mit-han-lab
how to train on ucf101 from scratch?
Does anyone train tsm on ucf101 only using imagenet pretrained resnet50? No Kinetics pretrained. I train tsm on ucf101 only using RGB stream, imagenet pretrained resnet50, with hyper-parameters same as TSN, only get 83.5 accuracy.
Depending on the length of the clips it also to change the amount of frames you sample. This can drastically improve (or decrease) the accuracy you obtain. In order to do so, set --num_segments to a higher value. Note that you might have to decrease batch_size (and the learning rate) when you do this.
As per your question in the title, "How to train from scratch" (with only imagenet), you can do this by not finetuning from an already provided checkpoint. Instead just start an entirely new training and use --pretrain='imagenet' which is also the default value for --pretrain.
Does anyone train tsm on ucf101 only using imagenet pretrained resnet50? No Kinetics pretrained. I train tsm on ucf101 only using RGB stream, imagenet pretrained resnet50, with hyper-parameters same as TSN, only get 83.5 accuracy.
would you please help me to train the model on UCF101. I am facing problem while training it
