search

soumith / inception.torch

Torch port of https://github.com/google/inception
Other
66 stars 20 forks source link

fix max pooling type checking #2

Closed deltheil closed 8 years ago

deltheil commented 8 years ago

Note: the same problem exists on convnet-benchmarks.

soumith commented 8 years ago

whoops!

soumith commented 8 years ago

btw Cedric, I haven't verified the outputs for this network yet. I just created it, made sure it runs without errors, that's all. I have to check that outputs are correct, using an example image.

deltheil commented 8 years ago

Sure. I made no such assumptions reading the README :)

deltheil commented 8 years ago

I haven't verified the outputs for this network yet. I just created it, made sure it runs without errors

In particular the LRN layers here and there have not been added to the network.

soumith commented 8 years ago

lol i'm such a mess.

deltheil commented 8 years ago

I have to check that outputs are correct, using an example image

@soumith I played a bit with it and the result (with the default dog.jpg) example sounds correct:

RESULTS (top-5):
----------------
score = 0.999909: n02110341 dalmatian, coach dog, carriage dog (41)
score = 0.000048: n03803284 muzzle (588)
score = 0.000021: n02109047 Great Dane (17)
score = 0.000007: n02100735 English setter (2)
score = 0.000002: n02108089 boxer (130)

See: https://github.com/Moodstocks/inception.torch/commit/1c8f208 (I can send another PR if you are OK with these changes).

soumith commented 8 years ago

@deltheil merged into master! thanks a ton.

Caffe if I remember, is BGR. I think the images have to be swapped. A sanity check would be on the color-specific classes, like an orange or something..

deltheil commented 8 years ago

Good point.

I made a quick check with an orange and got:

RESULTS (top-5):
----------------
score = 0.985520: n07747607 orange (319)
score = 0.014264: n07749582 lemon (320)
score = 0.000099: n03134739 croquet ball (756)
score = 0.000026: n07753592 banana (323)
score = 0.000026: n03530642 honeycomb (730)

Looking again at the notebook I can see that the image loading part works with imread from skimage, and the doc says it returns "an RGB-image as MxNx3".

So I have the feeling everything is OK - I was lucky :)

It might be the new spec of Caffe2?

soumith commented 8 years ago

nice, yea maybe Caffe2 indeed :)