ValueError：形状（无，无）和（无，无，无，43）不兼容

Question

I know that there were similar threads on this forum already, but though I checked them out I can't seem to find a solution.

I'm trying to use a VGG model for multi-classification of images. I'm following a tutorial from a book. I use the last layer from the VGG model as my input for the last sequentail layers.

My images are stored in a folder 'train', inside this folder there are 43 subfolders containing the images belonging to 43 classes. Each subfolder's name is a number from 0 to 42.

I use flow_from_directory() function to load the images, and then finally fit_generator() .

The last layer in my model is a dense layer model.add(Dense(43, activation='softmax'))

This is my code:

input_shape1 = (224, 224, 3)
vgg = vgg16.VGG16(include_top=False, weights='imagenet', input_shape=input_shape1)
output = vgg.layers[-2].output
output = keras.layers.Flatten()(output)
vgg_model = Model(vgg.input, output)
vgg_model.trainable = False
for layer in vgg_model.layers:
  layer.trainable = False
  vgg_model.summary()

input_shape = vgg_model.output_shape[1]
model = Sequential()
model.add(InputLayer(input_shape=(input_shape,)))
model.add(Dense(512, activation='relu', input_dim=input_shape))
model.add(Dropout(0.3)) 
model.add(Dense(512, activation='relu'))
model.add(Dropout(0.3))
model.add(Dense(43, activation='softmax'))

model.compile(optimizer=optimizers.RMSprop(lr=1e-4), loss='categorical_crossentropy', metrics=['accuracy'])

When I try to run the model with this line, I get the error:

epochs=100
history = model.fit_generator(train_ds, steps_per_epoch=1226, epochs=epochs, verbose=1)

WARNING:tensorflow:Model was constructed with shape (None, 100352) for input KerasTensor(type_spec=TensorSpec(shape=(None, 100352), dtype=tf.float32, name='input_4'), name='input_4', description="created by layer 'input_4'"), but it was called on an input with incompatible shape (None, None, None, None).

ValueError: Shapes (None, None) and (None, None, None, 43) are incompatible

I really have no idea where it is coming from. I tried experimenting with input shapes but with no luck.

EDIT

This is my image generator

train_datagen = ImageDataGenerator(
        rescale=1./255,
        validation_split=0.3,
        shear_range=0.2,
        zoom_range=0.2,
        horizontal_flip=True)
test_datagen = ImageDataGenerator(rescale=1./255)

train_ds = train_datagen.flow_from_directory(
  train_dir,
  seed=123,
  target_size=(224, 224),
  batch_size=32,
  class_mode='categorical')

Found 39209 images belonging to 43 classes. But I also specified validation split for this dataset.

EDIT 2

vgg_model.output_shape[0]
100352

The output shape of the model after adding the last layers is 43 though.

Also, I tried changing the loss function to sparse_categorical_crossentropy and got this error:

InvalidArgumentError:  Matrix size-incompatible: In[0]: [720000,3], In[1]: [8192,512]
     [[node sequential_8/dense_24/Tensordot/MatMul (defined at <ipython-input-37-e259535ec653>:2) ]] [Op:__inference_train_function_4462]

Somethig is wrong either with my model or with the way I'm loading the images, but I simply have no clue.

I'd really appreciate your help. Thanks!

Answer 1

I actually changed my Image Generator to flow_from_dataframe and it worked.

train_df = train_datagen.flow_from_dataframe(
  traindf,
  y_col='ClassId',
  x_col='Path',
  directory=None,
  subset='training',
  seed=123,
  target_size=(150, 150),
  batch_size=32,
  class_mode='categorical')

Answer 2

I was using MobileNetV3 and got same problem just like you. I tried this:

model = tf.keras.applications.MobileNetV3Large

You can find more info .