model package

Subpackages

• model.GCN package
  • Submodules
  • model.GCN.GCN module
  • model.GCN.Graph module
  • model.GCN.Graphsage module
  • model.GCN.non_local module
  • model.GCN.sparseGCN module
  • Module contents
• model.backbones package
  • Submodules
  • model.backbones.Attention module
  • model.backbones.resnet module
  • Module contents
• model.sync_batchnorm package
  • Submodules
  • model.sync_batchnorm.batchnorm module
  • model.sync_batchnorm.batchnorm_reimpl module
  • model.sync_batchnorm.comm module
  • model.sync_batchnorm.replicate module
  • model.sync_batchnorm.unittest module
  • Module contents

Submodules

model.BK module

model.make_mode_diffl module

model.make_model module

class model.make_model.Backbone(num_classes, cfg)

Bases: torch.nn.modules.module.Module

default use learning backbone network

set the configurations in /config

calling forward_gcn_cluster for forward function

other modfication

self.attention1 = GATConv(512,hidden,dropout=self.dropout,alpha=0.2,batch_size=8)

self.attention2 = GATConv(512, hidden, dropout=self.dropout, alpha=0.2,batch_size=8)

self.graphconv=GATConv(self.nheads*hidden,512,dropout=self.dropout,alpha=0.2,concat=False,batch_size=8)

forward(inputs, label=None, mode='train')

label is unused if self.cos_layer == 'no' Args:

inputs: input image batched label: labels batched mode: train or test

Returns:

classification probabilities batched.

forward_backbone_coattention(inputs, label=None)

cross attention fine-grained recognition, high-performance:88+ Args:

inputs: label:

Returns:

forward_gcn(inputs, label=None)

forward local connected GCNs

Args:

inputs: label:

Returns:

forward_gcn_cluster(inputs, label=None)

forward_pure(inputs, label=None)

forward_test(inputs, label=None)

load_param(trained_path)

reinit_param()

model.make_model.make_model(cfg, num_class)

model.make_model.weights_init_classifier(m)

model.make_model.weights_init_kaiming(m)

model.make_model_RGA module

model.make_model_vis module

Module contents