(
self, channels=None, kernel_size=3, gamma=2, beta=1, act_layer=None, gate_layer='sigmoid',
rd_ratio=1/8, rd_channels=None, rd_divisor=8, use_mlp=False)
| 58 | gate_layer: gating non-linearity to use |
| 59 | """ |
| 60 | def __init__( |
| 61 | self, channels=None, kernel_size=3, gamma=2, beta=1, act_layer=None, gate_layer='sigmoid', |
| 62 | rd_ratio=1/8, rd_channels=None, rd_divisor=8, use_mlp=False): |
| 63 | super(EcaModule, self).__init__() |
| 64 | if channels is not None: |
| 65 | t = int(abs(math.log(channels, 2) + beta) / gamma) |
| 66 | kernel_size = max(t if t % 2 else t + 1, 3) |
| 67 | assert kernel_size % 2 == 1 |
| 68 | padding = (kernel_size - 1) // 2 |
| 69 | if use_mlp: |
| 70 | # NOTE 'mlp' mode is a timm experiment, not in paper |
| 71 | assert channels is not None |
| 72 | if rd_channels is None: |
| 73 | rd_channels = make_divisible(channels * rd_ratio, divisor=rd_divisor) |
| 74 | act_layer = act_layer or nn.ReLU |
| 75 | self.conv = nn.Conv1d(1, rd_channels, kernel_size=1, padding=0, bias=True) |
| 76 | self.act = create_act_layer(act_layer) |
| 77 | self.conv2 = nn.Conv1d(rd_channels, 1, kernel_size=kernel_size, padding=padding, bias=True) |
| 78 | else: |
| 79 | self.conv = nn.Conv1d(1, 1, kernel_size=kernel_size, padding=padding, bias=False) |
| 80 | self.act = None |
| 81 | self.conv2 = None |
| 82 | self.gate = create_act_layer(gate_layer) |
| 83 | |
| 84 | def forward(self, x): |
| 85 | y = x.mean((2, 3)).view(x.shape[0], 1, -1) # view for 1d conv |
no test coverage detected