Gputill performance difference, tf.dataset and tf.In the tenor

new_shape = (self.bsz, -1) # The rest of the dimensions are automatically calculated
data_reshaped = tf.reshape(sliced_data, new_shape)
data_transposed = tf.transpose(data_reshaped) = data_transposed

why tf.dataset.tensor made by from.generation() has much higher gpu till than when i use tf tensor made by just tf.def like above example

Hi @wonjun_choi

Please provide reproducible code with dataset shape to replicate the error and find out the issue. Thank you.

I solved the problem, but I couldn’t answer. I’m sorry