Architecture

Backbone

Backbone modules.

class alonet.detr.backbone.Backbone(name, train_backbone, return_interm_layers, dilation, **kwargs)

Bases: alonet.detr.backbone.BackboneBase

ResNet backbone with frozen BatchNorm.

training: bool
class alonet.detr.backbone.BackboneBase(backbone, train_backbone, num_channels, return_interm_layers, aug_tensor_compatible=True)

Bases: torch.nn.modules.module.Module

Base class to define behavior of backbone

forward(frames, **kwargs)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool
class alonet.detr.backbone.FrozenBatchNorm2d(n)

Bases: torch.nn.modules.module.Module

BatchNorm2d where the batch statistics and the affine parameters are fixed.

Copy-paste from torchvision.misc.ops with added eps before rqsrt, without which any other models than torchvision.models.resnet[18,34,50,101] produce nans.

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool
class alonet.detr.backbone.Joiner(backbone, position_embedding)

Bases: torch.nn.modules.container.Sequential

A sequential wrapper for backbone and position embedding.

self.forward returns a tuple:

  • list of feature maps from backbone

  • list of position encoded feature maps

forward(frames, **kwargs)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool
alonet.detr.backbone.build_backbone()
alonet.detr.backbone.get_rank()
alonet.detr.backbone.get_world_size()
alonet.detr.backbone.is_dist_avail_and_initialized()
alonet.detr.backbone.is_main_process()

Transformer

DETR Transformer class.

Copy-paste from torch.nn.Transformer with modifications:

  • positional encodings are passed in MHattention

  • extra LN at the end of encoder is removed

  • decoder returns a stack of activations from all decoding layers

Then copy-past from detr official repository with modification to be usuable inside aloception along with deformable detr.

class alonet.detr.transformer.Transformer(d_model=512, nhead=8, encoder=None, decoder=None, decoder_layer=None, encoder_layer=None, num_encoder_layers=6, num_decoder_layers=6, dim_feedforward=2048, dropout=0.1, activation='relu', normalize_before=False, return_intermediate_dec=False)

Bases: torch.nn.modules.module.Module

forward(src, mask, query_embed, pos_embed, **kwargs)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool
class alonet.detr.transformer.TransformerDecoder(decoder_layer, num_layers, norm=None, return_intermediate=False)

Bases: torch.nn.modules.module.Module

decoder_forward(tgt, memory, tgt_mask=None, memory_mask=None, tgt_key_padding_mask=None, memory_key_padding_mask=None, pos=None, query_pos=None, **kwargs)
forward(tgt, memory, tgt_mask=None, memory_mask=None, tgt_key_padding_mask=None, memory_key_padding_mask=None, pos=None, query_pos=None, decoder_outputs=None, **kwargs)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

pre_process_tgt(tgt, query_pos, tgt_key_padding_mask, **kwargs)

Pre process decoder inputs

training: bool
class alonet.detr.transformer.TransformerDecoderLayer(d_model, n_heads=8, dim_feedforward=2048, dropout=0.1, activation='relu', normalize_before=False)

Bases: torch.nn.modules.module.Module

decoder_layer_forward(tgt, memory, tgt_mask=None, memory_mask=None, tgt_key_padding_mask=None, memory_key_padding_mask=None, pos=None, query_pos=None, **kwargs)
forward(tgt, memory, tgt_mask=None, memory_mask=None, tgt_key_padding_mask=None, memory_key_padding_mask=None, pos=None, query_pos=None, **kwargs)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

forward_post(tgt, memory, tgt_mask=None, memory_mask=None, tgt_key_padding_mask=None, memory_key_padding_mask=None, pos=None, query_pos=None)
forward_pre(tgt, memory, tgt_mask=None, memory_mask=None, tgt_key_padding_mask=None, memory_key_padding_mask=None, pos=None, query_pos=None)
pre_process_tgt(tgt, query_pos, tgt_key_padding_mask, **kwargs)

Pre process decoder inputs

training: bool
with_pos_embed(tensor, pos)
class alonet.detr.transformer.TransformerEncoder(encoder_layer, num_layers, norm=None)

Bases: torch.nn.modules.module.Module

forward(src, mask=None, src_key_padding_mask=None, pos=None, **kwargs)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool
class alonet.detr.transformer.TransformerEncoderLayer(d_model, nhead, dim_feedforward=2048, dropout=0.1, activation='relu', normalize_before=False)

Bases: torch.nn.modules.module.Module

forward(src, src_mask=None, src_key_padding_mask=None, pos=None, **kwargs)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

forward_post(src, src_mask=None, src_key_padding_mask=None, pos=None)
forward_pre(src, src_mask=None, src_key_padding_mask=None, pos=None)
training: bool
with_pos_embed(tensor, pos)
alonet.detr.transformer.build_transformer()