Checkpoint on SYNTHIA>Cityscapes

[kw01sg]

Is the pre-trained checkpoint for SYNTHIA to CityScapes available for download?

[lhoyer]

I have uploaded the checkpoint with the median performance for Synthia to Cityscapes to Google Drive: https://drive.google.com/file/d/1V9EpoTePjGq33B8MfombxEEcq9a2rBEt/view?usp=sharing

It achieves a mIoU for the 16 labeled classes in Synthia of 61.3. This differs from the mIoU in the log as it is divided by 19 instead of 16 there. The mIoU of 61.3 for this checkpoint slightly differs from the reported mIoU of 60.9 in Tab. 6 as the mean performance over three seeds is reported there.

[kw01sg]

Do you think you can provide the config file used to train this checkpoint? Would love to try and replicate the results. Have tried to make my own from the gta to cityscapes config file provided and I got an mIoU16 of 59.9, which is around 1.4 less (but still state of the art).

[lhoyer]

You can automatically generate the synthia training configs (and train with them) using:

python run_experimenty.py --exp 7

Please refer to https://github.com/lhoyer/DAFormer#training for more details.

[kw01sg]

  # -------------------------------------------------------------------------
  # Final DAFormer (Table 6)
  # -------------------------------------------------------------------------
  elif id == 7:
      seeds = [0, 1, 2]
      datasets = [
          # ('gta', 'cityscapes'),  # already run in exp 6
          ('synthia', 'cityscapes'),
      ]
      architecture, backbone = ('daformer_sepaspp', 'mitb5')
      uda = 'dacs_a999_fdthings'
      rcs_T = 0.01
      plcrop = True
      for (source, target), seed in \
              itertools.product(datasets, seeds):
          cfg = config_from_vars()
          cfgs.append(cfg)

From the source code, to obtain results from table 6 of the paper on synthia to cityscapes, can I confirm that i should be running experiment id 7?

[lhoyer]

Yes, that's right.

[kw01sg]

Got it. Thanks!