如何使用 3D 训练数据构建二元分类器
How to build a binary classifier with 3D training data
我有必须分类为 0 或 1 的数据。数据是从 .npz 文件加载的。它为我提供了训练、验证和测试数据。这是他们的样子：
x_train = [[[ 0 0 0 ... 0 1 4]
[ 0 0 0 ... 4 25 2]
[ 6 33 15 ... 33 0 0]
...
[ 0 23 4 ... 9 31 0]
[ 4 0 0 ... 0 0 12]
[ 5 0 0 ... 3 0 0]]
[[ 88 71 59 ... 61 62 62]
[ 74 88 73 ... 59 70 60]
[ 69 61 85 ... 60 58 82]
...
[ 68 85 58 ... 55 75 72]
[ 69 69 70 ... 81 76 83]
[ 74 68 76 ... 60 74 72]]
[[ 87 134 146 ... 108 116 157]
[108 117 144 ... 102 58 122]
[124 148 106 ... 97 135 146]
...
[ 96 153 111 ... 104 129 154]
[129 140 100 ... 74 114 97]
[119 115 160 ... 172 84 148]]
...
[[ 92 96 64 ... 69 83 83]
[ 85 44 89 ... 115 94 76]
[ 93 103 91 ... 92 81 75]
...
[ 16 109 81 ... 84 95 20]
[100 27 89 ... 66 107 48]
[ 24 67 144 ... 104 115 123]]
[[ 69 70 74 ... 72 73 75]
[ 72 72 76 ... 73 75 76]
[ 74 75 72 ... 72 69 73]
...
[ 72 72 69 ... 72 76 72]
[ 70 72 73 ... 72 76 67]
[ 69 72 72 ... 72 71 71]]
[[ 65 137 26 ... 134 57 174]
[ 91 76 123 ... 39 63 124]
[ 81 203 134 ... 192 63 143]
...
[ 1 102 96 ... 33 63 169]
[ 82 32 108 ... 151 75 151]
[ 12 97 164 ... 101 125 60]]]
y_train:
[0 0 0 ... 0 0 0]
这些是我的输入形状：
x_train.shape = (5000, 128, 128)
y_train.shape = (5000,)
如您所见，y 只是标签，x 只是 3D 数据。因为它是一个二元分类器，所以我想构建一个具有 3 个密集层的简单神经网络。这是我的：
model = Sequential()
model.add(Dense(12, input_dim = 8, activation='relu'))
model.add(Dense(8, activation='relu'))
model.add(Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
model.fit(x_train, y_train, epochs=150, batch_size=8)
但是由于我的输入，出现了这个错误：
ValueError: Input 0 of layer sequential_4 is incompatible with the layer: expected axis -1 of input shape to have value 8 but received input with shape (8, 128, 128)
我该如何解决这个问题？我的神经网络对于这类问题是不是太简单了？
您的代码存在多个问题。我尝试添加单独的部分来解释它们。请仔细阅读所有这些并尝试我在下面显示的代码示例。
1。将 samples/batch 通道作为输入维度传递
您正在将批处理通道作为密集层的输入形状传递。那是不正确的。相反，您需要做的是传递模型应该期望的每个样本的形状，在本例中为 (128,128)。该模型自动在前面添加一个通道，以便批次流经计算图为(None, 128, 128)，如下图model.summary()
2。密集层的二维输入。
您的每个样本（在本例中共有 5000 个样本）都是形状为 128,128 的二维矩阵。如果不先将其压平，致密层不能直接消耗它。 （或者使用不同的层来更适合于 2D/3D 输入，稍后讨论）。
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense, Flatten
model = Sequential()
model.add(Flatten(input_shape=(128,128)))
model.add(Dense(12, activation='relu'))
model.add(Dense(8, activation='relu'))
model.add(Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
model.summary()
Model: "sequential_6"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
flatten_3 (Flatten) (None, 16384) 0
_________________________________________________________________
dense_5 (Dense) (None, 12) 196620
_________________________________________________________________
dense_6 (Dense) (None, 8) 104
_________________________________________________________________
dense_7 (Dense) (None, 1) 9
=================================================================
Total params: 196,733
Trainable params: 196,733
Non-trainable params: 0
_________________________________________________________________
3。针对您的问题使用不同的体系结构。
"Is my NN too simplistic for this type of problem?"
这不是关于架构的复杂性，而是更多关于可以处理某种类型数据的层的类型。在这种情况下，您的图像具有单通道 (128,128)，这是一个 2D 输入。通常，彩色图像有 R、G、B 通道，最终以 (128,128,3) 形输入。
一般做法是为此使用 CNN 层。
下面显示了一个例子 -
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense, Flatten, Conv2D, MaxPooling2D, Reshape
model = Sequential()
model.add(Reshape((128,128,1), input_shape=(128,128)))
model.add(Conv2D(5, 5, activation='relu'))
model.add(MaxPooling2D((2,2)))
model.add(Conv2D(10, 5, activation='relu'))
model.add(MaxPooling2D((2,2)))
model.add(Conv2D(20, 5, activation='relu'))
model.add(MaxPooling2D((2,2)))
model.add(Conv2D(30, 5, activation='relu'))
model.add(MaxPooling2D((2,2)))
model.add(Flatten())
model.add(Dense(12, activation='relu'))
model.add(Dense(8, activation='relu'))
model.add(Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
model.summary()
Model: "sequential_13"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
reshape_5 (Reshape) (None, 128, 128, 1) 0
_________________________________________________________________
conv2d_16 (Conv2D) (None, 124, 124, 5) 130
_________________________________________________________________
max_pooling2d_16 (MaxPooling (None, 62, 62, 5) 0
_________________________________________________________________
conv2d_17 (Conv2D) (None, 58, 58, 10) 1260
_________________________________________________________________
max_pooling2d_17 (MaxPooling (None, 29, 29, 10) 0
_________________________________________________________________
conv2d_18 (Conv2D) (None, 25, 25, 20) 5020
_________________________________________________________________
max_pooling2d_18 (MaxPooling (None, 12, 12, 20) 0
_________________________________________________________________
conv2d_19 (Conv2D) (None, 8, 8, 30) 15030
_________________________________________________________________
max_pooling2d_19 (MaxPooling (None, 4, 4, 30) 0
_________________________________________________________________
flatten_9 (Flatten) (None, 480) 0
_________________________________________________________________
dense_23 (Dense) (None, 12) 5772
_________________________________________________________________
dense_24 (Dense) (None, 8) 104
_________________________________________________________________
dense_25 (Dense) (None, 1) 9
=================================================================
Total params: 27,325
Trainable params: 27,325
Non-trainable params: 0
_________________________________________________________________
要了解 Conv2D 层和 MaxPooling 层的作用，请查看我的 well-accepted answer on Data Science stack exchange, or check out this blog。但是下图是一种在脑海中形象化它正在做的事情的一般方法。
我有必须分类为 0 或 1 的数据。数据是从 .npz 文件加载的。它为我提供了训练、验证和测试数据。这是他们的样子：
x_train = [[[ 0 0 0 ... 0 1 4]
[ 0 0 0 ... 4 25 2]
[ 6 33 15 ... 33 0 0]
...
[ 0 23 4 ... 9 31 0]
[ 4 0 0 ... 0 0 12]
[ 5 0 0 ... 3 0 0]]
[[ 88 71 59 ... 61 62 62]
[ 74 88 73 ... 59 70 60]
[ 69 61 85 ... 60 58 82]
...
[ 68 85 58 ... 55 75 72]
[ 69 69 70 ... 81 76 83]
[ 74 68 76 ... 60 74 72]]
[[ 87 134 146 ... 108 116 157]
[108 117 144 ... 102 58 122]
[124 148 106 ... 97 135 146]
...
[ 96 153 111 ... 104 129 154]
[129 140 100 ... 74 114 97]
[119 115 160 ... 172 84 148]]
...
[[ 92 96 64 ... 69 83 83]
[ 85 44 89 ... 115 94 76]
[ 93 103 91 ... 92 81 75]
...
[ 16 109 81 ... 84 95 20]
[100 27 89 ... 66 107 48]
[ 24 67 144 ... 104 115 123]]
[[ 69 70 74 ... 72 73 75]
[ 72 72 76 ... 73 75 76]
[ 74 75 72 ... 72 69 73]
...
[ 72 72 69 ... 72 76 72]
[ 70 72 73 ... 72 76 67]
[ 69 72 72 ... 72 71 71]]
[[ 65 137 26 ... 134 57 174]
[ 91 76 123 ... 39 63 124]
[ 81 203 134 ... 192 63 143]
...
[ 1 102 96 ... 33 63 169]
[ 82 32 108 ... 151 75 151]
[ 12 97 164 ... 101 125 60]]]
y_train:
[0 0 0 ... 0 0 0]
这些是我的输入形状：
x_train.shape = (5000, 128, 128)
y_train.shape = (5000,)
如您所见，y 只是标签，x 只是 3D 数据。因为它是一个二元分类器，所以我想构建一个具有 3 个密集层的简单神经网络。这是我的：
model = Sequential()
model.add(Dense(12, input_dim = 8, activation='relu'))
model.add(Dense(8, activation='relu'))
model.add(Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
model.fit(x_train, y_train, epochs=150, batch_size=8)
但是由于我的输入，出现了这个错误：
ValueError: Input 0 of layer sequential_4 is incompatible with the layer: expected axis -1 of input shape to have value 8 but received input with shape (8, 128, 128)
我该如何解决这个问题？我的神经网络对于这类问题是不是太简单了？
您的代码存在多个问题。我尝试添加单独的部分来解释它们。请仔细阅读所有这些并尝试我在下面显示的代码示例。
1。将 samples/batch 通道作为输入维度传递
您正在将批处理通道作为密集层的输入形状传递。那是不正确的。相反，您需要做的是传递模型应该期望的每个样本的形状，在本例中为 (128,128)。该模型自动在前面添加一个通道，以便批次流经计算图为(None, 128, 128)，如下图model.summary()
2。密集层的二维输入。
您的每个样本（在本例中共有 5000 个样本）都是形状为 128,128 的二维矩阵。如果不先将其压平，致密层不能直接消耗它。 （或者使用不同的层来更适合于 2D/3D 输入，稍后讨论）。
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense, Flatten
model = Sequential()
model.add(Flatten(input_shape=(128,128)))
model.add(Dense(12, activation='relu'))
model.add(Dense(8, activation='relu'))
model.add(Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
model.summary()
Model: "sequential_6"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
flatten_3 (Flatten) (None, 16384) 0
_________________________________________________________________
dense_5 (Dense) (None, 12) 196620
_________________________________________________________________
dense_6 (Dense) (None, 8) 104
_________________________________________________________________
dense_7 (Dense) (None, 1) 9
=================================================================
Total params: 196,733
Trainable params: 196,733
Non-trainable params: 0
_________________________________________________________________
3。针对您的问题使用不同的体系结构。
"Is my NN too simplistic for this type of problem?"
这不是关于架构的复杂性，而是更多关于可以处理某种类型数据的层的类型。在这种情况下，您的图像具有单通道 (128,128)，这是一个 2D 输入。通常，彩色图像有 R、G、B 通道，最终以 (128,128,3) 形输入。
一般做法是为此使用 CNN 层。
下面显示了一个例子 -
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense, Flatten, Conv2D, MaxPooling2D, Reshape
model = Sequential()
model.add(Reshape((128,128,1), input_shape=(128,128)))
model.add(Conv2D(5, 5, activation='relu'))
model.add(MaxPooling2D((2,2)))
model.add(Conv2D(10, 5, activation='relu'))
model.add(MaxPooling2D((2,2)))
model.add(Conv2D(20, 5, activation='relu'))
model.add(MaxPooling2D((2,2)))
model.add(Conv2D(30, 5, activation='relu'))
model.add(MaxPooling2D((2,2)))
model.add(Flatten())
model.add(Dense(12, activation='relu'))
model.add(Dense(8, activation='relu'))
model.add(Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
model.summary()
Model: "sequential_13"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
reshape_5 (Reshape) (None, 128, 128, 1) 0
_________________________________________________________________
conv2d_16 (Conv2D) (None, 124, 124, 5) 130
_________________________________________________________________
max_pooling2d_16 (MaxPooling (None, 62, 62, 5) 0
_________________________________________________________________
conv2d_17 (Conv2D) (None, 58, 58, 10) 1260
_________________________________________________________________
max_pooling2d_17 (MaxPooling (None, 29, 29, 10) 0
_________________________________________________________________
conv2d_18 (Conv2D) (None, 25, 25, 20) 5020
_________________________________________________________________
max_pooling2d_18 (MaxPooling (None, 12, 12, 20) 0
_________________________________________________________________
conv2d_19 (Conv2D) (None, 8, 8, 30) 15030
_________________________________________________________________
max_pooling2d_19 (MaxPooling (None, 4, 4, 30) 0
_________________________________________________________________
flatten_9 (Flatten) (None, 480) 0
_________________________________________________________________
dense_23 (Dense) (None, 12) 5772
_________________________________________________________________
dense_24 (Dense) (None, 8) 104
_________________________________________________________________
dense_25 (Dense) (None, 1) 9
=================================================================
Total params: 27,325
Trainable params: 27,325
Non-trainable params: 0
_________________________________________________________________
要了解 Conv2D 层和 MaxPooling 层的作用，请查看我的 well-accepted answer on Data Science stack exchange, or check out this blog。但是下图是一种在脑海中形象化它正在做的事情的一般方法。