I updated the tree serialization format for the new scikit learn versions. I also updated the minimum requirement of scikit learn to 1.3 to ensure compatibility.
Fixes#555
For migration from scripts to console_scripts in setup.py,
the current long if __name__ == "__main__": section is a
blocker because the console_scripts requires to specify a
function as an entrypoint.
Move the logic into a main() function.
The eland_import_hub_model script supports uploading a local file where
the --hub-model-id argument is a file path. If the --es-model-id option is
not used the model Id is generated from the hub model id and when that
is a file path the path must be converted to a valid elasticsearch model id.
Closes#503
Note: I also had to fix the Sphinx version to 5.3.0 since, starting from 6.0, Sphinx suffers from a TypeError bug, which causes a CI failure.
Adds text_similarity task support. This is a cross-encoder transformer task where both sequences are given to the transformer at once.
According to 🤗 (or at least how the cross-encoder models are concerned) this is a sequence classification task with just one classification "label". But really, it isn't labeled at all and is more akin to a regression model.
related: elastic/elasticsearch#88439
Elasticsearch uses v1.11 of PyTorch. Models created with the latest PyTorch
release (v1.12) are not compatible with v1.11. This pins the PyTorch version
to 1.11 to prevent the incompatibility. The version of the Elasticsearch Python
client is now required to be >= Eland.
All users of Eland for importing NLP models should upgrade.
For many model types, we don't need to require the task requested. We can infer the task type based on the model configuration and architecture.
This commit makes the `task-type` parameter optional for the model up load script and adds logic for auto-detecting the task type based on the 🤗 model.
This switches our sklearn.DecisionTreeClassifier serialization logic to account for multi-valued leaves in the tree.
The key difference between our inference and DecisionTreeClassifier, is that we run a softMax over the leaf where sklearn simply normalizes the results.
This means that our "probabilities" returned will be different than sklearn.
This improves the user consumed functions and classes for PyTorch NLP model upload to Elasticsearch.
Previously it was difficult to wrap your own module for uploading to Elasticsearch.
This commit splits some classes out, adds new ones, and adds tests showing how to wrap some simple modules.
This adds some more definite types for our NLP tasks and tokenization configurations.
This is the first step in allowing users to more easily import their own transformer models via something other than hugging face.
## Changes
### Better logging
Switched from `print` statements to `logging` for a cleaner and more informative output - timestamps and log level are shown. The logging is now a bit more verbose, but it will help users to better understand what the script is doing.
### Add support for ES authentication using username/password or api key
Instead of being limited to passing credentials in the URL, there are now 2 additional methods:
- username/password using `--es-username` and `--es-password`
- API key using `--es-api-key`
Credentials can also be specified as environment variables with `ES_USERNAME`/`ES_PASSWORD` or `ES_API_KEY`
### Graceful handling of missing PyTorch requirements
In order to use the `eland_import_hub_model` script, PyTorch extras are required to be installed. If the user does not have the required packages installed, a helpful message is logged with a hint to install `eland[pytorch]` with `pip`.
### Graceful handling of already existing trained model
If a trained model with the same ID as the one we're trying to import already exists, and `--clear-previous` was not specified, we now log a clearer message about why the script can't proceed along with a hint to use the `--clear-previous` flag.
Prior to this change, we were letting the API exception seep through and the user was faced with a stack trace.
### `tqdm` added to main dependencies
If the user doesn't have `eland[pytorch]` extras installed, the first module to be reported as missing is `tqdm`. Since this module is [used in eland codebase](8294224e34/eland/ml/pytorch/_pytorch_model.py (L24)) directly, it makes sense to me to have it as part of the main set of requirements.
### Nit: Set tqdm unit to `parts` in `_pytorch_model.put_model`
The default unit is `it`, but `parts` better describes what the progress bar is tracking - uploading trained model definition parts.
We added the `pytorch` module which is type checked but was not in the
noxfile as such. This change also addresses type errors that arose after
adding type checking.
The padding parameter needs to be set on the tokenization call and not
in the constructor. Furthermore, the True value will only pad to the
largest input in a batch, however we don't trace with batches so this
value had no effect. The proper place to pass this parameter is in the
tokenization call itself and the proper value to use is "max_length"
which will pad the input to the maximum input size specified by the
model. Although we measure no functional or performance impact of this
setting, it has been suggested that this is a best practice.
See: https://huggingface.co/transformers/serialization.html#dummy-inputs-and-standard-lengths
