Oliver Adams
Oliver Adams
I spotted what I think is a bug in the Chapter 5 notebook. In the book, sections 5.19-5.22 do counterfactual modeling with a model where A -> M and A...
In `BiFPN.forward()` the feature weights are normalized: https://github.com/toandaominh1997/EfficientDet.Pytorch/blob/fbe56e58c9a2749520303d2d380427e5f01305ba/models/bifpn.py#L178-L180 In calculating the bottom up and top down features, it appears this normalization is done again unnecessarily: https://github.com/toandaominh1997/EfficientDet.Pytorch/blob/fbe56e58c9a2749520303d2d380427e5f01305ba/models/bifpn.py#L190 https://github.com/toandaominh1997/EfficientDet.Pytorch/blob/fbe56e58c9a2749520303d2d380427e5f01305ba/models/bifpn.py#L196
In this implementation as well as others I've found online, there are separate weights for each layer for the feature fusion: https://github.com/toandaominh1997/EfficientDet.Pytorch/blob/fbe56e58c9a2749520303d2d380427e5f01305ba/models/bifpn.py#L147 In [the paper](https://arxiv.org/pdf/1911.09070.pdf) it appears as though the...
https://github.com/persephone-tools/persephone/blob/da67e989639948ea1cfdf8befd8cc9090370a914/persephone/corpus_reader.py#L61-L84 This is an artificial limitation, since the remainder can always just be a smaller batch. It also causes a bug where if the number of training examples is less...
Wraps decode and writes the hypotheses to a text file, or latex source file.
Things to document: - [ ] The need to build Kaldi for `filterbank_and_pitch` features. - [x] Need for settings.ini / include a default. - [ ] Model training arguments. -...
The user selects their data via the web interface, it is then uploaded to the persephone server's `data/org` directory. The user should be able to select the data directory through...