Required broadcastable shapes

General Discussion
#1

During training, after 1 epoch i get following error

...
    File "/usr/local/lib/python3.10/dist-packages/keras/optimizers/optimizer_experimental/optimizer.py", line 259, in compute_gradients
      grads = tape.gradient(loss, var_list)
Node: 'gradient_tape/sparse_categorical_crossentropy/SparseSoftmaxCrossEntropyWithLogits/mul'
required broadcastable shapes
	 [[{{node gradient_tape/sparse_categorical_crossentropy/SparseSoftmaxCrossEntropyWithLogits/mul}}]] [Op:__inference_train_function_36362]

I am fitting my model during training using my custom dataset in batches,

input and output has the following shapes

img_shape:  (24, 64, 160, 3)
label_shape:  (24,)

my model has following input

Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 input_2 (InputLayer)           [(24, 64, 160, 3)]   0           []

and following output

dense (Dense)                  (24, 34)             34034       ['tf.compat.v1.squeeze_1[0][0]']

I am using sparse_categorical_crossentropy loss

#3

Hi @hlacik, Could you please provide the standalone code to reproduce issues. Thank You.

#4

Hello,

here is the link for google collab notebook with model/compile/fit

Google Colab Notebook

#6

Hello all,
Any updates on this? I am encountering a similar error. Thanks in advance!

#8

In my case, there was something wrong with the data generator, basically it ran through the dataset once and during the second epoch, the batch dimension had a different dimension, which caused the model output to be of different dimension which eventually screwed up the loss calculation. Hope that helps!

#10

basically same issue on my side - dataset batching.