Keras关于trainable的实验

Echo木

已于 2023-02-19 02:30:43 修改

阅读量879

点赞数 4

分类专栏：机器学习/TensorFlow/PyTorch 文章标签：机器学习 Keras

于 2021-01-28 20:36:59 首次发布

本文链接：https://blog.csdn.net/qq_43015237/article/details/113357939

版权

机器学习/TensorFlow/PyTorch 专栏收录该内容

6 篇文章 0 订阅

订阅专栏

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
# @Time   : 2021/1/28 17:20
# @Author : Jin Echo
"""
结论：
1.对已经编译后的模型设置不可训练trainable，设置之后还需要再编译才会生效，否则会有一个warning询问是否要进行编译
2.对组合网络，例如C = G+D，D编译后设置False，再C编译，则训练C时D参数不变；训练D时参数改变(D未编译)
此外，也对拼接的网络进行了实验
"""

from keras.models import Sequential, Model
from keras.layers import Input, Dense, Activation, Dropout, Subtract
import numpy as np

x = Input(shape=(10, ))
y = Dense(10)(x)
G = Model(inputs=x, outputs=y)                  # 110参数
# print('G.summary:')
# G.summary
G.compile(optimizer='rmsprop',
          loss='binary_crossentropy',
          metrics=['accuracy'])
# print('G.summary:')
# G.summary()

x = Input(shape=(10, ))
x_gen = Input(shape=(10, ))
output = Subtract()([x, x_gen])
output = Dense(1, activation='sigmoid')(output)
D = Model(inputs=[x, x_gen], outputs=output)        # 11参数
D.compile(optimizer='rmsprop',
          loss='binary_crossentropy',
          metrics=['accuracy'])
# print('D.summary:')
# D.summary()

x = Input(shape=(10, ))
x_gen = G(x)
pre = D([x, x_gen])
C = Model(inputs=x, outputs=pre)            # 121参数
C.compile(optimizer='rmsprop',
          loss='binary_crossentropy',
          metrics=['accuracy'])
# print('C.summary:')
# C.summary()


"""
# 实验1：D.fit
# 对D，尽管设置了False，但未编译，D仍然进行了训练
# (每轮的loss在下降，训练前后的参数改变都可证明)
# C与G、D参数同步，注意此时D、C的sumary结果很怪异，还会输出一个warning
D.trainable = False
G.summary()
D.summary()
C.summary()
print(G.get_weights())
print(D.get_weights())
print(C.get_weights())
print('--------------------------------------------------------------------------')
D.fit(x=[np.random.uniform(size=(2, 10)), np.random.uniform(size=(2, 10))],
      y=np.ones(shape=(2, 1)), epochs=3, verbose=2)
G.summary()
D.summary()
C.summary()
print(G.get_weights())
print(D.get_weights())
print(C.get_weights())
"""


"""
# 实验2：D.fit
# 对D，设置了False，并且D编译，D才不会进行训练)
# (每轮的loss不变，训练前后的参数不变都可证明)
# C与G、D参数同步，注意此时C的sumary结果很怪异，还会输出一个warning
D.trainable = False
D.compile(optimizer='rmsprop',
          loss='binary_crossentropy',
          metrics=['accuracy'])
G.summary()
D.summary()
C.summary()
print(G.get_weights())
print(D.get_weights())
print(C.get_weights())
print('--------------------------------------------------------------------------')
D.fit(x=[np.random.uniform(size=(2, 10)), np.random.uniform(size=(2, 10))],
      y=np.ones(shape=(2, 1)), epochs=3, verbose=2)
G.summary()
D.summary()
C.summary()
print(G.get_weights())
print(D.get_weights())
print(C.get_weights())
"""


"""
# 实验3：D.fit
# 对D，设置了False，并且C编译D不编译，D会进行训练
# (每轮的loss变化，训练前后D的参数变化都可证明)
# C与G、D参数同步，注意此时D的sumary结果很怪异，还会输出一个warning
D.trainable = False
C.compile(optimizer='rmsprop',
          loss='binary_crossentropy',
          metrics=['accuracy'])
G.summary()
D.summary()
C.summary()
print(G.get_weights())
print(D.get_weights())
print(C.get_weights())
print('--------------------------------------------------------------------------')
D.fit(x=[np.random.uniform(size=(2, 10)), np.random.uniform(size=(2, 10))],
      y=np.ones(shape=(2, 1)), epochs=3, verbose=2)
G.summary()
D.summary()
C.summary()
print(G.get_weights())
print(D.get_weights())
print(C.get_weights())
"""


"""
# 实验4：C.fit
# a.对D设置False，D编译，C并不编译
# a.结果：loss减小，G参数变化、D参数变化，C=G+D，训练前后C的参数与G和D同步变换，C.summary怪异出现warning(相当于False未设置成功)
# b.对D设置False，C编译，D并不编译
# b.结果：loss减小，G参数变化、D参数不变、C=G+D，训练前后C的参数与G和D同步变换，D.summary怪异出现warning
# c.对D设置False，D、C先后编译
# c.结果：参数变换情况与b相同，差别仅是无怪异无waring
# d.对D设置False，C、D先后编译
# d.结果：与C相同
D.trainable = False
C.compile(optimizer='rmsprop',
          loss='binary_crossentropy',
          metrics=['accuracy'])
D.compile(optimizer='rmsprop',
          loss='binary_crossentropy',
          metrics=['accuracy'])
G.summary()
D.summary()
C.summary()
print(G.get_weights())
print(D.get_weights())
print(C.get_weights())
print('-------------------------------------------------------------------------------------------------------')
C.fit(x=np.random.uniform(size=(2, 10)),
      y=np.ones(shape=(2, 1)), epochs=3, verbose=2)
G.summary()
D.summary()
C.summary()
print(G.get_weights())
print(D.get_weights())
print(C.get_weights())
"""