MCPcopy Create free account
hub / github.com/VCIP-RGBD/DFormer / forward_train

Method forward_train

mmseg/models/segmentors/encoder_decoder.py:115–143  ·  view source on GitHub ↗

Forward function for training. Args: img (Tensor): Input images. img_metas (list[dict]): List of image info dict where each dict has: 'img_shape', 'scale_factor', 'flip', and may also contain 'filename', 'ori_shape', 'pad_shape', and '

(self, img, img_metas, gt_semantic_seg)

Source from the content-addressed store, hash-verified

113 return seg_logit
114
115 def forward_train(self, img, img_metas, gt_semantic_seg):
116 """Forward function for training.
117
118 Args:
119 img (Tensor): Input images.
120 img_metas (list[dict]): List of image info dict where each dict
121 has: 'img_shape', 'scale_factor', 'flip', and may also contain
122 'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
123 For details on the values of these keys see
124 `mmseg/datasets/pipelines/formatting.py:Collect`.
125 gt_semantic_seg (Tensor): Semantic segmentation masks
126 used if the architecture supports semantic segmentation task.
127
128 Returns:
129 dict[str, Tensor]: a dictionary of loss components
130 """
131
132 x = self.extract_feat(img)
133
134 losses = dict()
135
136 loss_decode = self._decode_head_forward_train(x, img_metas, gt_semantic_seg)
137 losses.update(loss_decode)
138
139 if self.with_auxiliary_head:
140 loss_aux = self._auxiliary_head_forward_train(x, img_metas, gt_semantic_seg)
141 losses.update(loss_aux)
142
143 return losses
144
145 # TODO refactor
146 def slide_inference(self, img, img_meta, rescale):

Callers 2

Calls 4

extract_featMethod · 0.95
updateMethod · 0.80

Tested by

no test coverage detected