1906.07 Chainer NNモデルの中間層でデータを分岐・結合する方法

1906.07 Chainer NNモデルの中間層でデータを分岐・結合

chainer.functionsの「reshape」と「concat」を用いて結合できる模様。

• https://docs.chainer.org/en/stable/reference/generated/chainer.functions.reshape.html
• https://docs.chainer.org/en/stable/reference/generated/chainer.functions.concat.html#chainer.functions.concat

In [1]:

# sec: lib

import numpy as np

import chainer
import chainer.links as L
import chainer.functions as F

# sec: ver

chainer.print_runtime_info()

Platform: Windows-7-6.1.7601-SP1
Chainer: 5.3.0
NumPy: 1.16.2
CuPy: Not Available
iDeep: Not Available

NNモデルを定義

In [61]:

# sec: NNモデルの定義

class MyConvNet(chainer.Chain):

    def __init__(self):
        super(MyConvNet, self).__init__()

        # パラメータを持つ層の登録
        with self.init_scope():
            self.c1 = L.Convolution2D(None, 10, ksize=3, stride=2, pad=0)
            self.c2 = L.Convolution2D(None, 10, ksize=5, stride=2, pad=0)
            self.l1 = L.Linear(None, 100)
            self.l2 = L.Linear(None, 100)
            self.l3 = L.Linear(None, 40)

    def forward(self, x):
        # データを受け取った際のforward計算を書く
        
        # sec: conv
        
        print("x", x.shape, np.prod(x.shape[1:]))
        h1 = F.relu(self.c1(x))
        print("conv1", h1.shape, np.prod(h1.shape[1:]))
        h2 = F.relu(self.c2(x))
        print("conv2", h2.shape, np.prod(h2.shape[1:]))
        
        # sec: concat
        
        h1 = F.reshape(h1, (h1.shape[0], -1))
        print("reshape1", h1.shape)
        h2 = F.reshape(h2, (h2.shape[0], -1))
        print("reshape2", h2.shape)
        h = F.concat([h1, h2], axis=1)
        print("concat12", h.shape)
        
        # sec: fc
        
        h = F.relu(self.l1(h))
        print("fc1", h.shape)
        h = F.relu(self.l2(h))
        print("fc2", h.shape)
        h = self.l3(h)
        print("fc3", h.shape)
        return h

NNモデルをチェック

In [62]:

n_batch = 128
n_ch = 3
n_height = 20
n_width = 120

# sec: NNモデルの出力値を計算

model = MyConvNet()

x = np.zeros((n_batch, n_ch, n_height, n_width), dtype="float32")
y = model(x)
print("y", y.shape)

x (128, 3, 20, 120) 7200
conv1 (128, 10, 9, 59) 5310
conv2 (128, 10, 8, 58) 4640
reshape1 (128, 5310)
reshape2 (128, 4640)
concat12 (128, 9950)
fc1 (128, 100)
fc2 (128, 100)
fc3 (128, 40)
y (128, 40)

In [63]:

# sec: 勾配を計算

loss = F.softmax_cross_entropy(y, np.zeros(n_batch, dtype="int32"))
print(loss)

model.cleargrads()
loss.backward()

variable(3.68888)