Efficient-Image-Classification

This repository contains comparisons of different convolution neural networks for the CIFAR-10 data-set.ImageNet pretrained models have been fine tuned on CIFAR-10 dataset.The script automatically downloads the cifar 10 dataset. Clone this repo please follow the following steps

git clone https://github.com/L-A-Sandhu/Efficient-Image-Classification.git

The rest of the repository is divided as follows.

1. Requirements
2. Fine Tuning
3. Using Pretrained Model
4. Summary

Requirements

This repository requires

• tensorflow
• matplotlib
• scipy
• protobuf

For complete installation please follow the following steps

cd M0bile_Net/
conda create  -n <environment -name> python==3.7.4
conda activate <environment-name>
pip install -r requirements.txt
cd ../

Fine Tuning

This section disccusses the fine tunning method for Mobile Net and Inception net. The keras implementations of Mobile Net and Inception Net is used in this work. Their weights are trained on Image Net dataset. However, this work fine tuned the model on cifar10 dataset. please follow the following set of commands

cd < M0bile_Net or Inception_NET>

Fine Tune

python <Mobile-Net.py or Inception-Net.py >   --model_dir=<Location for saving model> --inp=<tune, train, test, resume > --b_s=< Batch size> --e=<epoch>
example command 
python Mobile-Net.py  --model_dir='./checkpoint/' --inp=tune --b_s=16 --e=100

Train

python <Mobile-Net.py or Inception-Net.py > --model_dir=<Location for saving model> --inp=<tune, train, test, resume > --b_s=< Batch size> --e=<epoch>
example command 
python Mobile-Net.py --model_dir='./checkpoint/' --inp=train --b_s=16 --e=100

Test

python <Mobile-Net.py or Inception-Net.py >  --model_dir=<Location for saving model> --inp=<tune, test, train, resume>
example command 
python Mobile-Net.py --model_dir='./checkpoint/' --inp=test

Resume Training

python <Mobile-Net.py or Inception-Net.py >  --model_dir=<Location for saving/loading model> --inp= <tune, train, test, resume,conv > --b_s=< Batch size> --e=<epoch>
example command 
python Mobile-Net.py  --model_dir='./checkpoint/' --inp=resume --b_s=16 --e=100

Convert to ONNX

python <Mobile-Net.py or Inception-Net.py >  --model_dir=<Location for loading saved check point> --onnx_dir=< Location to save onnx model> --inp= <tune, train, test, resume, conv > 
example command 
python Mobile-Net.py  --model_dir='./checkpoint/' --onnx_dir='./onnx/'--inp=conv 

Pretrained Model

In this work the models are trained on cifar10 dataset with batch size 128 and 50 epochs. You can download the pretrained weights and place them at ./Inception_NET/checkpoint/ or ./M0bile_Net/checkpoint// for inferene or resume training using the above mentioned commands. The pretrained weights for Mobile net and Inception Net can be be downloaded from the folllowing links respectively.

https://drive.google.com/file/d/1OCxDNDUbMJcoo8QbzB6hM4r4yqXOXpZU/view?usp=sharing
https://drive.google.com/file/d/144j9-G-v2x6YCTZ4u9_NDzXpVnVwT_kC/view?usp=sharing

Results and comparision

Test results and comparision for both models is shown in the following table

Model	Parameters	Acc.FineTune	Acc. Scratch	Latency(sec)	Size on Disk (MB)	Flops
Mobile-Net	3,743,718	0.852	0.76	0.0004	38.89	0.116 G
Inception-Net	22,115,894	0.811	0.73	0.0006	174.2	0.681 G

Confusion Matrix

Thee confusion matrix for Mobile net is shown shown below

	'airplane'	'automobile'	'bird'	'cat'	'deer'	'dog'	'frog'	'horse'	'ship'	'truck'
'airplane'	728	18	75	20	63	3	30	1	49	13
'automobile'	0	964	1	4	2	0	5	0	12	12
'bird'	10	0	898	23	31	5	30	1	2	0
'cat'	7	7	61	688	96	75	59	4	3	0
'deer'	0	1	35	17	912	11	19	4	1	0
'dog'	1	6	39	123	64	724	37	5	1	0
'frog'	0	1	16	19	11	3	949	0	1	0
'horse'	1	5	36	42	82	83	12	736	2	1
'ship'	5	10	17	8	11	4	9	1	932	3
'truck'	3	83	6	18	9	4	16	1	29	831

The confusion matrix for Inception net is shown below

	'airplane'	'automobile'	'bird'	'cat'	'deer'	'dog'	'frog'	'horse'	'ship'	'truck'
'airplane'	894	12	14	9	3	1	2	1	43	21
'automobile'	7	933	1	2	2	0	3	0	27	25
'bird'	103	7	742	63	23	19	30	5	6	2
'cat'	27	3	50	765	33	66	19	5	19	13
'deer'	21	3	82	67	769	14	13	5	20	6
'dog'	10	8	32	207	31	687	6	5	7	7
'frog'	18	4	42	55	18	14	828	0	19	2
'horse'	19	4	27	70	66	73	8	715	3	15
'ship'	29	8	6	5	1	1	1	1	933	15
'truck'	25	74	4	7	0	4	1	3	32	850