SVHNClassifier-PyTorch

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

If you're interested in C++ inference, move HERE

Results

Steps	GPU	Batch Size	Learning Rate	Patience	Decay Step	Decay Rate	Training Speed (FPS)	Accuracy
54000	GTX 1080 Ti	512	0.16	100	625	0.9	~1700	95.65%

Sample

$ python infer.py -c=./logs/model-54000.pth ./images/test-75.png
length: 2
digits: 7 5 10 10 10

$ python infer.py -c=./logs/model-54000.pth ./images/test-190.png
length: 3
digits: 1 9 0 10 10

Loss

Requirements

• Python 3.6

• torch 1.0

• torchvision 0.2.1

• visdom

  $ pip install visdom
  

• h5py

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

• protobuf

  $ pip install protobuf
  

• lmdb

  $ pip install lmdb
  

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.pth
   

6. Evaluate

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

7. Visualize

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

8. Infer

   $ python infer.py --checkpoint=./logs/model-100.pth ./images/test1.png
   

9. Clean

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