soumith / inception.torch Goto Github PK

Hi

sorry for bothering. In the example.lua I saw the following:
img:mul(255):clamp(0, 255):add(-117)
The mul(255) is to blow up the values from the range 0...1 to 0...255.
The add(-117) is to remove the mean of all the imagenet images I suppose.
I noticed that you do not divide by the std-deviation. Is this just a simplification for this example or not needed in general? If we should do the normalization, what value do you suggest to take (std_dev over all imagenet images)?

Regards, Felix

Hi @soumith,

Thanks for providing a torch version of google/inception. I made predictions on the same image using caffe2-based ipython notebook and example.lua in this repository. I encountered a difference in the prediction score:

Input image is an example image in ILSVRC12 validation set whose ID is 00048000

th example.lua 

RESULTS (top-5):
----------------
score = 0.260391: n07720875 bell pepper (735)
score = 0.195900: n03461385 grocery store, grocery, food market, market (703)
score = 0.097373: n07753275 pineapple, ananas (322)
score = 0.091731: n07747607 orange (319)
score = 0.073422: n07717410 acorn squash (741)

# inception.ipynb'
Top five predictions:
   735 (prob 0.7066) synset n07720875 bell pepper
   741 (prob 0.0866) synset n07717410 acorn squash
   703 (prob 0.0740) synset n03461385 grocery store, grocery, food market, market
   322 (prob 0.0556) synset n07753275 pineapple, ananas
   840 (prob 0.0183) synset n03482405 hamper

Is there any reason why I'm noticing the difference? Thanks

Is it possible (easy?) to wrap the Sequential container as an nngraph gModule without explicitly adding the input on each line? For instance, if you were to pass multiple inputs to the net?

Thanks.