ArtistCoherencyModel.py

from transformers import AutoTokenizer, AutoModelForSequenceClassification
from huggingface_hub import PyTorchModelHubMixin

import torch
import torch.nn as nn

from typing import Union

from FFNN import FFNN


class ArtistCoherencyModel(nn.Module, PyTorchModelHubMixin):
    def __init__(self, config: dict):
        super().__init__()

        coherency_model_repo_id = config["coherency_model_repo_id"]
        artist_model_repo_id = config["artist_model_repo_id"]
        ffnn_model_repo_id = config["ffnn_model_repo_id"]

        self.coherency_model_tokenizer = AutoTokenizer.from_pretrained(
            coherency_model_repo_id
        )
        self.artist_model_tokenizer = AutoTokenizer.from_pretrained(
            artist_model_repo_id
        )

        self.coherency_model = AutoModelForSequenceClassification.from_pretrained(
            coherency_model_repo_id
        )
        self.artist_model = AutoModelForSequenceClassification.from_pretrained(
            artist_model_repo_id
        )
        self.ffnn = FFNN.from_pretrained(ffnn_model_repo_id)

    def generate_artist_logits(self, song: str) -> torch.FloatTensor:
        inputs = self.artist_model_tokenizer(
            song, return_tensors="pt", max_length=512, truncation=True
        )
        with torch.no_grad():
            return self.artist_model(**inputs).logits

    def generate_coherency_logits(self, song: str) -> torch.FloatTensor:
        inputs = self.coherency_model_tokenizer(
            song, return_tensors="pt", max_length=512, truncation=True
        )
        with torch.no_grad():
            return self.coherency_model(**inputs).logits

    def generate_song_embedding(self, song: str) -> torch.FloatTensor:
        with torch.no_grad():
            artist_logits = self.generate_artist_logits(song)
            coherency_logits = self.generate_coherency_logits(song)
            return torch.hstack((artist_logits[0], coherency_logits[0]))

    def forward(self, song_or_embedding: Union[str, torch.Tensor]):
        if type(song_or_embedding) is str:
            song_or_embedding = self.generate_song_embedding(song_or_embedding)

        return self.ffnn(song_or_embedding)

    def generate_artist_coherency_logits(
        self, song_or_embedding: Union[str, torch.Tensor]
    ) -> torch.FloatTensor:
        with torch.no_grad():
            return self.forward(song_or_embedding)

    def generate_artist_coherency_score(
        self, artist_name: str, song_or_embedding: Union[str, torch.Tensor]
    ) -> float:
        coherent_index = self.ffnn.label2id[f"{artist_name}-coherent"]
        incoherent_index = self.ffnn.label2id[f"{artist_name}-incoherent"]
        logits = self.generate_artist_coherency_logits(song_or_embedding)
        coherent_score = logits[coherent_index]
        incoherent_score = logits[incoherent_index]
        score = (coherent_score + incoherent_score) * (
            coherent_score / incoherent_score
        )
        return float(score)

    def predict(
        self, song_or_embedding: Union[str, torch.Tensor], return_ids: bool = False
    ) -> Union[list[str], torch.Tensor]:
        if type(song_or_embedding) is str:
            song_or_embedding = self.generate_song_embedding(song_or_embedding)

        return self.ffnn.predict(song_or_embedding, return_ids=return_ids)