(self, attention, d_model, d_ff=None, moving_avg=25, dropout=0.1, activation="relu")
| 107 | Autoformer encoder layer with the progressive decomposition architecture |
| 108 | """ |
| 109 | def __init__(self, attention, d_model, d_ff=None, moving_avg=25, dropout=0.1, activation="relu"): |
| 110 | super(EncoderLayer, self).__init__() |
| 111 | d_ff = d_ff or 4 * d_model |
| 112 | self.attention = attention |
| 113 | self.conv1 = nn.Conv1d(in_channels=d_model, out_channels=d_ff, kernel_size=1, bias=False) |
| 114 | self.conv2 = nn.Conv1d(in_channels=d_ff, out_channels=d_model, kernel_size=1, bias=False) |
| 115 | |
| 116 | if isinstance(moving_avg, list): |
| 117 | self.decomp1 = series_decomp_multi(moving_avg) |
| 118 | self.decomp2 = series_decomp_multi(moving_avg) |
| 119 | else: |
| 120 | self.decomp1 = series_decomp(moving_avg) |
| 121 | self.decomp2 = series_decomp(moving_avg) |
| 122 | |
| 123 | self.dropout = nn.Dropout(dropout) |
| 124 | self.activation = F.relu if activation == "relu" else F.gelu |
| 125 | |
| 126 | def forward(self, x, attn_mask=None): |
| 127 | new_x, attn = self.attention( |
nothing calls this directly
no test coverage detected