[FFmpeg-devel] [GSOC] [PATCH] TensorFlow backend introduction for DNN module

[Pedro Arthur]

2018-06-03 17:05 GMT-03:00 Pedro Arthur <bygrandao at gmail.com>:
> 2018-06-03 16:55 GMT-03:00 Sergey Lavrushkin <dualfal at gmail.com>:
>>> My concern is when we add more models, currently we have to store 2
>>> models, one for the "native" implementation and one for the TF
>>> backend.
>>> There is also the case were one wants to update the weights for a
>>> model, it will be necessary to update both the native and TF data.
>>> Having duplicated data is much easier to get inconsistencies between
>>> implementations.
>>
>>
>> I understand the problem, but I am afraid that manual graph construction can
>> take a lot of time, especially if we add something more complicated than
>> srcnn,
>> and the second approach passing weights in placeholders will require to
>> add some logic for it in other parts of API besides model loading.
>> I am thinking of another way, that is to get weights for native model from
>> this binary
>> tf model, if they are stored there consistently, and not specify them as
>> float arrays.
>> But then for each new model we need to find offsets for each weights array.
> Ok, we can keep it until we find a reasonable solution.


If there are nothing else to be addressed I'll push it by tomorrow.