A walk with fastai2 - Tabular - Study Group and Online Lectures Megathread

fastai
122

Passing in empty list works !!

1 Like
123

Typo in 02_permutationā€¦ nb.
to = TabularPandas(ā€¦should be y_block=RegressionBlock()ā€¦) and not block_y =

124

They updated the syntax around a few days ago. Iā€™ll go make that adjustment.

Edit: this has been fixed @Srinivas

125

Thanks

126

Thanks for the reply. I did not see them in the arguments in the doc thus thought they are deprecated. My bad!

2 Likes
127

Hi @zdw2126062 I encountered the same error.
Does that mean having momentum does not have a positive effect on a tabular models?
Also, since momentum is not used will it be a good idea to use weight decay as an optimisation instead of momentum?

128

@AjayStark no see Sylvain comments about tabular_config. Iā€™ll try to show an example next week.

1 Like
129

Oh, thanks for pointing out. Got it ! :slight_smile:

130

After assigning momentum using config=tabular_config(), how to pass config into tabular_learner() ?

131

Just pass it in as config. IE:

tabular_learner(dls, layers, config=config)

1 Like
132

Ooops silly one. Didnā€™t see config in the arguments list.

And is it same as creating a Tabularmodel first with the ps and embed_p

133

You can pass it in as below:-

config = tabular_config(ps=[0.001,0.01], embed_p = 0.04)

learn = tabular_learner(dls,
layers=[100,50],
config = config,
metrics=[accuracy])

1 Like
134

Yeah got it, Thank you :slight_smile:

135

@muellerzr when is the lesson4 going to be?


Just came to know about this, how are you currently doing Zach?. Hope all itā€™s well and do take adequate rest as required.

136

As mentioned earlier they are now on Wednesdays. Iā€™m doing fine currently thank you for asking.

1 Like
137

@muellerzr thanks a lot for your passion when teaching! :heart_eyes:

I was watching your awesome video A walk with fastai2 - Tabular - Lesson 3 and I had a couple questions:

  1. Why is your test dataloader using df.iloc[:100]. Arenā€™t these rows used for training already? Or was this choice done just for the sake of illustration of how ClassConfusion works?

  2. When looking at the plots, in (<50K, >= 50K) the actual class was <50K and the predicted >=50K for the df.iloc[:100] that were wrongly classified. Is this statement correct? I just wanted to make sure.

image
image958Ɨ538 119 KB

  1. Where can I find reliable and understandable information for how fastai tabular embedding work? I understand how they are generated for NLP but to me this part is still very unclear for tabular data. :grimacing:

To everyone, any good courses you would recommend about explainability? I would like to learn more about how SHAP works and what the different plots mean.

138

Just for the sake of visualization :slight_smile:

Yup youā€™re correct!

See here, an article by Rachel back in 2018 (but nothingā€™s changed :wink: )

1 Like
139

Hereā€™s a Kaggle mini-series (I think thatā€™s what itā€™s called) that looks pretty good. Covers permutation importance along with SHAP and a few others:

1 Like
140

Thanks a lot! Quick doubt. How can I extract a row from here?

image
image1071Ɨ465 47 KB

I tried learn.dls.valid_ds.iloc[800,:] as I would in pandas but I get TypeError: only list-like objects are allowed to be passed to isin(), you passed a [slice]. Note that type(learn.dls.valid_ds) is not dataframe but fastai2.tabular.core.TabularPandas

1 Like
141

Promise Iā€™m not ignoring you :slight_smile: Iā€™ll be able to get back to you on this either Tuesday evening or Wednesday evening at the latest :slight_smile: Apologies!

Also in regards to this week, Iā€™m currently working on traveling home which will be on Wednesday when we would normally do class, so for everyoneā€™s ease itā€™ll be same time on Thursday. I appreciate your understanding and my apologies. Next week once we start NLP there shouldnā€™t be any issues whatsoever in terms of weird set backs thanks to covid (And those will be at our regular Wednesday time)

2 Likes