Dense layer probably produces InvalidArgumentError: Incompatible shapes: [0,2] vs. [32,2]

Question

The setup

I am using Python 3.6, TF 2.4.0 on an Azure DSVM STANDARD_NC6 (6 Cores, 56 GB RAM, 380 GB Disk) using 1 GPU

The parameters/model

I have training data: print(xtrain.shape) with shape (4599, 124, 124, 3) and ytrain | yval as categorical.

I use a classic generator

datagen = ImageDataGenerator(
    zoom_range=0.1,
    rotation_range=25,
    width_shift_range=0.1,
    height_shift_range=0.1,
    shear_range=0.15,
    horizontal_flip=True,
    fill_mode="nearest"
    )

datagen.fit(xtrain)

And my model is the base mobilenetv2 with own head:

baseModel = MobileNetV2(weights="imagenet", 
                        include_top=False, 
                        input_tensor=Input(shape=(224, 224,3)), 
                        #input_shape=(224, 224, 3),
                        )

headModel = baseModel.output
headModel = AveragePooling2D(pool_size=(7, 7))(headModel)
headModel = Flatten(name="flatten")(headModel)
headModel = Dense(64, activation="relu")(headModel)
headModel = Dropout(0.5)(headModel)
headModel = Dense(2, activation="softmax")(headModel)

model = Model(inputs=baseModel.input, outputs=headModel)

for layer in baseModel.layers:
    layer.trainable = False

model.compile(loss="mse", optimizer='adam', metrics=["accuracy"])

When I now fit the model

Batch_Size=1

h = model.fit(
datagen.flow(xtrain, ytrain, batch_size=Batch_Size),
steps_per_epoch=len(xtrain) // Batch_Size,
validation_data=(xval, yval),
validation_steps=len(xval) // Batch_Size,
epochs=EPOCHS,
callbacks=[model_checkpoint_callback, Board])

The Error

I get errors (all same but changing with batch size and loss function)

When I use batch_size=1 with loss=mse, categorical_crossentropy, or others, the model trains but throws following error at the epoch end

ValueError: Input 0 is incompatible with layer model_2: expected shape=(None, 224, 224, 3), found shape=(1, 124, 124, 3)

If I I use a batch_size above 1, e.g., 32 with loss=categorical_crossentropy the error is thrown befor training:

InvalidArgumentError: Incompatible shapes: [32] vs. [0] [[node Equal (defined at :12) ]] [Op:__inference_train_function_65107]

with loss=mse

InvalidArgumentError: Incompatible shapes: [0,2] vs. [32,2] [[node gradient_tape/mean_squared_error/BroadcastGradientArgs (defined at :12) ]] [Op:__inference_train_function_81958]

If I change the Hidden units of the last Dense Layer, the error changes to that. e.g.

...
headModel = Dense(5, activation="softmax")(headModel)

results in

InvalidArgumentError: Incompatible shapes: [0,5] vs. [32,2]

Apparently the correct input shape gets lost somwhere. Especially the batch size (second dimension is based on dense hidden units). Does anyone have an idea? thanks

I checked many answeres from this old thread on git : https://github.com/kuza55/keras-extras/issues/7 but could not find a solution there.

user_7370153 · Answer

The data you are feeding to the network must have the same shape than the input of the network. You are trying to feed a network that accept an image of dimensions 224x224x3 with data that has dimensions 124x124x3.

You can either:

load the mobilenet network with a compatible input dimension

baseModel = MobileNetV2(weights=None, 
                  include_top=False, 
                  input_tensor=Input(shape=(124, 124,3)), 
                  )

This approach has the downside of not being able to use pretrained weights.

reshape your input data to the size of the input of the model. In that case, resizing your 124xx124 images to 224x224. There is plenty of way to do it, but if you are willing to keep ImageDataGenerator, I suggest doing it before hand.

Dense layer probably produces InvalidArgumentError: Incompatible shapes: [0,2] vs. [32,2]

1 Answers