mead-baseline
mead-baseline copied to clipboard
mead-ml
Deep-Learning Model Exploration and Development for NLP
Recent versions of PyTorch (after 1.7.1) do not support the boolean comparison on export. It works if we change the definition of `MASK_FALSE` from a boolean to an int (`0`)
quickly mock up 1. supporting more losses in the classifier via kwargs and 2. what happens if the Y shape is dense. In that case the soft target argmax is...
our tril impl was using numpy -- according to the ONNX docs, this is not the right way during tracing, it should be tensor/pyt ops. PyTorch does have the proper...
Adds a prompt_toolkit based repl to make it easy to query an MLM. Also adds export of bias term to HF exporter. To change the method from sampling to greedy...
We should be able to use `tf.function` to make our new faster, eager models export into TF serving. I have just not gotten around to doing this because its tedious...
The methodology that was outline in the export.md is incredibly out-of-date. TensorFlow has official docker binaries now as well
- Adds an 8mi primitive called trainer to handle device specifics - Allow loaders to use "native" framework (ie tf's Dataset and pytorch's DataLoader) for batch delivery - Refactors TF...
While testing the dependency parser, I was using a combination of word, POS and `char-lstm` embeddings and it was failing with an autograph error. In most other cases when we...
So when you have an input tensor like `lengths` for a model that isn't used it often gets stripped out by onnx (or pytorch?). For example when you have a...
We have some files that have plural names like `baseline.vectorizers` and `eight_mile.${backend}.layers` and others that are singular like `baseline.reader` we should standardize.