📝 Deep Learning Lesson 1 Notes

Daniel · February 27, 2019, 1:12am

如何使用Colab

这是fast.ai v3 课程使用Colab的快速入门指南

如何你不是第一次使用，请参看returning to work 指南

注意: 这是免费服务但不是随时都提供，而且你需要额外工作保存你的实验。务必仔细阅读colab说明理解服务的局限性.

Step 1: Accessing Colab

首先，登陆谷歌账户here.
然后, 前往 Colab Welcome Page 并点击 ‘Github’. 在 ‘Enter a GitHub URL or search by organization or user’ 一栏填写 ‘fastai/course-v3’. 你将看到所有课程notebooks，点击一个你感兴趣的.

在你运行前，需要告知Colab你希望使用GPU. 你可以点击 ‘Runtime’ tab 选择 ‘Change runtime type’. 在随后打开的窗口下拉菜单里选择‘GPU’，或者在edit中选择Notebook settings中选择GPU 然后点击 ‘Save’.

Step 2: Configuring your notebook instance

在你开始使用notebook前, 你需要安装必须的软件包。你可以用以下代码，运行第一个Cell

!pip install numpy==1.15
# then restart the runtime, open a new cell to run the following
!curl -s https://course.fast.ai/setup/colab | bash

然后，你会看到一个跳出窗口说 ‘Warning: This notebook was not authored by Google’; 你选中’Reset all runtimes before running’, 然后点击‘Run Anyway’.

在新窗口中点击 ‘Yes’.

Step 3: Saving your notebook

如果你是从github打开notebook，你需要将实验存入Google Drive. 你可以点击 ‘File’ 再点击 ‘Save’. 你可以看到以下窗口，然后点击‘SAVE A COPY IN DRIVE’

这将打开一个新 tab 含有相同文件位于你的Google Drive. 如果你希望保存后继续工作，那么直接在这个新tab中工作就可以. 你的notebook将被保存在一个叫 Colab Notebooks的文件夹位于你的Google Drive中.

Step 4: Saving your data files

如果你希望改写你的文件，你需要允许你的Colaboratory instance读取和改写你的 Google Drive. 你只需要在每一个notebook的第一个Cell中写入以下代码

from google.colab import drive
drive.mount('/content/gdrive', force_remount=True)
root_dir = "/content/gdrive/My Drive/"
base_dir = root_dir + 'fastai-v3/'

现在，你的 Google Drive 变成了一个文件系统允许标准python读写文件. 不要忘记给所有notebook的根目录地址前加上 base_dir . 例如, 在 lesson2-download.ipynb 5th cell, 做一下修改

path = Path(base_dir + 'data/bears')
dest = path/folder
dest.mkdir(parents=True, exist_ok=True)

More help

更多问题可以在 fast.ai forum 提出。

Colab

mmilovec · February 28, 2019, 7:59pm

I have a problem on google cloud instance I cannot do git pull it say permission denied, not sure why, I am owner of the project and still not working. How to setup git on google cloud instance right so that it works and so that I can do git pull.
Thank you

arajendran · March 4, 2019, 10:29pm

Marijan, where are you calling git pull from? It works fine for me after changing into this directory:

cd tutorials/fastai/course-v3

Can you confirm that you are calling this from your fastai instance and not locally? (Do you see

jupyter@my-fastai-instance:

in your terminal?)

mmilovec · March 5, 2019, 6:17am

yes I can confirm that I am in cd tutorials/fastai/course-v3 and I can see jupyter@my-fastai-instance:

but when I go git pull it said permission denied.

Sharan99 · March 6, 2019, 6:03pm

@PoonamV Does that give me the freedom to first give the dimensions as 224 x 224 , find the learning rate corresponding to that from learn.lr_find() and then unfreeze the layers , change the dimension back again to whatever I want and then train the unfreezed layers? Because I feel that when the dimensions are reduced, a lot of information goes missing which may be very crucial for certain specific classifications…

rkirana · March 31, 2019, 10:18am

Very nicely done. Which tool did you use for the mindmap?

lbt · April 20, 2019, 3:29am

@leovcld, great job!!
I wonder whether you have also done mindmaps to the rest lesson 2-7!
If yes, could we share them too? thanks.

EaSara · May 11, 2019, 10:25pm

Thank you so much.

impiyush · May 15, 2019, 7:34pm

Very helpful. Thanks!!

cupofcalculus · May 23, 2019, 4:20pm

Two things:

create_cnn is now deprecated.
learn = create_cnn(data, models.resnet34, metrics=error_rate)
should now be
learn = cnn_learner(data, models.resnet34, metrics=error_rate)
doc(interp.plot_top_losses) throws an error.

DrC · July 4, 2019, 11:18pm

My lesson 1 notes with some added clarifications. Hope some will find it helpful.

amit_js · September 23, 2019, 4:17pm

If I click it’s not opening the link.

nbharatula · September 25, 2019, 10:58pm

I updated the link. Pls try now.

mhdmhdpor · November 7, 2019, 7:38pm

Always remember to do an update on the fast.ai library and course repo.

conda install -c fastai fastai for the library update

git pull for the course repo update.

چرخ گوشت صنعتی

feng3245 · November 30, 2019, 3:28pm

Have question regarind len(data.valid_ds)==len(losses)==len(idxs) on the first lesson. What is data.valid_ds?

feng3245 · December 1, 2019, 12:52am

These all show 0.00 probability… I think there might be a bug in the code or am I missing something

thunderkid · December 2, 2019, 7:35pm

Hi. I’m a bit confused about my resnet34 learning speeds. On a sagemaker p2.xlarge instance, running cell 13 of the notebook learn.fit_one_cycle(4), I’m finding each epoch takes about 1 minute. But in the lesson 1 video, it’s only shown as taking 29 seconds per epoch. I believe that the fastai docs suggest using this same p2.xlarge instance type, and Jeremy mentions a cost of $1/hour which is about right. If I upgrade to a p3.2xlarge instance I get times of around 32 seconds per epoch. Still slower than the lesson 1 video, and at a much higher cost of ~$4/hour. Any ideas why I should be seeing this performance gap?

deepanshMathur · January 9, 2020, 8:16pm

data is your ImageDataBunch object.
data.valid_ds is referring to the validation dataset that is present in your ImageDataBunch object.
Similarly data.train_ds refers to the training data set.

farbodee · January 29, 2020, 4:55am

Hi All
In regards to the “Results” section of the lesson 1 where we are checking the results,I was just wondering where we do (len(data.valid_ds)==len(losses)==len(idxs), if it comes true, does it mean all our training data sets items are in the top losses? if so doesn’t that mean our training was not worthy?

jwuphysics · January 29, 2020, 5:18am

No, it’s ensuring that each of these objects have the same length. If they had different lengths, then it means that there was a problem somewhere.