Help Needed: Fine-Tuning and Deploying a Model into OpenSearch

Plugins Machine Learning
, , ,
1

Hello everyone,

I’m currently working on fine-tuning a machine-learning model to enhance search capabilities in OpenSearch, but I’ve hit a snag. After successfully registering the model, I encountered errors during deployment as solving one error resulted in another. I would greatly appreciate any guidance or resources available online from where I can learn how to fine-tune the model and deploy it on OpenSearch

Here is my example Code:

from sentence_transformers import SentenceTransformer
model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
# Example dataset
train_examples = [
    ("Wooden dining table", "Oak coffee table", 0.5),
    ("Leather office chair", "Ergonomic mesh chair", 0.4),
    # Add more examples as needed
]

from sentence_transformers import InputExample

train_dataset = [InputExample(texts=[text1, text2], label=label) for text1, text2, label in train_examples]
from sentence_transformers import losses, evaluation, models, SentenceTransformer
from torch.utils.data import DataLoader

# Define a dataloader and a loss
train_dataloader = DataLoader(train_dataset, shuffle=True, batch_size=16)
train_loss = losses.CosineSimilarityLoss(model=model)

dev_examples = [
    InputExample(texts=['Wooden dining table', 'Oak coffee table'], label=0.5),
    InputExample(texts=['Leather office chair', 'Ergonomic mesh chair'], label=0.4),
    InputExample(texts=['Convertible sofa bed', 'Futon couch'], label=0.8),
]

# If available, add a development set for evaluation during training
dev_evaluator = evaluation.EmbeddingSimilarityEvaluator.from_input_examples(dev_examples, name='dev')

# Define training parameters
model.fit(train_objectives=[(train_dataloader, train_loss)],
          evaluator=dev_evaluator,
          epochs=1,
          warmup_steps=100,
          output_path='output/training_miniLM')
import torch

# Load the model and set to evaluation mode
model = SentenceTransformer('output/training_miniLM')
model.eval()

# Define the dummy input for the export
input_ids = torch.tensor([model.tokenizer.encode("Sample text for ONNX export")])

# Export to ONNX
torch.onnx.export(model=model._first_module().auto_model,
                  args=(input_ids,),
                  f='miniLM_finetuned.onnx',
                  input_names=['input_ids'],
                  output_names=['sentence_embedding'],
                  dynamic_axes={'input_ids': {0: 'batch_size'}, 'sentence_embedding': {0: 'batch_size'}},
                  opset_version=12)
2

After training the model, you can export to onnx or torchscript format and then you can upload this model to opensearch.

https://opensearch-project.github.io/opensearch-py-ml/examples/demo_tracing_model_torchscript_onnx.html

3

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.