Links from tonights discussion


#1

https://forums.fast.ai/t/visualizing-intermediate-layers-a-la-zeiler-and-fergus/28140

https://cloud.google.com/compute/docs/gpus/


(Nissan Dookeran) #2

The other blog post I did on the hummingbird classifier https://redditech.blog/2018/11/04/fast-ai-deep-learning-for-coders-week-2-experiment-trinidad-and-tobago-hummingbird-classifier/


#3

Finally! This was the callbacks post that got my excited…I can’t really remember why now I look at it.
https://forums.fast.ai/t/how-to-save-the-best-model-during-training-in-fastv1/28187/6


(YJ Park) #4

A dataset with bounding boxes just got released (a couple of days ago): https://storage.googleapis.com/openimages/web/index.html

Paper: https://arxiv.org/abs/1811.00982


#5

how does one become a ‘professional annotator’? asking for a friend.


#6

Send them to Amazon Turk


(James) #7

I mentioned Gabor filters as a conceptual stepping-stone to understand how conv nets work.

Gabor filters are a signal processing / image processing tool (dating back long before DNNs) used to detect oriented visual features in an image. Basically you have this collection of filters of different wavelengths & orientations, and you apply each filter in turn to your image, measuring the response to each filter independently. Each filter detects structures in the image of different size & orientation. So the collection of responses gives you a fairly thorough analysis of the spatial structure in the image.

There is a great paper by Zeiler & Fergus (that has been mentioned in lectures, IIRC?) where they show that the lowest layers of a conv net are basically Gabor-like filters (in colour) that the DNN learned in response to the training. These lowest layers are the feature detector layers in the conv net.


#8

Hey @jboy

I tried HOG descriptors. They work much better than ORB at identifying similar images. However, I think it’s a loosing battle trying to identify the right filters. My next step is to training a CNN on a diverse sample to learn a set of dynamic filters and then run each image through the network to generate embeddings. But first I have to label a few images :frowning: I will use the existing HOG descriptors to cluster and ease the load :slight_smile:


(YJ Park) #9

Tencent project released multi-label image database (~18mil images):


#10

https://forums.fast.ai/t/visualizing-intermediate-layers-a-la-zeiler-and-fergus/28140/21


#11

(YJ Park) #12

Very interesting. I enjoyed reading it as to how the absolute number of Draw and Away Win games are not much different but % predicted seems to be significantly different, which brings it to your (conspiracy?) theory - it makes sense to presume that the bookmakers hide the information signal in this way.


#13

Yes that is my conspiracy theory. I cannot bring myself to believe that despite ~ 1/3 games ending in draws people ignore this and continue to bet on home or away outcomes instead. It is soccer not AFL :slight_smile: