Method __init__

model/transformer_utils.py:423–429

(self, input_res, in_channels, patch_size, width, layers, heads, window_size)

class TransformerEncoder(TransformerBase):
    def __init__(self, input_res, in_channels, patch_size, width, layers, heads, window_size):
        self.input_res = input_res
        self.patch_size = patch_size
        token_len = (self.input_res[0] // patch_size) * (self.input_res[1] // patch_size)
        super().__init__(width, layers, heads, window_size, token_len, ResAttBlock)
        self.conv = nn.Conv2d(in_channels=in_channels, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False)
        self.positional_encoding = SinusoidalPositionalEncoding(max_len=token_len, d_model=width)

    def forward(self, x, condition=None):
        _, v = x.shape[:2]