Training

We use a system to automatically generate and train the configs:

python run_experiments.py --exp <ID>

More information about the available experiments and their assigned IDs, can be found in experiments.py. The generated configs will be stored in configs/generated/.

Warping Code Locations

Navigate to mmseg/models/uda/dacs.py

Init No-Warp for Target

self.warp_tgt = cfg.get('warp_tgt', True)

First No-Warp for Target

ema_logits = self.get_ema_model().encode_decode(
    target_img, target_img_metas,
    is_training=self.warp_tgt # NOTE: add non-warp for here
    )

Second No-Warp for Target

mix_losses = self.get_model().forward_train(
  mixed_img, img_metas, mixed_lbl, pseudo_weight, return_feat=True, 
  is_training=self.warp_tgt # NOTE: add no warp flag here!!!
  )

Navigate to mmseg/models/segmentors/encoder_decoder.py

Import Warping Modules

from ...transforms.fovea import build_grid_net, before_train_json, process_mmseg, read_seg_to_det

Build Warping Grid

self.grid_net = build_grid_net(warp_aug_lzu=warp_aug_lzu,
                                warp_fovea=warp_fovea,
                                warp_fovea_inst=warp_fovea_inst,
                                warp_fovea_mix=warp_fovea_mix,
                                warp_middle=warp_middle,
                                warp_scale=warp_scale,
                                warp_fovea_center=warp_fovea_center,
                                warp_fovea_inst_scale=warp_fovea_inst_scale,
                                warp_fovea_inst_scale_l2=warp_fovea_inst_scale_l2,
                                is_seg=is_seg,
                                bandwidth_scale=bandwidth_scale,
                                amplitude_scale=amplitude_scale,)

Warp Images and Unwarp Features

if (self.warp_aug_lzu is True) and (img_metas is not None):
    # print("self.warp_dataset is", self.warp_dataset)
    if any(src in img_metas[0]['filename'] for src in self.warp_dataset) and (is_training is True):
        # print(f"YES, RUNNING warping on {img_metas[0]['filename']}")
        x, img, img_metas = process_mmseg(img_metas,
                                            img,
                                            self.warp_aug_lzu,
                                            self.vanishing_point,
                                            self.grid_net,
                                            self.backbone,
                                            self.warp_debug,
                                            seg_to_det=self.seg_to_det,
                                            keep_grid=self.keep_grid
                                        )
        # print("images.shape", images.shape)

Checkpoints

Download checkpoints from [here]

Specific Configs

Click Here

Cityscapes -> DarkZurich

Experiments	Id
DAFormer	80
DAFormer + Sta. Prior	83
DAFormer + Geo. Prior	84
DAFormer + Ours	88

Cityscapes -> ACDC

Experiments	Id
DAFormer	90
DAFormer + Sta. Prior	93
DAFormer + Geo. Prior	94
DAFormer + Ours	98

Cityscapes -> Foggy Cityscapes

Experiments	Id
DAFormer	260
DAFormer + Ours	268

GTA -> Cityscapes

Experiments	Id
DAFormer	210
DAFormer + Ours	215

Synthia -> Cityscapes

Experiments	Id
DAFormer	220
DAFormer + Ours	228

ShenZheng2000/DAFormer