import numpy as np
import mxnet as mx
from mxnet import nd, autograd, gluon

model_ctx = mx.gpu(0)
# model_ctx = mx.cpu()

batch_size = 128 # Batch basina dusen ornek sayisi
epochs = 100 # Iterasyon sayisi

# Yuklenen verinin X ve y kismina uygulanir.
def transform_func(data, label):
    # Klasik MNIST verisi donusumu. Tip float yapilir ve
    # X degerleri 0 ile 1 arasinda sigdirilmak adina 255'e bolunur.
    return data.astype(np.float32)/255., label.astype(np.float32)

train_data = gluon.data.DataLoader(gluon.data.vision.MNIST(train=True,\
transform=transform_func), batch_size, shuffle=True)
test_data = gluon.data.DataLoader(gluon.data.vision.MNIST(train=False,\
transform=transform_func), batch_size, shuffle=True)

m = gluon.nn.Sequential() # Model
with m.name_scope(): # TF benzeri name_scope tanimi
    # Keras benzeri Sequential.add()
    m.add(gluon.nn.Dense(100,activation="relu"))
    m.add(gluon.nn.Dense(256,activation="relu"))
    m.add(gluon.nn.Dense(100,activation="relu"))
    m.add(gluon.nn.Dense(10))

m.collect_params().initialize(mx.init.Normal(sigma=0.2), ctx=model_ctx)
# Parametrelerin baslangic degerleri STD'si 0.2 olan bir Normal dagilimdan cekildi
loss = gluon.loss.SoftmaxCrossEntropyLoss() # Siniflandirma hata fonksiyonu
trainer = gluon.Trainer(m.collect_params(), "sgd", {"learning_rate": 0.01} )
# Parametreleri SGD kullanarak 0.01 LR ile egit

print m # Modeli yazdir
# Egitim dongusu
for e in range(epochs): # Her iterasyon icin
    for i, (data, label) in enumerate(train_data): # Her batch icin
        # X'i model contexte koy
        data = data.as_in_context(model_ctx).reshape([-1, 784])
        # y'i model contexte koy
        label = label.as_in_context(model_ctx)
        with autograd.record(): # Autograd icin name_scope
            output = m(data)
            l = loss(output, label) # Y ile net(X) arasindaki farki hesapla
            l.backward() # PyTorch stili Backward pass
    trainer.step(batch_size)

print(m) # Modeli yazdir

# Test Dongusu
acc = mx.metric.Accuracy() # Dogruluk metrigi
for i, (data, label) in enumerate(test_data): # Her batch icin
    data = data.as_in_context(model_ctx).reshape([-1, 784])
    label = label.as_in_context(model_ctx)
    output = m(data) # net(X) ciktisini hesapla
    preds = nd.argmax(output, axis=1)
    # Gelen olasilik degerlerinden en yuksek olanin indisini al
    # 3 sayi olsaydi 0 1 2 seklinde
    # olasiliklar [0.6, 0.4, 0.3] olsaydi
    # max=0.6
    # argmax = 0 olacakti.
    acc.update(preds=preds, labels=label) # Metrigi guncelle
print acc # Dogrulugu yazdir