patter
patter copied to clipboard
ryanleary
speech-to-text in pytorch
There may be a bug including more than 2 layers.
Librosa's implementation seems to cause issues despite being faster.
Current implementation is extremely inefficient. Use grouped convolutions.
Create script for automatically making required model changes required for onnx export to be successful and then serialize to disk.
Leverage github.com/nvidia/apex to add multi-gpu training
Use github.com/nvidia/apex to enable FP16 training on Volta
Does it sometimes make sense to save the model with the best CER and assume that the lower CER will help the beam decoder make a better decision (even if...
Sortagrad appears to be a pathological case for memory allocator (also possibly may be related to interactions with warp-ctc memory allocations) that causes OOM errors often very late in an...
Models should include most information required to be able to faithfully recreate them. To that end, metadata/information about the training environment should be persisted to ensure the mechanics of a...