def __init__(self):
"""
LeNet classifier for greyscale images.
"""
self.net = Net()
# -- Initialize layers ---
weight_filler = Filler(type="xavier")
bias_filler = Filler(type="constant", value=0.2)
conv_lr_mults = [1.0, 2.0]
conv_decay_mults = [1.0, 0.0]
self.layer_list = []
self.layer_list.append(Convolution(name="conv1", bottoms=["data"], param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults, kernel_size=5, stride=1, pad=0,
weight_filler=weight_filler, bias_filler=bias_filler, num_output=20))
self.layer_list.append(ReLU(name="relu1", bottoms=["conv1"], tops=["conv1"]))
self.layer_list.append(Pooling(name="pool1", bottoms=["conv1"], kernel_size=2, stride=2))
self.layer_list.append(Convolution(name="conv2", bottoms=["pool1"], param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults, kernel_size=5, stride=1, pad=0,
weight_filler=weight_filler, bias_filler=bias_filler, num_output=50))
self.layer_list.append(Pooling(name="pool2", bottoms=["conv2"], kernel_size=2, stride=2))
self.layer_list.append(InnerProduct(name='ip1', bottoms=['pool2'],
num_output=500, weight_filler=weight_filler))
self.layer_list.append(ReLU(name="relu2", bottoms=["ip1"], tops=["ip1"]))
self.layer_list.append(InnerProduct(name='ip2', bottoms=['pool2'],
num_output=10, weight_filler=weight_filler))
self.layer_list.append(SoftmaxWithLoss(name='softmax_loss', bottoms=['ip2', 'label']))
class LeNet:
def __init__(self):
"""
LeNet classifier for greyscale images.
"""
self.net = Net()
# -- Initialize layers ---
weight_filler = Filler(type="xavier")
bias_filler = Filler(type="constant", value=0.2)
conv_lr_mults = [1.0, 2.0]
conv_decay_mults = [1.0, 0.0]
self.layer_list = []
self.layer_list.append(Convolution(name="conv1", bottoms=["data"], param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults, kernel_size=5, stride=1, pad=0,
weight_filler=weight_filler, bias_filler=bias_filler, num_output=20))
self.layer_list.append(ReLU(name="relu1", bottoms=["conv1"], tops=["conv1"]))
self.layer_list.append(Pooling(name="pool1", bottoms=["conv1"], kernel_size=2, stride=2))
self.layer_list.append(Convolution(name="conv2", bottoms=["pool1"], param_lr_mults=conv_lr_mults,
param_decay_mults=conv_decay_mults, kernel_size=5, stride=1, pad=0,
weight_filler=weight_filler, bias_filler=bias_filler, num_output=50))
self.layer_list.append(Pooling(name="pool2", bottoms=["conv2"], kernel_size=2, stride=2))
self.layer_list.append(InnerProduct(name='ip1', bottoms=['pool2'],
num_output=500, weight_filler=weight_filler))
self.layer_list.append(ReLU(name="relu2", bottoms=["ip1"], tops=["ip1"]))
self.layer_list.append(InnerProduct(name='ip2', bottoms=['pool2'],
num_output=10, weight_filler=weight_filler))
self.layer_list.append(SoftmaxWithLoss(name='softmax_loss', bottoms=['ip2', 'label']))
def fit(self, batch_iter, max_iter = 10000):
training_manager = StepTraining(self.net, self.layer_list, batch_iter)
loss_list = training_manager.set_iter(max_iter)
return loss_list
def predict(self, img):
"""
Takes a 28 x 28 greyscale image as input and returns predicted label.
"""
self.net.reset_forward()
self.net.phase = "test"
image_blob = np.zeros((1, 1, 28, 28))
image_blob[0][0] = img
self.net.forward_layer(NumpyData(name = "data", data = image_blob))
for layer in self.layer_list:
if layer.deploy:
self.net.forward_layer(layer)
self.net.forward_layer(Softmax(name='softmax', bottoms=['ip2']))
softmax_prob = self.net.tops["softmax"].data[0]
return softmax_prob.argmax()