Share your work here ✅ (Part 2 2022)

Ifeanyi · April 14, 2023, 12:16am

Just went through your post and thought this medium article will help.
https://www.google.com/url?sa=t&source=web&rct=j&url=https://shubhamsaboo111.medium.com/prompt-engineering-the-career-of-future-2fb93f90f117&ved=2ahUKEwjPg9_ci6j-AhUFxgIHHTxVBBwQjjh6BAgVEAE&usg=AOvVaw1vOxNQ9Fw29dbxYAgnN20y

yfeng997 · May 2, 2023, 8:57pm

I used simple ReLU layers to model non-linear transformations, i.e. a 2d parabola. I find the visual results much more intuitive and convincing than a mathematics argument.

Take a look and let me know how you think!

github.com

yfeng997/SimpML/blob/main/relu.ipynb

{
 "cells": [
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 2D Parabola as a composition of ReLUs\n",
    "\n",
    "ReLU makes it possible for us to model non-linear function with linear/affine functions. Without ReLU units all matrix transformations collapse into a single matrix. \n",
    "\n",
    "This notebook explores a simple exercise to use a two layer neural network to model a 2D parabola. We visualize the original parabola together with each composing 'ReLU line' to demonstrate how ReLU effectively models non-linearity.\n",
    "\n",
    "*Bonus*: instead of using PyTorch backward propagation we derived our own gradients and used a vanilla gradient descent to update the model weights. Much credential goes to Jeremy's awesome [matrix calculus cheatsheet](https://arxiv.org/pdf/1802.01528.pdf). \n"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},

This file has been truncated. show original

quortil · May 25, 2023, 9:44pm

I created some interesting dreamlike animations using an image-to-image pipeline, by repeatedly feeding a zoomed-in version of the previous output. This is described in my blog Dreaming of the Giro – Science4Performance

ste · May 31, 2023, 9:23pm

Excited to announce my first public PyPI package developed with #nbdev: testcell.
The Python cell magic %%testcell executes a cell without polluting the notebook’s global namespace. This is useful whenever you want to test your code without having any of the local variables escape that cell.

Simple use case:

# Install
pip install testcell

# Import
ìmport testcell

# Annotate
%%testcell
a = "'a' is not polluting global scope"
a

# Verify
assert 'a' not in globals().keys()

nbdev use case

If you annotate your cell containing tests with %%testcell, all the functions and variable defined there will be forgotten after its execution.

%%testcell
def create_fixture():
  return ['Hamel','Stefano']

for n in create_fixture():
  test_eq(say_hello(n), f"Hello {n}!")

Links

PROJECT PAGE: GitHub - artste/testcell: testcell stops your testing code from polluting the global state
DOCUMENTATION: testcell
PYPI: testcell · PyPI
COLAB DEMO: testcell_demo.ipynb
KAGGLE SAMPLE NOTEBOOK: introducing_testcell | Kaggle

Blog post

quortil · June 14, 2023, 6:47pm

I’ve started making music videos on YouTube using Stable Diffusion

andobeno · June 29, 2023, 8:09am

Hello. Thank you for awesome lessons.

I appreciate re-implementing everything from scratch, I’m learning a lot. It really helps to understand but I still struggled with attention module, so I have re-implemented it in HTML/JS visual diagram I hope I did it right, it does give me the same output as nn.MultiheadAttention.

The diagram is here:
https://andobeno.github.io/daigram/multihead_attention.html

Hope it helps somebody.

quortil · August 7, 2023, 2:48pm

I had the privilege of presenting my Generative AI music videos at the PyData London MeetUp on 1 August 2023, sharing the stage with Paige Bailey, Lead Product Manager for Generative Models at Google DeepMind, and Peter Wang, CEO and co-founder of Anaconda

arun_arumugam · August 9, 2023, 4:46pm

mw00 · August 17, 2023, 7:31am

I had a look an recurrent neural networks (RNNs) after watching a StatQuest video so I wrote a blog post (part 1) about the topic implementing one from scratch. Part 2 will cover backpropagation in RNNs. Feedback is very appreciated

arun_arumugam · September 22, 2023, 2:19pm

quan.tran · October 3, 2023, 4:54am

Hi everyone! It’s been a while since I visited the forums, and cant wait to study the stable diffusion next

For the last couple of months I’ve been working on a few NLP tasks for my job, mainly text classification/regression and item matching via item name + description. Most of the models used are small BERT or GPT related model due to hardware constraints and fast inference time. During this time I have built an end-to-end tool to do

Text processing (data loading + metadatas format + text transformation + text augmentation + …), with HuggingFace as backend (using Arrow)
Model building (to handle any cases of classification/regression/multilabel …and to train + save model with ease.

I also include the ‘LM fine-tuning’ part (mentioned in Jeremy LM guide where you can finetune the LM model (right now it’s limited to masked LM or causal LM), or even train them from scratch if you want to, so that you have the best model for your downstream classifier finetuning task.

Here is the documentation (built with nbdev), which should include all the use-cases: that-nlp-library - Welcome to that-nlp-library

And it’s on pypi as well; you can install it with pip install that_nlp_library

I am looking forward to any feedback as this is my first major package, and several technical aspects of this package are inspired by fastai v2 and Jemery coding style.

joshfischer · November 25, 2023, 5:38pm

Here’s an example of how to enhance Stable Diffusion with LCM-LoRA.

usernotabuser · January 6, 2024, 9:51pm

Hi,

This is my revenge for banning my post with a picture of a dressed, beautiful woman Apparently, she must have not been to someone’s liking. However, I want to share my work, showing something that might be more in line with your very sophisticated tastes. It concerns lesson 9A, the technique used is the mixing of vectors embedding tokens in appropriate proportions. The picture depicts a multi-species monster:

Code:

# I suppose that one of the vector had been mixed up earlier, so this code not fully answer what it is.
text_input1 = tokenizer(["A anteater"], padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt")
text_input2 = tokenizer(["A leopard"], padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt")
with torch.no_grad():
    text_embeddings1 = text_encoder(text_input1.input_ids.to(torch_device))[0]
    text_embeddings2 = text_encoder(text_input2.input_ids.to(torch_device))[0]

# Biohacking ;)
mix_factor = 0.65
mixed_embeddings = (text_embeddings1*mix_factor + \
                   text_embeddings2*(1-mix_factor))

# Create!
generate_with_embs(mixed_embeddings)

Have a nice day or wonderful night

darkknightxi · February 3, 2024, 5:16am

What would your character look like if you were in GTA 5?

I played around with Jeremy’s and Johno’s notebooks in lesson 9 and implemented an image-to-image pipeline for textual inversion that lets you create your own GTA character. I used this concept dictionary for textual inversion.

Here’s what the Indian actor Shahrukh Khan would look like if he were in GTA 5:

darkknightxi · February 5, 2024, 12:09pm

I have deployed this on hugging face spaces. Check out the space and create our own GTA 5 character here.

darkknightxi · March 15, 2024, 7:35am

After procrastinating forever, I have finally made my website using GitHub pages and set up my blog using Quarto. I plan on writing an accompanying blog post with each lesson of part 2, mainly to solidify my understanding. Here’s the website: https://darkknightxi.github.io. Suggestions on how to improve and write better blogs are welcome.
Cheers!

arjanhada · April 16, 2024, 2:24pm

Protein design Convolutional VAE

Multiple Sequence Alignment generation code

squid · April 26, 2024, 3:05pm

3Blue1Brown recently dropped a wonderful explainer video on the Attention mechanism. I wrote an accompanying notebook that spells out the video in code: attention/attention.ipynb at main · SuchitG04/attention · GitHub. It’s a straightforward implementation that was a good tenor manipulation exercise. I plan to write a post that adds maybe the timestamps of the video near the corresponding code.

squid · May 8, 2024, 3:31pm

I’ve written a blog on implementing diffedit here. I had the chance to use some cool tricks (map , partial , and matplotlib animation) shown by Jeremy in lessons 10-12. Please check it out and provide feedback

lisa1 · May 10, 2024, 7:26pm

Hi all,
Playing around with the tree gif. Started with a notebook based on John’s.

Initially I had each frame conditioned only on the previous frame. This turned into very cartoonish images:

To keep the images more stable & more related to the input, I switched to conditioning on a 50-50 mix of a fixed image & the previous frame. Result is better:

From there I wanted to make it a bit less jumpy. I noticed that in some cases it didn’t seem to really be reaching prompt B when transitioning from prompt A to prompt B. So, after interpolating from prompt A to prompt B, I added in some extra frames which were conditioned on prompt B & the previous frame, to give the output frames a chance to fully “catch up” to prompt B.
I like the output a bit better, albeit it’s still not perfect.

Finally, I was curious if it was really necessary to use diffusion to generate a bunch of images with prompt interpolation – why not just generate one image per prompt and then interpolate directly between the latents?
In case anyone else was curious - this worked but the feeling of the output is a bit different. It gradually fades between seasons –