The examples are confusing because the parameters keep changing. First there is a vocabulary of 10 words, with each mapped to 4 numbers. Then it is 3 words, each mapped to 3 numbers. Then 50 words, each mapped to 8 numbers.
I've occasionally wondered what the best way to find the closest vector (embedding). It's possible to find the cosine similarity between two embeddings, but is there an efficient way to look up the closest embedding when you're searching through a large list? I wish there were something like a hash map function for embeddings.
I've also had a lot of luck with runpod which is probably priced similar, but has interesting combinations like dual 4090's for $1/hour which is enough to run Llama-2-70B quickly with 4-bit quantization. It supports Jupiter Notebooks like CoLab, though I prefer using a shell.
@@nathanbanks2354 That could be interesting. Right now I'm working on a complex project. I want to make reliable tutorial for hundreds of complex software like photoshop, blender, da vinci resolve etc... Unfortunately gpt 4 isn't very reliable for these tasks. I wanted to fine tune a model but I don't know if I should use llama2 7b or llama2 70b. I only considered the 7b for monetary concern but maybe it's not up to the task. Thanks for giving me these prices for the 70b! (even tho I don't know if it's enough for finetuning)
The examples are confusing because the parameters keep changing. First there is a vocabulary of 10 words, with each mapped to 4 numbers. Then it is 3 words, each mapped to 3 numbers. Then 50 words, each mapped to 8 numbers.
Well explained. Congrats!
How is nn.Embedding different from nn.Linear in this context?
I've occasionally wondered what the best way to find the closest vector (embedding). It's possible to find the cosine similarity between two embeddings, but is there an efficient way to look up the closest embedding when you're searching through a large list? I wish there were something like a hash map function for embeddings.
Thanks for the video!
On an unrelated note I saw your other video about google collab. Is it still 0.2$ an hour on a medium GPU? (Tesla t4)
That video is on my list to update, the pricing structure of CoLab has evolved greatly in the last few months.
I've also had a lot of luck with runpod which is probably priced similar, but has interesting combinations like dual 4090's for $1/hour which is enough to run Llama-2-70B quickly with 4-bit quantization. It supports Jupiter Notebooks like CoLab, though I prefer using a shell.
@@nathanbanks2354 That could be interesting. Right now I'm working on a complex project. I want to make reliable tutorial for hundreds of complex software like photoshop, blender, da vinci resolve etc... Unfortunately gpt 4 isn't very reliable for these tasks. I wanted to fine tune a model but I don't know if I should use llama2 7b or llama2 70b. I only considered the 7b for monetary concern but maybe it's not up to the task. Thanks for giving me these prices for the 70b! (even tho I don't know if it's enough for finetuning)
hey heaton, maybe you can fix the sound of the video
Get a life