src/factory.py

import sys 
sys.path.append('../')

from typing import Optional
from copy import deepcopy

from transformers import AutoModelForCausalLM, AutoTokenizer, AutoModel, Wav2Vec2FeatureExtractor, WhisperFeatureExtractor, WhisperModel
# from .modeling_whisper import WhisperModel
from my_laion_clap.CLAP.src.laion_clap.clap_module.htsat import create_htsat_model

import torch
import torchaudio
import torchaudio.transforms as T
import numpy as np
from torch import nn
import torchvision.transforms
from contextlib import suppress

try:
    from .flamingo import Flamingo
    from .flamingo_lm import FlamingoLMMixin
    from .utils import extend_instance
except:
    from flamingo import Flamingo
    from flamingo_lm import FlamingoLMMixin
    from utils import extend_instance

def int16_to_float32(x):
    return (x / 32767.0).astype(np.float32)

def float32_to_int16(x):
    x = np.clip(x, a_min=-1., a_max=1.)
    return (x * 32767.).astype(np.int16)

def int16_to_float32_torch(x):
    return (x / 32767.0).type(torch.float32)

def float32_to_int16_torch(x):
    x = torch.clamp(x, min=-1., max=1.)
    return (x * 32767.).type(torch.int16)

class CLAPAudioCfp:
    model_type: str = "HTSAT"
    model_name: str = "large"
    sample_rate: int = 16000
    audio_length: int = 1024
    window_size: int = 1024
    hop_size: int = 160
    fmin: int = 50
    fmax: int = 14000
    class_num: int = 527
    mel_bins: int = 64
    clip_samples: int = 160000


class CLAP(nn.Module):
    def __init__(self, clap_config):
        super(CLAP, self).__init__()

        self.clap_config = clap_config

        self.method = clap_config["method"]
        device_id = f'cuda:{torch.cuda.current_device()}'

        if ('finetune' in clap_config) and clap_config['finetune']:
            self.finetune = True 
            print('Finetuning CLAP encoder as well!')
        else:
            self.finetune = False 

        audio_cfg = CLAPAudioCfp()
        enable_fusion = True
        fusion_type = "aff_2d"
        self.nvclap = create_htsat_model(audio_cfg, enable_fusion, fusion_type)
        clap_state_dict = torch.load(clap_config["checkpoint"], map_location = 'cpu')
        clap_state_dict_copy = clap_state_dict['state_dict'].copy()
        for key in list(clap_state_dict['state_dict'].keys()):
            if 'audio' in key:
                clap_state_dict_copy[key.replace('module.audio_branch.','')] = clap_state_dict_copy[key]
                del clap_state_dict_copy[key]
            else:
                del clap_state_dict_copy[key]
        self.nvclap.load_state_dict(clap_state_dict_copy, strict = False)
        self.nvclap = self.nvclap.to(device_id)
        
        for param in self.nvclap.parameters():
            param.requires_grad = self.finetune

        if self.finetune:
            self.nvclap.train()
        else:
            self.nvclap.eval()

        print('loaded NVCLAP model: {}'.format(clap_config["checkpoint"]))
                
    def get_mel(self, audio_data):

        # mel shape: (n_mels, T)
        mel_tf = torchaudio.transforms.MelSpectrogram(
            sample_rate=16000,
            n_fft=1024,
            win_length=1024,
            hop_length=160,
            center=True,
            pad_mode="reflect",
            power=2.0,
            norm=None,
            onesided=True,
            n_mels=64,
            f_min=50,
            f_max=14000
        ).to(audio_data.device)
        
        mel = mel_tf(audio_data)

        # we use log mel spectrogram as input
        mel = torchaudio.transforms.AmplitudeToDB(top_db=None)(mel)

        return mel.T  # (T, n_mels)

    def get_audio_features(self, sample, audio_data, max_len, data_truncating, data_filling, require_grad=False):

        grad_fn = suppress if require_grad else torch.no_grad
        with grad_fn():
            if len(audio_data) > max_len:
                if data_truncating == "rand_trunc":
                    longer = torch.tensor([True])
                elif data_truncating == "fusion":
                    # fusion
                    mel = self.get_mel(audio_data)
                    # split to three parts
                    chunk_frames = max_len // 160 + 1  # the +1 related to how the spectrogram is computed
                    total_frames = mel.shape[0]
                    if chunk_frames == total_frames:
                        # there is a corner case where the audio length is
                        # larger than max_len but smaller than max_len+hop_size.
                        # In this case, we just use the whole audio.
                        mel_fusion = torch.stack([mel, mel, mel, mel], dim=0)
                        sample["mel_fusion"] = mel_fusion
                        longer = torch.tensor([False])
                    else:
                        ranges = np.array_split(list(range(0, total_frames - chunk_frames + 1)), 3)
                        if len(ranges[1]) == 0:
                            # if the audio is too short, we just use the first chunk
                            ranges[1] = [0]
                        if len(ranges[2]) == 0:
                            # if the audio is too short, we just use the first chunk
                            ranges[2] = [0]
                        # randomly choose index for each part
                        idx_front = np.random.choice(ranges[0])
                        idx_middle = np.random.choice(ranges[1])
                        idx_back = np.random.choice(ranges[2])
                        # select mel
                        mel_chunk_front = mel[idx_front:idx_front + chunk_frames, :]
                        mel_chunk_middle = mel[idx_middle:idx_middle + chunk_frames, :]
                        mel_chunk_back = mel[idx_back:idx_back + chunk_frames, :]

                        # shrink the mel
                        mel_shrink = torchvision.transforms.Resize(size=[chunk_frames, 64])(mel[None])[0]
                        # logging.info(f"mel_shrink.shape: {mel_shrink.shape}")

                        # stack
                        mel_fusion = torch.stack([mel_shrink, mel_chunk_front, mel_chunk_middle, mel_chunk_back], dim=0)
                        sample["mel_fusion"] = mel_fusion
                        longer = torch.tensor([True])
                else:
                    raise NotImplementedError(
                        f"data_truncating {data_truncating} not implemented"
                    )
                # random crop to max_len (for compatibility)
                overflow = len(audio_data) - max_len
                idx = np.random.randint(0, overflow + 1)
                audio_data = audio_data[idx: idx + max_len]

            else:  # padding if too short
                if len(audio_data) < max_len:  # do nothing if equal
                    if data_filling == "repeatpad":
                        n_repeat = int(max_len / len(audio_data))
                        audio_data = audio_data.repeat(n_repeat)
                        # audio_data = audio_data.unsqueeze(0).unsqueeze(0).unsqueeze(0)
                        # audio_data = F.interpolate(audio_data,size=max_len,mode="bicubic")[0,0,0]
                        audio_data = F.pad(
                            audio_data,
                            (0, max_len - len(audio_data)),
                            mode="constant",
                            value=0,
                        )
                    elif data_filling == "pad":
                        audio_data = F.pad(
                            audio_data,
                            (0, max_len - len(audio_data)),
                            mode="constant",
                            value=0,
                        )
                    elif data_filling == "repeat":
                        n_repeat = int(max_len / len(audio_data))
                        audio_data = audio_data.repeat(n_repeat + 1)[:max_len]
                    else:
                        raise NotImplementedError(
                            f"data_filling {data_filling} not implemented"
                        )
                if data_truncating == 'fusion':
                    mel = self.get_mel(audio_data)
                    mel_fusion = torch.stack([mel, mel, mel, mel], dim=0)
                    sample["mel_fusion"] = mel_fusion
                longer = torch.tensor([False])

        sample["longer"] = longer
        sample["waveform"] = audio_data

        return sample


    def load_audio(self, clips):

        # waveform, sr = torchaudio.load(filename)
        # waveform = torchaudio.functional.resample(waveform, orig_freq=self.clap_config['sampling_rate'], new_freq=16000)
        processed_clips = []
        for clip in clips:
            audio_data = int16_to_float32_torch(float32_to_int16_torch(clip))
            sample = self.get_audio_features({}, audio_data, 160000, "fusion", "repeatpad")
            processed_clips.append(sample)

        waveforms = {}
        waveforms["mel_fusion"] = torch.stack([item["mel_fusion"] for item in processed_clips], dim=0)
        waveforms["longer"] = torch.stack([item["longer"] for item in processed_clips], dim=0)
        waveforms["waveform"] = torch.stack([item["waveform"] for item in processed_clips], dim=0)

        return waveforms


    def forward(self, audio_clips):
        
        # It will handle various segments, 1 audio will have various segments [B X n_segments X time]
        # expand batch dimension during inference
        if len(audio_clips.shape) == 2:
            audio_clips = audio_clips.unsqueeze(0)
        assert len(audio_clips.shape) == 3

        audio_embeds = []
        for audio_clip in audio_clips:
            audio = self.load_audio(audio_clip)
            audio_embed = self.nvclap(audio) #.reshape(-1, self.clap_config["audio_embed_dim"])
            audio_embeds.append(audio_embed)

        audio_embeds = torch.stack(audio_embeds, dim=0)
        # audio_embeds.requires_grad = self.finetune

        return audio_embeds
    
class Whisper(nn.Module):

    def __init__(self, whisper_config):
        super(Whisper, self).__init__()

        self.whisper_config = whisper_config

        self.method = self.whisper_config["method"]
        device_id = f'cuda:{torch.cuda.current_device()}'

        if ('finetune' in self.whisper_config) and self.whisper_config['finetune']:
            self.finetune = True 
            print('Finetuning Whisper encoder as well!')
        else:
            self.finetune = False 

        self.whisper = WhisperModel.from_pretrained(self.whisper_config['path']).encoder
        self.whisper = self.whisper.to(device_id)

        self.wav_processor = WhisperFeatureExtractor.from_pretrained(self.whisper_config['path'])
        
        for param in self.whisper.parameters():
            param.requires_grad = self.finetune

        if self.finetune:
            self.whisper.train()
        else:
            self.whisper.eval()

        print('loaded Whisper model: {}'.format(self.whisper_config['path']))

    def load_audio(self, clips):

        device_id = f'cuda:{torch.cuda.current_device()}'
        sample = self.wav_processor(clips.cpu().numpy(), sampling_rate=self.whisper_config['sampling_rate'], return_tensors="pt")["input_features"].to(device_id)

        return sample

    def forward(self, audio_clips):

        # It will handle various segments, 1 audio will have various segments [batch X n_segments X time]
        if len(audio_clips.shape) == 2:
            audio_clips = audio_clips.unsqueeze(0)
        assert len(audio_clips.shape) == 3

        audio_embeds = []
        for audio_clip in audio_clips:
            audio = self.load_audio(audio_clip)
            audio_embed = self.whisper(audio).last_hidden_state #.reshape(-1, self.whisper_config["audio_embed_dim"])
            audio_embeds.append(audio_embed)

        audio_embeds = torch.stack(audio_embeds, dim=0)
        # audio_embeds.requires_grad = self.finetune

        return audio_embeds

class MERT(nn.Module):

    def __init__(self, mert_config):
        super(MERT, self).__init__()

        self.mert_config = mert_config

        self.method = mert_config["method"]
        device_id = f'cuda:{torch.cuda.current_device()}'

        if ('finetune' in mert_config) and mert_config['finetune']:
            self.finetune = True 
            print('Finetuning MERT encoder as well!')
        else:
            self.finetune = False

        self.mert = AutoModel.from_pretrained(mert_config['path'], trust_remote_code=True)
        self.mert = self.mert.to(device_id)
        self.resampler = T.Resample(16000, mert_config['sampling_rate']).to(device_id)

        self.wav_processor = Wav2Vec2FeatureExtractor.from_pretrained(mert_config['path'],trust_remote_code=True)
        
        for param in self.mert.parameters():
            param.requires_grad = self.finetune

        if self.finetune:
            self.mert.train()
        else:
            self.mert.eval()

        print('loaded MERT model: {}'.format(mert_config['path']))

    def load_audio(self, clips):
        device_id = f'cuda:{torch.cuda.current_device()}'
        clips = self.resampler(clips.float()).float()
        sample = self.wav_processor(clips, sampling_rate=self.mert_config['sampling_rate'], return_tensors="pt")["input_values"]
        if len(sample.shape) == 1:
            sample = sample.unsqueeze(0)
        return sample.to(device_id)

    def forward(self, audio_clips):

        # It will handle various segments, 1 audio will have various segments [batch X n_segments X time]
        if len(audio_clips.shape) == 2:
            audio_clips = audio_clips.unsqueeze(0)
        assert len(audio_clips.shape) == 3

        audio_embeds = []
        for audio_clip in audio_clips:
            audio = self.load_audio(audio_clip).to(torch.bfloat16) # all processing happens in float
            if len(audio.shape) > 2:
                audio = audio.squeeze(0)
            audio_embed = self.mert(audio, output_hidden_states=True).last_hidden_state #.reshape(-1, self.mert_config["audio_embed_dim"])
            audio_embeds.append(audio_embed)

        audio_embeds = torch.stack(audio_embeds, dim=0)
        audio_embeds.requires_grad = self.finetune

        return audio_embeds



def create_model_and_transforms(
    clap_config: dict,
    lang_encoder_path: str,
    tokenizer_path: str,
    audio_transformer_kwargs: dict,
    cross_attn_every_n_layers: int = 1,
    use_local_files: bool = False,
    decoder_layers_attr_name: str = None,
    freeze_lm_embeddings: bool = False,
    unfreeze_full_lm: bool = False,
    cache_dir: Optional[str] = None,
    **flamingo_kwargs,
):
    clap = CLAP(clap_config)

    text_tokenizer = AutoTokenizer.from_pretrained(
        tokenizer_path,
        local_files_only=use_local_files,
        trust_remote_code=True,
        cache_dir=cache_dir,
    )
    text_tokenizer.add_special_tokens(
        {"additional_special_tokens": ["<audio>", "<|endofchunk|>", "<|PAD_TOKEN|>"]}
    )

    text_tokenizer.pad_token = None
    text_tokenizer.pad_token_id = None

    text_tokenizer.pad_token = "<|PAD_TOKEN|>"
    text_tokenizer.pad_token_id = text_tokenizer.encode("<|PAD_TOKEN|>")[-1]

    if text_tokenizer.sep_token is None:
        text_tokenizer.add_special_tokens({"sep_token": "<SEP>"})

    lang_encoder = AutoModelForCausalLM.from_pretrained(
        lang_encoder_path,
        local_files_only=use_local_files,
        trust_remote_code=True,
        cache_dir=cache_dir,
    )

    extend_instance(lang_encoder, FlamingoLMMixin)

    if decoder_layers_attr_name is None:
        decoder_layers_attr_name = _infer_decoder_layers_attr_name(lang_encoder)
    lang_encoder.set_decoder_layers_attr_name(decoder_layers_attr_name)
    lang_encoder.resize_token_embeddings(len(text_tokenizer))
    
    if ('finetune' in clap_config) and clap_config['finetune']:
        unfreeze_clap = True 
    else:
        unfreeze_clap = False 

    model = Flamingo(
        clap,
        unfreeze_clap,
        lang_encoder,
        text_tokenizer.encode("<|endofchunk|>")[-1],
        text_tokenizer.encode("<audio>")[-1],
        text_tokenizer.sep_token_id,
        clap_embed_dim = clap_config["audio_embed_dim"],
        audio_transformer_kwargs=audio_transformer_kwargs, 
        cross_attn_every_n_layers=cross_attn_every_n_layers,
        **flamingo_kwargs,
    )

    model.requires_grad_(False)
    assert sum(p.numel() for p in model.parameters() if p.requires_grad) == 0

    model.audio_transformer_clap.requires_grad_(True)
    
    model.lang_encoder.gated_cross_attn_layers_sound.requires_grad_(True)

    if not freeze_lm_embeddings:
        model.lang_encoder.get_input_embeddings().requires_grad_(True)
    
    if unfreeze_full_lm:
        model.lang_encoder.requires_grad_(True)

    if unfreeze_clap:
        model.clap.requires_grad_(True)


    print("Flamingo model initialized with {:,} trainable parameters (audio transformer has {:,}, LM has {:,})".format(
        sum(p.numel() for p in model.parameters() if p.requires_grad),
        sum(p.numel() for p in model.audio_transformer_clap.parameters() if p.requires_grad),
        sum(p.numel() for p in model.lang_encoder.parameters() if p.requires_grad),
    ))

    return model, text_tokenizer


def _infer_decoder_layers_attr_name(model):
    for k in __KNOWN_DECODER_LAYERS_ATTR_NAMES:
        if k.lower() in model.__class__.__name__.lower():
            return __KNOWN_DECODER_LAYERS_ATTR_NAMES[k]

    raise ValueError(
        f"We require the attribute name for the nn.ModuleList in the decoder storing the transformer block layers. Please supply this string manually."
    )


__KNOWN_DECODER_LAYERS_ATTR_NAMES = {
    "opt": "model.decoder.layers",
    "gptj": "transformer.h",
    "gpt-j": "transformer.h",
    "pythia": "gpt_neox.layers",
    "llama": "model.layers",
    "gptneoxforcausallm": "gpt_neox.layers",
    "mpt": "transformer.blocks",
    "mosaicgpt": "transformer.blocks",
    "qwen": "model.layers",
}


if __name__ == '__main__':
    import torch
    torch.set_printoptions(profile="full")  # only in debug mode
    import sys 
    sys.path.append('../')
    import os
    os.environ["TOKENIZERS_PARALLELISM"] = "false"
    import yaml
    import argparse

    parser = argparse.ArgumentParser()
    parser.add_argument('-c', '--config', type=str, default='../configs/config.yaml', help='yaml config path')
    args = parser.parse_args()

    config = yaml.load(open(args.config), Loader=yaml.FullLoader)

    data_config = config['data_config']
    model_config = config["model_config"]
    clap_config = config["clap_config"]

    model, tokenizer = create_model_and_transforms(
        **model_config,
        clap_config=clap_config,
        use_local_files=False,
        gradient_checkpointing=True,
        freeze_lm_embeddings=True
    )
    model = model.cuda()

    from data.data import AudioTextData, DataCollator
    from torch.utils.data import DataLoader

    batch_size = 8
    trainset = AudioTextData(
        **data_config, clap_config=clap_config, tokenizer=tokenizer,
        epoch=1, force_reblend=True
    )
    data_collator = DataCollator(tokenizer)
    trainloader = DataLoader(trainset, batch_size=batch_size, shuffle=True, collate_fn=data_collator, num_workers=4)

    for step, batch in enumerate(trainloader):
        audio_clips = batch["audio_clips"].cuda()
        audio_embed_mask = batch["audio_embed_mask"].cuda()
        input_ids = batch["input_ids"].cuda()
        attention_mask = batch["attention_mask"].cuda()

        print('batch {}:'.format(step+1), audio_clips.shape, audio_embed_mask.shape, input_ids.shape, attention_mask.shape)

        labels = input_ids.clone()

        labels[labels == tokenizer.pad_token_id] = -100
        labels[:, :2] = -100
        labels[labels == tokenizer.encode("<audio>")[-1]] = -100

        sep_locations = labels == tokenizer.sep_token_id
        endofchunk_token_id = tokenizer.encode("<|endofchunk|>")[-1]
        eoc_locations = labels == endofchunk_token_id

        if not all(sep_locations.sum(dim=1) == eoc_locations.sum(dim=1)):
            print("Warning: sep loc {} but eoc loc {}".format(sep_locations.sum(dim=1), eoc_locations.sum(dim=1)))
            
            for input_id in labels:
                input_id[input_id==-100] = tokenizer.encode("-")[-1]
                print(input_id, '\n', tokenizer.decode(input_id))

        for i in range(labels.shape[0]):
            shouldmask = True
            for j in range(labels.shape[1]):
                if shouldmask and (labels[i][j] != tokenizer.eos_token_id):
                    masked_value = -100
                else:
                    masked_value = labels[i][j]

                if labels[i][j] == tokenizer.sep_token_id:
                    shouldmask = False
                elif labels[i][j] == endofchunk_token_id:
                    shouldmask = True
                
                labels[i][j] = masked_value

            if labels[i][-1] not in [-100, tokenizer.eos_token_id, tokenizer.pad_token_id, endofchunk_token_id]:
                debug_masked_labels_in_the_end = []
                for j in range(labels.shape[1]-1, -1, -1):
                    if labels[i][j] not in [-100, tokenizer.eos_token_id, endofchunk_token_id]:
                        debug_masked_labels_in_the_end.insert(0, deepcopy(labels[i][j].item()))
                        labels[i][j] = -100
                    else:
                        break
                        
                print('hit max_token and masking ids from the end:', \
                    tokenizer.decode(torch.LongTensor(debug_masked_labels_in_the_end).to(labels.device))
                )

        if step == 50:
            break