happycoding
happycoding
Same question. A huge gap between the reproduced and the reported results.
Same situation. Beit-Base I only obtain 53.0, and Beit-Large 55.9, following the official implementation of MMSegmentation. It is hard to reproduce the results mentioned in the paper.
@whai362 Thanks for your reply. Is the finetuned result 85.69 from the model with output strides 2 or 4?
@BowieHsu 请问我如何利用您Pretrain的模型跳过批pretrain那一步呢？？请问exp/sgd/checkpoint里头是pretrain过程当中的模型吗？但是我将您的模型放进去他说formar不对
@BowieHsu 那个finetune的json文件里头只有一个finetune_model, 似乎EXP/SGD里头需要有一个checkpoint文件存在，但是我没有经过pretrain所以没有，您的模型里头似乎也只有3个文件，请问这个如何解决呢？
@BowieHsu 十分感谢!好人一生平安. 还解决了一些其他的问题(gpu什么的...)终于跑起来了
@BowieHsu 谢谢!我目前用的是默认参数,但是训练起来很慢,7个小时训练了6%,感觉很慢阿qwq 请问您训练大概用了多久呢? 我目前集群申请的16core cpu\1个gpu和32gb内存以及10g硬盘
Thanks for your prompt reply! I have rerun the experiments with your suggested configuration (setting the weights of --lambda-memory-pixel and --lambda-memory-region to 0.01), and the results are as follows. [deeplabv3_resnet101_resnet18_log_cirkd_20220720.txt](https://github.com/winycg/CIRKD/files/9155697/deeplabv3_resnet101_resnet18_log_cirkd_20220720.txt)...
@zmghub I also did not successfully reproduce the results mentioned in the paper. Following the default configuration, I got 73.0 on the model implemented with the proposed method. Have you...
@YukangWang, I have the same request. Could you provide the EfficientNet implementation? Thanks in advance!