SVHNClassifier-PyTorch

A PyTorch implementation of Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks

Results

Accuracy

Accuracy 95.32% on test dataset after 721,000 steps

Requirements

• Python 2.7

• PyTorch

• h5py

  In Ubuntu:
  $ sudo apt-get install libhdf5-dev
  $ sudo pip install h5py

• Protocol Buffers 3

• LMDB

• Visdom

Setup

1. Clone the source code

   $ git clone https://github.com/potterhsu/SVHNClassifier-PyTorch
   $ cd SVHNClassifier-PyTorch

2. Download SVHN Dataset format 1

3. Extract to data folder, now your folder structure should be like below:

   SVHNClassifier
       - data
           - extra
               - 1.png 
               - 2.png
               - ...
               - digitStruct.mat
           - test
               - 1.png 
               - 2.png
               - ...
               - digitStruct.mat
           - train
               - 1.png 
               - 2.png
               - ...
               - digitStruct.mat

Usage

1. (Optional) Take a glance at original images with bounding boxes

   Open `draw_bbox.ipynb` in Jupyter

2. Convert to LMDB format

   $ python convert_to_lmdb.py --data_dir ../data

3. (Optional) Test for reading LMDBs

   Open `read_lmdb_sample.ipynb` in Jupyter

4. Train

   $ python train.py --data_dir ../data --logdir ./logs

5. Retrain if you need

   $ python train.py --data_dir ./data --logdir ./logs_retrain --restore_checkpoint ./logs/model-100.tar

6. Evaluate

   $ python eval.py --data_dir ./data ./logs/model-100.tar

7. Visualize

   $ python -m visdom.server
   $ python visualize.py --logdir ./logs

8. Clean

   $ rm -rf ./logs
   or
   $ rm -rf ./logs_retrain