tools/get_model.py

import torch
import hydra
from hydra import compose, initialize
from hydra.utils import instantiate
from lightning import LightningModule
from loguru import logger
from omegaconf import OmegaConf
import sys
sys.path.insert(0,'/workspace/user_code/kuachen/projects/v2s')

from fish_speech.models.v2s_tts.pretrain_model import V2S_TTS_Pretrain_Model
from fish_speech.models.v2s_tts.flow_matching_dit import ConditionalCFM
from fish_speech.models.v2s_tts.model.backbones.dit import DiT_Style
from fish_speech.models.v2s_tts.transformer.encoder import ConformerEncoder
from fish_speech.models.v2s_tts.style_bank import StyleBankExtractor
from omegaconf import DictConfig

class CFMParams:
    def __init__(self):
        self.sigma_min = 1e-06
        self.solver = "euler"
        self.t_scheduler = "cosine"
        self.training_cfg_rate = 0.2
        self.inference_cfg_rate = 0.7
        self.reg_loss_type = "l1"


def load_model(config_name, checkpoint_path, device="cpu"):
    hydra.core.global_hydra.GlobalHydra.instance().clear()
    with initialize(version_base="1.3", config_path="../fish_speech/configs"):
        cfg = compose(config_name=config_name)

    model: LightningModule = instantiate(cfg.model)
    state_dict = torch.load(
        checkpoint_path,
        map_location=model.device,
    )

    if "state_dict" in state_dict:
        state_dict = state_dict["state_dict"]

    model.load_state_dict(state_dict, strict=False)
    model.eval()
    model.to(device)
    logger.info("Restored model from checkpoint")

    return model

def get_pretrain_model(checkpoint_path):
    # 初始化 encoder
    encoder = ConformerEncoder(
        output_size=512,
        attention_heads=8,
        linear_units=2048,
        num_blocks=6,
        dropout_rate=0.1,
        positional_dropout_rate=0.1,
        attention_dropout_rate=0.1,
        normalize_before=True,
        input_layer='linear',
        pos_enc_layer_type='rel_pos_espnet',
        selfattention_layer_type='rel_selfattn',
        input_size=512,
        use_cnn_module=False,
        macaron_style=False
    )

    # 初始化 style_qformer
    style_qformer = StyleBankExtractor(
        dim_in=1024,
        n_layers=4,
        n_emb=32,
        d_model=64,
        nhead=4
    )

    # 初始化 estimator
    estimator = DiT_Style(
        dim=1024,
        depth=22,
        heads=16,
        ff_mult=2,
        conv_layers=4,
        mel_dim=80,
        style_dim=64
    )

    cfm_params = CFMParams()

    # 初始化 decoder
    decoder = ConditionalCFM(
        in_channels=160,
        n_spks=0,
        spk_emb_dim=80,
        cfm_params=cfm_params,
        estimator=estimator
    )

    # 初始化模型
    model = V2S_TTS_Pretrain_Model(
        input_size=512,
        output_size=80,
        output_type='mel',
        vocab_size=500,
        spk_dim=192,
        sll_checkpoint='checkpoints/wavlm_large.pt',
        output_layer=6,
        decoder=decoder,
        encoder=encoder,
        style_qformer=style_qformer
    )
    state_dict = torch.load(checkpoint_path)['state_dict']
    new_params = {}
    for k,v in state_dict.items():
        if k.startswith('generator'):
            new_k = k[10:]
            new_params[new_k] = v
    # Load the new parameters into the model
    model.load_state_dict(new_params, strict=True)
    model.eval()
    return model


def get_pretrain_model_32_dim32(checkpoint_path):
    # 初始化 encoder
    encoder = ConformerEncoder(
        output_size=512,
        attention_heads=8,
        linear_units=2048,
        num_blocks=6,
        dropout_rate=0.1,
        positional_dropout_rate=0.1,
        attention_dropout_rate=0.1,
        normalize_before=True,
        input_layer='linear',
        pos_enc_layer_type='rel_pos_espnet',
        selfattention_layer_type='rel_selfattn',
        input_size=512,
        use_cnn_module=False,
        macaron_style=False
    )

    # 初始化 style_qformer
    style_qformer = StyleBankExtractor(
        dim_in=1024,
        n_layers=4,
        n_emb=32,
        d_model=32,
        nhead=4
    )

    # 初始化 estimator
    estimator = DiT_Style(
        dim=1024,
        depth=22,
        heads=16,
        ff_mult=2,
        conv_layers=4,
        mel_dim=80,
        style_dim=32
    )

    cfm_params = CFMParams()

    # 初始化 decoder
    decoder = ConditionalCFM(
        in_channels=160,
        n_spks=0,
        spk_emb_dim=80,
        cfm_params=cfm_params,
        estimator=estimator
    )

    # 初始化模型
    model = V2S_TTS_Pretrain_Model(
        input_size=512,
        output_size=80,
        output_type='mel',
        vocab_size=500,
        spk_dim=192,
        sll_checkpoint='checkpoints/wavlm_large.pt',
        output_layer=6,
        decoder=decoder,
        encoder=encoder,
        style_qformer=style_qformer
    )
    state_dict = torch.load(checkpoint_path)['state_dict']
    new_params = {}
    for k,v in state_dict.items():
        if k.startswith('generator'):
            new_k = k[10:]
            new_params[new_k] = v
    # Load the new parameters into the model
    model.load_state_dict(new_params, strict=True)
    model.eval()
    return model

# ckpt_path = '/workspace/user_code/kuachen/projects/v2s/results/v2s_tts_pretrain_32_dim64/step_000336000.ckpt'
# config_name = 'v2s_tts_pretrain_32_dim64.yaml'
# model = load_model(config_name,ckpt_path)
# print(model.generator)
# model = get_pretrain_model(ckpt_path)

# state_dict = torch.load(ckpt_path)['state_dict']
# print(state_dict)