def __init__(self,
n_channels,
n_attention_heads):
"""Initializes a new TransformerBlock instance.
Args:
n_channels: The number of input and output channels.
n_attention_heads: The number of attention heads to use.
"""
super().__init__()
self._attn = pg_nn.MaskedAttention(
in_channels=n_channels,
embed_channels=n_channels,
out_channels=n_channels,
n_heads=n_attention_heads,
is_causal=False)
self._out = nn.Sequential(
nn.Conv2d(
in_channels=n_channels,
out_channels=4*n_channels,
kernel_size=1),
nn.GELU(),
nn.Conv2d (
in_channels=4*n_channels,
out_channels=n_channels,
kernel_size=1))
def __init__(
self,
n_channels,
input_img_channels=1,
n_residual_blocks=2,
attention_key_channels=4,
attention_value_channels=32,
):
"""Initializes a new PixelSnailBlock instance.
Args:
n_channels: Number of input and output channels.
input_img_channels: The number of channels in the original input_img. Used
for the positional encoding channels and the extra channels for the key
and value convolutions in the attention block.
n_residual_blocks: Number of residual blocks.
attention_key_channels: Number of channels (dims) for the attention key.
attention_value_channels: Number of channels (dims) for the attention value.
"""
super().__init__()
def conv(in_channels):
return nn.Conv2d(in_channels,
out_channels=n_channels,
kernel_size=1)
self._residual = nn.Sequential(
*[ResidualBlock(n_channels) for _ in range(n_residual_blocks)])
self._attention = pg_nn.MaskedAttention(
in_channels=n_channels + 2 * input_img_channels,
embed_channels=attention_key_channels,
out_channels=attention_value_channels,
is_causal=True,
extra_input_channels=input_img_channels,
)
self._residual_out = conv(n_channels)
self._attention_out = conv(attention_value_channels)
self._out = conv(n_channels)
