Of less importance, I think dropping keras and going native pytorch should be a high priority (I have seen recent keras versions have memory leaks with pytorch). But as you said this could also have been done in pytorch.
Originally posted by @matham in #589 (comment)
Out of interest, are there any other reasons to motivate this? The reason we continued to use keras is to try and mitigate the effects of needing to migrate frameworks every few years. cellfinder has already gone from caffe -> NiftyNet -> Keras (TensorFlow) -> Keras (pytorch) in the last 10 years.
Originally posted by @adamltyson in #589 (comment)
Of less importance, I think dropping keras and going native pytorch should be a high priority (I have seen recent keras versions have memory leaks with pytorch). But as you said this could also have been done in pytorch.
Originally posted by @matham in #589 (comment)
Out of interest, are there any other reasons to motivate this? The reason we continued to use keras is to try and mitigate the effects of needing to migrate frameworks every few years. cellfinder has already gone from caffe -> NiftyNet -> Keras (TensorFlow) -> Keras (pytorch) in the last 10 years.
Originally posted by @adamltyson in #589 (comment)