未提供梯度/'NoneType' 对象在尝试拟合多输出模型时不可调用
No gradients are provided / 'NoneType' object is not callable when trying to fit a multi-output model
我是机器学习的新手，在这个问题上遇到了很多困难。我正在使用带有 tensorflow 版本 2.3.1 的 Kaggle 笔记本。我需要用面部图像训练模型并预测多个属性、皱纹、雀斑、头发颜色等、头发厚度和眼镜，因此是多输出模型。当我尝试 model.fit 时，我首先收到“未提供渐变”的错误消息。在 运行 上，没有任何更改的相同代码给我错误“NoneType 对象不可调用”。我被困在这里一个多星期了，到目前为止，互联网上还没有解决方案能够解决这个问题，所以我在这里包含了尽可能多的细节。关于该问题的一些辅助信息，皱纹和雀斑的值为 0 或 1，而其他输出的值范围为 0 到 3、0 到 5 或 0 到 9。这是代码。
正在设置 CNN：
IMAGE_SIZE = 100
def base_hidden_layers(input_shape):
model = Conv2D(64, kernel_size = (3, 3), padding = 'same', activation= 'relu')(input_shape)
model = BatchNormalization(axis=-1)(model)
model = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(model)
model = Dropout(0.25)(model)
model = Conv2D(128, kernel_size = (3, 3), padding = 'same', activation= 'relu')(model)
model = BatchNormalization(axis=-1)(model)
model = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(model)
model = Dropout(0.25)(model)
model = Conv2D(128, kernel_size = (3, 3), padding = 'same', activation= 'relu')(model)
model = BatchNormalization(axis=-1)(model)
model = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(model)
model = Dropout(0.25)(model)
return model
def wrinkle_layers(input_shape):
model = base_hidden_layers(input_shape)
model = Flatten()(model)
model = Dense(1024, activation = 'relu')(model)
model = BatchNormalization()(model)
model = Dropout(0.5)(model)
model = Dense(2, activation = 'sigmoid', name = 'wrinkles')(model)
return model
def freckles_layers(input_shape):
model = base_hidden_layers(input_shape)
model = Flatten()(model)
model = Dense(1024, activation = 'relu')(model)
model = BatchNormalization()(model)
model = Dropout(0.5)(model)
model = Dense(2, activation = 'sigmoid', name = 'freckles')(model)
return model
def glasses_layers(input_shape):
model = base_hidden_layers(input_shape)
model = Flatten()(model)
model = Dense(1024, activation = 'relu')(model)
model = BatchNormalization()(model)
model = Dropout(0.5)(model)
model = Dense(3, activation = 'softmax', name = 'glasses')(model)
return model
def hair_color_layers(input_shape):
model = base_hidden_layers(input_shape)
model = Flatten()(model)
model = Dense(1024, activation = 'relu')(model)
model = BatchNormalization()(model)
model = Dropout(0.5)(model)
model = Dense(9, activation = 'softmax', name = 'hair_color')(model)
return model
def hair_top_layers(input_shape):
model = base_hidden_layers(input_shape)
model = Flatten()(model)
model = Dense(1024, activation = 'relu')(model)
model = BatchNormalization()(model)
model = Dropout(0.5)(model)
model = Dense(4, activation = 'softmax', name = 'hair_top')(model)
return model
def neural_net_model(image_size):
shape = (image_size, image_size, 3)
shape = Input(shape=shape)
wrinkle_branch = wrinkle_layers(shape)
freckles_branch = freckles_layers(shape)
glasses_branch = glasses_layers(shape)
hair_col_branch = hair_color_layers(shape)
hair_top_branch = hair_top_layers(shape)
model = Model(inputs=shape, outputs= [wrinkle_branch, freckles_branch, glasses_branch, hair_col_branch, hair_top_branch], name='face_net')
return model
model = neural_net_model(IMAGE_SIZE)
编译模型，设置损失函数：
model.compile(optimizer=Adam(lr=1e-4, decay= 1e-4 / 100),
loss = {'wrinkles': 'binary_crossentropy',
'freckles': 'binary_crossentropy',
'glasses': 'categorical_crossentropy',
'hair_color': 'categorical_crossentropy',
'hair_top': 'categorical_crossentropy'}, metrics=['accuracy'])
model.summary()
model.summary() 的输出：
Model: "face_net"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 100, 100, 3) 0
__________________________________________________________________________________________________
conv2d (Conv2D) (None, 100, 100, 64) 1792 input_1[0][0]
__________________________________________________________________________________________________
conv2d_3 (Conv2D) (None, 100, 100, 64) 1792 input_1[0][0]
__________________________________________________________________________________________________
conv2d_6 (Conv2D) (None, 100, 100, 64) 1792 input_1[0][0]
__________________________________________________________________________________________________
conv2d_9 (Conv2D) (None, 100, 100, 64) 1792 input_1[0][0]
__________________________________________________________________________________________________
conv2d_12 (Conv2D) (None, 100, 100, 64) 1792 input_1[0][0]
__________________________________________________________________________________________________
batch_normalization (BatchNorma (None, 100, 100, 64) 256 conv2d[0][0]
__________________________________________________________________________________________________
batch_normalization_4 (BatchNor (None, 100, 100, 64) 256 conv2d_3[0][0]
__________________________________________________________________________________________________
batch_normalization_8 (BatchNor (None, 100, 100, 64) 256 conv2d_6[0][0]
__________________________________________________________________________________________________
batch_normalization_12 (BatchNo (None, 100, 100, 64) 256 conv2d_9[0][0]
__________________________________________________________________________________________________
batch_normalization_16 (BatchNo (None, 100, 100, 64) 256 conv2d_12[0][0]
__________________________________________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 50, 50, 64) 0 batch_normalization[0][0]
__________________________________________________________________________________________________
max_pooling2d_3 (MaxPooling2D) (None, 50, 50, 64) 0 batch_normalization_4[0][0]
__________________________________________________________________________________________________
max_pooling2d_6 (MaxPooling2D) (None, 50, 50, 64) 0 batch_normalization_8[0][0]
__________________________________________________________________________________________________
max_pooling2d_9 (MaxPooling2D) (None, 50, 50, 64) 0 batch_normalization_12[0][0]
__________________________________________________________________________________________________
max_pooling2d_12 (MaxPooling2D) (None, 50, 50, 64) 0 batch_normalization_16[0][0]
__________________________________________________________________________________________________
dropout (Dropout) (None, 50, 50, 64) 0 max_pooling2d[0][0]
__________________________________________________________________________________________________
dropout_4 (Dropout) (None, 50, 50, 64) 0 max_pooling2d_3[0][0]
__________________________________________________________________________________________________
dropout_8 (Dropout) (None, 50, 50, 64) 0 max_pooling2d_6[0][0]
__________________________________________________________________________________________________
dropout_12 (Dropout) (None, 50, 50, 64) 0 max_pooling2d_9[0][0]
__________________________________________________________________________________________________
dropout_16 (Dropout) (None, 50, 50, 64) 0 max_pooling2d_12[0][0]
__________________________________________________________________________________________________
conv2d_1 (Conv2D) (None, 50, 50, 128) 73856 dropout[0][0]
__________________________________________________________________________________________________
conv2d_4 (Conv2D) (None, 50, 50, 128) 73856 dropout_4[0][0]
__________________________________________________________________________________________________
conv2d_7 (Conv2D) (None, 50, 50, 128) 73856 dropout_8[0][0]
__________________________________________________________________________________________________
conv2d_10 (Conv2D) (None, 50, 50, 128) 73856 dropout_12[0][0]
__________________________________________________________________________________________________
conv2d_13 (Conv2D) (None, 50, 50, 128) 73856 dropout_16[0][0]
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 50, 50, 128) 512 conv2d_1[0][0]
__________________________________________________________________________________________________
batch_normalization_5 (BatchNor (None, 50, 50, 128) 512 conv2d_4[0][0]
__________________________________________________________________________________________________
batch_normalization_9 (BatchNor (None, 50, 50, 128) 512 conv2d_7[0][0]
__________________________________________________________________________________________________
batch_normalization_13 (BatchNo (None, 50, 50, 128) 512 conv2d_10[0][0]
__________________________________________________________________________________________________
batch_normalization_17 (BatchNo (None, 50, 50, 128) 512 conv2d_13[0][0]
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D) (None, 25, 25, 128) 0 batch_normalization_1[0][0]
__________________________________________________________________________________________________
max_pooling2d_4 (MaxPooling2D) (None, 25, 25, 128) 0 batch_normalization_5[0][0]
__________________________________________________________________________________________________
max_pooling2d_7 (MaxPooling2D) (None, 25, 25, 128) 0 batch_normalization_9[0][0]
__________________________________________________________________________________________________
max_pooling2d_10 (MaxPooling2D) (None, 25, 25, 128) 0 batch_normalization_13[0][0]
__________________________________________________________________________________________________
max_pooling2d_13 (MaxPooling2D) (None, 25, 25, 128) 0 batch_normalization_17[0][0]
__________________________________________________________________________________________________
dropout_1 (Dropout) (None, 25, 25, 128) 0 max_pooling2d_1[0][0]
__________________________________________________________________________________________________
dropout_5 (Dropout) (None, 25, 25, 128) 0 max_pooling2d_4[0][0]
__________________________________________________________________________________________________
dropout_9 (Dropout) (None, 25, 25, 128) 0 max_pooling2d_7[0][0]
__________________________________________________________________________________________________
dropout_13 (Dropout) (None, 25, 25, 128) 0 max_pooling2d_10[0][0]
__________________________________________________________________________________________________
dropout_17 (Dropout) (None, 25, 25, 128) 0 max_pooling2d_13[0][0]
__________________________________________________________________________________________________
conv2d_2 (Conv2D) (None, 25, 25, 128) 147584 dropout_1[0][0]
__________________________________________________________________________________________________
conv2d_5 (Conv2D) (None, 25, 25, 128) 147584 dropout_5[0][0]
__________________________________________________________________________________________________
conv2d_8 (Conv2D) (None, 25, 25, 128) 147584 dropout_9[0][0]
__________________________________________________________________________________________________
conv2d_11 (Conv2D) (None, 25, 25, 128) 147584 dropout_13[0][0]
__________________________________________________________________________________________________
conv2d_14 (Conv2D) (None, 25, 25, 128) 147584 dropout_17[0][0]
__________________________________________________________________________________________________
batch_normalization_2 (BatchNor (None, 25, 25, 128) 512 conv2d_2[0][0]
__________________________________________________________________________________________________
batch_normalization_6 (BatchNor (None, 25, 25, 128) 512 conv2d_5[0][0]
__________________________________________________________________________________________________
batch_normalization_10 (BatchNo (None, 25, 25, 128) 512 conv2d_8[0][0]
__________________________________________________________________________________________________
batch_normalization_14 (BatchNo (None, 25, 25, 128) 512 conv2d_11[0][0]
__________________________________________________________________________________________________
batch_normalization_18 (BatchNo (None, 25, 25, 128) 512 conv2d_14[0][0]
__________________________________________________________________________________________________
max_pooling2d_2 (MaxPooling2D) (None, 12, 12, 128) 0 batch_normalization_2[0][0]
__________________________________________________________________________________________________
max_pooling2d_5 (MaxPooling2D) (None, 12, 12, 128) 0 batch_normalization_6[0][0]
__________________________________________________________________________________________________
max_pooling2d_8 (MaxPooling2D) (None, 12, 12, 128) 0 batch_normalization_10[0][0]
__________________________________________________________________________________________________
max_pooling2d_11 (MaxPooling2D) (None, 12, 12, 128) 0 batch_normalization_14[0][0]
__________________________________________________________________________________________________
max_pooling2d_14 (MaxPooling2D) (None, 12, 12, 128) 0 batch_normalization_18[0][0]
__________________________________________________________________________________________________
dropout_2 (Dropout) (None, 12, 12, 128) 0 max_pooling2d_2[0][0]
__________________________________________________________________________________________________
dropout_6 (Dropout) (None, 12, 12, 128) 0 max_pooling2d_5[0][0]
__________________________________________________________________________________________________
dropout_10 (Dropout) (None, 12, 12, 128) 0 max_pooling2d_8[0][0]
__________________________________________________________________________________________________
dropout_14 (Dropout) (None, 12, 12, 128) 0 max_pooling2d_11[0][0]
__________________________________________________________________________________________________
dropout_18 (Dropout) (None, 12, 12, 128) 0 max_pooling2d_14[0][0]
__________________________________________________________________________________________________
flatten (Flatten) (None, 18432) 0 dropout_2[0][0]
__________________________________________________________________________________________________
flatten_1 (Flatten) (None, 18432) 0 dropout_6[0][0]
__________________________________________________________________________________________________
flatten_2 (Flatten) (None, 18432) 0 dropout_10[0][0]
__________________________________________________________________________________________________
flatten_3 (Flatten) (None, 18432) 0 dropout_14[0][0]
__________________________________________________________________________________________________
flatten_4 (Flatten) (None, 18432) 0 dropout_18[0][0]
__________________________________________________________________________________________________
dense (Dense) (None, 1024) 18875392 flatten[0][0]
__________________________________________________________________________________________________
dense_1 (Dense) (None, 1024) 18875392 flatten_1[0][0]
__________________________________________________________________________________________________
dense_2 (Dense) (None, 1024) 18875392 flatten_2[0][0]
__________________________________________________________________________________________________
dense_3 (Dense) (None, 1024) 18875392 flatten_3[0][0]
__________________________________________________________________________________________________
dense_4 (Dense) (None, 1024) 18875392 flatten_4[0][0]
__________________________________________________________________________________________________
batch_normalization_3 (BatchNor (None, 1024) 4096 dense[0][0]
__________________________________________________________________________________________________
batch_normalization_7 (BatchNor (None, 1024) 4096 dense_1[0][0]
__________________________________________________________________________________________________
batch_normalization_11 (BatchNo (None, 1024) 4096 dense_2[0][0]
__________________________________________________________________________________________________
batch_normalization_15 (BatchNo (None, 1024) 4096 dense_3[0][0]
__________________________________________________________________________________________________
batch_normalization_19 (BatchNo (None, 1024) 4096 dense_4[0][0]
__________________________________________________________________________________________________
dropout_3 (Dropout) (None, 1024) 0 batch_normalization_3[0][0]
__________________________________________________________________________________________________
dropout_7 (Dropout) (None, 1024) 0 batch_normalization_7[0][0]
__________________________________________________________________________________________________
dropout_11 (Dropout) (None, 1024) 0 batch_normalization_11[0][0]
__________________________________________________________________________________________________
dropout_15 (Dropout) (None, 1024) 0 batch_normalization_15[0][0]
__________________________________________________________________________________________________
dropout_19 (Dropout) (None, 1024) 0 batch_normalization_19[0][0]
__________________________________________________________________________________________________
wrinkles (Dense) (None, 2) 2050 dropout_3[0][0]
__________________________________________________________________________________________________
freckles (Dense) (None, 2) 2050 dropout_7[0][0]
__________________________________________________________________________________________________
glasses (Dense) (None, 3) 3075 dropout_11[0][0]
__________________________________________________________________________________________________
hair_color (Dense) (None, 9) 9225 dropout_15[0][0]
__________________________________________________________________________________________________
hair_top (Dense) (None, 4) 4100 dropout_19[0][0]
==================================================================================================
Total params: 95,540,500
Trainable params: 95,527,060
Non-trainable params: 13,440
训练集和验证集的形状：
print(valid_images.shape)
print(valid_labels.shape)
print(train_images.shape)
print(train_labels.shape)
print(type(valid_labels))
输出：
(319, 100, 100, 3)
(319, 5)
(1272, 100, 100, 3)
(1272, 5)
<class 'numpy.ndarray'>
以上shape for input表示：（图片数量，图片长度，图片宽度，图片深度（RGB））
输出方式：（图像数，输出列数）
打印训练输入和输出：
train_images
输出：
array([[[[0.41176471, 0.49019608, 0.44705882],
[0.40784314, 0.49411765, 0.44705882],
[0.41176471, 0.50588235, 0.45490196],
...,
[0.77647059, 0.81568627, 0.81960784],
[0.7372549 , 0.76862745, 0.78823529],
[0.57254902, 0.6 , 0.63921569]]]])
train_labels
输出：
array([[0, 0, 0, 2, 2],
[0, 1, 0, 0, 2],
[0, 0, 0, 0, 2],
...,
[0, 0, 0, 1, 2],
[0, 0, 0, 2, 2],
[1, 0, 0, 2, 2]])
拟合模型：
history = model.fit((train_images, train_labels),
epochs = 100,
validation_data=(valid_images,
valid_labels))
plt.plot(history.history['accuracy'], label='accuracy')
plt.plot(history.history['val_accuracy'], label = 'val_accuracy')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.show()
test_loss, test_acc = model.evaluate(test_images, test_labels)
运行model.fit 上的错误堆栈跟踪，这是笔记本启动后的第一次：
Epoch 1/100
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-34-0dd91e39ee8e> in <module>
5 epochs = 100,
6 validation_data=(valid_images,
----> 7 valid_labels))
8
9 plt.plot(history.history['accuracy'], label='accuracy')
/opt/conda/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py in _method_wrapper(self, *args, **kwargs)
106 def _method_wrapper(self, *args, **kwargs):
107 if not self._in_multi_worker_mode(): # pylint: disable=protected-access
--> 108 return method(self, *args, **kwargs)
109
110 # Running inside `run_distribute_coordinator` already.
/opt/conda/
...
ValueError: in user code:
/opt/conda/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py:806 train_function *
return ...
/opt/conda/lib/python3.7/site-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py:1271 _filter_grads
([v.name for _, v in grads_and_vars],))
ValueError: No gradients provided for any variable: ['conv2d/kernel:0', 'conv2d/bias:0', 'conv2d_3/kernel:0', ... 'hair_top/kernel:0', 'hair_top/bias:0'].
运行 代码再次出错：
Epoch 1/100
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-38-356b6cd04e18> in <module>
2 epochs = 100,
3 validation_data=(valid_images,
----> 4 valid_labels))
5
6 plt.plot(history.history['accuracy'], label='accuracy')
/opt/conda/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py in _method_wrapper(self, *args, **kwargs)
106 def _method_wrapper(self, *args, **kwargs):
107 if not self._in_multi_worker_mode(): # pylint: disable=protected-access
--> 108 return method(self, *args, **kwargs)
109
110 # Running inside `run_distribute_coordinator` already.
...
TypeError: 'NoneType' object is not callable
只添加了堆栈跟踪的重要部分，因为添加整个堆栈跟踪会使这个 post 超过 3000 个字符。
显然我遵循的教程不准确。由于我的 NN 有 5 个分支，因为它需要进行 5 次预测，我的 model.fit 函数需要更改，每个 Nn 分支需要映射到相应的标签。
错误代码：
history = model.fit((train_images, train_labels),
epochs = 100,
validation_data=(valid_images,
valid_labels))
正确代码：
history = model.fit(x=train_images, y={'wrinkles': train_wrinkles, 'freckles': train_freckles, 'glasses': train_glasses,
'hair_color': train_hair_col, 'hair_top': train_hair_top},
epochs = 20,
validation_data=(valid_images,
{'wrinkles': valid_wrinkles, 'freckles': valid_freckles, 'glasses': valid_glasses,
'hair_color': valid_hair_col, 'hair_top': valid_hair_top}))
在错误代码中，train_labels和valid_labels包含一个包含所有标签的数组对于 5 个预测。在正确的代码中，每个预测的标签是单独传递的。
此外，这些标签中的每一个都需要在训练前进行单热编码。这就是我对标签进行编码的方式。
from tensorflow.keras.utils import to_categorical
wrinkles = to_categorical(wrinkles, num_classes=2)
请注意，您需要根据 类 的数量在 num_classes 参数中输入不同的数字。
我是机器学习的新手，在这个问题上遇到了很多困难。我正在使用带有 tensorflow 版本 2.3.1 的 Kaggle 笔记本。我需要用面部图像训练模型并预测多个属性、皱纹、雀斑、头发颜色等、头发厚度和眼镜，因此是多输出模型。当我尝试 model.fit 时，我首先收到“未提供渐变”的错误消息。在 运行 上，没有任何更改的相同代码给我错误“NoneType 对象不可调用”。我被困在这里一个多星期了，到目前为止，互联网上还没有解决方案能够解决这个问题，所以我在这里包含了尽可能多的细节。关于该问题的一些辅助信息，皱纹和雀斑的值为 0 或 1，而其他输出的值范围为 0 到 3、0 到 5 或 0 到 9。这是代码。
正在设置 CNN：
IMAGE_SIZE = 100
def base_hidden_layers(input_shape):
model = Conv2D(64, kernel_size = (3, 3), padding = 'same', activation= 'relu')(input_shape)
model = BatchNormalization(axis=-1)(model)
model = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(model)
model = Dropout(0.25)(model)
model = Conv2D(128, kernel_size = (3, 3), padding = 'same', activation= 'relu')(model)
model = BatchNormalization(axis=-1)(model)
model = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(model)
model = Dropout(0.25)(model)
model = Conv2D(128, kernel_size = (3, 3), padding = 'same', activation= 'relu')(model)
model = BatchNormalization(axis=-1)(model)
model = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(model)
model = Dropout(0.25)(model)
return model
def wrinkle_layers(input_shape):
model = base_hidden_layers(input_shape)
model = Flatten()(model)
model = Dense(1024, activation = 'relu')(model)
model = BatchNormalization()(model)
model = Dropout(0.5)(model)
model = Dense(2, activation = 'sigmoid', name = 'wrinkles')(model)
return model
def freckles_layers(input_shape):
model = base_hidden_layers(input_shape)
model = Flatten()(model)
model = Dense(1024, activation = 'relu')(model)
model = BatchNormalization()(model)
model = Dropout(0.5)(model)
model = Dense(2, activation = 'sigmoid', name = 'freckles')(model)
return model
def glasses_layers(input_shape):
model = base_hidden_layers(input_shape)
model = Flatten()(model)
model = Dense(1024, activation = 'relu')(model)
model = BatchNormalization()(model)
model = Dropout(0.5)(model)
model = Dense(3, activation = 'softmax', name = 'glasses')(model)
return model
def hair_color_layers(input_shape):
model = base_hidden_layers(input_shape)
model = Flatten()(model)
model = Dense(1024, activation = 'relu')(model)
model = BatchNormalization()(model)
model = Dropout(0.5)(model)
model = Dense(9, activation = 'softmax', name = 'hair_color')(model)
return model
def hair_top_layers(input_shape):
model = base_hidden_layers(input_shape)
model = Flatten()(model)
model = Dense(1024, activation = 'relu')(model)
model = BatchNormalization()(model)
model = Dropout(0.5)(model)
model = Dense(4, activation = 'softmax', name = 'hair_top')(model)
return model
def neural_net_model(image_size):
shape = (image_size, image_size, 3)
shape = Input(shape=shape)
wrinkle_branch = wrinkle_layers(shape)
freckles_branch = freckles_layers(shape)
glasses_branch = glasses_layers(shape)
hair_col_branch = hair_color_layers(shape)
hair_top_branch = hair_top_layers(shape)
model = Model(inputs=shape, outputs= [wrinkle_branch, freckles_branch, glasses_branch, hair_col_branch, hair_top_branch], name='face_net')
return model
model = neural_net_model(IMAGE_SIZE)
编译模型，设置损失函数：
model.compile(optimizer=Adam(lr=1e-4, decay= 1e-4 / 100),
loss = {'wrinkles': 'binary_crossentropy',
'freckles': 'binary_crossentropy',
'glasses': 'categorical_crossentropy',
'hair_color': 'categorical_crossentropy',
'hair_top': 'categorical_crossentropy'}, metrics=['accuracy'])
model.summary()
model.summary() 的输出：
Model: "face_net"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 100, 100, 3) 0
__________________________________________________________________________________________________
conv2d (Conv2D) (None, 100, 100, 64) 1792 input_1[0][0]
__________________________________________________________________________________________________
conv2d_3 (Conv2D) (None, 100, 100, 64) 1792 input_1[0][0]
__________________________________________________________________________________________________
conv2d_6 (Conv2D) (None, 100, 100, 64) 1792 input_1[0][0]
__________________________________________________________________________________________________
conv2d_9 (Conv2D) (None, 100, 100, 64) 1792 input_1[0][0]
__________________________________________________________________________________________________
conv2d_12 (Conv2D) (None, 100, 100, 64) 1792 input_1[0][0]
__________________________________________________________________________________________________
batch_normalization (BatchNorma (None, 100, 100, 64) 256 conv2d[0][0]
__________________________________________________________________________________________________
batch_normalization_4 (BatchNor (None, 100, 100, 64) 256 conv2d_3[0][0]
__________________________________________________________________________________________________
batch_normalization_8 (BatchNor (None, 100, 100, 64) 256 conv2d_6[0][0]
__________________________________________________________________________________________________
batch_normalization_12 (BatchNo (None, 100, 100, 64) 256 conv2d_9[0][0]
__________________________________________________________________________________________________
batch_normalization_16 (BatchNo (None, 100, 100, 64) 256 conv2d_12[0][0]
__________________________________________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 50, 50, 64) 0 batch_normalization[0][0]
__________________________________________________________________________________________________
max_pooling2d_3 (MaxPooling2D) (None, 50, 50, 64) 0 batch_normalization_4[0][0]
__________________________________________________________________________________________________
max_pooling2d_6 (MaxPooling2D) (None, 50, 50, 64) 0 batch_normalization_8[0][0]
__________________________________________________________________________________________________
max_pooling2d_9 (MaxPooling2D) (None, 50, 50, 64) 0 batch_normalization_12[0][0]
__________________________________________________________________________________________________
max_pooling2d_12 (MaxPooling2D) (None, 50, 50, 64) 0 batch_normalization_16[0][0]
__________________________________________________________________________________________________
dropout (Dropout) (None, 50, 50, 64) 0 max_pooling2d[0][0]
__________________________________________________________________________________________________
dropout_4 (Dropout) (None, 50, 50, 64) 0 max_pooling2d_3[0][0]
__________________________________________________________________________________________________
dropout_8 (Dropout) (None, 50, 50, 64) 0 max_pooling2d_6[0][0]
__________________________________________________________________________________________________
dropout_12 (Dropout) (None, 50, 50, 64) 0 max_pooling2d_9[0][0]
__________________________________________________________________________________________________
dropout_16 (Dropout) (None, 50, 50, 64) 0 max_pooling2d_12[0][0]
__________________________________________________________________________________________________
conv2d_1 (Conv2D) (None, 50, 50, 128) 73856 dropout[0][0]
__________________________________________________________________________________________________
conv2d_4 (Conv2D) (None, 50, 50, 128) 73856 dropout_4[0][0]
__________________________________________________________________________________________________
conv2d_7 (Conv2D) (None, 50, 50, 128) 73856 dropout_8[0][0]
__________________________________________________________________________________________________
conv2d_10 (Conv2D) (None, 50, 50, 128) 73856 dropout_12[0][0]
__________________________________________________________________________________________________
conv2d_13 (Conv2D) (None, 50, 50, 128) 73856 dropout_16[0][0]
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 50, 50, 128) 512 conv2d_1[0][0]
__________________________________________________________________________________________________
batch_normalization_5 (BatchNor (None, 50, 50, 128) 512 conv2d_4[0][0]
__________________________________________________________________________________________________
batch_normalization_9 (BatchNor (None, 50, 50, 128) 512 conv2d_7[0][0]
__________________________________________________________________________________________________
batch_normalization_13 (BatchNo (None, 50, 50, 128) 512 conv2d_10[0][0]
__________________________________________________________________________________________________
batch_normalization_17 (BatchNo (None, 50, 50, 128) 512 conv2d_13[0][0]
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D) (None, 25, 25, 128) 0 batch_normalization_1[0][0]
__________________________________________________________________________________________________
max_pooling2d_4 (MaxPooling2D) (None, 25, 25, 128) 0 batch_normalization_5[0][0]
__________________________________________________________________________________________________
max_pooling2d_7 (MaxPooling2D) (None, 25, 25, 128) 0 batch_normalization_9[0][0]
__________________________________________________________________________________________________
max_pooling2d_10 (MaxPooling2D) (None, 25, 25, 128) 0 batch_normalization_13[0][0]
__________________________________________________________________________________________________
max_pooling2d_13 (MaxPooling2D) (None, 25, 25, 128) 0 batch_normalization_17[0][0]
__________________________________________________________________________________________________
dropout_1 (Dropout) (None, 25, 25, 128) 0 max_pooling2d_1[0][0]
__________________________________________________________________________________________________
dropout_5 (Dropout) (None, 25, 25, 128) 0 max_pooling2d_4[0][0]
__________________________________________________________________________________________________
dropout_9 (Dropout) (None, 25, 25, 128) 0 max_pooling2d_7[0][0]
__________________________________________________________________________________________________
dropout_13 (Dropout) (None, 25, 25, 128) 0 max_pooling2d_10[0][0]
__________________________________________________________________________________________________
dropout_17 (Dropout) (None, 25, 25, 128) 0 max_pooling2d_13[0][0]
__________________________________________________________________________________________________
conv2d_2 (Conv2D) (None, 25, 25, 128) 147584 dropout_1[0][0]
__________________________________________________________________________________________________
conv2d_5 (Conv2D) (None, 25, 25, 128) 147584 dropout_5[0][0]
__________________________________________________________________________________________________
conv2d_8 (Conv2D) (None, 25, 25, 128) 147584 dropout_9[0][0]
__________________________________________________________________________________________________
conv2d_11 (Conv2D) (None, 25, 25, 128) 147584 dropout_13[0][0]
__________________________________________________________________________________________________
conv2d_14 (Conv2D) (None, 25, 25, 128) 147584 dropout_17[0][0]
__________________________________________________________________________________________________
batch_normalization_2 (BatchNor (None, 25, 25, 128) 512 conv2d_2[0][0]
__________________________________________________________________________________________________
batch_normalization_6 (BatchNor (None, 25, 25, 128) 512 conv2d_5[0][0]
__________________________________________________________________________________________________
batch_normalization_10 (BatchNo (None, 25, 25, 128) 512 conv2d_8[0][0]
__________________________________________________________________________________________________
batch_normalization_14 (BatchNo (None, 25, 25, 128) 512 conv2d_11[0][0]
__________________________________________________________________________________________________
batch_normalization_18 (BatchNo (None, 25, 25, 128) 512 conv2d_14[0][0]
__________________________________________________________________________________________________
max_pooling2d_2 (MaxPooling2D) (None, 12, 12, 128) 0 batch_normalization_2[0][0]
__________________________________________________________________________________________________
max_pooling2d_5 (MaxPooling2D) (None, 12, 12, 128) 0 batch_normalization_6[0][0]
__________________________________________________________________________________________________
max_pooling2d_8 (MaxPooling2D) (None, 12, 12, 128) 0 batch_normalization_10[0][0]
__________________________________________________________________________________________________
max_pooling2d_11 (MaxPooling2D) (None, 12, 12, 128) 0 batch_normalization_14[0][0]
__________________________________________________________________________________________________
max_pooling2d_14 (MaxPooling2D) (None, 12, 12, 128) 0 batch_normalization_18[0][0]
__________________________________________________________________________________________________
dropout_2 (Dropout) (None, 12, 12, 128) 0 max_pooling2d_2[0][0]
__________________________________________________________________________________________________
dropout_6 (Dropout) (None, 12, 12, 128) 0 max_pooling2d_5[0][0]
__________________________________________________________________________________________________
dropout_10 (Dropout) (None, 12, 12, 128) 0 max_pooling2d_8[0][0]
__________________________________________________________________________________________________
dropout_14 (Dropout) (None, 12, 12, 128) 0 max_pooling2d_11[0][0]
__________________________________________________________________________________________________
dropout_18 (Dropout) (None, 12, 12, 128) 0 max_pooling2d_14[0][0]
__________________________________________________________________________________________________
flatten (Flatten) (None, 18432) 0 dropout_2[0][0]
__________________________________________________________________________________________________
flatten_1 (Flatten) (None, 18432) 0 dropout_6[0][0]
__________________________________________________________________________________________________
flatten_2 (Flatten) (None, 18432) 0 dropout_10[0][0]
__________________________________________________________________________________________________
flatten_3 (Flatten) (None, 18432) 0 dropout_14[0][0]
__________________________________________________________________________________________________
flatten_4 (Flatten) (None, 18432) 0 dropout_18[0][0]
__________________________________________________________________________________________________
dense (Dense) (None, 1024) 18875392 flatten[0][0]
__________________________________________________________________________________________________
dense_1 (Dense) (None, 1024) 18875392 flatten_1[0][0]
__________________________________________________________________________________________________
dense_2 (Dense) (None, 1024) 18875392 flatten_2[0][0]
__________________________________________________________________________________________________
dense_3 (Dense) (None, 1024) 18875392 flatten_3[0][0]
__________________________________________________________________________________________________
dense_4 (Dense) (None, 1024) 18875392 flatten_4[0][0]
__________________________________________________________________________________________________
batch_normalization_3 (BatchNor (None, 1024) 4096 dense[0][0]
__________________________________________________________________________________________________
batch_normalization_7 (BatchNor (None, 1024) 4096 dense_1[0][0]
__________________________________________________________________________________________________
batch_normalization_11 (BatchNo (None, 1024) 4096 dense_2[0][0]
__________________________________________________________________________________________________
batch_normalization_15 (BatchNo (None, 1024) 4096 dense_3[0][0]
__________________________________________________________________________________________________
batch_normalization_19 (BatchNo (None, 1024) 4096 dense_4[0][0]
__________________________________________________________________________________________________
dropout_3 (Dropout) (None, 1024) 0 batch_normalization_3[0][0]
__________________________________________________________________________________________________
dropout_7 (Dropout) (None, 1024) 0 batch_normalization_7[0][0]
__________________________________________________________________________________________________
dropout_11 (Dropout) (None, 1024) 0 batch_normalization_11[0][0]
__________________________________________________________________________________________________
dropout_15 (Dropout) (None, 1024) 0 batch_normalization_15[0][0]
__________________________________________________________________________________________________
dropout_19 (Dropout) (None, 1024) 0 batch_normalization_19[0][0]
__________________________________________________________________________________________________
wrinkles (Dense) (None, 2) 2050 dropout_3[0][0]
__________________________________________________________________________________________________
freckles (Dense) (None, 2) 2050 dropout_7[0][0]
__________________________________________________________________________________________________
glasses (Dense) (None, 3) 3075 dropout_11[0][0]
__________________________________________________________________________________________________
hair_color (Dense) (None, 9) 9225 dropout_15[0][0]
__________________________________________________________________________________________________
hair_top (Dense) (None, 4) 4100 dropout_19[0][0]
==================================================================================================
Total params: 95,540,500
Trainable params: 95,527,060
Non-trainable params: 13,440
训练集和验证集的形状：
print(valid_images.shape)
print(valid_labels.shape)
print(train_images.shape)
print(train_labels.shape)
print(type(valid_labels))
输出：
(319, 100, 100, 3)
(319, 5)
(1272, 100, 100, 3)
(1272, 5)
<class 'numpy.ndarray'>
以上shape for input表示：（图片数量，图片长度，图片宽度，图片深度（RGB））
输出方式：（图像数，输出列数）
打印训练输入和输出：
train_images
输出：
array([[[[0.41176471, 0.49019608, 0.44705882],
[0.40784314, 0.49411765, 0.44705882],
[0.41176471, 0.50588235, 0.45490196],
...,
[0.77647059, 0.81568627, 0.81960784],
[0.7372549 , 0.76862745, 0.78823529],
[0.57254902, 0.6 , 0.63921569]]]])
train_labels
输出：
array([[0, 0, 0, 2, 2],
[0, 1, 0, 0, 2],
[0, 0, 0, 0, 2],
...,
[0, 0, 0, 1, 2],
[0, 0, 0, 2, 2],
[1, 0, 0, 2, 2]])
拟合模型：
history = model.fit((train_images, train_labels),
epochs = 100,
validation_data=(valid_images,
valid_labels))
plt.plot(history.history['accuracy'], label='accuracy')
plt.plot(history.history['val_accuracy'], label = 'val_accuracy')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.show()
test_loss, test_acc = model.evaluate(test_images, test_labels)
运行model.fit 上的错误堆栈跟踪，这是笔记本启动后的第一次：
Epoch 1/100
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-34-0dd91e39ee8e> in <module>
5 epochs = 100,
6 validation_data=(valid_images,
----> 7 valid_labels))
8
9 plt.plot(history.history['accuracy'], label='accuracy')
/opt/conda/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py in _method_wrapper(self, *args, **kwargs)
106 def _method_wrapper(self, *args, **kwargs):
107 if not self._in_multi_worker_mode(): # pylint: disable=protected-access
--> 108 return method(self, *args, **kwargs)
109
110 # Running inside `run_distribute_coordinator` already.
/opt/conda/
...
ValueError: in user code:
/opt/conda/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py:806 train_function *
return ...
/opt/conda/lib/python3.7/site-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py:1271 _filter_grads
([v.name for _, v in grads_and_vars],))
ValueError: No gradients provided for any variable: ['conv2d/kernel:0', 'conv2d/bias:0', 'conv2d_3/kernel:0', ... 'hair_top/kernel:0', 'hair_top/bias:0'].
运行 代码再次出错：
Epoch 1/100
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-38-356b6cd04e18> in <module>
2 epochs = 100,
3 validation_data=(valid_images,
----> 4 valid_labels))
5
6 plt.plot(history.history['accuracy'], label='accuracy')
/opt/conda/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py in _method_wrapper(self, *args, **kwargs)
106 def _method_wrapper(self, *args, **kwargs):
107 if not self._in_multi_worker_mode(): # pylint: disable=protected-access
--> 108 return method(self, *args, **kwargs)
109
110 # Running inside `run_distribute_coordinator` already.
...
TypeError: 'NoneType' object is not callable
只添加了堆栈跟踪的重要部分，因为添加整个堆栈跟踪会使这个 post 超过 3000 个字符。
显然我遵循的教程不准确。由于我的 NN 有 5 个分支，因为它需要进行 5 次预测，我的 model.fit 函数需要更改，每个 Nn 分支需要映射到相应的标签。
错误代码：
history = model.fit((train_images, train_labels),
epochs = 100,
validation_data=(valid_images,
valid_labels))
正确代码：
history = model.fit(x=train_images, y={'wrinkles': train_wrinkles, 'freckles': train_freckles, 'glasses': train_glasses,
'hair_color': train_hair_col, 'hair_top': train_hair_top},
epochs = 20,
validation_data=(valid_images,
{'wrinkles': valid_wrinkles, 'freckles': valid_freckles, 'glasses': valid_glasses,
'hair_color': valid_hair_col, 'hair_top': valid_hair_top}))
在错误代码中，train_labels和valid_labels包含一个包含所有标签的数组对于 5 个预测。在正确的代码中，每个预测的标签是单独传递的。
此外，这些标签中的每一个都需要在训练前进行单热编码。这就是我对标签进行编码的方式。
from tensorflow.keras.utils import to_categorical
wrinkles = to_categorical(wrinkles, num_classes=2)
请注意，您需要根据 类 的数量在 num_classes 参数中输入不同的数字。