Extension that evaluates tensors for a given datastream. Typical use case is when you want to generate predictions for a test set. If indicated, predictions are pickled and saved. I chose pickle over numpy.save because it allowed me to map the tensor names to arrays - sadly, using the tensor itself as key resulted in KeyErrors.
It's possible that MonitoredQuantity or DatasetEvaluator could be used instead, but @dwf warned me in this issue that concatenating after every step (using e.g. a Concatenate aggregation scheme) was probably going to be more memory-intensive than concatenating once at the end of the callback.
Extension that evaluates tensors for a given datastream. Typical use case is when you want to generate predictions for a test set. If indicated, predictions are pickled and saved. I chose pickle over numpy.save because it allowed me to map the tensor names to arrays - sadly, using the tensor itself as key resulted in KeyErrors.
It's possible that
MonitoredQuantity
orDatasetEvaluator
could be used instead, but @dwf warned me in this issue that concatenating after every step (using e.g. a Concatenate aggregation scheme) was probably going to be more memory-intensive than concatenating once at the end of the callback.