Awesome Open Source
Awesome Open Source

EfficientNet-Gluon

EfficientNet Gluon implementation

ImageNet experiments

Requirements

Python 3.7 or later with packages:

  • mxnet >= 1.5.0
  • gluoncv >= 0.6.0
  • nvidia-dali >= 0.19.0

Usage

Prepare ImageNet dataset

  1. Download and extract dataset following this tutorial:
    https://gluon-cv.mxnet.io/build/examples_datasets/imagenet.html
  2. Create mxnet-record files following this turorial:
    https://gluon-cv.mxnet.io/build/examples_datasets/recordio.html#imagerecord-file-for-imagenet

Clone this repo

git clone https://github.com/mnikitin/EfficientNet.git
cd EfficientNet/train_imagenet

Train your model

Example of training efficientnet-b0 with nvidia-dali data loader using 4 gpus:

IMAGENET_RECORD_ROOT='path/to/imagenet/record/files'
MODEL='efficientnet-b0'
python3 train_dali.py --rec-train $IMAGENET_RECORD_ROOT/train --rec-val $IMAGENET_RECORD_ROOT/val --input-size 224 --batch-size 64 --num-gpus 4 --num-epochs 80 --lr 0.1 --lr-decay-epoch 40,60 --save-dir params-$MODEL --logging-file params-$MODEL/log.txt --save-frequency 5 --mode hybrid --model $MODEL

Results

Code in this repo was used to train efficientnet-b0 and efficientnet-lite0 models. Pretrained params are avaliable (18.8 mb in total = 13.7 mb for extractor + 5.1 mb for classifier).

err-top1 err-top5 pretrained params
efficientnet-b0 0.335842 0.128043 dropbox link
efficientnet-lite0 0.305316 0.106322 dropbox link

Note that due to limited computational resources obtained results are worse than in the original paper. Moreover, efficientnet-lite0 was trained using more gpus and bigger batch size, so in spite of simpler architecture (relu6 instead of swish) its results are better than for efficientnet-b0 model. Anyway, I believe provided pretrained params can serve as a good initialization for your task.

That's how efficientnet-b0 and efficientnet-lite0 were trained exactly:

MODEL='efficientnet-b0'
python3 train_dali.py --rec-train $IMAGENET_RECORD_ROOT/train --rec-val $IMAGENET_RECORD_ROOT/val --input-size 224 --batch-size 56 --num-gpus 4 --num-epochs 50 --lr 0.1 --lr-decay-epoch 20,30,40 --save-dir params-$MODEL --logging-file params-$MODEL/log.txt --save-frequency 5 --mode hybrid --model $MODEL
MODEL='efficientnet-lite0'
python3 train_dali.py --rec-train $IMAGENET_RECORD_ROOT/train --rec-val $IMAGENET_RECORD_ROOT/val --input-size 224 --batch-size 72 --num-gpus 6 --num-epochs 60 --lr 0.1 --lr-decay-epoch 20,35,50 --save-dir params-$MODEL --logging-file params-$MODEL/log.txt --save-frequency 5 --mode hybrid --model $MODEL

Get A Weekly Email With Trending Projects For These Topics
No Spam. Unsubscribe easily at any time.
python (53,199
deep-learning (3,906
mxnet (106
efficientnet (31
gluon (28