iamrjkunal / flixstock

Flixstock Classification Assignment

Table of Contents

• Prerequisites
• Usage Guide
• To Do Left

Prerequisites

• Create python3.6 environment and install the required packages

• Installation Guide

  pip install -r requirements.txt

Usage Guide

Training

This is the training stage of the project. Steps

1. First download xception weights using below command:

  wget https://github.com/fchollet/deep-learning-models/releases/download/v0.4/xception_weights_tf_dim_ordering_tf_kernels_notop.h5 -P pretrained_weights/

2. Then move folder containing images and their label csv files to data directory

Usage:

python train.py --data_dir data/ \
                --ckpt_dir ckpt/ \
                 --imagenet_weights pretrained_weights/xception_weights_tf_dim_ordering_tf_kernels_notop.h5 \
                 --test_split 0.1 \
                 --random_state 30 \
                 --num_epochs 3000 \
                 --batch_size 64 \
                 --lr 0.0001 \
                 --stoploss 0.6 \
                 --restore_model=True

Testing

This module is used to test the performance of the trained model on test dataset generated above.

Usage:

python train.py --data_dir data/ \
                --ckpt_dir ckpt/ \
                 --imagenet_weights pretrained_weights/xception_weights_tf_dim_ordering_tf_kernels_notop.h5 \
                 --batch_size 16

Custom Testing

This module is used to test the performance of trained model on any general dataset which can be of different distribution also. Steps

1. move all the images to custom_data/test_images directory and after model get run output.csv will be generated in custom_data directory

Usage:

python train.py --test_data_dir custom_data/test_images/ \
                --output_dir custom_data/ \
                --ckpt_dir ckpt/

To Do Left

1. Adding proper callbacks to make training process efficient.
2. Since it is a single neural net for all the 3 independent classes so model will take a lot time to converge although implemented loss function works good but takes time. Some more advanced loss function can be used.