| 6 | |
| 7 | |
| 8 | class CNN2(pl.LightningModule): |
| 9 | def __init__(self, num_features=40, num_classes=3, temp=249): |
| 10 | super().__init__() |
| 11 | |
| 12 | # Convolution 1 |
| 13 | self.conv1 = nn.Conv2d(in_channels=1, out_channels=16, kernel_size=(10, 42), padding=(0, 2)) |
| 14 | self.bn1 = nn.BatchNorm2d(16) |
| 15 | self.prelu1 = nn.PReLU() |
| 16 | |
| 17 | # Convolution 2 |
| 18 | self.conv2 = nn.Conv1d(in_channels=16, out_channels=16, kernel_size=(10,)) # 3 |
| 19 | self.bn2 = nn.BatchNorm1d(16) |
| 20 | self.prelu2 = nn.PReLU() |
| 21 | |
| 22 | # Convolution 3 |
| 23 | self.conv3 = nn.Conv1d(in_channels=16, out_channels=32, kernel_size=(8,)) # 1 |
| 24 | self.bn3 = nn.BatchNorm1d(32) |
| 25 | self.prelu3 = nn.PReLU() |
| 26 | |
| 27 | # Convolution 4 |
| 28 | self.conv4 = nn.Conv1d(in_channels=32, out_channels=32, kernel_size=(6,)) # 1 |
| 29 | self.bn4 = nn.BatchNorm1d(32) |
| 30 | self.prelu4 = nn.PReLU() |
| 31 | |
| 32 | # Convolution 5 |
| 33 | self.conv5 = nn.Conv1d(in_channels=32, out_channels=32, kernel_size=(4,)) # 1 |
| 34 | self.bn5 = nn.BatchNorm1d(32) |
| 35 | self.prelu5 = nn.PReLU() |
| 36 | |
| 37 | # Fully connected 1 |
| 38 | self.fc1 = nn.Linear(temp*32, 32) |
| 39 | self.prelu6 = nn.PReLU() |
| 40 | |
| 41 | # Fully connected 2 |
| 42 | self.fc2 = nn.Linear(32, num_classes) |
| 43 | |
| 44 | def forward(self, x): |
| 45 | # Convolution 1 |
| 46 | out = self.conv1(x) |
| 47 | # print('After convolution1:', out.shape) |
| 48 | |
| 49 | out = self.bn1(out) |
| 50 | # print('After bn1:', out.shape) |
| 51 | |
| 52 | out = self.prelu1(out) |
| 53 | out = out.reshape(out.shape[0], out.shape[1], -1) |
| 54 | # print('After prelu1:', out.shape) |
| 55 | |
| 56 | # Convolution 2 |
| 57 | out = self.conv2(out) |
| 58 | out = self.bn2(out) |
| 59 | out = self.prelu2(out) |
| 60 | # print('After convolution2, bn2, prelu2:', out.shape) |
| 61 | |
| 62 | # Convolution 3 |
| 63 | out = self.conv3(out) |
| 64 | out = self.bn3(out) |
| 65 | out = self.prelu3(out) |