search

irfanICMLL / structure_knowledge_distillation

The official code for the paper 'Structured Knowledge Distillation for Semantic Segmentation'. (CVPR 2019 ORAL) and extension to other tasks.
BSD 2-Clause "Simplified" License
694 stars 104 forks source link

Training from scratch reached the performance similar with that when initializing from ImageNet #68

Open Shawn207 opened 2 years ago

Shawn207 commented 2 years ago

Hi there, I was trying to reproduce the results reported here on PSPNet-Resnet18. I actually trained it from stratch( --is_student_load_imgnet = False, and empty in ./ckpt/save_path/Student), so student is not loaded with any weights, but I achieved 70% acc with batch_size = 7, which is even a bit smaller than original setting. When I tried with 6000 steps, the acc is 72.7%. The behavior is actually similar with what is reported when initializing with pretrained imagenet model. Do you have any idea on that? Thanks !

AiladReb commented 8 months ago

@Shawn207 Could you please tell me how did you train the teacher ?