request for cityscapes related configs and pretrain weight

[Cloveryww]

Hi,
Thanks for your great work firstly and could you please release configs and pretrain weights of cityscapes dataset?
I have seen part of files about cityscapes, but still lack of some configs file.

[GuoleiSun]

Thanks for your interest. Initially, we did not release the configs and pre-trained weights for cityscapes because video data of cityscapes is extremely large (>300 GB). Nevertheless, we will release the files for cityscapes later.

[Cloveryww]

Hi，
I have some question about CFFM when running in cityscapes dataset.

1. When I want to train CFFM in cityscapes dataset, since training CFFM in cityscapes is a semi-supervised fasion as its sparse annotations, after I change the codes about dataset and loss compute(listed bellow), is any other important code needed adjustment?
   loss compute:
   

   VSS-CFFM/mmseg/models/decode_heads/decode_head.py

   Lines 752 to 760 in d93f8a0

   	if seg_logit.shape[1]==seg_label.shape[1]+1: # k+1
   	seg_logit_ori=seg_logit[:,:-1]
   	batch_size, num_clips, _, h ,w=seg_logit_ori.shape
   	seg_logit_ori=seg_logit_ori.reshape(batch_size*num_clips,-1,h,w)
   	seg_logit_lastframe=seg_logit[:,-1]
   	batch_size, num_clips, _, h ,w=seg_label.shape
   	seg_label_ori=seg_label.reshape(batch_size*num_clips,-1,h,w)
   	seg_label_lastframe=seg_label[:,-1]

   
   seg label in dataset:
   

   VSS-CFFM/mmseg/datasets/custom.py

   Lines 1060 to 1078 in d93f8a0

   	try:
   	img_anns=[]
   	for ii in dilation_used:
   	img_info_one={}
   	filename=img_info['filename']
   	seg_map=img_info['ann']['seg_map']
   	value_i_splits=filename.split('_')
   	im_name_new = "_".join(
   	value_i_splits[:-2] + [(str(int(value_i_splits[-2]) + ii)).rjust(6, "0")] + value_i_splits[-1:])
   	# value_i_splits=seg_map.split('_')
   	# seg_map_new = "_".join(
   	# value_i_splits[:-2] + [(str(int(value_i_splits[-2]) - ii)).rjust(6, "0")] + value_i_splits[-1:])
   	img_info_one['filename']=im_name_new
   	img_info_one['ann']=dict(seg_map=seg_map)
   	ann_info_one=img_info_one['ann']
   	img_anns.append([img_info_one, ann_info_one])
   	if not os.path.isfile(self.img_dir+'/'+im_name_new):
   	assert False

2. If the model run in cityscapes with a size of 512x1024, the params about core module of CFFM(such as bellow codes) is the same as params in VSPW? If not, how to adjust them? for example "input_resolution"?
   

   VSS-CFFM/mmseg/models/decode_heads/cffm_head.py

   Lines 80 to 106 in d93f8a0

   	self.decoder_focal=BasicLayer3d3(dim=embedding_dim,
   	input_resolution=(60,
   	60),
   	depth=depths,
   	num_heads=8,
   	window_size=7,
   	mlp_ratio=4.,
   	qkv_bias=True,
   	qk_scale=None,
   	drop=0.,
   	attn_drop=0.,
   	drop_path=0.,
   	norm_layer=nn.LayerNorm,
   	pool_method='fc',
   	downsample=None,
   	focal_level=2,
   	focal_window=5,
   	expand_size=3,
   	expand_layer="all",
   	use_conv_embed=False,
   	use_shift=False,
   	use_pre_norm=False,
   	use_checkpoint=False,
   	use_layerscale=False,
   	layerscale_value=1e-4,
   	focal_l_clips=[7,4,2],
   	focal_kernel_clips=[7,5,3])

3. Does the current version of codes not support efficient inference on video data(reach the fps reported in paper), or I did not find the corresponding codes?

Sincerely look forward to your reply, thank you!!!

[ydhongHIT]

Thanks for your interest. Initially, we did not release the configs and pre-trained weights for cityscapes because video data of cityscapes is extremely large (>300 GB). Nevertheless, we will release the files for cityscapes later.

Hi, when are you going to provide the config files for cityscapes? And it is very kind to provide a brief instruction for dataset preparation. Thanks