Відео

Excellent video. Why do we want a different post-embedding vector for the same token in the decoder versus the encoder? reference 12:34

Awesome! Thank you Niels!

the best, thank u so much

what's the tool you used for plotting these figures

i really wanted this exact content and i found you, thank you.

Great video ! Quick question : Does AutoTokenizer work same as the tiktoken lib? Or maybe is it the same and it loads the titoken.get...("gpt2") as the tokenizer? Also do we need to torch.softmax before the argmax ? Weird because everybody tells to do it but nobody really normalizes the output.

awesome!

Hi, do we also apply masking during inference?

thanks for sharing, NR God

Never seen such a video best explanation line by line

is there any way to get the whole graph you've drawn?😀

The best explanatios , your channel is a gem ❤

Awesome job explaining how Inference works. This clarified my confusion about most videos which largely discuss only pre-training. 🙏

Do you have a patreon or some other way to support this work?

Thank you for this tutorial 😀! I still have a question about the benefit of the "json2token" function, can't we simply transform the ground truth dictionaries into a string directly? And feed them to the tokenizer ? Why should we need the special tokens <s_{k}> ?

Great Explanation!!

Hey this is a cool video. Do you have any resources to get started on HF contribution?

Great video! waiting for the benefits of using past_key_values and transformer tools on fine tuning

Are your notes from this video available anywhere online? Really liked the video and would love to add your notes to my personal study notes as well

I am new to this, I am just trying to understand if this during Inference or Training. I guess it is during Inference. please correct me

Thanks for the awesome tutorial, can this be run in colab T4 GPUs (16GB)?

Thanks for your video.

Hi sir, I have trained the model in the exact way you showed. While prompting, it takes me around 15 mins to get one response. Is the inference speed the same for you? If yes, how to optimize this?

Very informative video, thanks for it. I was trying to do some stuff like this myself, can you help me out in knowing how can i finetune a pretrained model into an instruct version but also adding a new language to it. I have looked at the tokenization part and thankfully the tokens are not too sperated that it would cause much trouble. However i wanna know that is it possible that the new layers can learn the language as well as the skill to follow the instructions. Thanks again

Great video, very comprehensible explanation of a complex subject.

can you train my dataset?

Superb content however if I want to add specific key elements to be extracted from custom data how will i do it pls reply

Best video i've seen so far about finetuning with huggingface. I appreciate that you also take time to explain some concepts behind. Other youtubers are just reading out loud their notebooks...

I didn't find a lot of resources that include both drawings of the process, as well as code examples / snippets that demonstrate the drawings practically. Thank you, this helps me a lot :)

@Niels Rogge it is amazing content and i have been following you since the time of donut tutorial keep going buddy. I have one question is it possible to get the snippets extractions also from the invoices? please let me know and reply

You're a beast mate. Thanks!

amazing video, great explanation

no more runpod?

Hi, why attention mask is added to the attn weights instead of multiplied (1h00:11)? if you add the attention weight with zero the weights will not be ignored

Thanks for making this super in-depth tutorial Niels! And very nice dataset/task you chose, I like it.

banger video as always 🔥

Is this model robust against incomplete labels/ground truth ? Let say you have for some images all details and for others not so many

Excellent overview of how the encoder-decoder work together. Thanks.

Thank you very much for the explanation, Niels. It was excellent. I have just one question regarding 'conditioning the decoder' during inference: How exactly does it work? Does it operate in the same way it does during training, i.e., the encoder hidden states are projected into queries, keys, and values, and then the dot products between the decoder and encoder hidden states are computed to generate the new hidden states? It seems like a lot of calculations for me, and in this way, the text generation process would be very slow, wouldn't it?

Thank you so much for the amazing tutorial!

This is awesome, thanks for sharing, I'm only 5 minutes in right now but will definitely follow it all later. These types of tutorials are gold-dust, calm, concise and not just for entertainment.

Thank you for this excellent presentation. In future can you do a video on how to take two take two different HF models and merge them , like this one .

Amazingg. Such a comprehensive and detailed video. Loved it

Wow, Look who's back 🔥

superb video. I will recommend this video in my next lab presentation.

Fantastic breakdown, thank you Niels

谢谢你，讲得很好，之前只是大概了解，现在是更清楚其中的细节了。非常感谢，爱来自瓷器

I learned a lot thank you.

Transformers are '"COMPLICATED" ? Not really after this video. Thanks.

Just woderful. My search ends.