File size: 5,525 Bytes
d9a7330
 
 
7019cb6
d9a7330
 
 
eb39bd3
d9a7330
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c90e89d
d9a7330
 
eb39bd3
 
 
d9a7330
 
eb39bd3
d9a7330
 
 
 
 
 
 
eb39bd3
d9a7330
 
 
 
 
 
 
 
 
 
 
 
 
 
eb39bd3
d9a7330
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5d104f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eb39bd3
5d104f9
 
 
285840d
 
f340743
 
 
 
 
 
 
285840d
f340743
5d104f9
285840d
5d104f9
285840d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5d104f9
 
eb39bd3
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
import os
import torch
import random
import spaces
import numpy as np
import gradio as gr
import soundfile as sf
from accelerate import Accelerator
from transformers import T5Tokenizer, T5EncoderModel
from diffusers import DDIMScheduler
from src.models.conditioners import MaskDiT
from src.modules.autoencoder_wrapper import Autoencoder
from src.inference import inference
from src.utils import load_yaml_with_includes


# Load model and configs
def load_models(config_name, ckpt_path, vae_path, device):
    params = load_yaml_with_includes(config_name)

    # Load codec model
    autoencoder = Autoencoder(ckpt_path=vae_path,
                              model_type=params['autoencoder']['name'],
                              quantization_first=params['autoencoder']['q_first']).to(device)
    autoencoder.eval()

    # Load text encoder
    tokenizer = T5Tokenizer.from_pretrained(params['text_encoder']['model'])
    text_encoder = T5EncoderModel.from_pretrained(params['text_encoder']['model']).to(device)
    text_encoder.eval()

    # Load main U-Net model
    unet = MaskDiT(**params['model']).to(device)
    unet.load_state_dict(torch.load(ckpt_path, map_location='cpu')['model'])
    unet.eval()

    accelerator = Accelerator(mixed_precision="fp16")
    unet = accelerator.prepare(unet)

    # Load noise scheduler
    noise_scheduler = DDIMScheduler(**params['diff'])

    latents = torch.randn((1, 128, 128), device=device)
    noise = torch.randn_like(latents)
    timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (1,), device=device)
    _ = noise_scheduler.add_noise(latents, noise, timesteps)

    return autoencoder, unet, tokenizer, text_encoder, noise_scheduler, params


MAX_SEED = np.iinfo(np.int32).max

# Model and config paths
config_name = 'ckpts/ezaudio-xl.yml'
ckpt_path = 'ckpts/s3/ezaudio_s3_xl.pt'
vae_path = 'ckpts/vae/1m.pt'
save_path = 'output/'
os.makedirs(save_path, exist_ok=True)

device = 'cuda' if torch.cuda.is_available() else 'cpu'

autoencoder, unet, tokenizer, text_encoder, noise_scheduler, params = load_models(config_name, ckpt_path, vae_path,
                                                                                  device)


@spaces.GPU
def generate_audio(text, length,
                   guidance_scale, guidance_rescale, ddim_steps, eta,
                   random_seed, randomize_seed):
    neg_text = None
    length = length * params['autoencoder']['latent_sr']

    if randomize_seed:
        random_seed = random.randint(0, MAX_SEED)

    pred = inference(autoencoder, unet, None, None,
                     tokenizer, text_encoder,
                     params, noise_scheduler,
                     text, neg_text,
                     length,
                     guidance_scale, guidance_rescale,
                     ddim_steps, eta, random_seed,
                     device)

    pred = pred.cpu().numpy().squeeze(0).squeeze(0)
    # output_file = f"{save_path}/{text}.wav"
    # sf.write(output_file, pred, samplerate=params['autoencoder']['sr'])

    return params['autoencoder']['sr'], pred


# Examples (if needed for the demo)
examples = [
    "the sound of rain falling softly",
    "a dog barking in the distance",
    "light guitar music is playing",
]

# CSS styling (optional)
css = """
#col-container {
    margin: 0 auto;
    max-width: 1280px;
}
"""

# Gradio Blocks layout
with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown("""
        # EzAudio: High-quality Text-to-Audio Generator
        Generate audio from text using a diffusion transformer. Adjust advanced settings for more control.
        """)

    # Basic Input: Text prompt
    with gr.Row():
        text_input = gr.Textbox(
            label="Text Prompt",
            show_label=True,
            max_lines=2,
            placeholder="Enter your prompt",
            container=True,
            value="a dog barking in the distance",
            scale=4
        )
        # Run button
        run_button = gr.Button("Generate", scale=1)

    # Output Component
    result = gr.Audio(label="Result", type="numpy")

    # Advanced settings in an Accordion
    with gr.Accordion("Advanced Settings", open=False):
        # Audio Length
        length_input = gr.Slider(minimum=1, maximum=10, step=1, value=10, label="Audio Length (in seconds)")
        guidance_scale = gr.Slider(minimum=1.0, maximum=10, step=0.1, value=5.0, label="Guidance Scale")
        guidance_rescale = gr.Slider(minimum=0.0, maximum=1, step=0.05, value=0.75, label="Guidance Rescale")
        ddim_steps = gr.Slider(minimum=25, maximum=200, step=5, value=50, label="DDIM Steps")
        eta = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=1.0, label="Eta")
        seed = gr.Slider(minimum=0, maximum=100, step=1, value=0, label="Seed")
        randomize_seed = gr.Checkbox(label="Randomize Seed (Disable Seed)", value=True)

    # Examples block
    gr.Examples(
        examples=examples,
        inputs=[text_input]
    )

    # Define the trigger and input-output linking
    run_button.click(
        fn=generate_audio,
        inputs=[text_input, length_input, guidance_scale, guidance_rescale, ddim_steps, eta, seed, randomize_seed],
        outputs=[result]
    )
    text_input.submit(fn=generate_audio,
        inputs=[text_input, length_input, guidance_scale, guidance_rescale, ddim_steps, eta, seed, randomize_seed],
        outputs=[result]
    )

# Launch the Gradio demo
demo.launch()