# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
"""Compression models or wrapper around existing models.
Also defines the main interface that a model must follow to be usable as an audio tokenizer.
"""

from abc import ABC, abstractmethod
import logging
from pathlib import Path
import typing as tp

from einops import rearrange
import numpy as np
import torch
from torch import nn
from transformers import EncodecModel as HFEncodecModel


logger = logging.getLogger()


class CompressionModel(ABC, nn.Module):
    """Base API for all compression models that aim at being used as audio tokenizers
    with a language model.
    """

    

    @abstractmethod
    def decode(self, codes: torch.Tensor, scale: tp.Optional[torch.Tensor] = None):
        """See `EncodecModel.decode`."""
        ...

    @abstractmethod
    def decode_latent(self, codes: torch.Tensor):
        """Decode from the discrete codes to continuous latent space."""
        ...

    @property
    @abstractmethod
    def channels(self) -> int:
        ...

    @property
    @abstractmethod
    def frame_rate(self) -> float:
        ...

    @property
    @abstractmethod
    def sample_rate(self) -> int:
        ...

    @property
    @abstractmethod
    def cardinality(self) -> int:
        ...

    @property
    @abstractmethod
    def num_codebooks(self) -> int:
        ...

    @property
    @abstractmethod
    def total_codebooks(self) -> int:
        ...

    @abstractmethod
    def set_num_codebooks(self, n: int):
        """Set the active number of codebooks used by the quantizer."""
        ...

    


class EncodecModel(CompressionModel):
    """Encodec model operating on the raw waveform.

    Args:
        encoder (nn.Module): Encoder network.
        decoder (nn.Module): Decoder network.
        quantizer (qt.BaseQuantizer): Quantizer network.
        frame_rate (int): Frame rate for the latent representation.
        sample_rate (int): Audio sample rate.
        channels (int): Number of audio channels.
        causal (bool): Whether to use a causal version of the model.
        renormalize (bool): Whether to renormalize the audio before running the model.
    """
    # we need assignment to override the property in the abstract class,
    # I couldn't find a better way...
    frame_rate: float = 0
    sample_rate: int = 0
    channels: int = 0

    def __init__(self,
                 decoder=None,
                 quantizer=None,
                 frame_rate=None,
                 sample_rate=None,
                 channels=None,
                 causal=False,
                 renormalize=False):
        super().__init__()
        
        self.decoder = decoder
        self.quantizer = quantizer
        self.frame_rate = frame_rate
        self.sample_rate = sample_rate
        self.channels = channels
        self.renormalize = renormalize
        self.causal = causal
        if self.causal:
            # we force disabling here to avoid handling linear overlap of segments
            # as supported in original EnCodec codebase.
            assert not self.renormalize, 'Causal model does not support renormalize'

    @property
    def total_codebooks(self):
        """Total number of quantizer codebooks available."""
        return self.quantizer.total_codebooks

    @property
    def num_codebooks(self):
        """Active number of codebooks used by the quantizer."""
        return self.quantizer.num_codebooks

    def set_num_codebooks(self, n: int):
        """Set the active number of codebooks used by the quantizer."""
        self.quantizer.set_num_codebooks(n)

    @property
    def cardinality(self):
        """Cardinality of each codebook."""
        return self.quantizer.bins

    def preprocess(self, x: torch.Tensor) -> tp.Tuple[torch.Tensor, tp.Optional[torch.Tensor]]:
        scale: tp.Optional[torch.Tensor]
        if self.renormalize:
            mono = x.mean(dim=1, keepdim=True)
            volume = mono.pow(2).mean(dim=2, keepdim=True).sqrt()
            scale = 1e-8 + volume
            x = x / scale
            scale = scale.view(-1, 1)
        else:
            scale = None
        return x, scale

    def postprocess(self,
                    x: torch.Tensor,
                    scale: tp.Optional[torch.Tensor] = None) -> torch.Tensor:
        if scale is not None:
            assert self.renormalize
            x = x * scale.view(-1, 1, 1)
        return x

    def decode(self, codes: torch.Tensor, scale: tp.Optional[torch.Tensor] = None):
        # B,K,T -> B,C,T
        emb = self.decode_latent(codes)
        
        out = self.decoder(emb)
        
        out = self.postprocess(out, scale)
        
        return out

    def decode_latent(self, codes: torch.Tensor):
        """Decode from the discrete codes to continuous latent space."""
        return self.quantizer.decode(codes)