Видео

Typo in calculation at 5:06, as i=0 for first 2 items of sin and cos for "boy", you have as 1.

The division by 2 relates to splitting the embedding dimensions for sine and cosine computations and is independent of the number of sentences or words.

this is amazing, thanks!!

First Multihead Attention is masked, and the second is normal multihead attention. I think you forgot to mention this is masked cause the decoder behaves as an autoencoder which sends the output back into the output. During training as we will need to mimic this behaviour of inference/testing in training , therefore attention is basically the attention with mask. Great videos overall!

hi, Thank u for starting RF , would u also provide code :) and examples.?

wonderful explanation. Keep doing.

best video, ever

Thank you for the explanation

Hi is this the cpu version? If so, where do we get the gpu version and could you just explain it in a bit?? Thanks

Here I have a observation, in the input data number of features are 3 and we have number of training samples 2. That's why the input_data matrix's shape is 2 by 3. And also the number of the neurons in input MLP layer = number of features of the input data i.e. 3. Plz make me correct, if my speculation is wrong. And thanks for the interactive video.

i jut found your channel and i can not wait to watch all of your videos. this is awesome thanks

Why is cosine of 0 = 0?? I think there is an error

Thanks a lot! Definitely cleared up a lot of things

Thanks for the details, keep going 👏

goat

Brother your videos are great! Very helpful thank you

All videos are great. Thanks a lot. Can you start some series or videos on diffusion models specially class conditioning and latent diffusion models (or other score matching models) ?

Sorry, I don't understand you lesson. 1. Why activation layer to 3 neurons on second layer (each of them has three connections from previously) have matrix 3x2. 2. Why you disable neurons at 1col [2,3rows] and after disable on 2col [3row]. If you have p=0.5, I am assume you should disable two-three neurons on each step of training, leave only 1-2 neuron enabled.

Thanks for details. I am learning Transformer Architecture, Attention mechanism and Math hidden under them. I am from Moscow / Russia.

maybe a better microphone

you bro you skip that when model do the test~~

Good Explanation !

but in fully connected layer a 1D vector is passed so the 3*2 matrix will be flattened out right? so how are we defining the shape of weights and bias it should be (6*1)

Very nice

Got rid of the jargon , straight to the point , great tutorial

😊😊❤amazing tutorial man

thanks!

Love these hand writing style explanations ❤

Thanks for the video! Really do me a big favor.

Explanation is clear... great job.. but the audio is little bad

can you explain tgt_key_padding_mask parameter in nn.TransformerDecoderLayer.

Great explanation, do you have videos on layer norm, instance norm and group norm?

thank you for the clear math. I still couldn't get how the word order is preserved though. Is there any visual representation or mathematical illustration of how the positions are preserved?

this video started off well, but it would ahve been better if it showed the implied second line of python code, explicitly

great job explaining this concept!

your videos are game changing for sure thank you very much please you are a life saver

Most underrated channel ever

i don'r get why we divide two

very good explaine thanks can you do vit transformers plz

Great job on explaining. Love your content!

naji dadin video dinnan bansako kai bahause bane amma muryanka kaman na bahause ...Allah yakara basira

Excellent! Thank you so much! The best explanation ever for embedding layer. We can find it anywhere else in the web!

Thank you! I've been trying to understand that math unsuccessfully for a long time.....seen lots of videos, but somehow yours explained best

Very good videos, in my opinion The best i ve ever seen about The math perspective of transformers

Amazing

May I ask what device you used to record this video?

Thank you! I didn't understand why the bias isn't of dimension 1 and this sorted it out for me

Tysm :)

Very good man, Keep posting ☺️