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

Accuracy 93.45% on test dataset after about 14 hours

Training	Test

digit "10" means no digits

• Python 2.7

• Tensorflow

• h5py

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

1. Clone the source code

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

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
   

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

   Open `draw_bbox.ipynb` in Jupyter
   

2. Convert to TFRecords format

   $ python convert_to_tfrecords.py --data_dir ./data
   

3. (Optional) Test for reading TFRecords files

   Open `read_tfrecords_sample.ipynb` in Jupyter
   Open `donkey_sample.ipynb` in Jupyter
   

4. Train

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

5. Retrain if you need

   $ python train.py --data_dir ./data --train_logdir ./logs/train2 --restore_checkpoint ./logs/train/latest.ckpt
   

6. Evaluate

   $ python eval.py --data_dir ./data --checkpoint_dir ./logs/train --eval_logdir ./logs/eval
   

7. Visualize

   $ tensorboard --logdir ./logs
   

8. (Optional) Try to make an inference

   Open `inference_sample.ipynb` in Jupyter
   Open `inference_outside_sample.ipynb` in Jupyter
   $ python inference.py --image /path/to/image.jpg --restore_checkpoint ./logs/train/latest.ckpt
   

9. Clean

   $ rm -rf ./logs
   or
   $ rm -rf ./logs/train2
   or
   $ rm -rf ./logs/eval