geowatch.tasks.fusion.architectures.segmenter_decoder module

Adapted from https://github.com/rst

geowatch.tasks.fusion.architectures.segmenter_decoder.trunc_normal_(tensor: Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0) → Tensor

Fills the input Tensor with values drawn from a truncated normal distribution. The values are effectively drawn from the normal distribution \(\mathcal{N}(\text{mean}, \text{std}^2)\) with values outside \([a, b]\) redrawn until they are within the bounds. The method used for generating the random values works best when \(a \leq \text{mean} \leq b\). :Parameters: * tensor – an n-dimensional torch.Tensor

• mean – the mean of the normal distribution

• std – the standard deviation of the normal distribution

• a – the minimum cutoff value

• b – the maximum cutoff value

Examples

>>> w = torch.empty(3, 5)
>>> nn.init.trunc_normal_(w)

geowatch.tasks.fusion.architectures.segmenter_decoder.drop_path(x, drop_prob: float = 0.0, training: bool = False)

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

This is the same as the DropConnect impl I created for EfficientNet, etc networks, however, the original name is misleading as ‘Drop Connect’ is a different form of dropout in a separate paper… See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 … I’ve opted for changing the layer and argument names to ‘drop path’ rather than mix DropConnect as a layer name and use ‘survival rate’ as the argument.

class geowatch.tasks.fusion.architectures.segmenter_decoder.DropPath(drop_prob=None)

Bases: Module

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

forward(x)

geowatch.tasks.fusion.architectures.segmenter_decoder.init_weights(m)

class geowatch.tasks.fusion.architectures.segmenter_decoder.FeedForward(dim, hidden_dim, dropout, out_dim=None)

Bases: Module

property unwrapped

forward(x)

class geowatch.tasks.fusion.architectures.segmenter_decoder.Attention(dim, heads, dropout)

Bases: Module

property unwrapped

forward(x, mask=None)

class geowatch.tasks.fusion.architectures.segmenter_decoder.Block(dim, heads, mlp_dim, dropout, drop_path)

Bases: Module

forward(x, mask=None, return_attention=False)

class geowatch.tasks.fusion.architectures.segmenter_decoder.DecoderLinear(n_cls, patch_size, d_encoder)

Bases: Module

no_weight_decay()

forward(x, im_size)

class geowatch.tasks.fusion.architectures.segmenter_decoder.MaskTransformerDecoder(n_cls, d_model=192, n_layers=2, d_encoder='auto', n_heads='auto', d_ff='auto', drop_path_rate=0.0, dropout=0.1)

Bases: Module

This was originally the MaskTransformer and is being modified for the arbitrary output size design.

Example

>>> from geowatch.tasks.fusion.architectures.segmenter_decoder import *  # NOQA
>>> self = MaskTransformerDecoder(8, n_layers=8)
>>> # Batch size, number of tokens, and token dimension size
>>> B, N, D = 1, 10, self.d_model
>>> x = torch.rand(B, N, D)
>>> y = self.forward(x)
>>> assert y.shape[0:2] == x.shape[0:2], 'only in case where x is token size'
>>> assert y.shape[2] == self.n_cls, 'should always be true'

no_weight_decay()

forward(x)

get_attention_map(x, layer_id)