Skip to content
Branch: master
Find file History
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


This folder contains the MXNet implementation of the inception-v3 model.

Prepare Dataset

The original directory of MXNet image classification contains scripts for all MNIST, cifar10 and imagenet datasets. In our benchmark, we use only the imagenet1K dataset.

To use our benchmark, first prepare the dataset according to the following steps:

1, Download and decompress the imagenet1K 2012 dataset. Note that you need to sign up for an account at to download the dataset. After decompressing, your dataset directory should look like this:

$ DATA_DIR=.... # put your directory here
$ ls $DATA_DIR

2, Generate the data format of RecordIO:

python dataset/ --list True --recursive True imagenet1k $DATA_DIR
python dataset/ --resize 480 --quality 95 --num-thread 16 imagenet1k $DATA_DIR

You should have a file named imagenet1k_train.rec generated, the size of this file is about 137G.

Run Training

We choose the following hyper-parameters for the inception-v3 models:

Learning rate: 0.1 for 30 epochs, 0.01 for 30 epochs, 0.001 for 40 epochs
momentum: 0.9
weight decay: 0.0001
optimizer: sgd

We leave other hyper parameters as default.

Change the --data-train option in scripts/, then start the training by:

cd scripts

Change the --batch-size option in scripts/ if you want to use a different mini-batch size.

You can’t perform that action at this time.