stephenrawls opened a new issue #14019: SequenceLast undefined behavior URL: https://github.com/apache/incubator-mxnet/issues/14019 SequenceLast currently causes undefined behavior if `use_sequence_length` is true and any of the values in the `sequence_length` array are 0. Proof: Run this program with cuda-memcheck ``` import mxnet as mx data = mx.nd.array([[0,0],[0,0]], ctx=mx.gpu()) seq_lens = mx.nd.array([0, 0], ctx=mx.gpu()) res = mx.nd.SequenceLast(data, sequence_length=seq_lens, use_sequence_length=True, axis=0) print("res = %s" % str(res.as_in_context(mx.cpu()))) ``` Line in code where sequence_length index is used w/o checking if it is valid: https://github.com/apache/incubator-mxnet/blob/master/src/operator/sequence_last-inl.h#L73-L75 On the one hand, the documentation does clearly state that: ``` Parameter sequence_length is used to handle variable-length sequences. sequence_length should be an input array of positive ints of dimension [batch_size] ``` And 0 is not a positive integer so it fails this pre-condition. On the other hand, both of SequenceReverse and SequenceMask handle 0-valued sequence_length's safely. Also, the Take operator, which SequenceLast even mentions as an almost equivalent alternative, also handles out-of-band indices safely. The argument I have for why you would want to pass in a 0 sequence length (or perhaps even an invalid sequence length) -- In a bucketing executor approach where the user only supports a fixed number of shapes, it is entirely likely that my batch size is, say 16, but I only have 10 actual elements to fill into my input array. The other 6 elements are all padding, and my code currently just zeros out all padding without regard to if it is an input element that will be used as a "sequence_length" parameter. In the interest of not having undefined behavior (which led to non-deterministic crashes in some of my code), I would like to propose making SequenceLast safe when any of the sequence_lengths is 0. If there is interest in this I can put in a PR. Please let me know. Thanks, Stephen
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
