Can skip "Overview" video
Final goal: making domain expert able to use deep learning Part 1 best practises Part 2 join research 7:23 we need GPU 9:28 amazon P2 instances
47:47-> 59:48
maybe need to separate dogs and cats in different directories 54:18 57:25 split into training set and validation set (1000 cats and 1000 dogs) 58:22 sample directory (8 cats/dogs in train set, and 4 cats/dogs in test set)
http://files.fast.ai/models/ (file vgg16.h5 )
source code from https://github.com/fastai/courses/tree/master/deeplearning1/nbs
59:48->1:03.29 start lesson 1 but a lot of blabla until summarized here
1:03.29 -> 1:04.49
Explain:
%matplotlib inline
tells jupyter notebook do display matplot lib graphs inside the window
Explain why do we have a path variable (switch from sample to train data)
1:04.49-> 1:06.50 -> "blabla use anaconda"
1:06.50 -> 1:08.07 numpy
Import micellaneaous libs (utils.py) import utils
1:09.26 -> 1:14.06 pretrained with imagenet (worth a look) images cathegories the images only contain one thing, which is not real case use
1:14.06 -> 1:15.15 how do download the pretrained weigths
1:15.15 -> 1:17.05 presentation of vgg vgg is not the most powerful but it is an easy to reuse base for image recognition
1:17.05 -> 1:18.49 7 lines of code introduction (not really useful)
1:19.02 -> 1:21.28 answer to the question why do we use pretrained models
1:21.54 -> 1:24.32 vgg object based on keras based on theano (transform python into GPU code) theano is based on cuda/cudnn keras can also convert to tensorflow code
1:24.32 -> 1:26.30 tensorflow vs theano blabla about do we need lots of data and lots of GPU 1:26.30 -> 1:27.11 use tensorflow to manage multi GPU
1:27.11 -> 1:28.04 how to configure keras backend 1:28.04 -> 1:28.53 .theanorc configuration file to tell theano to use GPU
1:29.29 -> 1:30.28 (useless blabla) vgg object creation, lines of code behind the scenes
1:30.28 -> 1.30:48 minibatch 1.31:14 -> 1:32.07 GPU are efficient when we parallelize execution not all of it at once because of GPU intern memory
1:32.07 -> 1:33.24 get_batches construct the mini batch from the content of the directory example : grab 4 images and 4 labels
1:33.24 -> 1:34.08 use vgg to tell us what it sees in the images (before finetuning)
1:34.08 -> 1:34.41 how sure vgg is
1:34.41 -> 1:37.00 turn probabilities into finetuning