sentiment-analyzer-coreml

This is a CoreML version of the DistilBERT sentiment analysis model, converted from the Hugging Face model distilbert-base-uncased-finetuned-sst-2-english.

Model Details

• Original Model: distilbert-base-uncased-finetuned-sst-2-english
• Task: Sentiment Analysis
• Framework: CoreML
• Input: Text (tokenized as input_ids and attention_mask)
• Output: Logits for sentiment classification (2 classes: negative, positive)

Usage

Python (CoreML)

import coremltools as ct

# Load the model
model = ct.models.MLModel("sentiment_analyzer.mlpackage")

# Get model spec
spec = model.get_spec()
print("Model type:", spec.WhichOneof('Type'))

# Make predictions (you'll need to tokenize your input first)
# The model expects input_ids and attention_mask as inputs

Swift (iOS/macOS)

import CoreML

// Load the model
guard let model = try? MLModel(contentsOf: URL(fileURLWithPath: "sentiment_analyzer.mlpackage")) else { return }

// Make predictions
// You'll need to convert your text to the required input format

Input Format

The model expects two inputs:

• input_ids: Tokenized input text (shape: [1, sequence_length])
• attention_mask: Attention mask (shape: [1, sequence_length])

Output Format

The model outputs logits for sentiment classification:

• Shape: [1, 2] (batch_size, num_classes)
• Classes: [negative, positive]

Conversion Notes

This model was converted using coremltools from the original PyTorch model. The conversion process involved:

1. Loading the Hugging Face model
2. Wrapping it to return only logits (tensor output)
3. Tracing with PyTorch JIT
4. Converting to CoreML format

Requirements

• iOS 15+ / macOS 12+ (for ML Program format)
• CoreML framework