louiscaixuran / mymnist

readme.md

Predict a integer image

Test data is stored in testdata directory

Dataset is the mnist standard train and test set

##Normal way

Step1

Use train.py to train and store the two layer net model in two_layer_net.pkl

$python3 train.py
train acc, test acc | 0.10441666666666667, 0.1028
train acc, test acc | 0.8023833333333333, 0.8073
train acc, test acc | 0.8756666666666667, 0.8785
train acc, test acc | 0.8978666666666667, 0.8999
train acc, test acc | 0.9068666666666667, 0.9102
train acc, test acc | 0.9144333333333333, 0.9148
train acc, test acc | 0.9203166666666667, 0.9218
train acc, test acc | 0.9235666666666666, 0.9244
train acc, test acc | 0.92845, 0.9296
train acc, test acc | 0.9311166666666667, 0.9317
train acc, test acc | 0.9345833333333333, 0.9351
train acc, test acc | 0.93695, 0.9374
train acc, test acc | 0.94025, 0.9396
train acc, test acc | 0.9419333333333333, 0.9396
train acc, test acc | 0.9441833333333334, 0.9425
train acc, test acc | 0.9457833333333333, 0.9445
train acc, test acc | 0.9484833333333333, 0.9477

Step2

Use two_layer_test.py to test the accuracy of the model

$python3 two_layer_test.py
Accuracy:0.9466

with the get_image function you can also test your own image

CNN

###Step1 Use train_conv.py to train and store CNN net model in params.pkl

$python3 train_conv.py
train loss:2.300001846087919
=== epoch:1, train acc:0.212, test acc:0.191 ===
train loss:2.29748375539359
train loss:2.294312634680256
train loss:2.286364878660862
train loss:2.2797246011860945
train loss:2.2627870681081244
train loss:2.250901554926544
train loss:2.2377228652672003
train loss:2.1955270072011173
train loss:2.1901397325996363
train loss:2.161590794937312
train loss:2.0857797169219743
train loss:2.0658089204986436
train loss:2.0255680630766495
train loss:1.9911901577369833
...

finally loss can come to less than 0.001

###Step2 Use test.py to test your own image in testdata

$ python3 test.py
5
[[ -3669.03271876  -2723.0830266   -1336.10245562  -4087.55747946
-1673.66868792   1515.17049045    675.42958532 -11025.10787864
609.69700127 -10413.63506436]]

pytorch

net

Rewrite the two_layer net with pytorch. convert the definition of linear function and the predict function with the pytorch keep the accuracy function to show if it works in train_pytorch function, change the numpy array into tensorarray and use optimizer to do the loss and backwards