Digit_Recognizer_Kaggle

Tensorflow implemention of a CNN model for Kaggle competition: Digit Recognizer.

Result

99.175% is achieved after 100k training steps. Better results can be achieved by tuning hyperparameters.

How to run

1. Split train and validation dataset.

   python help.py

2. Training, validation and inference.

   Modify the code bellow based your own setting.

   hps = net.HParams(data_size=34000,
   
                     batch_size=100,
   
                     num_classes=10,
   
                     lrn_rate=0.01,
   
                     mode='validation',
   
                     weight_decay_rate=0.04)

   then run:

   python ./train.py  --batch_count=80  --num_gpus=1

   batch_count is the numbers of batch your are going to run in validation or infer mode, basically: batch_count = hps.data_size / hps.batch_size.