gluonネットワークの可視化
4759 ワード
1,
栗を挙げます.
print(net)を利用する.Netネットワークの具体的な各層のパラメータが表示されます.2,mx.viz.plot_network(net).view()を利用する.これはpdf画像を具体的に出すことができますが、symプログラミングに移行しなければなりません.どうやってここに?栗を挙げます.
import mxnet as mx
from mxnet.gluon import nn
class Net(nn.HybridBlock):
def __init__(self, **kwargs):
super(Net, self).__init__(**kwargs)
with self.name_scope():
self.rgb_conv1 = nn.Conv2D(channels=128, kernel_size=3, strides=1, padding=1, activation='relu')
self.rgb_conv2 = nn.Conv2D(channels=128, kernel_size=3, strides=1, padding=1, activation='relu')
self.tdf_conv1 = nn.Conv3D(channels=128, kernel_size=3, strides=1, padding=1, activation='relu')
self.tdf_conv2 = nn.Conv3D(channels=128, kernel_size=3, strides=1, padding=1, activation='relu')
self.dense_fusion1 = nn.Dense(1024, activation='relu')
self.dense_fusion2 = nn.Dense(2048, activation='relu')
self.dense_fusion3 = nn.Dense(512, activation='relu')
self.dense_prediction1 = nn.Dense(2048, activation='relu')
self.dense_prediction2 = nn.Dense(2)
self.maxpool2D = nn.MaxPool2D(pool_size=3, strides=2)
self.maxpool3D = nn.MaxPool3D(pool_size=3, strides=2)
def CNNBlock(self, F, rgb, tdf):
rgb_conv = self.rgb_conv1(rgb)
rgb_conv = self.maxpool2D(rgb_conv)
rgb_conv = self.rgb_conv2(rgb_conv)
tdf_conv = self.tdf_conv1(tdf)
tdf_conv = self.maxpool3D(tdf_conv)
tdf_conv = self.tdf_conv2(tdf_conv)
flatten = nn.Flatten()
rgb_conv = flatten(rgb_conv)
tdf_conv = flatten(tdf_conv)
fc = F.concat(rgb_conv, tdf_conv, dim=1)
fc = self.dense_fusion1(fc)
fc = self.dense_fusion2(fc)
fc = self.dense_fusion3(fc)
return fc
def hybrid_forward(self, F, rgb1, tdf1, rgb2, tdf2):
out1 = self.CNNBlock(F, rgb1, tdf1)
out2 = self.CNNBlock(F, rgb2, tdf2)
out = F.concat(out1, out2, dim=1)
out = self.dense_prediction2(self.dense_prediction1(out))
return out
def getFeature(self, img, depth):
return self.CNNBlock(img, depth)
net = Net()
#
print(net)
#
mx.viz.plot_network(net(mx.sym.var("data1"), mx.sym.var("data2"), mx.sym.var("data3"), mx.sym.var("data4"))).view()