Fortnightly Summary
- Improvements in the training pipeline. Primarily adding GPU queuing abilities, removing auxiliary heads causing extra computation loads. (Perhaps enabling mixed precision training and adding the ability to transfer to more gpus during training.)
- A discovery that MMsegmentation uses its own pretrained ResNet model as backbone for Encoder-Decoder structure to conduct further finetuning on downstream segmentation tasks on different datasets. Under the same hyperparameters, same model with random initialization could only achieve half the performance of the official released model.
若有收获，就点个赞吧
0 人点赞
