diodiogod-TTS-Audio-Suite/utils/voicefixer_bundled/tools/pytorch_util.py

import torch
import torch.nn as nn
import numpy as np


def check_cuda_availability(cuda):
    if cuda and not torch.cuda.is_available():
        raise RuntimeError("Error: You set cuda=True but no cuda device found.")


def try_tensor_cuda(tensor, cuda):
    if cuda and torch.cuda.is_available():
        return tensor.cuda()
    else:
        return tensor.cpu()


def to_log(input):
    assert torch.sum(input < 0) == 0, (
        str(input) + " has negative values counts " + str(torch.sum(input < 0))
    )
    return torch.log10(torch.clip(input, min=1e-8))


def from_log(input):
    input = torch.clip(input, min=-np.inf, max=5)
    return 10**input


def move_data_to_device(x, device):
    if "float" in str(x.dtype):
        x = torch.Tensor(x)
    elif "int" in str(x.dtype):
        x = torch.LongTensor(x)
    else:
        return x
    return x.to(device)


def tensor2numpy(tensor):
    if "cuda" in str(tensor.device):
        return tensor.detach().cpu().numpy()
    else:
        return tensor.detach().numpy()


def count_parameters(model):
    for p in model.parameters():
        if p.requires_grad:
            print(p.shape)
    return sum(p.numel() for p in model.parameters() if p.requires_grad)


def count_flops(model, audio_length):
    multiply_adds = False
    list_conv2d = []

    def conv2d_hook(self, input, output):
        batch_size, input_channels, input_height, input_width = input[0].size()
        output_channels, output_height, output_width = output[0].size()

        kernel_ops = (
            self.kernel_size[0]
            * self.kernel_size[1]
            * (self.in_channels / self.groups)
            * (2 if multiply_adds else 1)
        )
        bias_ops = 1 if self.bias is not None else 0

        params = output_channels * (kernel_ops + bias_ops)
        flops = batch_size * params * output_height * output_width

        list_conv2d.append(flops)

    list_conv1d = []

    def conv1d_hook(self, input, output):
        batch_size, input_channels, input_length = input[0].size()
        output_channels, output_length = output[0].size()

        kernel_ops = (
            self.kernel_size[0]
            * (self.in_channels / self.groups)
            * (2 if multiply_adds else 1)
        )
        bias_ops = 1 if self.bias is not None else 0

        params = output_channels * (kernel_ops + bias_ops)
        flops = batch_size * params * output_length

        list_conv1d.append(flops)

    list_linear = []

    def linear_hook(self, input, output):
        batch_size = input[0].size(0) if input[0].dim() == 2 else 1

        weight_ops = self.weight.nelement() * (2 if multiply_adds else 1)
        bias_ops = self.bias.nelement()

        flops = batch_size * (weight_ops + bias_ops)
        list_linear.append(flops)

    list_bn = []

    def bn_hook(self, input, output):
        list_bn.append(input[0].nelement())

    list_relu = []

    def relu_hook(self, input, output):
        list_relu.append(input[0].nelement())

    list_pooling2d = []

    def pooling2d_hook(self, input, output):
        batch_size, input_channels, input_height, input_width = input[0].size()
        output_channels, output_height, output_width = output[0].size()

        kernel_ops = self.kernel_size * self.kernel_size
        bias_ops = 0
        params = output_channels * (kernel_ops + bias_ops)
        flops = batch_size * params * output_height * output_width

        list_pooling2d.append(flops)

    list_pooling1d = []

    def pooling1d_hook(self, input, output):
        batch_size, input_channels, input_length = input[0].size()
        output_channels, output_length = output[0].size()

        kernel_ops = self.kernel_size
        bias_ops = 0
        params = output_channels * (kernel_ops + bias_ops)
        flops = batch_size * params * output_length

        list_pooling2d.append(flops)

    def foo(net):
        childrens = list(net.children())
        if not childrens:
            if isinstance(net, nn.Conv2d):
                net.register_forward_hook(conv2d_hook)
            elif isinstance(net, nn.ConvTranspose2d):
                net.register_forward_hook(conv2d_hook)
            elif isinstance(net, nn.Conv1d):
                net.register_forward_hook(conv1d_hook)
            elif isinstance(net, nn.Linear):
                net.register_forward_hook(linear_hook)
            elif isinstance(net, nn.BatchNorm2d) or isinstance(net, nn.BatchNorm1d):
                net.register_forward_hook(bn_hook)
            elif isinstance(net, nn.ReLU):
                net.register_forward_hook(relu_hook)
            elif isinstance(net, nn.AvgPool2d) or isinstance(net, nn.MaxPool2d):
                net.register_forward_hook(pooling2d_hook)
            elif isinstance(net, nn.AvgPool1d) or isinstance(net, nn.MaxPool1d):
                net.register_forward_hook(pooling1d_hook)
            else:
                print("Warning: flop of module {} is not counted!".format(net))
            return
        for c in childrens:
            foo(c)

    foo(model)

    input = torch.rand(1, audio_length, 2)
    out = model(input)

    total_flops = (
        sum(list_conv2d)
        + sum(list_conv1d)
        + sum(list_linear)
        + sum(list_bn)
        + sum(list_relu)
        + sum(list_pooling2d)
        + sum(list_pooling1d)
    )

    return total_flops