Open tensorflow kernel for Cdiscount’s Image Classification Challenge

Open cxflow-tensorflow kernel for Cdiscount’s Image Classification Challenge Kaggle competition.

Start training on multiple GPUs with tensorflow right away!

Works on Linux with Python 3.5+.

Features:

• CLI data download
• Data validation with SHA256 hash
• Simple data visualization
• Train-Valid splitting
• Low memory footprint data streams
• GPU-CPU parallelism
• Base VGG-like convnet
• Multi-GPU training with a single argument!
• TensorBoard training tracking
• Model prediction and submission

Quick start

Install tensorflow and 7z.

Clone repo and install the requirements

git clone https://github.com/Cognexa/cdiscount-kernel && cd cdiscount-kernel
pip3 install -r requirements.txt --user

Download dataset with kaggle-cli (this may take a while, 3 hours in my case)

# requires >57Gb of free space
KG_USER="<YOUR KAGGLE USERNAME" KG_PASS="<YOUR KAGGLE PASSWORD>" cxflow dataset download cdc

Validate your download and see the example data:

# in the root directory (cdiscount-kernel)
cxflow dataset validate cdc
cxflow dataset show cdc
# now see the newly created visual directory

Create a random validation split with 10% of the data and start training:

cxflow dataset split cdc
cxflow train cdc model.n_gpus=<NUMBER OF GPUS TO USE>

Observe the training with TensorBoard (note: a summary is written only after each epoch)

tensorboard --logdir=log

Obtain predictions for submission:

cxflow predict log/<DIR> log/<DIR> model.restore_model_name=<CHECKPOINT NAME>

With only one checkpoint in the output directory, this simplifies to:

cxflow predict log/<DIR>

UPDATE [LB 0.65]

important: update cxflow and cxflow-tensorflow with pip3 install cxflow cxflow-tensorflow --user --upgrade

Main features:

• XCeption net (https://arxiv.org/abs/1610.02357)
• Fast random data access

Resize the data to dataset.size with (this may take a few hours)

cxflow dataset resize cdc/xception.yaml
cxflow dataset split cdc/xception.yaml

Run the training with

cxflow train cdc/xception.yaml

Training procedure that reached 0.65:

• Train with original size, LR 0.0001, 4 middle flow repeats until stalled
• Fine-tune with 128x128, LR 0.0001, 0.5 dropout, 0.00001 weight decay until stalled
• Fine-tune as above but with LR 0.00001 (10x smaller)

Tips:

• Use small images right away
• The final GlobalAveragePooling may be a bottleneck
• Net does not overfit so far, no augmentations needed

About

This kernel is written in cxflow-tensorflow, a plugin for cxflow framework. Make sure you check it out!

A simple submission script will be added soon, stay tuned!