import spaces
import gradio as gr
import torch
import torchaudio
import librosa
import torch.nn as nn
from modules.commons import build_model, load_checkpoint, recursive_munch
import yaml
from hf_utils import load_custom_model_from_hf
import numpy as np
from pydub import AudioSegment

# =========================================================
# Device
# =========================================================
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# =========================================================
# Load Seed-VC DiT model (non-f0)
# =========================================================
dit_checkpoint_path, dit_config_path = load_custom_model_from_hf(
    "Plachta/Seed-VC",
    "DiT_seed_v2_uvit_whisper_small_wavenet_bigvgan_pruned.pth",
    "config_dit_mel_seed_uvit_whisper_small_wavenet.yml"
)

config = yaml.safe_load(open(dit_config_path, 'r'))
model_params = recursive_munch(config['model_params'])
model = build_model(model_params, stage='DiT')
hop_length = config['preprocess_params']['spect_params']['hop_length']
sr = config['preprocess_params']['sr']

model, _, _, _ = load_checkpoint(
    model, None, dit_checkpoint_path,
    load_only_params=True, ignore_modules=[],
    is_distributed=False
)
for key in model:
    model[key].eval()
    model[key].to(device)

# Cache setup
model.cfm.estimator.setup_caches(max_batch_size=1, max_seq_length=8192)

# =========================================================
# Speaker embedding: ECAPA (SpeechBrain) replacement
# - This reduces CN accent bias vs campplus_cn_common
# - Fallback to original CAMPPlus if SpeechBrain not available
# =========================================================
USE_ECAPA = True
spk_encoder = None

try:
    from speechbrain.pretrained import EncoderClassifier
    spk_encoder = EncoderClassifier.from_hparams(
        source="speechbrain/spkrec-ecapa-voxceleb",
        run_opts={"device": str(device)}
    )
    spk_encoder.eval()
except Exception as e:
    # If SpeechBrain isn't installed/available, fallback to CAMPPlus
    USE_ECAPA = False
    spk_encoder = None
    print("[WARN] SpeechBrain ECAPA not available. Falling back to CAMPPlus. Error:", str(e))

# CAMPPlus fallback (original)
campplus_model = None
if not USE_ECAPA:
    from modules.campplus.DTDNN import CAMPPlus
    campplus_ckpt_path = load_custom_model_from_hf(
        "funasr/campplus",
        "campplus_cn_common.bin",
        config_filename=None
    )
    campplus_model = CAMPPlus(feat_dim=80, embedding_size=192)
    campplus_model.load_state_dict(torch.load(campplus_ckpt_path, map_location="cpu"))
    campplus_model.eval()
    campplus_model.to(device)

# A small projection to map ECAPA embedding dim -> expected style dim
# We build it lazily at first inference once we know ECAPA dim.
style_proj = None
STYLE_DIM_EXPECTED = 192  # CAMPPlus embedding_size used originally in this app

# =========================================================
# Vocoder (BigVGAN)
# =========================================================
from modules.bigvgan import bigvgan

bigvgan_model = bigvgan.BigVGAN.from_pretrained(
    'nvidia/bigvgan_v2_22khz_80band_256x',
    use_cuda_kernel=False
)
bigvgan_model.remove_weight_norm()
bigvgan_model = bigvgan_model.eval().to(device)

# =========================================================
# Codec (FAcodec)
# =========================================================
ckpt_path, config_path = load_custom_model_from_hf("Plachta/FAcodec", 'pytorch_model.bin', 'config.yml')
codec_config = yaml.safe_load(open(config_path))
codec_model_params = recursive_munch(codec_config['model_params'])
codec_encoder = build_model(codec_model_params, stage="codec")

ckpt_params = torch.load(ckpt_path, map_location="cpu")
for key in codec_encoder:
    codec_encoder[key].load_state_dict(ckpt_params[key], strict=False)
_ = [codec_encoder[key].eval() for key in codec_encoder]
_ = [codec_encoder[key].to(device) for key in codec_encoder]

# =========================================================
# Whisper encoder (content)
# =========================================================
from transformers import AutoFeatureExtractor, WhisperModel

whisper_name = (
    model_params.speech_tokenizer.whisper_name
    if hasattr(model_params.speech_tokenizer, 'whisper_name')
    else "openai/whisper-small"
)
whisper_model = WhisperModel.from_pretrained(whisper_name, torch_dtype=torch.float16).to(device)
del whisper_model.decoder
whisper_feature_extractor = AutoFeatureExtractor.from_pretrained(whisper_name)

# =========================================================
# Mel Spectrogram
# =========================================================
mel_fn_args = {
    "n_fft": config['preprocess_params']['spect_params']['n_fft'],
    "win_size": config['preprocess_params']['spect_params']['win_length'],
    "hop_size": config['preprocess_params']['spect_params']['hop_length'],
    "num_mels": config['preprocess_params']['spect_params']['n_mels'],
    "sampling_rate": sr,
    "fmin": 0,
    "fmax": None,
    "center": False
}
from modules.audio import mel_spectrogram
to_mel = lambda x: mel_spectrogram(x, **mel_fn_args)

# =========================================================
# Load Seed-VC DiT model (f0 conditioned)
# =========================================================
dit_checkpoint_path, dit_config_path = load_custom_model_from_hf(
    "Plachta/Seed-VC",
    "DiT_seed_v2_uvit_whisper_base_f0_44k_bigvgan_pruned_ft_ema.pth",
    "config_dit_mel_seed_uvit_whisper_base_f0_44k.yml"
)

config_f0 = yaml.safe_load(open(dit_config_path, 'r'))
model_params_f0 = recursive_munch(config_f0['model_params'])
model_f0 = build_model(model_params_f0, stage='DiT')
hop_length_f0 = config_f0['preprocess_params']['spect_params']['hop_length']
sr_f0 = config_f0['preprocess_params']['sr']

model_f0, _, _, _ = load_checkpoint(
    model_f0, None, dit_checkpoint_path,
    load_only_params=True, ignore_modules=[],
    is_distributed=False
)
for key in model_f0:
    model_f0[key].eval()
    model_f0[key].to(device)

model_f0.cfm.estimator.setup_caches(max_batch_size=1, max_seq_length=8192)

# F0 extractor
from modules.rmvpe import RMVPE

model_path = load_custom_model_from_hf("lj1995/VoiceConversionWebUI", "rmvpe.pt", None)
rmvpe = RMVPE(model_path, is_half=False, device=device)

mel_fn_args_f0 = {
    "n_fft": config_f0['preprocess_params']['spect_params']['n_fft'],
    "win_size": config_f0['preprocess_params']['spect_params']['win_length'],
    "hop_size": config_f0['preprocess_params']['spect_params']['hop_length'],
    "num_mels": config_f0['preprocess_params']['spect_params']['n_mels'],
    "sampling_rate": sr_f0,
    "fmin": 0,
    "fmax": None,
    "center": False
}
to_mel_f0 = lambda x: mel_spectrogram(x, **mel_fn_args_f0)

bigvgan_44k_model = bigvgan.BigVGAN.from_pretrained(
    'nvidia/bigvgan_v2_44khz_128band_512x',
    use_cuda_kernel=False
)
bigvgan_44k_model.remove_weight_norm()
bigvgan_44k_model = bigvgan_44k_model.eval().to(device)

# =========================================================
# Helpers
# =========================================================
def adjust_f0_semitones(f0_sequence, n_semitones):
    factor = 2 ** (n_semitones / 12)
    return f0_sequence * factor

def crossfade(chunk1, chunk2, overlap):
    fade_out = np.cos(np.linspace(0, np.pi / 2, overlap)) ** 2
    fade_in = np.cos(np.linspace(np.pi / 2, 0, overlap)) ** 2
    chunk2[:overlap] = chunk2[:overlap] * fade_in + chunk1[-overlap:] * fade_out
    return chunk2

# Streaming and chunk params
bitrate = "320k"
overlap_frame_len = 16

def get_style_embedding(ref_waves_16k: torch.Tensor) -> torch.Tensor:
    """
    ref_waves_16k: (B, T) float tensor @ 16k
    returns: style2 (B, STYLE_DIM_EXPECTED)
    """
    global style_proj

    if USE_ECAPA and spk_encoder is not None:
        with torch.no_grad():
            # SpeechBrain ECAPA returns (B, 1, D) or (B, D) depending on version
            emb = spk_encoder.encode_batch(ref_waves_16k)
            if emb.dim() == 3:
                emb = emb.squeeze(1)  # (B, D)
            style2 = emb.to(device)

        # Project to expected style dim if needed
        if style2.size(-1) != STYLE_DIM_EXPECTED:
            if style_proj is None:
                style_proj = nn.Linear(style2.size(-1), STYLE_DIM_EXPECTED).to(device)
                style_proj.eval()
            with torch.no_grad():
                style2 = style_proj(style2)
        return style2

    # Fallback: CAMPPlus
    feat2 = torchaudio.compliance.kaldi.fbank(
        ref_waves_16k,
        num_mel_bins=80,
        dither=0,
        sample_frequency=16000
    )
    feat2 = feat2 - feat2.mean(dim=0, keepdim=True)
    style2 = campplus_model(feat2.unsqueeze(0))
    return style2

# =========================================================
# Voice Conversion
# =========================================================
@spaces.GPU
@torch.no_grad()
@torch.inference_mode()
def voice_conversion(source, target, diffusion_steps, length_adjust, inference_cfg_rate,
                     f0_condition, auto_f0_adjust, pitch_shift):

    inference_module = model if not f0_condition else model_f0
    mel_fn = to_mel if not f0_condition else to_mel_f0
    bigvgan_fn = bigvgan_model if not f0_condition else bigvgan_44k_model
    sr_local = 22050 if not f0_condition else 44100
    hop_local = 256 if not f0_condition else 512

    max_context_window = sr_local // hop_local * 30
    overlap_wave_len = overlap_frame_len * hop_local

    # Load audio
    source_audio = librosa.load(source, sr=sr_local)[0]
    ref_audio = librosa.load(target, sr=sr_local)[0]

    source_audio = torch.tensor(source_audio).unsqueeze(0).float().to(device)
    ref_audio = torch.tensor(ref_audio[:sr_local * 25]).unsqueeze(0).float().to(device)

    # Resample for whisper and speaker embedding
    ref_waves_16k = torchaudio.functional.resample(ref_audio, sr_local, 16000)
    converted_waves_16k = torchaudio.functional.resample(source_audio, sr_local, 16000)

    # Whisper content encoding (S_alt)
    if converted_waves_16k.size(-1) <= 16000 * 30:
        alt_inputs = whisper_feature_extractor(
            [converted_waves_16k.squeeze(0).cpu().numpy()],
            return_tensors="pt",
            return_attention_mask=True,
            sampling_rate=16000
        )
        alt_input_features = whisper_model._mask_input_features(
            alt_inputs.input_features, attention_mask=alt_inputs.attention_mask
        ).to(device)

        alt_outputs = whisper_model.encoder(
            alt_input_features.to(whisper_model.encoder.dtype),
            head_mask=None,
            output_attentions=False,
            output_hidden_states=False,
            return_dict=True,
        )
        S_alt = alt_outputs.last_hidden_state.to(torch.float32)
        S_alt = S_alt[:, :converted_waves_16k.size(-1) // 320 + 1]
    else:
        overlapping_time = 5
        S_alt_list = []
        buffer = None
        traversed_time = 0
        while traversed_time < converted_waves_16k.size(-1):
            if buffer is None:
                chunk = converted_waves_16k[:, traversed_time:traversed_time + 16000 * 30]
            else:
                chunk = torch.cat(
                    [buffer, converted_waves_16k[:, traversed_time:traversed_time + 16000 * (30 - overlapping_time)]],
                    dim=-1
                )
            alt_inputs = whisper_feature_extractor(
                [chunk.squeeze(0).cpu().numpy()],
                return_tensors="pt",
                return_attention_mask=True,
                sampling_rate=16000
            )
            alt_input_features = whisper_model._mask_input_features(
                alt_inputs.input_features, attention_mask=alt_inputs.attention_mask
            ).to(device)

            alt_outputs = whisper_model.encoder(
                alt_input_features.to(whisper_model.encoder.dtype),
                head_mask=None,
                output_attentions=False,
                output_hidden_states=False,
                return_dict=True,
            )
            S_alt_chunk = alt_outputs.last_hidden_state.to(torch.float32)
            S_alt_chunk = S_alt_chunk[:, :chunk.size(-1) // 320 + 1]
            if traversed_time == 0:
                S_alt_list.append(S_alt_chunk)
            else:
                S_alt_list.append(S_alt_chunk[:, 50 * overlapping_time:])
            buffer = chunk[:, -16000 * overlapping_time:]
            traversed_time += 30 * 16000 if traversed_time == 0 else chunk.size(-1) - 16000 * overlapping_time

        S_alt = torch.cat(S_alt_list, dim=1)

    # Whisper prompt (S_ori)
    ori_waves_16k = torchaudio.functional.resample(ref_audio, sr_local, 16000)
    ori_inputs = whisper_feature_extractor(
        [ori_waves_16k.squeeze(0).cpu().numpy()],
        return_tensors="pt",
        return_attention_mask=True
    )
    ori_input_features = whisper_model._mask_input_features(
        ori_inputs.input_features, attention_mask=ori_inputs.attention_mask
    ).to(device)

    ori_outputs = whisper_model.encoder(
        ori_input_features.to(whisper_model.encoder.dtype),
        head_mask=None,
        output_attentions=False,
        output_hidden_states=False,
        return_dict=True,
    )
    S_ori = ori_outputs.last_hidden_state.to(torch.float32)
    S_ori = S_ori[:, :ori_waves_16k.size(-1) // 320 + 1]

    mel = mel_fn(source_audio.to(device).float())
    mel2 = mel_fn(ref_audio.to(device).float())

    target_lengths = torch.LongTensor([int(mel.size(2) * length_adjust)]).to(mel.device)
    target2_lengths = torch.LongTensor([mel2.size(2)]).to(mel2.device)

    # Speaker embedding (ECAPA or fallback)
    style2 = get_style_embedding(ref_waves_16k)

    # f0 handling
    if f0_condition:
        F0_ori = rmvpe.infer_from_audio(ori_waves_16k[0], thred=0.5)
        F0_alt = rmvpe.infer_from_audio(converted_waves_16k[0], thred=0.5)

        F0_ori = torch.from_numpy(F0_ori).to(device)[None]
        F0_alt = torch.from_numpy(F0_alt).to(device)[None]

        voiced_F0_ori = F0_ori[F0_ori > 1]
        voiced_F0_alt = F0_alt[F0_alt > 1]

        log_f0_alt = torch.log(F0_alt + 1e-5)
        voiced_log_f0_ori = torch.log(voiced_F0_ori + 1e-5)
        voiced_log_f0_alt = torch.log(voiced_F0_alt + 1e-5)
        median_log_f0_ori = torch.median(voiced_log_f0_ori)
        median_log_f0_alt = torch.median(voiced_log_f0_alt)

        shifted_log_f0_alt = log_f0_alt.clone()
        if auto_f0_adjust and voiced_F0_alt.numel() > 0 and voiced_F0_ori.numel() > 0:
            shifted_log_f0_alt[F0_alt > 1] = log_f0_alt[F0_alt > 1] - median_log_f0_alt + median_log_f0_ori

        shifted_f0_alt = torch.exp(shifted_log_f0_alt)
        if pitch_shift != 0:
            shifted_f0_alt[F0_alt > 1] = adjust_f0_semitones(shifted_f0_alt[F0_alt > 1], pitch_shift)
    else:
        F0_ori = None
        shifted_f0_alt = None

    # Length regulation
    cond, _, _, _, _ = inference_module.length_regulator(
        S_alt, ylens=target_lengths, n_quantizers=3, f0=shifted_f0_alt
    )
    prompt_condition, _, _, _, _ = inference_module.length_regulator(
        S_ori, ylens=target2_lengths, n_quantizers=3, f0=F0_ori
    )

    max_source_window = max_context_window - mel2.size(2)

    processed_frames = 0
    generated_wave_chunks = []
    previous_chunk = None

    while processed_frames < cond.size(1):
        chunk_cond = cond[:, processed_frames:processed_frames + max_source_window]
        is_last_chunk = processed_frames + max_source_window >= cond.size(1)

        cat_condition = torch.cat([prompt_condition, chunk_cond], dim=1)

        with torch.autocast(device_type='cuda', dtype=torch.float16) if device.type == "cuda" else torch.no_grad():
            vc_target = inference_module.cfm.inference(
                cat_condition,
                torch.LongTensor([cat_condition.size(1)]).to(mel2.device),
                mel2, style2, None, diffusion_steps,
                inference_cfg_rate=inference_cfg_rate
            )
            vc_target = vc_target[:, :, mel2.size(-1):]

        vc_wave = bigvgan_fn(vc_target.float())[0]

        if processed_frames == 0:
            if is_last_chunk:
                output_wave = vc_wave[0].cpu().numpy()
                generated_wave_chunks.append(output_wave)
                output_i16 = (output_wave * 32768.0).astype(np.int16)

                mp3_bytes = AudioSegment(
                    output_i16.tobytes(),
                    frame_rate=sr_local,
                    sample_width=output_i16.dtype.itemsize,
                    channels=1
                ).export(format="mp3", bitrate=bitrate).read()
                yield mp3_bytes, (sr_local, np.concatenate(generated_wave_chunks))
                break

            output_wave = vc_wave[0, :-overlap_wave_len].cpu().numpy()
            generated_wave_chunks.append(output_wave)
            previous_chunk = vc_wave[0, -overlap_wave_len:]
            processed_frames += vc_target.size(2) - overlap_frame_len

            output_i16 = (output_wave * 32768.0).astype(np.int16)
            mp3_bytes = AudioSegment(
                output_i16.tobytes(),
                frame_rate=sr_local,
                sample_width=output_i16.dtype.itemsize,
                channels=1
            ).export(format="mp3", bitrate=bitrate).read()
            yield mp3_bytes, None

        elif is_last_chunk:
            output_wave = crossfade(previous_chunk.cpu().numpy(), vc_wave[0].cpu().numpy(), overlap_wave_len)
            generated_wave_chunks.append(output_wave)
            processed_frames += vc_target.size(2) - overlap_frame_len

            output_i16 = (output_wave * 32768.0).astype(np.int16)
            mp3_bytes = AudioSegment(
                output_i16.tobytes(),
                frame_rate=sr_local,
                sample_width=output_i16.dtype.itemsize,
                channels=1
            ).export(format="mp3", bitrate=bitrate).read()
            yield mp3_bytes, (sr_local, np.concatenate(generated_wave_chunks))
            break

        else:
            output_wave = crossfade(previous_chunk.cpu().numpy(), vc_wave[0, :-overlap_wave_len].cpu().numpy(), overlap_wave_len)
            generated_wave_chunks.append(output_wave)
            previous_chunk = vc_wave[0, -overlap_wave_len:]
            processed_frames += vc_target.size(2) - overlap_frame_len

            output_i16 = (output_wave * 32768.0).astype(np.int16)
            mp3_bytes = AudioSegment(
                output_i16.tobytes(),
                frame_rate=sr_local,
                sample_width=output_i16.dtype.itemsize,
                channels=1
            ).export(format="mp3", bitrate=bitrate).read()
            yield mp3_bytes, None

# =========================================================
# Gradio UI
# =========================================================
if __name__ == "__main__":
    description = (
        "State-of-the-Art zero-shot voice conversion/singing voice conversion. "
        "For local deployment please check GitHub repository for details and updates.<br>"
        "Note: reference audio will be clipped to 25s if longer.<br>"
        "If total duration exceeds 30s, source audio will be processed in chunks.<br>"
        "<br>"
        "Hindi tip: Use Hindi SOURCE + Hindi REFERENCE for best Hindi output. "
        "This app converts voice (audio→audio), it does not do text-to-speech."
    )

    inputs = [
        gr.Audio(type="filepath", label="Source Audio / 源音频"),
        gr.Audio(type="filepath", label="Reference Audio / 参考音频"),
        gr.Slider(minimum=1, maximum=200, value=25, step=1,
                  label="Diffusion Steps / 扩散步数",
                  info="25 by default, 50~100 for best quality / 默认为 25，50~100 为最佳质量"),
        gr.Slider(minimum=0.5, maximum=2.0, step=0.1, value=1.0,
                  label="Length Adjust / 长度调整",
                  info="<1.0 speed-up, >1.0 slow-down / <1.0 加速，>1.0 减速"),
        gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=0.7,
                  label="Inference CFG Rate", info="subtle influence / 有微小影响"),
        gr.Checkbox(label="Use F0 conditioned model / 启用F0输入", value=False,
                    info="Must set to true for singing voice conversion / 歌声转换时必须勾选"),
        gr.Checkbox(label="Auto F0 adjust / 自动F0调整", value=True,
                    info="Roughly adjust F0 to match target voice. Only when F0 model is used."),
        gr.Slider(label='Pitch shift / 音调变换', minimum=-24, maximum=24, step=1, value=0,
                  info="Semitones. Only when F0 model is used / 半音，仅F0模型生效"),
    ]

    examples = [
        ["examples/source/yae_0.wav", "examples/reference/dingzhen_0.wav", 25, 1.0, 0.7, False, True, 0],
        ["examples/source/jay_0.wav", "examples/reference/azuma_0.wav", 25, 1.0, 0.7, False, True, 0],
        ["examples/source/Wiz Khalifa,Charlie Puth - See You Again [vocals]_[cut_28sec].wav",
         "examples/reference/kobe_0.wav", 50, 1.0, 0.7, True, False, -6],
        ["examples/source/TECHNOPOLIS - 2085 [vocals]_[cut_14sec].wav",
         "examples/reference/trump_0.wav", 50, 1.0, 0.7, True, False, -12],
    ]

    outputs = [
        gr.Audio(label="Stream Output Audio / 流式输出", streaming=True, format='mp3'),
        gr.Audio(label="Full Output Audio / 完整输出", streaming=False, format='wav')
    ]

    gr.Interface(
        fn=voice_conversion,
        description=description,
        inputs=inputs,
        outputs=outputs,
        title="Seed Voice Conversion (ECAPA speaker embedding)",
        examples=examples,
        cache_examples=False
    ).launch()