Network training

[xiechunhong]

@Tamuel how can I train this network using my own dataset?

[Tamuel]

@xiechunhong First, you have to convert your own dataset into tfrecord by utils/dataset_util.py. You can simply modify def make_***_tfrecord function in dataset_util.py to convert your dataset into tfrecord. And then, change model_dir, train_data, test_data in __init__.py and run. You can change other variables to train your network like batch_size, max_iter, etc.

[xiechunhong]

@Tamuel Thank you for your reply, Tamuel! Your work is great!

[nieyan]

@Tamuel Great work! And can you tell me what dataset do you use to train? And the lane in demo has been postprocess or the raw model output?

[Tamuel]

@nieyan Thank you for your comment. In demo, I just overlapped raw model output on original image. There are no other postprocessing conducted. And I pretrain network by BDD100k dataset (https://bair.berkeley.edu/blog/2018/05/30/bdd/) which have information of lane by vertices. So, I make sparse lane segmentation dataset from those vertices by OpenCV and pretrain with it. And then, fine-tune network by KAIST dataset (https://sites.google.com/site/highwaydrivingdataset/) which finely labeled. If you want to increase performance of network more than this, you can use other huge dataset like Mapillary dataset (https://www.mapillary.com/dataset/vistas?pKey=H1P0sKnFsYu1MkfcjGUZTg).

[nieyan]

@Tamuel Many thanks for the prompt reply!

[nieyan]

@Tamuel By the way, can you explain the process of BDD100k lane vertices to lane mask? Or it's not dense lane mask while training ? I am confused of sparse lane segmentation.
Sparse lane segmentation is train only via the lane vertices during loss build?

[uzair-mehmood]

@Tamuel, did you apply augmentations, other than those applied in init?

[xuxiuzhi2627]

@Tamuel how can I train this network using my own dataset?

Hi，Have you implemented this code yet?

[xiechunhong]

@xxzcpsmemeda Just follow Tamuel's comments, the training can be done.

[fsxy1063200037]

Hi！I get this problem, how did you solve it

tensorflow.python.framework.errors_impl.InvalidArgumentError: Unsuccessful TensorSliceReader constructor: Failed to get matching files on ./init_checkpoints/resnet_v2_50/resnet_v2_50.ckpt: Not found: FindFirstFile failed for: ./init_checkpoints/resnet_v2_50 : ϵͳ�Ҳ���ָ����·����
; No such process