Base on this video https://www.youtube.com/watch?v=IPBSB1HLNLo&feature=youtu.be
https://medium.com/@hiromi_suenaga/deep-learning-2-part-1-lesson-1-602f73869197 http://forums.fast.ai/t/wiki-lesson-1/9398
blabla until 6:21 Can use cresle
6:21 -> 12:57 There is a paperspace setup we can run just like this curl http://files.fast.ai/setup/paperspace | bash This command installs everything
Then here is how to run the notebook cd fastai git pull jupyter notebook
https://github.com/fastai/fastai/blob/master/courses/dl1/lesson1.ipynb
12:57 -> 14.49 blabla about how to use jupyter notebook use python 3
14.49 -> 16:11 how to set up the data set cresle vs paperspace setup
16:11 expected content of the folder !ls {PATH} expected output: models sample test1 tmp train valid blabla until 18:36
18:36 -> 18:44 (just follow the notebook) plot the first cat
18:44 -> 20:00 how is composed the input image (rgb, pixels)
20:00 -> 20:27 context (not really useful) dataset comes from kaggle, the accuracy used to be 80 %
20:27 -> 21:32 the three lines of code to do training run the code and explain the output loss function training set/validation set + accuracy
21:32 -> 24:13 blabla we are better than 2013 and it takes 17seconds blabla we are using research papers blabla we are using pytorch
24:13 -> 25:33 Get into the model how is labeled the training set pytorch returns the log of the probability instead of the probability itself
25:33 -> 25:57 blabla learn to use numpy
25:57 -> 27:47 plot the correct images
27:47 -> 28:57 blabla we plot things because we want to understand where our model is wrong 28:25 -> 28:57 some images need data augmentation to be recognized because they are not correctly shaped
28:57 -> 30:53 blabla try with new images yourself
30:53 blabla education is top down blabla fast.ai has a different approach blabla we are going to solve more and more problems blabla zzzzz 33:57 -> 39:25 blabla course workflow 39:25 -> 41:28 blabla try the code before knowing the theory 41:28 -> 44:44 blabla where the notebook spends time + where we can use image classifiers 44:44 blabla deep learning is a kind of machine learning
48:50 what is a neural network
49:48 presentation of gradient descent in neural network we have one global minimum gpu are good to run this algorithm
53:12 -> 57:00 blabla google invest in deep learning microsoft skype translate
57:00 -> 59:14 blabla
59:14 -> 1:02.12 presentation of convolution
1:02.12 -> 1:04.27 non linear layer
1:04.27-> 1:08.00 principle of gradient descent what is the learning rate
1:08.00 -> 1:11.36 displaying convolution filters layers
1:11.27 -> 1:11.57 30 seconds of blabla
1:11.57 -> 1:19.28 how to set the learning rate
1:19.28 -> 1:20.40 how to choose the number of epochs run until accuracy gets worse (overfitting)
1:20.40 -> 1:21.57 things to try
1:21.57 -> blabla jupyter trics (tab = autocompletion) (shift+tab = function parameters) (shift+tab *2 = documentation) (shift+tab *3 = documentation in a pop up window) (?? + function = source code of the function)