Sockeye is a fast and scalable deep learning library based on Apache MXNet. The Sockeye code base has unique advantages from MXNet. For example, Sockeye combines declarative and imperative programming styles through symbolic and imperative MXNet APIs; models can be trained in parallel on multiple GPUs.
Currently Sockeye has released versions 3.1.18 & 3.1.29, which bring the following changes:
Added from Khandelwal et al., 2021 kNN-MT model,
Installation: Refer to the faiss documentation, it is recommended to install via conda.
Building a faiss index from a sockeye model requires two steps:
- Generate decoder state:
sockeye-generate-decoder-states -m [model] --source [src] --target [tgt] --output-dir [output dir]
- Build the index:
sockeye-knn -i [input_dir] -o [output_dir] -t [faiss_index_signature]where input_dir is the same as
sockeye-generate-decoder-statesThe output_dir in the command is the same.
Faiss index signature reference: refer here
- Run inference using built-in indexes:
sockeye-translate ... --knn-index [index_dir] --knn-lambda [interpolation_weight]where index_dir is the same as
sockeye-knnThe output_dir in the command is the same.
- Running sockeye-evaluate no longer applies text tokenization to TER (same behavior as other metrics).
- Turned on type checking for all sockeye modules except test_utils, and resolved type issues arising from it.
- Refactor code in various modules without changing user-level behavior.
Update announcement: https://github.com/awslabs/sockeye/releases/tag/3.1.29
#Sockeye #released #fast #scalable #deep #learning #library #News Fast Delivery