Load Custom Models from The Local Filesystem

11.8 Load Custom Models from The Local Filesystem

Custom models that have been fine-tuned from pre-trained Hugging Face models and reside on the local filesystem can be converted to the ONNX format using the new local_model_path setting. This setting can also be useful in cases where networking is unavailable and a pre-trained model has previously been downloaded.

The model directory for a custom model must follow the following format:

<basedir>/<model_file_subfolders>

basedir: This is the base directory for the model sub folders and is passed as the local_model_path settings parameter. This parameter is described in the Generate the Augmented ONNX File section. It can be an absolute or relative path. If relative, it will be relative to where Python was started.
model_file_subfolders: These nested folders are created based on the model's name. For each '/' character in the model's name, you should create a new subdirectory.

Example 11-1 Local Directory for Custom Model Files

For example, if the model name is mymodel, only one directory named mymodel is expected. However, if the model name is user/mymodel, you should create a parent directory called user with a subdirectory named mymodel. All model files should be placed in the final subdirectory (mymodel in this case).


user
    mymodel

Generate the Augmented ONNX File

When initializing the EmbeddingModel object, use the settings parameter to pass a new property local_model_path, which identifies the basedir for the model. For example, if the basedir for the model on the local filesystem is $HOME/sample-model, and there is a custom model named my-model, you can achieve this with the following Python code:

from oml.utils import EmbeddingModel
import os

em = EmbeddingModel('my-model',settings={'local_model_path':f'{os.environ["HOME"]}/sample-model'})
em.export2file('my-onnx-model','.')

At the end of this process, you should find a file named my-onnx-model.onnx in the current directory.

Note:

The above code is the same for exporting to the database except that export2db is used instead of export2file. For more information, see ONNX Pipeline Models: Text Embedding. The code is compatible with all downloaded pretrained models, preconfigured models, and templates.

Error and Warning Conditions

Table 11-3 Error and Warning Conditions

Error ( or Warning) Conditions	Error ( or Warning) Message	Notes
force_download specified	force_download cannot be true when specifying 'local_model_path'	force_download is mutually exclusive with this feature so it cannot be specified when local_model_path has a value.
local_model_path is not a directory	'{local_path}' specified by 'local_model_path' must be an existing directory.	`{local_path}` is the value specified in `local_model_path` . This message is returned when the value of local_model_path is not an existing directory.
Missing model config file	Expected local model folder {model_path} to contain the file config.json but it was not found.	`{model_path}` is the full path to the local model folder. This message is returned when the local model folder does not contain a file named config.json.
Missing model files	Expected local model folder {model_path} to at least contain one of model.safetensors or pytorch_model.bin files. Neither were found.	`{model_path}` is the full path to the local model folder. This message is returned when the local model folder does not contain either a pytorch_mode.bin file OR at least one file that does not end in .safetensors.
preconfigured model used (Warning)	Warning: checksum validation is disabled when loading preconfigured models from a local path.	This warning is shown when a preconfigured model is loaded from a local path.

Parent topic: Import Pretrained Models in ONNX Format