ouya-bytes / kaggle_learning

use the theano to solve the kaggle

kaggle_learning -- Digit Recognizer

---

tips

there are two files that can be run,Al_cnn.py and Digit*.py

process

---

Code that created by theano is based LeNet,I am not choose the model that corresponds to the min-cost.Model is trained on the cpu, so I cancel the validation during the each train epoch, and this make the code runs about less time than train-validation. first, I use the raw dataset, accuracy=0.97886 then, I think data should be 1-0, so I make the value who is larger than zero be 1,others is zero, accuracy=0.97700 and I scala the data to [0, 1], accuracy=0.98043; when I use the all train data to train the model, accuracy=0.98371

Improve

Change the structure of the LeNet、learning_rate、 add normalizer