import os
import random
import json
import librosa
import numpy as np
import soundfile as sf
from torch.utils import data
from suno_utils.audio import Audio
from suno_utils.utils.text import read_jsonl


class CleanedYTMSDDataset(data.Dataset):
    def __init__(
        self,
        data_path="/app/suno/data/audio_mono_24khz/cleaned_ytmsd",
        split="train",
        input_length_s=30.0,
        sample_rate=24000,
        num_samples=-1,
    ):
        assert split in ["train", "valid"]
        self.data_path = data_path
        self.split = split
        self.input_length_s = input_length_s
        self.sample_rate = sample_rate
        self.num_samples = num_samples

        # load files
        self.metadata = read_jsonl(os.path.join(data_path, "%s.jsonl" % split))
        print("%d files are available for %s set" % (len(self.metadata), split))

    def concatenate_tags(self, tags):
        if self.split == "train":
            random.shuffle(tags)
        tags = [tag.lower() for tag in tags if len(tag.split(" ")) < 5]
        concatenated_tags = ", ".join(tags)

        # add [CLS] token
        concatenated_tags = "[CLS]" + concatenated_tags
        return concatenated_tags

    def __getitem__(self, index):
        # read data
        metadata = self.metadata[index]
        tags = metadata["clean_tags"]

        # load audio
        audio_path = metadata["filepath"]
        audio = Audio.from_file(audio_path, sample_rate=self.sample_rate)
        if self.split == "train":
            # random crop
            try:
                start_ms = random.randint(
                    0, (audio.duration_ms - int(1000 * self.input_length_s) - 1)
                )
                start_s = start_ms / 1000
            except ValueError as e:
                start_s = 0.0
            wav = audio.get_segment(
                from_s=start_s, to_s=start_s + self.input_length_s
            ).array_float

        elif self.split == "valid":
            # crop first 30s
            wav = audio.get_segment(from_s=0.0, to_s=self.input_length_s).array_float

        if len(wav) < int(self.sample_rate * self.input_length_s):
            pad = int(self.sample_rate * self.input_length_s) - len(wav)
            wav = np.pad(wav, (0, pad), mode="constant", constant_values=0)

        # load tag labels
        concatenated_tags = self.concatenate_tags(tags)

        return wav, concatenated_tags

    def __len__(self):
        if self.num_samples > 0:
            return self.num_samples
        else:
            return len(self.metadata)