No retrain.yml for segmentation tasks
bachml opened this issue · comments
Hi,
There is no retrain.yml available in the configs folder. I am wondering how to retrain the searched models for cityscapes.
Also, when I train the search model with the seg_cityscapes.yml, there will be an issue comes up when evaluating the model:
Traceback (most recent call last):
File "train.py", line 655, in
main()
File "train.py", line 651, in main
train_val_test()
File "train.py", line 507, in train_val_test
get_prune_weights(model_eval_wrapper), prune_threshold) # get mask for all bn weights (depth-wise)
File "/mnt/lustre/zengguohang/HR-NAS/utils/prune.py", line 274, in cal_mask_network_slimming_by_threshold
weights = torch.cat(bn_weights_abs)
RuntimeError: There were no tensor arguments to this function (e.g., you passed an empty list of Tensors), but no fallback function is registered for schema aten::_cat. This usually means that this function requires a non-empty list of Tensors. Available functions are [CPU, CUDA, QuantizedCPU, BackendSelect, Named, AutogradOther, AutogradCPU, AutogradCUDA, AutogradXLA, AutogradNestedTensor, UNKNOWN_TENSOR_TYPE_ID, AutogradPrivateUse1, AutogradPrivateUse2, AutogradPrivateUse3, Tracer, Autocast, Batched, VmapMode]
Hi,
- The searched segmentation model can be directly used for inference. Retraining the model provides almost no gains. You can still try retraining by using the searched network parameters and setting:
bn_prune_filter: ~, rho: 0.0,
- It seems that the weights for pruning are not collected. Did you use the default setting for training? Can the model be trained but not evaluated?
Regards,
Mingyu