输入形状的预期轴 -1 的值为 28，但收到的输入形状为 (None, 28, 28, 5)

Question

I'm new to keras. I'm trying to train a model which focuses on batchnormalization. My code is

batchnorm_model = Sequential()
batchnorm_model.add(Dense(50, input_shape=(X_train.shape[1],), activation='relu', kernel_initializer='normal')) 
batchnorm_model.add(BatchNormalization())
batchnorm_model.add(Dense(50, activation='relu', kernel_initializer='normal')) 
batchnorm_model.add(BatchNormalization())
batchnorm_model.add(Dense(2))
# Compile your model with sgd
batchnorm_model.compile(optimizer='sgd', loss='categorical_crossentropy', metrics=['accuracy'])
h2_callback = batchnorm_model.fit(X_train, train_labels, validation_data=(X_test, test_labels), epochs=10, verbose = 0)

And my X_train is

print(X_train.shape)
(7000, 28, 28, 5)

And my error is

ValueError: in user code:

File "/usr/local/lib/python3.8/dist-packages/keras/engine/training.py", line 1051, in train_function  *
    return step_function(self, iterator)
File "/usr/local/lib/python3.8/dist-packages/keras/engine/training.py", line 1040, in step_function  **
    outputs = model.distribute_strategy.run(run_step, args=(data,))
File "/usr/local/lib/python3.8/dist-packages/keras/engine/training.py", line 1030, in run_step  **
    outputs = model.train_step(data)
File "/usr/local/lib/python3.8/dist-packages/keras/engine/training.py", line 889, in train_step
    y_pred = self(x, training=True)
File "/usr/local/lib/python3.8/dist-packages/keras/utils/traceback_utils.py", line 67, in error_handler
    raise e.with_traceback(filtered_tb) from None
File "/usr/local/lib/python3.8/dist-packages/keras/engine/input_spec.py", line 248, in assert_input_compatibility
    raise ValueError(

ValueError: Exception encountered when calling layer "sequential_14" (type Sequential).

Input 0 of layer "dense_48" is incompatible with the layer: expected axis -1 of input shape to have value 28, but received input with shape (None, 28, 28, 5)

Call arguments received by layer "sequential_14" (type Sequential):
  • inputs=tf.Tensor(shape=(None, 28, 28, 5), dtype=float32)
  • training=True
  • mask=None

Do I need to reshape each image in X_train to (-1,28,28,1)? The process of dealing with X_train is below:

 width = 28
height = 28
dim = (width, height)
from google.colab.patches import cv2_imshow
from skimage.io import imread
from skimage.io import imshow
all_images = []
for id in new_id:
   PIC = '/content/new/' + id
   im = cv2.imread(PIC)
   resized = cv2.resize(im, dim, interpolation = cv2.INTER_AREA)/255
   indices = np.dstack(np.indices(resized.shape[:2]))
   data = np.concatenate((resized, indices), axis=-1)
   all_images.append(data)

...processing labels data

 X_train, X_test, y_train, y_test = train_test_split(all_images, all_labels, 
   test_size=0.3)
 X_train = np.array(X_train,dtype="float32")
 X_test = np.array(X_test,dtype="float32")

Answer 1

You should have a Flatten layer before your first Dense layer. And if you don't use one-hot encoding, but provide the labels as integers, you should use SparseCategoricalCrossentropy as loss, with from_logits=True because there is no activation in your last Dense layer.

batchnorm_model = Sequential()
batchnorm_model.add(Flatten(input_shape=(X_train.shape[1],)))
batchnorm_model.add(Dense(50, activation='relu', kernel_initializer='normal')) 
batchnorm_model.add(BatchNormalization())
batchnorm_model.add(Dense(50, activation='relu', kernel_initializer='normal')) 
batchnorm_model.add(BatchNormalization())
batchnorm_model.add(Dense(2))
# Compile your model with sgd
batchnorm_model.compile(optimizer='sgd', loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True), metrics=['accuracy'])
h2_callback = batchnorm_model.fit(X_train, train_labels, validation_data=(X_test, test_labels), epochs=10, verbose = 0)

Answer 2

try to match image channels with the data input layer. The image channels is important you may try to use 'grayscales' or 'rgb' in color_mode from the original image or simply feed it to the model ( 28, 28, 5 ) it is the same when you working with features extraction image that had more than one channels frequency responses.

Problem:

1. From you question the line batchnorm_model.add(Dense(50, input_shape=(X_train.shape[1],), activation='relu', kernel_initializer='normal')) indicated mismatched input
2. Error message: Input 0 of layer "dense_48" is incompatible with the layer: expected axis -1 of input shape to have value 28, but received input with shape (None, 28, 28, 5)

You should

1. Match the input shape to input_shape=(X_train.shape[1], X_train.shape[2], X_train.shape[3]) or
2. Convert them into 'rgb' or 'grayscale' format then use an image generator.
3. The number of channels does not often indicate data type when displayed but the information is the frequency response for the image sees the image in example.

Sample: Image to a dataset, custom_image_preprocess function to convert to the target format.

"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
: Class / Function
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
def custom_image_preprocess( image ):

    random_lotation_layer = tf.keras.layers.RandomRotation(
                            factor=(-0.2, 0.3),
                            fill_mode='nearest',
                            interpolation='nearest',
                            seed=None,
                            fill_value=0.0,
                        )
                    
    """""""""""""""""""""""""""""""""""""""""""""""""""""""""
    : Image conversion function / sample ( you can applied feature extraction example MFCC as in the example
    """""""""""""""""""""""""""""""""""""""""""""""""""""""""
    image = tf.experimental.numpy.dstack( [image, tf.zeros([IMG_WIDTH, IMG_HEIGHT, IMG_CHANNELS])] )
    image = image[:,:,0:IMG_CHANNELS]

    return  image

"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
: DataSet
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
train_image_generator = ImageDataGenerator(rescale=1. / 255, vertical_flip=True, horizontal_flip=True, preprocessing_function=custom_image_preprocess,) 
train_data_gen = train_image_generator.flow_from_directory(batch_size=BATCH_SIZE,
    directory=train_dir,
    shuffle=True,
    target_size=(IMG_WIDTH, IMG_HEIGHT),
    class_mode='binary',
    color_mode='rgb',
    seed=seed_1,)
    
test_image_generator = ImageDataGenerator(rescale=1. / 255, vertical_flip=True, horizontal_flip=True, preprocessing_function=custom_image_preprocess,)
test_data_gen = test_image_generator.flow_from_directory(batch_size=BATCH_SIZE,
    directory=test_dir,
    shuffle=False,
    target_size=(IMG_WIDTH, IMG_HEIGHT),
    class_mode='binary',
    color_mode='rgb',
    seed=seed_2,)

"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
: Model Initialize
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
base_model = tf.keras.applications.Xception( weights='imagenet', input_shape=(IMG_WIDTH, IMG_HEIGHT, IMG_CHANNELS), include_top=False)  
base_model.trainable = False
inputs = tf.keras.Input(shape=(IMG_WIDTH, IMG_HEIGHT, IMG_CHANNELS))

x = tf.keras.applications.xception.preprocess_input(inputs)
x = base_model(x, training=False)
x = tf.keras.layers.GlobalAveragePooling2D()(x)
x = tf.keras.layers.Dropout(0.2)(x)  
outputs = tf.keras.layers.Dense(1)(x)
model = tf.keras.Model(inputs, outputs)

Output: Complie with the environments.

Epoch 1/10
2022-12-09 01:07:34.157717: I tensorflow/stream_executor/cuda/cuda_dnn.cc:368] Loaded cuDNN version 8100
 44/321 [===>..........................] - ETA: 1:16 - loss: 0.4925 - binary_accuracy: 0.1321