class CNNMR:
def __init__(self):
self.conv1 = Conv(3, 8, 1)
self.relu1 = Relu()
self.pool1 = Pool()
self.conv2 = Conv(3, 16, 8)
self.relu2 = Relu()
self.pool2 = Pool()
#self.fcl1 = FCL(12 * 6 * 16, 120)
#self.fcl2 = FCL(120, 80)
#self.relu3 = Relu()
#self.relu4 = Relu()
#self.relu = Relu()
#self.fcl1 = FCL(256, 256)
#self.relu1 = Relu()
self.fcl = Reg(6 * 6 * 16, 2)
self.soft = Softmax()
def normalize_batch(self, batch):
batch[1] = (batch[1] / 255) - 0.5
if batch[1].ndim != 4:
batch[1] = batch[1][:, :, np.newaxis, :]
#print(batch[1].shape)
return batch
def forward(self, batch):
batch = self.normalize_batch(batch)
out = self.apply(batch[1])
loss = np.sum(self.soft.cross_entropy_loss(batch[0]))
acc = np.abs(self.soft.acc(batch[0]))
return out, loss, acc
def apply(self, batch):
out = self.conv1.apply(batch)
out = self.relu1.apply(out)
out = self.pool1.apply(out)
out = self.conv2.apply(out)
out = self.relu2.apply(out)
out = self.pool2.apply(out)
# out = self.fcl1.apply(out)
# out = self.relu3.apply(out)
# out = self.fcl2.apply(out)
# out = self.relu4.apply(out)
# out = self.reg.apply(out)
out = self.fcl.apply(out)
out = self.soft.apply(out)
return out
def train(self, batch, lr=0.05):
out, loss, acc = self.forward(batch)
# gradient = self.reg.backprop(batch[0], lr)
# gradient = self.relu4.backprop(gradient)
# gradient = self.fcl2.backprop(gradient, lr)
# gradient = self.relu3.backprop(gradient)
# gradient = self.fcl1.backprop(gradient, lr)
gradient = self.soft.backprop(batch[0])
gradient = self.fcl.backprop(gradient, lr)
gradient = self.pool2.backprop(gradient)
gradient = self.relu2.backprop(gradient)
gradient = self.conv2.backprop(gradient, lr)
gradient = self.pool1.backprop(gradient)
gradient = self.relu1.backprop(gradient)
gradient = self.conv1.backprop(gradient, lr)
#gradient = self.conv.backprop(gradient, lr)
return out, loss, acc
