search

yjxiong / temporal-segment-networks

Code & Models for Temporal Segment Networks (TSN) in ECCV 2016
BSD 2-Clause "Simplified" License
1.53k stars 477 forks source link

some questions about training #229

Closed YanYan0716 closed 6 years ago

YanYan0716 commented 6 years ago

I am a new about this field, and have a question when i was training the network. I just have one GPU, so when i was training,i set the batch_size=12 ,in order to meet the condition about "batch_sizeiter_sizen_gpu=128", i set iter_size=11, so when i was training ,the log showed "Gradient clipping: scaling down gradients (L2 norm 68.3038 > 40) by scale factor 0.585619" . i want to know, if it will make a big influence about the result. thanks a lot and i am looking forward to your reply if you can give me some advice

yjxiong commented 6 years ago

It is fine as long as you can get a good final accuracy. But generally, I would recommend getting more GPUs as training with 1GPU will be very slow.

Mohamad73 commented 5 years ago

how did u change the iteration size

YanYan0716 commented 5 years ago

I only used one GPU ,i set iters more than four times，the result is the same as the paper shown

Mohamad73 commented 5 years ago

I meant from which file? I could not find the file where you can change the iter_size. Thank you for your help

YanYan0716 commented 5 years ago

about temporal-segment-networks-master/models/ucf101/tsn_bn_inception_flow_solver.prototxt , you can change max_iter=....

yjxiong commented 5 years ago

@Mohamad73 It is in the solver file. The option is named iter_size