File size: 8,153 Bytes
052cf68
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
from prefigure.prefigure import get_all_args, push_wandb_config
import json
import os
import re
import torch
import torchaudio
# import pytorch_lightning as pl
import lightning as L
from lightning.pytorch.callbacks import Timer, ModelCheckpoint, BasePredictionWriter
from lightning.pytorch.callbacks import Callback
from lightning.pytorch.tuner import Tuner
from lightning.pytorch import seed_everything
import random
from datetime import datetime

from ThinkSound.data.datamodule import DataModule
from ThinkSound.models import create_model_from_config
from ThinkSound.models.utils import load_ckpt_state_dict, remove_weight_norm_from_model
from ThinkSound.training import create_training_wrapper_from_config, create_demo_callback_from_config
from ThinkSound.training.utils import copy_state_dict
from huggingface_hub import hf_hub_download

class ExceptionCallback(Callback):
    def on_exception(self, trainer, module, err):
        print(f'{type(err).__name__}: {err}')

class ModelConfigEmbedderCallback(Callback):
    def __init__(self, model_config):
        self.model_config = model_config

    def on_save_checkpoint(self, trainer, pl_module, checkpoint):
        checkpoint["model_config"] = self.model_config

class CustomWriter(BasePredictionWriter):

    def __init__(self, output_dir, write_interval='batch', batch_size=32):
        super().__init__(write_interval)
        self.output_dir = output_dir
        self.batch_size = batch_size

    def write_on_batch_end(self, trainer, pl_module, predictions, batch_indices, batch, batch_idx, dataloader_idx):

        audios = predictions
        ids = [item['id'] for item in batch[1]]
        current_date = datetime.now()

        formatted_date = current_date.strftime('%m%d')
        os.makedirs(os.path.join(self.output_dir, f'{formatted_date}_batch_size{self.batch_size}'),exist_ok=True)
        for audio, id in zip(audios, ids):
            save_path = os.path.join(self.output_dir, f'{formatted_date}_batch_size{self.batch_size}', f'{id}.wav')
            torchaudio.save(save_path, audio, 44100)

def main():

    args = get_all_args()


    # args.pretransform_ckpt_path = hf_hub_download(
    #     repo_id="liuhuadai/ThinkSound",
    #     filename="vae.ckpt"
    # )

    args.pretransform_ckpt_path = "./ckpts/vae.ckpt"


    seed = 10086

    # Set a different seed for each process if using SLURM
    if os.environ.get("SLURM_PROCID") is not None:
        seed += int(os.environ.get("SLURM_PROCID"))

    # random.seed(seed)
    # torch.manual_seed(seed)
    seed_everything(seed, workers=True)

    #Get JSON config from args.model_config
    with open(args.model_config) as f:
        model_config = json.load(f)

    with open(args.dataset_config) as f:
        dataset_config = json.load(f)
        
    for td in dataset_config["test_datasets"]:
        td["path"] = args.results_dir

    # train_dl = create_dataloader_from_config(
    #     dataset_config, 
    #     batch_size=args.batch_size, 
    #     num_workers=args.num_workers,
    #     sample_rate=model_config["sample_rate"],
    #     sample_size=model_config["sample_size"],
    #     audio_channels=model_config.get("audio_channels", 2),
    # )


    duration=(float)(args.duration_sec)
    
    dm = DataModule(
        dataset_config, 
        batch_size=args.batch_size,
        test_batch_size=args.test_batch_size,
        num_workers=args.num_workers,
        sample_rate=model_config["sample_rate"],
        sample_size=(float)(args.duration_sec) * model_config["sample_rate"],
        audio_channels=model_config.get("audio_channels", 2),
        latent_length=round(44100/64/32*duration),
    )
    
    model_config["sample_size"] = duration * model_config["sample_rate"]
    model_config["model"]["diffusion"]["config"]["sync_seq_len"] = 24*int(duration)
    model_config["model"]["diffusion"]["config"]["clip_seq_len"] = 8*int(duration)
    model_config["model"]["diffusion"]["config"]["latent_seq_len"] = round(44100/64/32*duration)

    model = create_model_from_config(model_config)

    ## speed by torch.compile
    if args.compile:
        model = torch.compile(model)
        
    if args.pretrained_ckpt_path:
        copy_state_dict(model, load_ckpt_state_dict(args.pretrained_ckpt_path,prefix='diffusion.')) # autoencoder.  diffusion.
    
    if args.remove_pretransform_weight_norm == "pre_load":
        remove_weight_norm_from_model(model.pretransform)
    # import ipdb
    # ipdb.set_trace()
    if args.pretransform_ckpt_path:
        load_vae_state = load_ckpt_state_dict(args.pretransform_ckpt_path, prefix='autoencoder.') 
        # new_state_dict = {k.replace("autoencoder.", ""): v for k, v in load_vae_state.items() if k.startswith("autoencoder.")}
        model.pretransform.load_state_dict(load_vae_state)
    
    # Remove weight_norm from the pretransform if specified
    if args.remove_pretransform_weight_norm == "post_load":
        remove_weight_norm_from_model(model.pretransform)

    training_wrapper = create_training_wrapper_from_config(model_config, model)

    # wandb_logger = L.pytorch.loggers.WandbLogger(project=args.name)
    # wandb_logger.watch(training_wrapper)

    exc_callback = ExceptionCallback()

    # if args.save_dir and isinstance(wandb_logger.experiment.id, str):
    #     checkpoint_dir = os.path.join(args.save_dir, wandb_logger.experiment.project, wandb_logger.experiment.id, "checkpoints")
    # else:
    #     checkpoint_dir = None

    # ckpt_callback = ModelCheckpoint(every_n_train_steps=args.checkpoint_every, dirpath=checkpoint_dir, monitor='val_loss', mode='min', save_top_k=10)
    save_model_config_callback = ModelConfigEmbedderCallback(model_config)
    audio_dir = args.results_dir
    pred_writer = CustomWriter(output_dir=audio_dir, write_interval="batch", batch_size=args.test_batch_size)
    timer = Timer(duration="00:15:00:00")
    demo_callback = create_demo_callback_from_config(model_config, demo_dl=dm)

    #Combine args and config dicts
    args_dict = vars(args)
    args_dict.update({"model_config": model_config})
    args_dict.update({"dataset_config": dataset_config})
    # push_wandb_config(wandb_logger, args_dict)

    #Set multi-GPU strategy if specified
    if args.strategy:
        if args.strategy == "deepspeed":
            from pytorch_lightning.strategies import DeepSpeedStrategy
            strategy = DeepSpeedStrategy(stage=2, 
                                        contiguous_gradients=True, 
                                        overlap_comm=True, 
                                        reduce_scatter=True, 
                                        reduce_bucket_size=5e8, 
                                        allgather_bucket_size=5e8,
                                        load_full_weights=True
                                        )
        else:
            strategy = args.strategy
    else:
        strategy = 'ddp_find_unused_parameters_true' if args.num_gpus > 1 else "auto" 

    trainer = L.Trainer(
        devices=args.num_gpus,
        accelerator="gpu",
        num_nodes = args.num_nodes,
        strategy=strategy,
        precision=args.precision,
        accumulate_grad_batches=args.accum_batches, 
        callbacks=[demo_callback, exc_callback, save_model_config_callback, timer, pred_writer],
        log_every_n_steps=1,
        max_epochs=1000,
        default_root_dir=args.save_dir,
        gradient_clip_val=args.gradient_clip_val,
        reload_dataloaders_every_n_epochs = 0,
        check_val_every_n_epoch=2,
    )

    

    # ckpt_path = hf_hub_download(
    #     repo_id="liuhuadai/ThinkSound",
    #     filename="thinksound.ckpt"
    # )
    ckpt_path = 'ckpts/thinksound.ckpt'



    current_date = datetime.now()
    formatted_date = current_date.strftime('%m%d')

    audio_dir = f'{formatted_date}_step68k_batch_size'+str(args.test_batch_size)
    metrics_path = os.path.join(args.ckpt_dir, 'audios',audio_dir,'cache',"output_metrics.json")
    # if os.path.exists(metrics_path): continue

    trainer.predict(training_wrapper, dm, return_predictions=False,ckpt_path=ckpt_path)

if __name__ == '__main__':
    main()