers from pretrained EfficientNetB0 and ResNet18 to construct lightweight CNN model with less than 1M #params.
Assessed on Trash Annotations in Context(TACO) Dataset sampled for 6 classes with 20,851 images.
Compared performance with lightweight models generated with Optuna's Neural Architecture Search(NAS) constituted with same convolutional blocks.

Quickstart

Installation

# clone the repository
git clone https://github.com/snoop2head/elimnet

# fetch image dataset and unzip
!wget -cq https://aistages-prod-server-public.s3.amazonaws.com/app/Competitions/000081/data/data.zip
!unzip ./data.zip -d ./

Train

# finetune on the dataset with pretrained model
python train.py --model ./model/efficientnet_b0.yaml

# finetune on the dataset with ElimNet
python train.py --model ./model/efficientnet_b0_elim_3.yaml

Inference

# inference with the lastest ran model
python inference.py --model_dir ./exp/latest/

Performance

Performance is compared with (1) original pretrained model and (2) Optuna NAS constructed models with no pretrained weights.

Indicates that top convolutional layers eliminated pretrained CNN models outperforms empty Optuna NAS models generated with same convolutional blocks.
Suggests that eliminating top convolutional layers creates lightweight model that shows similar(or better) classifcation performance with original pretrained model.
Reduces parameters to 7%(or less) of its original parameters while maintaining(or improving) its performance. Saves inference time by 20% or more by eliminating top convolutional layters.

ELimNet vs Pretrained Models (Train)

[100 epochs]	# of Parameters	# of Layers	Train	Validation	Test F1
Pretrained EfficientNet B0	4.0M	352	Loss: 0.43 Acc: 81.23% F1: 0.84	Loss: 0.469 Acc: 82.17% F1: 0.76	0.7493
EfficientNet B0 Elim 2	0.9M	245	Loss:0.652 Acc: 87.22% F1: 0.84	Loss: 0.622 Acc: 87.22% F1: 0.77	0.7603
EfficientNet B0 Elim 3	0.30M	181	Loss: 0.602 Acc: 78.17% F1: 0.74	Loss: 0.661 Acc: 77.41% F1: 0.74	0.7349

Resnet18	11.17M	69	Loss: 0.578 Acc: 78.90% F1: 0.76	Loss: 0.700 Acc: 76.17% F1: 0.719	-
Resnet18 Elim 2	0.68M	37	Loss: 0.447 Acc: 83.73% F1: 0.71	Loss: 0.712 Acc: 75.42% F1: 0.71	-

ELimNet vs Pretrained Models (Inference)

	# of Parameters	# of Layers	CPU times (sec)	CUDA time (sec)	Test Inference Time (sec)
Pretrained EfficientNet B0	4.0M	352	3.9s	4.0s	105.7s
EfficientNet B0 Elim 2	0.9M	245	4.1s	13.0s	83.4s
EfficientNet B0 Elim 3	0.30M	181	3.0s	9.0s	73.5s

Resnet18	11.17M	69	-	-	-
Resnet18 Elim 2	0.68M	37	-	-	-

ELimNet vs Empty Optuna NAS Models (Train)

[100 epochs]	# of Parameters	# of Layers	Train	Valid	Test F1
Empty MobileNet V3	4.2M	227	Loss 0.925 Acc: 65.18% F1: 0.58	Loss 0.993 Acc: 62.83% F1: 0.56	-
Empty EfficientNet B0	1.3M	352	Loss 0.867 Acc: 67.28% F1: 0.61	Loss 0.898 Acc: 66.80% F1: 0.61	0.6337

Empty DWConv & InvertedResidualv3 NAS	0.08M	66	-	Loss: 0.766 Acc: 71.71% F1: 0.68	0.6740
Empty MBConv NAS	0.33M	141	Loss: 0.786 Acc: 70.72% F1: 0.66	Loss: 0.866 Acc: 68.09% F1: 0.62	0.6245

Resnet18 Elim 2	0.68M	37	Loss: 0.447 Acc: 83.73% F1: 0.71	Loss: 0.712 Acc: 75.42% F1: 0.71	-
EfficientNet B0 Elim 3	0.30M	181	Loss: 0.602 Acc: 78.17% F1: 0.74	Loss: 0.661 Acc: 77.41% F1: 0.74	0.7603

ELimNet vs Empty Optuna NAS Models (Inference)

	# of Parameters	# of Layers	CPU times (sec)	CUDA time (sec)	Test Inference Time (sec)
Empty MobileNet V3	4.2M	227	4	13	-
Empty EfficientNet B0	1.3M	352	3.780	3.782	68.4s

Empty DWConv & InvertedResidualv3 NAS	0.08M	66	1	3.5	61.1s
Empty MBConv NAS	0.33M	141	2.14	7.201	67.1s

Resnet18 Elim 2	0.68M	37	-	-	-
EfficientNet B0 Elim 3	0.30M	181	3.0s	9s	73.5s

Background & WiP

Background

NLP tasks are usually downstream tasks of finetuning large pretrained transformers models(i.e. BERT, RoBERTa, XLNet).
Removing top transformers layers may yield 40% reduction in size while preserving up to 98.2% of the performance.
Likewise, for computer vision's classification task, removing convolutional top layers from pretrained models are tested.

Work in Progress

Will test the performance of replacing convolutional blocks with pretrained weights with a single convolutional layer without pretrained weights.
Will add ResNet18's inference time data and compare with Optuna's NAS constructed lightweight model.
Will test on pretrained MobileNetV3, MnasNet on torchvision with elimination based lightweight model architecture search.
Will be applied on other small datasets such as Fashion MNIST dataset and Plant Village dataset.

Others

"Empty" stands for model with no pretrained weights.
"EfficientNet B0 Elim 2" means 2 convolutional blocks have been eliminated from pretrained EfficientNet B0. Number next to "Elim" annotates how many convolutional blocks have been removed.
Table's performance illustrates best performance out of 100 epochs of finetuning on TACO Dataset.

🔪 Elimination based Lightweight Neural Net with Pretrained Weights

Related tags

Overview

ELimNet

Quickstart

Installation

Train

Inference

Performance

ELimNet vs Pretrained Models (Train)

ELimNet vs Pretrained Models (Inference)

ELimNet vs Empty Optuna NAS Models (Train)

ELimNet vs Empty Optuna NAS Models (Inference)

Background & WiP

Background

Work in Progress

Others

Authors

Owner

snoop2head

PerfFuzz: Automatically Generate Pathological Inputs for C/C++ programs

🌊 Online machine learning in Python

Offline Reinforcement Learning with Implicit Q-Learning

A simple but complete full-attention transformer with a set of promising experimental features from various papers

Transformer - Transformer in PyTorch

Code for NeurIPS2021 submission "A Surrogate Objective Framework for Prediction+Programming with Soft Constraints"

The official re-implementation of the Neurips 2021 paper, "Targeted Neural Dynamical Modeling".

FB-tCNN for SSVEP Recognition

PyTorch Implementation of our paper Explain Me the Painting: Multi-Topic Knowledgeable Art Description Generation

Used to record WKU's utility bills on a regular basis.

Towards Long-Form Video Understanding

Code for 'Blockwise Sequential Model Learning for Partially Observable Reinforcement Learning' (AAAI 2022)

This repository is related to an Arabic tutorial, within the tutorial we discuss the common data structure and algorithms and their worst and best case for each, then implement the code using Python.

Additional functionality for use with fastai’s medical imaging module

Implement of "Training deep neural networks via direct loss minimization" in PyTorch for 0-1 loss

Summary Explorer is a tool to visually explore the state-of-the-art in text summarization.

Alias-Free Generative Adversarial Networks (StyleGAN3) Official PyTorch implementation

Deep Learning tutorials in jupyter notebooks.

MINIROCKET: A Very Fast (Almost) Deterministic Transform for Time Series Classification

Code for our NeurIPS 2021 paper 'Exploiting the Intrinsic Neighborhood Structure for Source-free Domain Adaptation'