def different_layer_num():
"""
不同的网络层结构的对照试验
第一个是采用1层卷积 + 1层全连接的结构的情况
第二个是采用2层卷积 + 2层全连接的结构的情况
"""
conv1_shape = [3, 3, 1, 20] # -------关键参数--卷积层1
fc1_shape = [14 * 14 * 20, 10] # -------关键参数--全连接层1
learning_rate = 0.001
keep_prob = 0.7
max_iteration = 20
result1 = _1_1_model(conv1_shape,
fc1_shape,
learning_rate=learning_rate,
keep=keep_prob,
max_iteration=max_iteration,
save_result=True)
conv1_shape = [5, 5, 1, 20] # -------关键参数--卷积层1
conv2_shape = [3, 3, 20, 50] # -------关键参数--卷积层2
fc1_shape = [7 * 7 * 50, 500] # -------关键参数--全连接层1
fc2_shape = [500, 10] # -------关键参数--全连接层2
learning_rate = 0.001
keep_prob = 0.7
max_iteration = 20
result2 = _2_2_model(conv1_shape,
conv2_shape,
fc1_shape,
fc2_shape,
learning_rate=learning_rate,
keep=keep_prob,
max_iteration=max_iteration,
save_result=True)
plot(result1, result2, "1+1", "2+2", "different_layer_num")
def different_dropout():
"""
不同的keep_prob值的对照试验
采用2层卷积 + 2层全连接的结构
第一个是keep_prob = 0.7的情况
第二个是keep_prob = 1.0的情况
"""
conv1_shape = [5, 5, 1, 20]
conv2_shape = [5, 5, 20, 50]
fc1_shape = [7 * 7 * 50, 500]
fc2_shape = [500, 10]
learning_rate = 0.001
keep_prob = 0.7 # -------关键参数--0.7
max_iteration = 20
result1 = _2_2_model(conv1_shape,
conv2_shape,
fc1_shape,
fc2_shape,
learning_rate=learning_rate,
keep=keep_prob,
max_iteration=max_iteration,
save_result=True)
conv1_shape = [5, 5, 1, 20]
conv2_shape = [5, 5, 20, 50]
fc1_shape = [7 * 7 * 50, 500]
fc2_shape = [500, 10]
learning_rate = 0.001
keep_prob = 1.0 # -------关键参数--1.0
max_iteration = 20
result2 = _2_2_model(conv1_shape,
conv2_shape,
fc1_shape,
fc2_shape,
learning_rate=learning_rate,
keep=keep_prob,
max_iteration=max_iteration,
save_result=True)
plot(result1, result2, "dropout_0.7", "dropout_1.0", "different_dropout")
def different_learning_rate():
"""
不同的学习速率大小的对照试验
第一个是采用 0.001 的速率的情况
第二个是采用 0.0005 的速率的情况
"""
conv1_shape = [5, 5, 1, 20]
conv2_shape = [5, 5, 20, 50]
fc1_shape = [7 * 7 * 50, 500]
fc2_shape = [500, 10]
learning_rate = 0.001 # -------关键参数--学习率，0.001
keep_prob = 0.7
max_iteration = 20
result1 = _2_2_model(conv1_shape,
conv2_shape,
fc1_shape,
fc2_shape,
learning_rate=learning_rate,
keep=keep_prob,
max_iteration=max_iteration,
save_result=True)
conv1_shape = [5, 5, 1, 20]
conv2_shape = [5, 5, 20, 50]
fc1_shape = [7 * 7 * 50, 500]
fc2_shape = [500, 10]
learning_rate = 0.0005 # -------关键参数--学习率，0.0005
keep_prob = 0.7
max_iteration = 20
result2 = _2_2_model(conv1_shape,
conv2_shape,
fc1_shape,
fc2_shape,
learning_rate=learning_rate,
keep=keep_prob,
max_iteration=max_iteration,
save_result=True)
plot(result1, result2, "rate0.001", "rate0.0005",
"different_learning_rate")
def different_filter_size():
"""
不同的卷积核大小的对照试验
第一个是采用 3 * 3 的卷积核的情况
第二个是采用 5 * 5 的卷积核的情况
"""
conv1_shape = [3, 3, 1, 20] # -------关键参数--卷积核，3 * 3
conv2_shape = [3, 3, 20, 50] # -------关键参数--卷积核，3 * 3
fc1_shape = [7 * 7 * 50, 500]
fc2_shape = [500, 10]
learning_rate = 0.001
keep_prob = 0.7
max_iteration = 20
result1 = _2_2_model(conv1_shape,
conv2_shape,
fc1_shape,
fc2_shape,
learning_rate=learning_rate,
keep=keep_prob,
max_iteration=max_iteration,
save_result=True)
conv1_shape = [5, 5, 1, 20] # -------关键参数--卷积核，5 * 5
conv2_shape = [5, 5, 20, 50] # -------关键参数--卷积核，5 * 5
fc1_shape = [7 * 7 * 50, 500]
fc2_shape = [500, 10]
learning_rate = 0.001
keep_prob = 0.7
max_iteration = 20
result2 = _2_2_model(conv1_shape,
conv2_shape,
fc1_shape,
fc2_shape,
learning_rate=learning_rate,
keep=keep_prob,
max_iteration=max_iteration,
save_result=True)
plot(result1, result2, "3*3 filter", "5*5 filter", "different_filter_size")
def custom():
"""
这里自定义参数模型
"""
conv1_shape = [5, 5, 1, 20]
conv2_shape = [5, 5, 20, 50]
fc1_shape = [7 * 7 * 50, 500]
fc2_shape = [500, 10]
learning_rate = 0.001
keep_prob = 0.7
max_iteration = 2
result = _2_2_model(conv1_shape,
conv2_shape,
fc1_shape,
fc2_shape,
learning_rate=learning_rate,
keep=keep_prob,
max_iteration=max_iteration,
save_result=True)
plot_custom(result, model_name="test", title="test")