ValueError: Dimension size must be evenly divisible by 10 but is 1 for '{{node training/Reshape}} = Reshape[T=DT_FLOAT, Tshape=DT_INT32](loss/mul, training/Reshape/shape)' with input shapes: [], [2] and with input tensors computed as partial shapes: input[1] = [10,?]

Hello,
I am trying to train with differential privacy using tensorflow_privacy.DPKerasSGDOptimizer() here is my code

E = 1
B = 100
N = 2
dataset = pickle.load( open(path, "rb" ) )
data = deepcopy(dataset)
list_cli = list(dataset.keys())

for client in list_cli:
     data[client]['x_train'] = dataset[client]['x_train'][:int((len(dataset[client]['x_train'])//batch_length)*batch_length)]
     data[client]['y_train'] = dataset[client]['y_train'][:int((len(dataset[client]['y_train'])//batch_length)*batch_length)]
     data[client]['x_test'] = dataset[client]['x_test'][:int((len(dataset[client]['x_test'])//B)*B)]
     data[client]['y_test'] = dataset[client]['y_test'][:int((len(dataset[client]['y_test'])//B)*B)]

I get the following error :

ValueError: Dimension size must be evenly divisible by 10 but is 1 for '{{node training/Reshape}} = Reshape[T=DT_FLOAT, Tshape=DT_INT32](loss/mul, training/Reshape/shape)' with input shapes: [], [2] and with input tensors computed as partial shapes: input[1] = [10,?]

Note the when I use

model.compile(optimizer="Adam", loss='categorical_crossentropy', metrics=['accuracy'])

The code executes without errors.

Do you have any ideas ?

Thanks

Start with the model.fit call and tensorflow_privacy.DPKerasSGDOptimizer configuration shown in the report, especially num_microbatches=10 and batch_size=B. Reproduce the reshape error with the reported training setup and compare it with the Adam configuration. Done means identifying the incompatible configuration or missing condition and documenting or correcting it with a focused regression check.