🍵 Tea Grading

A pytorch model for grading tea quality based on a sequence of procedures.

📖 Project Structure

- tea_grading/
    - load_training.py
    - train.py
    - main.db
    - schemas.sqlite3-query
    - tokens.json

• load_training.py: Script to load training data.
• train.py: Script to train the model.
• main.db: SQLite database containing the training data.
• schemas.sqlite3-query: SQL queries for the database schema.
• tokens.json: JSON file containing tokens for the model.

🏠 Architecture

The model is a sequence autoencoder with three parts: tokenizer, encoder, and decoder.

🔧 Tokenizer

• Normalizes step parameters (param1, param2, param3) using global min-max scaling.
• Converts normalized parameters into fixed-length decimal-string tokens.
• Combines each operation token with its parameter tokens into a flat token sequence.
• Wraps the sequence with special <START> and <END> tokens.

💻 Encoder

• Embeds tokens and projects parameter triplets separately.
• Adds token and parameter embeddings element-wise.
• Feeds combined embeddings into a GRU to produce a latent representation.

🔒 Decoder

• Takes the token sequence (excluding last token) and latent vector as input.
• Uses the shared token embedding and a GRU to generate output sequences.
• Applies a linear layer to produce logits for next-token prediction.

🔮 Output

The semantic output is in the hidden state of the encoder(ProcEncoder), which captures the sequence of operations and their parameters. The three components contains

1. Phenol / 500g (mg)
2. Caffeine / 500g (mg)
3. Price / 500g (RMB)

✍️ License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.