General course chat

(Emmanuel) #469

Did you find out ?


(Atte Juvonen) #470

I didn’t find assignments anywhere, no.


(Mike Moloch) #471

Somewhere around the end of the first lecture, Jeremy says “I expect you to do this” … and mentions that you should try to do the first notebook with your own dataset.

You can miss it very easily if you zone out just for 3 seconds. I only caught it the third time around.

None of the class notes mention this (that I know of.)

I’m not sure about the 2nd lecture, I have watched it at 1.5x but I think I better watch it at normal speed without distractions. :smile:

DL1V3LEC1 Home work


(Gregor) #472

Hi, does anyone know some good data sources where I can practice my lesson 1 chops from fastai v3 course? I finished lesson 1 and would like to experiment with it a bit “in the wild” before immediately moving on to Lesson 2.


1 Like

(Kerem Turgutlu) #473

Hi Gregor,

Kaggle is a great place to start, Why ?

  • Free 6 hours GPU Kernel
  • library is already in the docker image you will be using
  • You will create a fresh notebook from scratch which will allow you to practice what you’ve learned so far
  • Lastly, Kaggle is the biggest open source dataset repo by my knowledge. So you can just attach datasets which are already in their server
1 Like

(Kerem Turgutlu) #474

Does anyone know when we can apply for Part 2 of V3 ?


(Amit Kayal) #475

I am using a tabular learner model as shared below…

valid_idx = range(len(df_train_revised)-3800, len(df_train_revised))
data = TabularDataBunch.from_df(path, df_train_revised, dep_var, valid_idx=valid_idx, procs=procs, cat_names=cat_names)

and getting following error which seems to be suggest that split between validation and training set not perfect and has unknown labels. Vision model ensures this while splitting training and validation model but it seems that tabular model does not have any such automatic one. Is there any way i can avoid such error?

/mnt/disks/user/anaconda3/lib/python3.6/site-packages/fastai/ UserWarning: You are labelling your items with CategoryList.
Your valid set contained the folowing unknown labels, the corresponding items have been discarded.
920, 30, 20, 180, 125…
if getattr(ds, ‘warn’, False): warn(ds.warn)


(Amit Kayal) #476

I am exploring tabular learner and found following highlighted ones. Can anyone please help me to understand this statement? Does this “embedding sizes for each of our categorical variables” means no of unique categorical values for any categorical dependent variable?

To use that function, we just need to specify the embedding sizes for each of our categorical variables

learn = tabular_learner(data, layers=[200,100], emb_szs={‘native-country’: 10}, metrics=accuracy)
learn.fit_one_cycle(1, 1e-2)


(Pierre Guillou) #477

Hello @martijnd, @zhacker

Your solution works well in order to import a densenet model.
Just one new thing in fastai 1.0.43: densenet models are included in fa,stai/vision/models/

from torchvision.models import densenet121,densenet169,densenet201,densenet161

So, learn = create_cnn(data, models.densenet121, metrics=[accuracy]) works now.

1 Like

(Rahul Indra) #478

Hi, how can I learn how open_mask() function in working in lesson3-camvid part ? I am sure most people have figured this out. I am bit confused with this function. How this function does magic to images ?


(ALI) #479

Hi guys I am an experienced coder in Python and starting out deep learning with this course. However I am quite used to working in IDEs like PyCharm where everything just works and all I have to do is pull out a blank file and start coding and run it. Now transitioning from that to this new world of installing modules in command line interfaces and working on Jupyter notebooks is intimidating. I am ok doing the notebooks but i want to follow up by doing it all myself from scratch in a PyCharm empty file but I cant get fastai to work with pycharm it never installs, along with the many other prerequisits it needs like Pytorch etc. I just want a good IDE where I can do deep learning with fastai library and not have to install modules using command lines etc. Any help in this regards is much appreciated. Cheers!


(Greg McKenzie) #480

check out spyder IDE, it’s part of anaconda that fastai relies on with conda



Hi Ali,
I spent a number of days trying to get fastai working in terminal on a Mac as that is the machine I currently have. I hope to build my own GPU in three or four months.

I had so many issues installing it not using Anaconda.
After 3 days I installed it in its own virtual environment using Anaconda.
Now by switching to this environment I can use any IDE I like.
Unfortunately I don’t use Pycharm as I like simpler IDE’s.
I am currently using SublimeText, and Atom. I do most of my coding in Jupyter Notebook and transfer the finished code to my IDE to create working programs.

I ran the lesson1-pets.ipynb as code on my Macbook Pro and the and it took 7hrs 22mins to complete. (Don’t know if it used the onboard GPU yet! )

So I use for the noteboook exercises and training.

I only found out about FastAI last week so I am pretty happy about getting the environment running on my Macbook Pro and the cloud.

Can’t wait till the remote course starts as I am based in England.
Hope this helps.


(Amit Kayal) #482

Do we have any update on this please…


(Amit Kayal) #483

I am using tabular data for prediction and wanted to to know if there are any guideline for emb_szs for categorical columns…

learn = tabular_learner(data, layers=[200,100], emb_szs={‘native-country’: 10}, metrics=accuracy)
learn.fit_one_cycle(1, 1e-2)


(Edward Ross) #484

They will have fastai live for part 2 and details will be released in the next few weeks:


(Amit Kayal) #485

This is great. Would be waiting for this


(Sahil KARKHANIS) #486

There is no guideline as such. But in the previous version Jeremy had kept a threshold of 50 i guess. Wherein if the categories exceed 50, limit it to 50 if less make them as per the appropriate value.


(Ariel Gamiño) #487

Any announcements on the Part 2 online course? Don’t see any mention yet. Can’t wait.

1 Like

(Amit Kayal) #488

I was looking at the example of tabular learner shared by Jeremy and found that he has considered Y value as log but I am not sue why he has taken log value. Can’t we have RMSE as loss function here?

max_log_y = np.log(np.max(train_df[‘Sales’])*1.2)
y_range = torch.tensor([0, max_log_y], device=defaults.device)