Kai Li

I want to only use the key frames for video classification task. Thanks @dukebw

Same question

OMG, I even didn't notice the cnmem flag before... I am also using Ubuntu14.04 + The latest version of theano and lasagne + TITIAN X + cudnn v5. when cnmem...

Hey guys, I aggressively change cnmem=0.8, ``` [global] floatX = float32 device = gpu2 [dnn.conv] algo_fwd = time_once algo_bwd_data = time_once algo_bwd_filter = time_once [lib] cnmem = .8 ``` The...

add `allow_gc = 0` to the `[global]` section can still speed up a bit. ``` [global] allow_gc = 0 floatX = float32 device = gpu2 [dnn.conv] algo_fwd = time_once algo_bwd_data...

the same question

I have the same problem

@xiaolonw that's make sense, thank you xiaolong

Hi, @benanne Thank you very much, I tried as you said and remove the `LocalResponseNormalization2DLayer`s. The running speed become twice as fast. Now the time for 20 iterations become 27...

Thank you very much, @benanne I will check it out~