app.py

import torch
import torchaudio
import gradio as gr
from PIL import Image
import torchvision.transforms as transforms
import torchaudio.transforms as T
from torch import nn, optim
import torchvision.transforms as transforms
from torch.utils.data import Dataset, DataLoader
from PIL import Image
import os
import numpy as np

# Set device to 'cpu' or 'cuda' if available
device = torch.device('cpu')

# Parameters
sample_rate = 44100  # 44.1kHz stereo sounds
n_fft = 4096  # FFT size
hop_length = 2048  # Hop length for STFT
duration = 5  # Duration of the sound files (5 seconds)
n_channels = 2  # Stereo sound
output_time_frames = duration * sample_rate // hop_length  # Number of time frames in the spectrogram

stft_transform = T.Spectrogram(n_fft=n_fft, hop_length=hop_length, win_length=n_fft)

image_transform = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])  # Normalize to [-1, 1]
])

# Image Encoder (for the Generator)
class ImageEncoder(nn.Module):
    def __init__(self):
        super(ImageEncoder, self).__init__()
        self.encoder = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=4, stride=2, padding=1),
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.Conv2d(64, 128, kernel_size=4, stride=2, padding=1),
            nn.BatchNorm2d(128),
            nn.ReLU(),
            nn.Conv2d(128, 256, kernel_size=4, stride=2, padding=1),
            nn.BatchNorm2d(256),
            nn.ReLU(),
            nn.Conv2d(256, 512, kernel_size=4, stride=2, padding=1),
            nn.BatchNorm2d(512),
            nn.ReLU()
        )
        self.fc = nn.Linear(512 * 16 * 16, 512)

    def forward(self, x):
        x = self.encoder(x)
        x = x.view(x.size(0), -1)
        return self.fc(x)


# Sound Decoder (for the Generator)
class SoundDecoder(nn.Module):
    def __init__(self, output_time_frames):
        super(SoundDecoder, self).__init__()
        self.fc = nn.Linear(512, 512 * 8 * 8)

        self.decoder = nn.Sequential(
            nn.ConvTranspose2d(512, 256, kernel_size=4, stride=2, padding=1),
            nn.BatchNorm2d(256),
            nn.ReLU(),
            nn.ConvTranspose2d(256, 128, kernel_size=4, stride=2, padding=1),
            nn.BatchNorm2d(128),
            nn.ReLU(),
            nn.ConvTranspose2d(128, 64, kernel_size=4, stride=2, padding=1),
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.ConvTranspose2d(64, n_channels, kernel_size=4, stride=2, padding=1),
        )

        # Modify the upsample to exactly match the real spectrogram size (108 time frames)
        self.upsample = nn.Upsample(size=(n_fft // 2 + 1, 108), mode='bilinear', align_corners=True)

    def forward(self, x):
        x = self.fc(x)
        x = x.view(x.size(0), 512, 8, 8)
        x = self.decoder(x)
        x = self.upsample(x)
        # Debugging shape
        print(f'Generated spectrogram shape: {x.shape}')
        return x

# Generator model
class Generator(nn.Module):
    def __init__(self, output_time_frames):
        super(Generator, self).__init__()
        self.encoder = ImageEncoder()
        self.decoder = SoundDecoder(output_time_frames)

    def forward(self, img):
        # Debugging: Image encoder
        encoded_features = self.encoder(img)
        print(f"Encoded features shape (from Image Encoder): {encoded_features.shape}")

        # Debugging: Sound decoder
        generated_spectrogram = self.decoder(encoded_features)
        print(f"Generated spectrogram shape (from Sound Decoder): {generated_spectrogram.shape}")

        return generated_spectrogram


# Function to generate and save audio from a test image using the pre-trained GAN model
def test_model(generator, test_img_path, output_audio_path, device):
    # Load and preprocess test image
    test_img = Image.open(test_img_path).convert('RGB')
    test_img = image_transform(test_img).unsqueeze(0).to(device)  # Add batch dimension

    # Generate sound spectrogram from the image
    with torch.no_grad():  # Disable gradient calculation for inference
        generated_spectrogram = generator(test_img)

    # Debugging: Check generated spectrogram shape
    print(f"Generated spectrogram shape: {generated_spectrogram.shape}")

    # Convert the generated spectrogram to audio
    generated_audio = spectrogram_to_audio(generated_spectrogram.squeeze(0).cpu())  # Remove batch dimension

# Load the pre-trained GAN model
def load_gan_model(generator, model_path, device):
    generator.load_state_dict(torch.load(model_path, map_location=device))
    generator.eval()  # Set the model to evaluation mode
    return generator


def magnitude_to_complex_spectrogram(magnitude_spectrogram):
    # Clip values to avoid extreme values or potential invalid inputs
    magnitude_spectrogram = torch.clamp(magnitude_spectrogram, min=1e-10, max=1e5)

    zero_phase = torch.zeros_like(magnitude_spectrogram)
    complex_spectrogram = torch.stack([magnitude_spectrogram, zero_phase], dim=-1)

    # Check for NaNs in the complex spectrogram
    if torch.isnan(complex_spectrogram).any():
        raise ValueError("Complex spectrogram contains NaN values.")

    return complex_spectrogram


def spectrogram_to_audio(magnitude_spectrogram):
    # Perform inverse log scaling to undo any log scaling
    magnitude_spectrogram = torch.expm1(magnitude_spectrogram)

    # Convert magnitude-only spectrogram to complex format (real part and zero imaginary)
    zero_phase = torch.zeros_like(magnitude_spectrogram)
    complex_spectrogram = torch.stack([magnitude_spectrogram, zero_phase], dim=-1)

    # Inverse STFT to convert the spectrogram back to time-domain audio
    audio = torch.istft(complex_spectrogram, n_fft=n_fft, hop_length=hop_length)

    # Handle NaNs or Infs in the audio and replace them with zeros
    audio = torch.nan_to_num(audio, nan=0.0, posinf=0.0, neginf=0.0)

    # Normalize the audio to the range [-1, 1]
    if torch.max(torch.abs(audio)) != 0:
        audio = audio / torch.max(torch.abs(audio))

    # Clip the audio to ensure no values fall outside the range [-1, 1]
    audio = torch.clamp(audio, min=-1, max=1)

    # Convert to 16-bit PCM format by scaling and casting to int16
    audio = (audio * 32767).short()

    # Ensure the audio is in the valid range for int16 [-32768, 32767]
    audio = torch.clamp(audio, min=-32768, max=32767)

    # Convert the audio to a NumPy array of int16
    audio_numpy = audio.cpu().numpy().astype(np.int16)

    return audio_numpy




def generate_audio_from_image(image):
    test_img = image_transform(image).unsqueeze(0).to(device)  # Preprocess the image

    # Generate a sound spectrogram from the image using the loaded generator
    with torch.no_grad():
        generated_spectrogram = generator(test_img)

    # Convert the generated spectrogram to time-domain audio
    generated_audio_numpy = spectrogram_to_audio(generated_spectrogram.squeeze(0).cpu())

    # Return the sample rate and the audio in the correct format for Gradio
    return (sample_rate, generated_audio_numpy)



# Gradio Interface
def main():
    global generator  # Declare the generator object globally
    # Instantiate your Generator model
    generator = Generator(output_time_frames).to(device)

    # Load the pre-trained model
    model_path = './gan_model.pth'  # Change this path
    generator = load_gan_model(generator, model_path, device)

    # Gradio interface: allow users to upload an image and generate audio
    iface = gr.Interface(fn=generate_audio_from_image, inputs=gr.Image(type="pil"), outputs=gr.Audio(type="numpy", label="Generated Audio"))
    iface.launch()

if __name__ == "__main__":
    main()