Novora/CodeClassifier-v1-Tiny

Introduction

Novora Code Classifier v1 Tiny, is a tiny Text Classification model, which classifies given code text input under 1 of 31 different classes (programming languages).

This model is designed to be able to run on CPU, but optimally runs on GPUs.

Info

• 1 of 31 classes output
• 512 token input dimension
• 64 hidden dimensions
• 2 linear layers
• The snowflake-arctic-embed-xs model is used as the embeddings model.
• Dataset split into 80% training set, 20% testing set.
• The combined test and training data is around 1000 chunks per programming language, the data is 31,100 chunks (entries) as 512 tokens per chunk, being a snippet of the code.
• Picked from the 18th epoch out of 20 done.

Architecture

The CodeClassifier-v1-Tiny model employs a neural network architecture optimized for text classification tasks, specifically for classifying programming languages from code snippets. This model includes:

• Bidirectional LSTM Feature Extractor: This bidirectional LSTM layer processes input embeddings, effectively capturing contextual relationships in both forward and reverse directions within the code snippets.

• Fully Connected Layers: The network includes two linear layers. The first projects the pooled features into a hidden feature space, and the second linear layer maps these to the output classes, which correspond to different programming languages. A dropout layer with a rate of 0.5 between these layers helps mitigate overfitting.

The model's bidirectional nature and architectural components make it adept at understanding the syntax and structure crucial for code classification.

Testing/Training Datasets

I have put here the samples entered into the training/testing pipeline, its a very small amount.

Language	Testing Count	Training Count
Ada	20	80
Assembly	20	80
C	20	80
C#	20	80
C++	20	80
COBOL	14	55
Common Lisp	20	80
Dart	20	80
Erlang	20	80
F#	20	80
Go	20	80
Haskell	20	80
Java	20	80
JavaScript	20	80
Julia	20	80
Kotlin	20	80
Lua	20	80
MATLAB	20	80
PHP	20	80
Perl	20	80
Prolog	1	4
Python	20	80
R	20	80
Ruby	20	80
Rust	20	80
SQL	20	80
Scala	20	80
Swift	20	80
TypeScript	20	80

Example Code

import torch.nn as nn
import torch.nn.functional as F

class CodeClassifier(nn.Module):
    def __init__(self, num_classes, embedding_dim, hidden_dim, num_layers, bidirectional=False):
        super(CodeClassifier, self).__init__()
        self.feature_extractor = nn.LSTM(embedding_dim, hidden_dim, num_layers, batch_first=True, bidirectional=bidirectional)
        self.dropout = nn.Dropout(0.5)  # Reintroduce dropout
        self.fc1 = nn.Linear(hidden_dim * (2 if bidirectional else 1), hidden_dim)  # Intermediate layer
        self.fc2 = nn.Linear(hidden_dim, num_classes)  # Output layer

    def forward(self, x):
        x = x.unsqueeze(1)  # Add sequence dimension
        x, _ = self.feature_extractor(x)
        x = x.squeeze(1)  # Remove sequence dimension
        x = self.fc1(x)
        x = self.dropout(x)  # Apply dropout
        x = self.fc2(x)
        return x

import torch
from transformers import AutoTokenizer, AutoModel
from pathlib import Path

def infer(text, model_path, embedding_model_name):
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    
    # Load tokenizer and embedding model
    tokenizer = AutoTokenizer.from_pretrained(embedding_model_name)
    embedding_model = AutoModel.from_pretrained(embedding_model_name).to(device)
    embedding_model.eval()

    # Prepare inputs
    inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)
    inputs = {k: v.to(device) for k, v in inputs.items()}
    
    # Generate embeddings
    with torch.no_grad():
        embeddings = embedding_model(**inputs)[0][:, 0]

    # Load classifier model
    model = CodeClassifier(num_classes=31, embedding_dim=embeddings.size(-1), hidden_dim=64, num_layers=2, bidirectional=True)
    model.load_state_dict(torch.load(model_path, map_location=device))
    model = model.to(device)
    model.eval()

    # Predict class
    with torch.no_grad():
        output = model(embeddings)
        _, predicted = torch.max(output, dim=1)

    # Language labels
    languages = [
        'Ada', 'Assembly', 'C', 'C#', 'C++', 'COBOL', 'Common Lisp', 'Dart', 'Erlang', 'F#',
        'Fortran', 'Go', 'Haskell', 'Java', 'JavaScript', 'Julia', 'Kotlin', 'Lua', 'MATLAB',
        'Objective-C', 'PHP', 'Perl', 'Prolog', 'Python', 'R', 'Ruby', 'Rust', 'SQL', 'Scala',
        'Swift', 'TypeScript'
    ]
    
    return languages[predicted.item()]

# Example usage
if __name__ == "__main__":
    example_text = "print('Hello, world!')"  # Replace with actual text for inference
    model_file_path = Path("./model.safetensors")
    predicted_language = infer(example_text, model_file_path, "Snowflake/snowflake-arctic-embed-xs")
    print(f"Predicted programming language: {predicted_language}")