Introduction to character level CNN in text classification with PyTorch Implementation

[@chaeheumpark1987]

6:58 Is there something else that contributes to the output shape (1,1,136) other than the kernel size and input size?? Please correct me if I am wrong but to my understanding the output shape should be (1,1,138). Thx for the great tutorial !! Saved a lot of hassle

[@tejanmehndiratta]

yes it should be 138

[@AhmedBesbes]

yes it is (1,1, 138) indeed. thanks for pointing the mistake :)

[@gohjiayi5521]

Very helpful video, thank you for sharing!

[@sehlikhouloud4301]

Thanks for sharing !

[@daksh6752]

Great video, subscribed!

[@AhmedBesbes]

Thanks! Stay tuned for more content.

[@bawadhruv1]

Hi Ahmed, whenever i try to run train.py i am always encounted with 'tuple' object has no attribute 'cuda' kindly help please

[@fabianmeyer506]

Great Video :)

[@theDrewDag]

Content is perfect. One problem? Too few videos. The community needs your knowledge to be shared more!!

[@AhmedBesbes]

Thanks a lot, I usually tend to publish more things on my blog since it takes less time to produce you can have a look at my articles here: ahmedbesbes.com/ Point of inquiry though. Is there a specific topic you'd like to see covered in a video?

[@julianwittek7663]

The shape of the kernel that you presented at 5:28 confused me a little bit. It looks like your kernel is of the shape (1, 3), but it is actually of the shape (70, 3). 1D convolution does not necessarily mean that our kernel is one-dimensional. Otherwise the result of the convolution would be the same for texts of the same length.

[@AhmedBesbes]

yes you're right, this a mistake I've meaning to fix in the video. thank you!

[@lacmedical7370]

trop fort !

[@John-jd2tu]

Thanks for sharing! Is it possible sharing the training data also?

[@AhmedBesbes]

Xu Jiahua sorry, due to privacy reasons I cannot share the dataset. However, you could build one very easily by scraping text reviews that have a number of stars (or a rating) associated to them. The rating will in fact allow you to define the labels.

[@ysupriyadi123]

​@@AhmedBesbes Can you give some example of train data?

[@abdulhameedalshemary1223]

Hi Ahmed, thanks for sharing My question is that you said scalar product bw kernel and input data(after encoding) the values of the kernel how do we get them. is it random as in class MLP.

[@kumarshwetaketu1165]

Is there any alternative to one hot embedding of characters?

[@AhmedBesbes]

You can learn a dense embedding but this adds an additional layer to your network and eventually more training time. you can also look into Elmo character embeddings

[@danishmuzafar5572]

sir plz help i have project Depression detection using emotion artificial Intelligence with the help of deep learning

[@AhmedBesbes]

Hello! Given a labeled dataset, you can apply this model for a classification problem :)

[@souravbarik8470]

Why does character level embedding performs badly compared to word-level embedding using CNN for text classification?

[@AhmedBesbes]

it actually depends on the type of data you're working with. Whenever I work with user-generated content (with typos and abbreviations for example) character-based models proved to be quite efficient

[@valerysalov8208]

but bert performers better right?

[@AhmedBesbes]

I've done this video way before Bert was released. There are however great advantages in character-based CNNs; they are lightweight and can be surprisingly good to pick on things like typos or abbreviations

[@lordnaive]

hi ahmed, how to refactor files properly?

[@AhmedBesbes]

What do you mean?

[@lordnaive]

@@AhmedBesbes I see many. People put different functions in different python files. How to decide which functionalities go where? For example. vid2vid by NVIDIA

[@AhmedBesbes]

@@lordnaive I usually put functions that can be used across many python files in a utility script that I call utils.py

[@lordnaive]

@@AhmedBesbes alright

[@RiyaSadrani]

Hello ! very helpful video ! subscribed ! Instead of classification, I have a regression problem where the input is string. Can you guide ? I will just ping you the same in linked in. It will be great if you can help.

[@AhmedBesbes]

If you need to build a regressor instead of a classifier, you'll have to change the ouptut layer and the loss (use MSE for example instead of logloss)

[@ravibhatnagar9732]

Jayuibhatnagar