These are chat archives for beniz/deepdetect

26th
May 2017
alkollo
@alkollo
May 26 2017 11:41
@beniz, I do not want to crop images, sorry I made a confusion between both as I thought setting the size in input would crop the images, not resize them. Anyway the error remain. If I want to use resize other than 224 I Get an error (mentioned above). the transform param dictionnarie is right because I did not setup a crop size anyway.
alkollo
@alkollo
May 26 2017 11:56

hello,

I'm using resnet 50 on 220000 images in 140 classes.
Using classification image I get:

iterations 13000.0 : loss 4.64893245697 : prec 0.00747848763786 : top-1 0.011547921791 : top-5 0.0575311627131
iterations 13013.0 : loss 4.64168167114 : prec 0.00869452733306 : top-1 0.0139242089465 : top-5 0.0661608371201
iterations 13386.0 : loss 4.58001661301 : prec 0.00869452733306 : top-1 0.0139242089465 : top-5 0.0661608371201
iterations 13758.0 : loss 4.60921859741 : prec 0.00869452733306 : top-1 0.0139242089465 : top-5 0.0661608371201
iterations 14000.0 : loss 4.52883768082 : prec 0.00869452733306 : top-1 0.0139242089465 : top-5 0.0661608371201
iterations 14023.0 : loss 4.61971855164 : prec 0.00727645985379 : top-1 0.0132988702214 : top-5 0.0506107474882
iterations 17050.0 : loss 4.51852464676 : prec 0.00993728717994 : top-1 0.0170925918206 : top-5 0.0623671155209
iterations 17423.0 : loss 4.58611488342 : prec 0.00993728717994 : top-1 0.0170925918206 : top-5 0.0623671155209
iterations 17796.0 : loss 4.57050991058 : prec 0.00993728717994 : top-1 0.0170925918206 : top-5 0.0623671155209
iterations 18000.0 : loss 4.57423448563 : prec 0.00993728717994 : top-1 0.0170925918206 : top-5 0.0623671155209
iterations 84000.0 : loss 3.80068969727 : prec 0.0114324296926 : top-1 0.0130904239797 : top-5 0.0545712260808
iterations 84227.0 : loss 3.74527812004 : prec 0.0077129340257 : top-1 0.00658690123817 : top-5 0.0318088964856
iterations 84528.0 : loss 3.81403040886 : prec 0.0077129340257 : top-1 0.00658690123817 : top-5 0.0318088964856

here are my calls:
https://0bin.net/paste/V3grlDLMdurai7Z4#nAEU1yfKjZDRJUBANDCO+1cqUC6p2mMVEfRk-WwHq6X

after 85k iterations I still have very low numbers and a loss that is not decreasing a lot.

Emmanuel Benazera
@beniz
May 26 2017 12:50
I think you want to stick to 299x299 as input size otherwise you'd need to modify the convolution kernel sizes.
alkollo
@alkollo
May 26 2017 13:00
ok but if I change the input size other than 224 I got an error when testing: "Cannot share param 0 weights from layer..."
Emmanuel Benazera
@beniz
May 26 2017 13:01
this doesn't mean anything if you are not using a pre-trained model
unless this happens at testing time, when sharing the trained weights
alkollo
@alkollo
May 26 2017 13:06
mmm I see hat you mean, I will clean my model dir and start from a blank dir to be sure
Emmanuel Benazera
@beniz
May 26 2017 13:15
I'll run a test in case something has changed anyways...
alkollo
@alkollo
May 26 2017 13:18
ok thanks
Emmanuel Benazera
@beniz
May 26 2017 13:29
I've just tested on full Imagenet, the few first hundreds iterations, and everything seems well, loss decreasing etc...
but I can replicate your issue at test time
alkollo
@alkollo
May 26 2017 13:40
you can or you can't ?
Emmanuel Benazera
@beniz
May 26 2017 13:50
it's because the template should not be defining the crop_size. It was ignored in the past but recent changes have modified this behavior.
alkollo
@alkollo
May 26 2017 13:52
ho
Emmanuel Benazera
@beniz
May 26 2017 13:52
We'll push a fix, but the easiest way for you to fix this is to remove the crop_size line from resnet_50.prototxt, then issue the PUT call without the template:resnet_50 line in it, keep the modified prototxt files.
You'll have to reduce your batch_size, since the network will be working in 300x300 directly now.
alkollo
@alkollo
May 26 2017 13:59
I will try that, thx
sunatthegilddotcom
@sunatthegilddotcom
May 26 2017 14:17
@alkollo my 10k precision is .428
alkollo
@alkollo
May 26 2017 14:19
@sunatthegilddotcom . Thx It seems I have a problem so
alkollo
@alkollo
May 26 2017 16:10
@sunatthegilddotcom wich image set do you use ? Ilsvrc2012 ? could you also share your train params please ? thank you