def __init__(self, input_dim=(1, 28, 28),
conv_param_1={'filter_num': 16, 'filter_size': 3, 'pad': 1, 'stride': 1},
conv_param_2={'filter_num': 16, 'filter_size': 3, 'pad': 1, 'stride': 1},
conv_param_3={'filter_num': 32, 'filter_size': 3, 'pad': 1, 'stride': 1},
conv_param_4={'filter_num': 32, 'filter_size': 3, 'pad': 2, 'stride': 1},
conv_param_5={'filter_num': 64, 'filter_size': 3, 'pad': 1, 'stride': 1},
conv_param_6={'filter_num': 64, 'filter_size': 3, 'pad': 1, 'stride': 1},
hidden_size=50, output_size=10):
pre_node_nums = np.array([1*3*3, 16*3*3, 16*3*3, 32*3*3, 32*3*3, 64*3*3, 64*4*4, hidden_size])
weight_init_scale = np.sqrt(2.0 / pre_node_nums)
# weights init
self.params = {}
pre_channel_num = input_dim[0]
for idx, conv_param in enumerate([conv_param_1, conv_param_2, conv_param_3,
conv_param_4, conv_param_5, conv_param_6]):
self.params['w'+str(idx+1)] = weight_init_scale[idx] *\
np.random.randn(
conv_param['filter_num'],
pre_channel_num, conv_param['filter_size'],
conv_param['filter_size'])
self.params['b'+str(idx+1)] = np.zeros(conv_param['filter_num'])
pre_channel_num = conv_param['filter_num']
self.params['w7'] = weight_init_scale[6] * np.random.randn(64*4*4, hidden_size)
self.params['b7'] = np.zeros(hidden_size)
self.params['w8'] = weight_init_scale[7] * np.random.randn(hidden_size, output_size)
self.params['b8'] = np.zeros(output_size)
# gen layers
self.layers = []
self.layers.append(Convolution(self.params['w1'], self.params['b1'], conv_param_1['stride'],
conv_param_1['pad']))
self.layers.append(Relu())
self.layers.append(Convolution(self.params['w2'], self.params['b2'], conv_param_2['stride'],
conv_param_2['pad']))
self.layers.append(Relu())
self.layers.append(Pooling(pool_h=2, pool_w=2, stride=2))
self.layers.append(Convolution(self.params['w3'], self.params['b3'], conv_param_3['stride'],
conv_param_3['pad']))
self.layers.append(Relu())
self.layers.append(Convolution(self.params['w4'], self.params['b4'], conv_param_4['stride'],
conv_param_4['pad']))
self.layers.append(Relu())
self.layers.append(Pooling(pool_h=2, pool_w=2, stride=2))
self.layers.append(Convolution(self.params['w5'], self.params['b5'], conv_param_5['stride'],
conv_param_5['pad']))
self.layers.append(Relu())
self.layers.append(Convolution(self.params['w6'], self.params['b6'], conv_param_6['stride'],
conv_param_6['pad']))
self.layers.append(Relu())
self.layers.append(Pooling(pool_h=2, pool_w=2, stride=2))
self.layers.append(Affine(self.params['w7'], self.params['b7']))
self.layers.append(Relu())
self.layers.append(Dropout(0.5))
self.layers.append(Affine(self.params['w8'], self.params['b8']))
self.layers.append(Dropout(0.5))
self.last_layer = SoftmaxWithLoss()
def __init__(self, input_dim=(1, 28, 28),
conv_param_1 = {'filter_num':16, 'filter_size':3, 'pad':1, 'stride':1},
conv_param_2 = {'filter_num':16, 'filter_size':3, 'pad':1, 'stride':1},
conv_param_3 = {'filter_num':32, 'filter_size':3, 'pad':1, 'stride':1},
conv_param_4 = {'filter_num':32, 'filter_size':3, 'pad':2, 'stride':1},
conv_param_5 = {'filter_num':64, 'filter_size':3, 'pad':1, 'stride':1},
conv_param_6 = {'filter_num':64, 'filter_size':3, 'pad':1, 'stride':1},
hidden_size=50, output_size=10):
# 重みの初期化===========
# 各層のニューロンひとつあたりが、前層のニューロンといくつのつながりがあるか（TODO:自動で計算する）
pre_node_nums = np.array([1*3*3, 16*3*3, 16*3*3, 32*3*3, 32*3*3, 64*3*3, 64*4*4, hidden_size])
weight_init_scales = np.sqrt(2.0 / pre_node_nums) # ReLUを使う場合に推奨される初期値
self.params = {}
pre_channel_num = input_dim[0]
for idx, conv_param in enumerate([conv_param_1, conv_param_2, conv_param_3, conv_param_4, conv_param_5, conv_param_6]):
self.params['W' + str(idx+1)] = weight_init_scales[idx] * np.random.randn(conv_param['filter_num'], pre_channel_num, conv_param['filter_size'], conv_param['filter_size'])
self.params['b' + str(idx+1)] = np.zeros(conv_param['filter_num'])
pre_channel_num = conv_param['filter_num']
self.params['W7'] = weight_init_scales[6] * np.random.randn(64*4*4, hidden_size)
self.params['b7'] = np.zeros(hidden_size)
self.params['W8'] = weight_init_scales[7] * np.random.randn(hidden_size, output_size)
self.params['b8'] = np.zeros(output_size)
# レイヤの生成===========
self.layers = []
self.layers.append(Convolution(self.params['W1'], self.params['b1'],
conv_param_1['stride'], conv_param_1['pad']))
self.layers.append(Relu())
self.layers.append(Convolution(self.params['W2'], self.params['b2'],
conv_param_2['stride'], conv_param_2['pad']))
self.layers.append(Relu())
self.layers.append(Pooling(pool_h=2, pool_w=2, stride=2))
self.layers.append(Convolution(self.params['W3'], self.params['b3'],
conv_param_3['stride'], conv_param_3['pad']))
self.layers.append(Relu())
self.layers.append(Convolution(self.params['W4'], self.params['b4'],
conv_param_4['stride'], conv_param_4['pad']))
self.layers.append(Relu())
self.layers.append(Pooling(pool_h=2, pool_w=2, stride=2))
self.layers.append(Convolution(self.params['W5'], self.params['b5'],
conv_param_5['stride'], conv_param_5['pad']))
self.layers.append(Relu())
self.layers.append(Convolution(self.params['W6'], self.params['b6'],
conv_param_6['stride'], conv_param_6['pad']))
self.layers.append(Relu())
self.layers.append(Pooling(pool_h=2, pool_w=2, stride=2))
self.layers.append(Affine(self.params['W7'], self.params['b7']))
self.layers.append(Relu())
self.layers.append(Dropout(0.5))
self.layers.append(Affine(self.params['W8'], self.params['b8']))
self.layers.append(Dropout(0.5))
self.last_layer = SoftmaxWithLoss()
def __init__(self,
input_size,
hidden_size_list,
output_size,
activation="relu",
weight_init_std="relu",
weight_decay_lambda=0,
use_dropout=False,
dropout_ratio=0.5,
use_batchnorm=False):
"""
:param input_size: 输入的大小
:param hidden_size_list: 隐藏层的神经元数量列表
:param output_size: 输出的大小
:param activation: "relu" or "sigmoid"
:param weight_init_std: 指定权重的标准差，
指定"relu" 或者 "he" 是定为"He"的初始值
指定"sigmoid" 或者 "xavier" 是定为"Xauver"的初始值
:param weight_decay_lambda: Weight Decay(L2范数)的强度
:param use_dropout: 是否使用Dropout
:param dropout_ratio: Dropout比例
:param use_batchnorm: 是否只用Batch Normalization
"""
self.input_size = input_size
self.output_size = output_size
self.hidden_size_list = hidden_size_list
self.hidden_layer_num = len(hidden_size_list)
self.use_dropout = use_dropout
self.weight_decay_lambda = weight_decay_lambda
self.use_batchnorm = use_batchnorm
self.params = {}
# 初始化权值
self.__init_weight(weight_init_std)
# 生成层
activation_layer = {"sigmoid": Sigmoid, "relu": ReLU}
self.layers = OrderedDict()
for idx in range(1, self.hidden_layer_num + 1):
self.layers["Affine" + str(idx)] = Affine(
self.params["W" + str(idx)], self.params["b" + str(idx)])
if self.use_batchnorm:
self.params["gamma" + str(idx)] = np.ones(
hidden_size_list[idx - 1])
self.params["beta" + str(idx)] = np.zeros(
hidden_size_list[idx - 1])
self.layers['BatchNorm' + str(idx)] = BatchNormalization(
self.params['gamma' + str(idx)],
self.params['beta' + str(idx)])
self.layers["Activation_function" +
str(idx)] = activation_layer[activation]()
if self.use_dropout:
self.layers["Dropout" + str(idx)] = Dropout(dropout_ratio)
idx = self.hidden_layer_num + 1
self.layers["Affine" + str(idx)] = Affine(self.params["W" + str(idx)],
self.params["b" + str(idx)])
self.last_layer = SoftmaxWithLoss()
def __init__(
self,
input_size,
hidden_size_list,
output_size,
activation="relu",
weight_init_std="relu",
weight_decay_lambda=0,
use_dropout=False,
dropout_ratio=0.5,
use_batchnorm=False,
):
self.input_size = input_size
self.hidden_size_list = hidden_size_list
self.hidden_layer_num = len(hidden_size_list)
self.output_size = output_size
self.weight_decay_lambda = weight_decay_lambda
self.use_dropout = use_dropout
self.use_batchnorm = use_batchnorm
self.params = {}
# Weight initialization
self.__init_weight(weight_init_std)
# Create layers
activation_layer = {"sigmoid": Sigmoid, "relu": Relu}
self.layers = OrderedDict()
for idx in range(1, self.hidden_layer_num + 1):
self.layers["Affine" + str(idx)] = Affine(
self.params["W" + str(idx)], self.params["b" + str(idx)]
)
if self.use_batchnorm:
self.params["gamma" + str(idx)] = np.ones(hidden_size_list[idx - 1])
self.params["beta" + str(idx)] = np.zeros(hidden_size_list[idx - 1])
self.layers["BatchNorm" + str(idx)] = BatchNormalization(
self.params["gamma" + str(idx)], self.params["beta" + str(idx)]
)
self.layers["Activation_function" + str(idx)] = activation_layer[activation]()
if self.use_dropout:
self.layers["Dropout" + str(idx)] = Dropout(dropout_ratio)
idx = self.hidden_layer_num + 1
self.layers["Affine" + str(idx)] = Affine(
self.params["W" + str(idx)], self.params["b" + str(idx)]
)
self.last_layer = SoftmaxLoss()
def __init__(self, input_dim=(1, 28, 28),
conv_param={'filter_num':30, 'filter_size':5, 'pad':0, 'stride':1},
hidden_size=100, output_size=10, weight_init_std=0.01):
"""
input_size : 入力の配列形状(チャンネル数、画像の高さ、画像の幅)
conv_param : 畳み込みの条件, dict形式 例、{'filter_num':30, 'filter_size':5, 'pad':0, 'stride':1}
hidden_size : 隠れ層のノード数
output_size : 出力層のノード数
weight_init_std ： 重みWを初期化する際に用いる標準偏差
"""
filter_num = conv_param['filter_num']
filter_size = conv_param['filter_size']
filter_pad = conv_param['pad']
filter_stride = conv_param['stride']
input_size = input_dim[1]
conv_output_size = (input_size - filter_size + 2*filter_pad) / filter_stride + 1
pool_output_size = int(filter_num * (conv_output_size/2) * (conv_output_size/2))
# 重みの初期化
self.params = {}
std = weight_init_std
self.params['W1_1'] = std * np.random.randn(16, 1, 3,3)
self.params['b1_1'] = np.zeros(16)
self.params['W1_3'] = std * np.random.randn(16, 16, 3,3)
self.params['b1_3'] = np.zeros(16)
self.params['W2_1'] = std * np.random.randn(32, 16, 3,3)
self.params['b2_1'] = np.zeros(32)
self.params['W2_3'] = std * np.random.randn(32, 32, 3,3)
self.params['b2_3'] = np.zeros(32)
self.params['W3_1'] = std * np.random.randn(64, 32, 3,3)
self.params['b3_1'] = np.zeros(64)
self.params['W3_3'] = std * np.random.randn(64, 64, 3,3)
self.params['b3_3'] = np.zeros(64)
self.params['W4_1'] = std * np.random.randn(64*4*4, hidden_size)
self.params['b4_1'] = np.zeros(hidden_size)
self.params['W5_1'] = std * np.random.randn(hidden_size, output_size)
self.params['b5_1'] = np.zeros(output_size)
# レイヤの生成
self.layers = OrderedDict()
self.layers['Conv1_1'] = Convolution(self.params['W1_1'], self.params['b1_1'],1,1)
self.layers['ReLU1_2'] = ReLU()
self.layers['Conv1_3'] = Convolution(self.params['W1_3'], self.params['b1_3'],1,1)
self.layers['ReLU1_4'] = ReLU()
self.layers['Pool1_5'] = MaxPooling(pool_h=2, pool_w=2, stride=2)
self.layers['Conv2_1'] = Convolution(self.params['W2_1'], self.params['b2_1'],1,1)
self.layers['ReLU2_2'] = ReLU()
self.layers['Conv2_3'] = Convolution(self.params['W2_3'], self.params['b2_3'],1,2)
self.layers['ReLU2_4'] = ReLU()
self.layers['Pool2_5'] = MaxPooling(pool_h=2, pool_w=2, stride=2)
self.layers['Conv3_1'] = Convolution(self.params['W3_1'], self.params['b3_1'],1,1)
self.layers['ReLU3_2'] = ReLU()
self.layers['Conv3_3'] = Convolution(self.params['W3_3'], self.params['b3_3'],1,1)
self.layers['ReLU3_4'] = ReLU()
self.layers['Pool3_5'] = MaxPooling(pool_h=2, pool_w=2, stride=2)
self.layers['Affine4_1'] = Affine(self.params['W4_1'], self.params['b4_1'])
self.layers['ReLU4_2'] = ReLU()
self.layers['Dropout4_3'] = Dropout(0.5)
self.layers['Affine5_1'] = Affine(self.params['W5_1'], self.params['b5_1'])
self.layers['Dropout5_2'] = Dropout(0.5)
self.last_layer = SoftmaxWithLoss()
def __init__(self,
input_dim=(1, 28, 28),
use_conv2=True,
use_affine2=True,
conv_param={
'filter_num': 128,
'filter_size': 3,
'pad': 1,
'stride': 1
},
pool_param={
'pool_size': 2,
'pad': 1,
'stride': 2
},
conv_param2={
'filter_num2': 128,
'filter_size2': 3,
'pad2': 1,
'stride2': 1
},
pool_param2={
'pool_size2': 2,
'pad2': 1,
'stride2': 2
},
hidden_size=128,
hidden_size2=128,
output_size=15,
weight_init_std=0.01,
use_batchnorm_C1=False,
use_batchnorm_C2=False,
use_batchnorm_A1=False,
use_batchnorm_A2=False,
use_dropout_A1=False,
dropout_ratio_A1=0.5,
use_dropout_A2=False,
dropout_ratio_A2=0.5,
use_succession=False,
data_num=1,
prediction_mode=False):
filter_num = conv_param['filter_num']
filter_size = conv_param['filter_size']
filter_pad = conv_param['pad']
filter_stride = conv_param['stride']
pool_size = pool_param['pool_size']
pool_pad = pool_param['pad']
pool_stride = pool_param['stride']
filter_num2 = conv_param2['filter_num2']
filter_size2 = conv_param2['filter_size2']
filter_pad2 = conv_param2['pad2']
filter_stride2 = conv_param2['stride2']
pool_size2 = pool_param2['pool_size2']
pool_pad2 = pool_param2['pad2']
pool_stride2 = pool_param2['stride2']
input_size = input_dim[1]
conv_output_size = (input_size + 2 * filter_pad - filter_size
) // filter_stride + 1 # 畳み込み後のサイズ(H,W共通)
pool_output_size = (conv_output_size + 2 * pool_pad -
pool_size) // pool_stride + 1 # プーリング後のサイズ(H,W共通)
pool_output_pixel = filter_num * pool_output_size * pool_output_size # プーリング後のピクセル総数
input_size2 = pool_output_size
conv_output_size2 = (input_size2 + 2 * filter_pad2 - filter_size2
) // filter_stride2 + 1 # 畳み込み後のサイズ(H,W共通)
pool_output_size2 = (conv_output_size2 + 2 * pool_pad2 - pool_size2
) // pool_stride2 + 1 # プーリング後のサイズ(H,W共通)
pool_output_pixel2 = filter_num2 * pool_output_size2 * pool_output_size2 # プーリング後のピクセル総数
self.use_conv2 = use_conv2
self.use_affine2 = use_affine2
self.use_batchnorm_C1 = use_batchnorm_C1
self.use_batchnorm_C2 = use_batchnorm_C2
self.use_batchnorm_A1 = use_batchnorm_A1
self.use_batchnorm_A2 = use_batchnorm_A2
self.use_dropout_A1 = use_dropout_A1
self.use_dropout_A2 = use_dropout_A2
self.dropout_ratio_A1 = dropout_ratio_A1
self.dropout_ratio_A2 = dropout_ratio_A2
self.use_succession = use_succession
self.data_num = data_num
self.prediction_mode = prediction_mode
# if W1 == []:
self.params = {}
self.paramsB = {}
std = weight_init_std
if self.use_succession:
#----------重みをpickleから代入--------------
with open("params_" + str(self.data_num) + ".pickle", "rb") as f:
params_s = pickle.load(f)
with open("params_BN" + str(self.data_num) + ".pickle", "rb") as f:
params_BN = pickle.load(f)
# self.params = {}
# self.paramsB = {}
self.params['W1'] = params_s['W1'] # W1は畳み込みフィルターの重みになる
self.params['b1'] = params_s['b1']
if self.use_batchnorm_C1:
self.paramsB["BC1_moving_mean"] = params_BN["BC1_moving_mean"]
self.paramsB["BC1_moving_var"] = params_BN["BC1_moving_var"]
if self.use_conv2:
self.params['W1_2'] = params_s['W1_2']
self.params['b1_2'] = params_s['b1_2']
if self.use_batchnorm_C2:
self.paramsB["BC2_moving_mean"] = params_BN[
"BC2_moving_mean"]
self.paramsB["BC2_moving_var"] = params_BN[
"BC2_moving_var"]
self.params['W2'] = params_s['W2']
self.params['b2'] = params_s['b2']
if self.use_batchnorm_A1:
self.paramsB["BA1_moving_mean"] = params_BN["BA1_moving_mean"]
self.paramsB["BA1_moving_var"] = params_BN["BA1_moving_var"]
if self.use_affine2:
self.params['W2_2'] = params_s['W2_2']
self.params['b2_2'] = params_s['b2_2']
if self.use_batchnorm_A2:
self.paramsB["BA2_moving_mean"] = params_BN[
"BA2_moving_mean"]
self.paramsB["BA2_moving_var"] = params_BN[
"BA2_moving_var"]
self.params['W3'] = params_s['W3']
self.params['b3'] = params_s['b3']
#----------重みをpickleから代入--------------
else:
# 重みの初期化
#----第１層Conv----
self.params['W1'] = std * np.random.randn(
filter_num, input_dim[0], filter_size,
filter_size) # W1は畳み込みフィルターの重みになる
self.params['b1'] = np.zeros(filter_num) #b1は畳み込みフィルターのバイアスになる
#----第２層Conv----
if self.use_conv2:
self.params['W1_2'] = std * np.random.randn(
filter_num2, filter_num, filter_size2,
filter_size2) #-----追加------
self.params['b1_2'] = np.zeros(filter_num2) #-----追加------
#----第３層Affine----
self.params['W2'] = std * np.random.randn(
pool_output_pixel2, hidden_size)
else:
self.params['W2'] = std * np.random.randn(
pool_output_pixel, hidden_size)
self.params['b2'] = np.zeros(hidden_size)
#----第４層Affine----
if self.use_affine2:
self.params['W2_2'] = std * np.random.randn(
hidden_size, hidden_size2) #-----追加------
self.params['b2_2'] = np.zeros(hidden_size2) #-----追加------
#----第５層出力----
self.params['W3'] = std * np.random.randn(
hidden_size2, output_size) #--変更--
else:
self.params['W3'] = std * np.random.randn(
hidden_size, output_size) #--変更--
self.params['b3'] = np.zeros(output_size)
# レイヤの生成
self.layers = OrderedDict()
#----第１層Conv----
self.layers['Conv1'] = Convolution(
self.params['W1'], self.params['b1'], conv_param['stride'],
conv_param['pad']) # W1が畳み込みフィルターの重み, b1が畳み込みフィルターのバイアスになる
if self.use_batchnorm_C1:
print(conv_output_size)
print(conv_output_size ^ 2)
batch_num = conv_output_size * conv_output_size
if self.prediction_mode:
self.layers['BatchNormalization_C1'] = BatchNormalization(
np.ones(batch_num, filter_num),
np.zeros(filter_num),
moving_mean=self.paramsB["BC1_moving_mean"],
moving_var=self.paramsB["BC1_moving_var"])
else:
self.layers['BatchNormalization_C1'] = BatchNormalization(
np.ones(batch_num),
np.zeros(batch_num),
DataNum=self.data_num,
LayerNum="C1")
self.paramsB["BC1_moving_mean"] = self.layers[
'BatchNormalization_C1'].moving_mean
self.paramsB["BC1_moving_var"] = self.layers[
'BatchNormalization_C1'].moving_var
self.layers['ReLU1'] = ReLU()
self.layers['Pool1'] = MaxPooling(pool_h=pool_size,
pool_w=pool_size,
stride=pool_stride,
pad=pool_pad)
#----第２層Conv----
if self.use_conv2:
self.layers['Conv1_2'] = Convolution(
self.params['W1_2'], self.params['b1_2'],
conv_param2['stride2'], conv_param2['pad2']) #-----追加------
if self.use_batchnorm_C2:
batch_num2 = conv_output_size2 * conv_output_size2 * filter_num2
if self.prediction_mode:
self.layers['BatchNormalization_C2'] = BatchNormalization(
np.ones(batch_num),
np.zeros(batch_num),
moving_mean=self.paramsB["BC2_moving_mean"],
moving_var=self.paramsB["BC12moving_var"])
else:
self.layers['BatchNormalization_C2'] = BatchNormalization(
np.ones(batch_num),
np.zeros(batch_num),
DataNum=self.data_num,
LayerNum="C2")
self.paramsB["BC2_moving_mean"] = self.layers[
'BatchNormalization_C2'].moving_mean
self.paramsB["BC2_moving_var"] = self.layers[
'BatchNormalization_C2'].moving_var
self.layers['ReLU1_2'] = ReLU() #-----追加------
self.layers['Pool1_2'] = MaxPooling(pool_h=pool_size2,
pool_w=pool_size2,
stride=pool_stride2,
pad=pool_pad2) #-----追加------
#----第３層Affine----
self.layers['Affine1'] = Affine(self.params['W2'], self.params['b2'])
if self.use_batchnorm_A1:
if self.prediction_mode:
self.layers['BatchNormalization_A1'] = BatchNormalization(
np.ones(hidden_size),
np.zeros(hidden_size),
moving_mean=self.paramsB["BA1_moving_mean"],
moving_var=self.paramsB["BA1_moving_var"])
else:
self.layers['BatchNormalization_A1'] = BatchNormalization(
np.ones(hidden_size),
np.zeros(hidden_size),
DataNum=self.data_num,
LayerNum="A1")
self.paramsB["BA1_moving_mean"] = self.layers[
'BatchNormalization_A1'].moving_mean
self.paramsB["BA1_moving_var"] = self.layers[
'BatchNormalization_A1'].moving_var
if self.use_dropout_A1:
self.layers['DropoutA1'] = Dropout(self.dropout_ratio_A1)
self.layers['ReLU2'] = ReLU()
# ----第４層Affine----
if self.use_affine2:
self.layers['Affine2'] = Affine(
self.params['W2_2'], self.params['b2_2']) #-----追加------
if self.use_batchnorm_A2:
if self.prediction_mode:
self.layers['BatchNormalization_A2'] = BatchNormalization(
np.ones(hidden_size2),
np.zeros(hidden_size2),
moving_mean=self.paramsB["BA2_moving_mean"],
moving_var=self.paramsB["BA2_moving_var"])
else:
self.layers['BatchNormalization_A2'] = BatchNormalization(
np.ones(hidden_size2),
np.zeros(hidden_size2),
DataNum=self.data_num,
LayerNum="A2")
self.paramsB["BA2_moving_mean"] = self.layers[
'BatchNormalization_A2'].moving_mean
self.paramsB["BA2_moving_var"] = self.layers[
'BatchNormalization_A2'].moving_var
if self.use_dropout_A2:
self.layers['DropoutA2'] = Dropout(self.dropout_ratio_A2)
self.layers['ReLU3'] = ReLU() #-----追加------
#----第５層出力----
self.layers['Affine3'] = Affine(self.params['W3'], self.params['b3'])
self.last_layer = SoftmaxWithLoss()
def __init__(self,
input_dim=(1, 28, 28),
conv_param_1=None,
conv_param_2=None,
conv_param_3=None,
conv_param_4=None,
conv_param_5=None,
conv_param_6=None,
hidden_size=50,
output_size=10):
# 第一个卷积层输入1x28x28，输出16x28x28
if conv_param_1 is None:
conv_param_1 = {
'filter_num': 16,
'filter_size': 3,
'pad': 1,
'stride': 1
}
# 第二个卷积层输入16x28x28，输出16x28x28
if conv_param_2 is None:
conv_param_2 = {
'filter_num': 16,
'filter_size': 3,
'pad': 1,
'stride': 1
}
# 第二个卷积层之后接最大池化层，池化层大小为2x2,步长为2，即高、宽减半
# 第三个卷积层输入16x14x14，输出32x14x14
if conv_param_3 is None:
conv_param_3 = {
'filter_num': 32,
'filter_size': 3,
'pad': 1,
'stride': 1
}
# 第四个卷积层输入32x14x14，但由于pad2个，因此输出32x16x16
if conv_param_4 is None:
conv_param_4 = {
'filter_num': 32,
'filter_size': 3,
'pad': 2,
'stride': 1
}
# 第四个卷积层之后接最大池化层，池化层大小为2x2,步长为2，即高、宽减半
# 第五个卷积层输入32x8x8，输出64x8x8
if conv_param_5 is None:
conv_param_5 = {
'filter_num': 64,
'filter_size': 3,
'pad': 1,
'stride': 1
}
# 第五个卷积层输入64x8x8，输出64x8x8
if conv_param_6 is None:
conv_param_6 = {
'filter_num': 64,
'filter_size': 3,
'pad': 1,
'stride': 1
}
"""
卷积层的每个节点只与前一层的filter_size个节点连接，
即本层卷积层的卷积核 高x宽有多少，就和前一层的多少个节点连接。
如果有多个通道，那还要乘上通道数(深度)
这里的所有卷积层都用3x3的大小
各层输出如下：
卷积层1： 16 28 28
卷积层2 | 池化层1： 16 28 28 | 16 14 14
卷积层3： 32 14 14
卷积层4 | 池化层2： 32 16 16 | 32 8 8
卷积层5： 64 8 8
卷积层6： 64 8 8 | 64 4 4
"""
pre_node_nums = np.array([
1 * 3 * 3, # 卷积层1：前一层(输入层)通道数(深度)为1
16 * 3 * 3, # 卷积层2：前一层(卷积层1)通道数(深度)为16
16 * 3 * 3, # 卷积层3：前一层(卷积层2)通道数(深度)为16
32 * 3 * 3, # 卷积层4：前一层(卷积层3)通道数(深度)为32
32 * 3 * 3, # 卷积层5：前一层(卷积层4)通道数(深度)为32
64 * 3 * 3, # 卷积层6：前一层(卷积层5)通道数(深度)为64
# 隐藏层：前一层(池化层)，池化层接全连接层需要拉直成一维数组，
# 因此隐藏层与前一层(池化层)的连接数为池化层的输出节点总数
64 * 4 * 4,
# 输出层：前一层(隐藏层)，全连接与前一层全部节点相连，即隐藏层大小
hidden_size
])
# 权重初始化时的标准差。由于使用ReLU激活函数，因此使用He初始化方式
weight_init_scales = np.sqrt(2.0 / pre_node_nums)
"""初始化权重参数和偏置"""
self.params = {}
pre_channel_num = input_dim[0] # 记录上一层的通道数(即滤波器的通道数)
for idx, conv_param in enumerate([
conv_param_1, conv_param_2, conv_param_3, conv_param_4,
conv_param_5, conv_param_6
]):
# 卷积层滤波器的形状：滤波器个数、通道数、高度、宽度
self.params['W'+str(idx+1)] = weight_init_scales[idx] *\
np.random.randn(
conv_param['filter_num'],
pre_channel_num,
conv_param['filter_size'],
conv_param['filter_size'])
self.params['b' + str(idx + 1)] = np.zeros(
conv_param['filter_num'])
pre_channel_num = conv_param['filter_num'] # 更新上一层的通道数
self.params['W7'] = weight_init_scales[6] * np.random.randn(
64 * 4 * 4, hidden_size)
self.params['b7'] = np.zeros(hidden_size)
self.params['W8'] = weight_init_scales[7] * np.random.randn(
hidden_size, output_size)
self.params['b8'] = np.zeros(output_size)
"""
构造神经网络：
书上没有用到之前用的有序字典，其实我觉得很好用，所以就实现了有序字典版本
Conv1->ReLU1->Conv2->ReLU2->Pool1->
Conv3->ReLU3->Conv4->ReLU4->Pool2->
Conv5->ReLU5->Conv6->ReLU6->Pool3->
Affine1(Hidden Layer1)->ReLU7->Dropout1->
Affine2(Output Layer1)->Dropout2------->SoftmaxWithLoss
"""
self.layers = OrderedDict()
self.layers['Conv1'] = Convolution(self.params['W1'],
self.params['b1'],
stride=conv_param_1['stride'],
pad=conv_param_1['pad'])
self.layers['ReLU1'] = ReLU()
self.layers['Conv2'] = Convolution(self.params['W2'],
self.params['b2'],
stride=conv_param_2['stride'],
pad=conv_param_2['pad'])
self.layers['ReLU2'] = ReLU()
self.layers['Pool1'] = Pooling(pool_h=2, pool_w=2, stride=2, pad=0)
self.layers['Conv3'] = Convolution(self.params['W3'],
self.params['b3'],
stride=conv_param_3['stride'],
pad=conv_param_3['pad'])
self.layers['ReLU3'] = ReLU()
self.layers['Conv4'] = Convolution(self.params['W4'],
self.params['b4'],
stride=conv_param_4['stride'],
pad=conv_param_4['pad'])
self.layers['ReLU4'] = ReLU()
self.layers['Pool2'] = Pooling(pool_h=2, pool_w=2, stride=2, pad=0)
self.layers['Conv5'] = Convolution(self.params['W5'],
self.params['b5'],
stride=conv_param_5['stride'],
pad=conv_param_5['pad'])
self.layers['ReLU5'] = ReLU()
self.layers['Conv6'] = Convolution(self.params['W6'],
self.params['b6'],
stride=conv_param_6['stride'],
pad=conv_param_6['pad'])
self.layers['ReLU6'] = ReLU()
self.layers['Pool3'] = Pooling(pool_h=2, pool_w=2, stride=2, pad=0)
self.layers['Affine1'] = Affine(self.params['W7'], self.params['b7'])
self.layers['ReLU7'] = ReLU()
self.layers['Dropout1'] = Dropout(dropout_ratio=0.5)
self.layers['Affine2'] = Affine(self.params['W8'], self.params['b8'])
self.layers['Dropout2'] = Dropout(dropout_ratio=0.5)
self.last_layer = SoftmaxWithLoss()