Beat-Transformer/code/ablation_models/tcn.py

import torch
import torch.nn as nn


class residual_block(nn.Module):
    def __init__(self, i, in_channels, num_filter, kernel_size, dropout):
        super(residual_block, self).__init__()
        self.res = nn.Conv1d(in_channels=in_channels, out_channels=num_filter, kernel_size=1, padding='same')
        self.conv1 = nn.Conv1d(in_channels=in_channels, out_channels=num_filter, kernel_size=kernel_size, dilation=i, padding='same')
        self.conv2 = nn.Conv1d(in_channels=in_channels, out_channels=num_filter, kernel_size=kernel_size, dilation=i*2, padding='same')
        self.elu = nn.ELU()
        self.spatial_dropout = nn.Dropout2d(p=dropout)
        self.conv_final = nn.Conv1d(in_channels=num_filter * 2, out_channels=num_filter, kernel_size=1, padding='same')

    def forward(self, x):
        #x: (B, F, T)
        x_res = self.res(x)
        x_1 = self.conv1(x)
        x_2 = self.conv2(x)
        x = torch.cat([x_1, x_2], dim=1)
        x = self.elu(x).unsqueeze(-1) #(B, F, T, 1)
        x = self.spatial_dropout(x).squeeze(-1) #(B, F, T)
        x = self.conv_final(x)
        return x + x_res, x


class TCN(nn.Module):
    def __init__(self, num_layers=11, dropout=.1, kernel_size=5, n_token=2):
        super(TCN, self).__init__()
        self.nlayers = num_layers

        self.conv1 = nn.Conv2d(in_channels=1, out_channels=20, kernel_size=3, stride=1, padding=(2, 0))
        self.maxpool1 = nn.MaxPool2d(kernel_size=(1, 3), stride=(1, 3))
        self.dropout1 = nn.Dropout(p=dropout)

        self.conv2 = nn.Conv2d(in_channels=20, out_channels=20, kernel_size=(1, 12), stride=1, padding=0)
        self.maxpool2 = nn.MaxPool2d(kernel_size=(1, 3), stride=(1, 3))
        self.dropout2 = nn.Dropout(p=dropout)

        self.conv3 = nn.Conv2d(in_channels=20, out_channels=20, kernel_size=3, stride=1, padding=0)
        self.maxpool3 = nn.MaxPool2d(kernel_size=(1, 3), stride=(1, 3))
        self.dropout3 = nn.Dropout(p=dropout)

        self.tcn_layers = nn.ModuleDict({})
        for layer in range(num_layers):
            self.tcn_layers[f'TCN_layer_{layer}'] = residual_block(i=2**layer, in_channels=20, num_filter=20, kernel_size=kernel_size, dropout=dropout)

        self.out_linear = nn.Linear(20, n_token)

        self.dropout_t = nn.Dropout(p=.5)
        self.out_linear_t = nn.Linear(20, 300)

        
    def forward(self, x):
        # x: spectrogram of size (B, T, mel_bin)
        x = x.unsqueeze(1)  #(B, 1, T, mel_bin)
        x = self.conv1(x)
        x = self.maxpool1(x)
        x = self.dropout1(x)

        x = self.conv2(x)
        x = self.maxpool2(x)
        x = self.dropout2(x)

        x = self.conv3(x)
        x = self.maxpool3(x)
        x = self.dropout3(x)    #(B, 20, T, 1)
        x = x.squeeze(-1) #(B, 20, T)

        t = []
        for layer in range(self.nlayers):
            x, skip = self.tcn_layers[f'TCN_layer_{layer}'](x)  #x: B, 20, T; skip: B, 20, T
            t.append(skip)

        x = torch.relu(x).transpose(-2, -1)
        x = self.out_linear(x)

        t = torch.stack(t, axis=-1).sum(dim=-1)
        t = torch.relu(t)
        t = self.dropout_t(t)
        t = t.mean(dim=-1) #(batch, 20)
        t = self.out_linear_t(t)

        return x, t


if __name__ == '__main__':
    from spectrogram_dataset import audioDataset
    from torch.utils.data import DataLoader

    DEVICE = 'cuda:2'
    model = TCN(num_layers=11, dropout=.15)
    model.to(DEVICE)
    model.eval()
    dataset = audioDataset(data_to_load=['smc'],
                            data_path = "/data1/zhaojw/dataset/madmom_data_100fps.npz", 
                            annotation_path = "/data1/zhaojw/dataset/beat_annotation.npz",
                            fps = 100,
                            sample_size = None,
                            downsample_size=1,
                            hop_size = 128,
                            fold = 0,
                            num_folds = 8)
    
    print(len(dataset.train_set), len(dataset.val_set), len(dataset.test_set))
    train_data = DataLoader(dataset.train_set, batch_size=1, shuffle=True)
    val_data = DataLoader(dataset.val_set, batch_size=1, shuffle=False)

    #for i, (key, instr, data, mask, beat, downbeat) in enumerate(val_data):
    for i, (key, data, beat, downbeat, tempo) in enumerate(val_data):
        print(key)

        data = data.float().cuda(DEVICE)#B, time, mel
        print(f'data: {data.shape}')
        print(f'beat: {beat.shape}')
        print(f'downbeat: {downbeat.shape}')
        print(f'tempo: {tempo.shape}')

        x, t = model(data)
        print(f'out: {x.shape}, {t.shape}')
        break