Skip to content

Instantly share code, notes, and snippets.

View morrisalp's full-sized avatar

Morris Alper morrisalp

36 followers · 4 following
View GitHub Profile
@morrisalp
morrisalp / prompt_mixing.py
Created April 4, 2024 11:47
example of prompt mixing (different prompts at different denoising steps) with SD2.1
import torch
from diffusers import StableDiffusionPipeline
from matplotlib import pyplot as plt
device = "cuda" if torch.cuda.is_available() else "cpu"
pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16)
pipe = pipe.to(device)
prompts = ["a hamburger"] * 8 + ["a birthday cake"] * 12
@morrisalp
morrisalp / sdxl_with_grad.py
Created January 14, 2024 18:24
Diffusers SDXL pipeline with gradients (overriding no_grad), tested with diffusers v25.0. Use output_type="latent" when calling pipeline to get latents with gradients.
from diffusers import StableDiffusionXLPipeline
class CustomPipeline(StableDiffusionXLPipeline):
@classmethod
def from_pretrained(cls, *args, **kwargs):
self = super().from_pretrained(*args, **kwargs)
assert self.watermark is None # watermarking currently not supported
def postprocess_no_grad(image, *args, **kwargs):
@morrisalp
morrisalp / commons_data.py
Created September 10, 2023 16:43
simple API calls to get lists of files and (sub)categories from WikiMedia Commons
import requests
import json
from tqdm.auto import tqdm
import pandas as pd
def kwargs2url(**kwargs):
url = 'https://commons.wikimedia.org/w/api.php?'
for k, v in kwargs.items():
url += f'{k}={v}&'
return url.rstrip('&')
@morrisalp
morrisalp / surprise_ml-100k.py
Last active May 14, 2021 07:49
loading MovieLens 100K dataset with Surprise & Pandas
import surprise
import pandas as pd
data = surprise.Dataset.load_builtin('ml-100k')
ddir = surprise.get_dataset_dir()
item_data = pd.read_csv(f'{ddir}/ml-100k/ml-100k/u.item',
sep='|',
header=None,
encoding='ISO-8859-1',
@morrisalp
morrisalp / search-pip.sh
Created February 26, 2021 07:42
Minimal Bash command for searching for Pip packages, now that 'pip search' no longer works. Usage: `./search-pip.sh pyngrok`
#!/bin/sh
wget -O - https://pypi.org/simple/ 2>/dev/null | grep "$1" | grep -Po "(?<=>).*(?=<)"
@morrisalp
morrisalp / bandit_demo.py
Last active July 6, 2022 07:35
Demo for illustrating multi-armed bandit algorithms. The methods of the class Agent are placeholders and may be changed.
import numpy as np
class Agent:
def __init__(self, n_machines=10):
self.t = 0 # current time step
self.n_machines = n_machines
def pick_machine(self):
# returns index of machine
@morrisalp
morrisalp / flask_caching_demo.py
Created August 23, 2020 17:59
demo of flask_caching - calculating pi with ζ(2)
from flask import Flask
from flask_caching import Cache
app = Flask(__name__)
app.config.from_mapping({"CACHE_TYPE": "simple"})
cache = Cache(app)
def approximate_pi(n):
output = 0
for i in range(1, n):
@morrisalp
morrisalp / top_k_categorical_accuracy.py
Created July 24, 2020 07:46
top K categorical accuracy for numpy arrays (sklearn predict_proba outputs)
import numpy as np
def top_k_categorical_accuracy(y_true, y_pred_proba, k=1):
return np.equal(np.argsort(y_pred_proba)[:, -k:], y_true[:, None]).any(axis=1).mean()
@morrisalp
morrisalp / simple_bert.py
Last active June 17, 2020 00:53
minimal example of getting BERT embeddings for sentence, using TF 2.0 + Tensorflow Hub + HuggingFace tokenizers library
import tensorflow as tf
import tensorflow_hub as hub
from tokenizers import BertWordPieceTokenizer
from tensorflow.keras.layers import Input
from tensorflow.keras.models import Model
import numpy as np
class BERTPreprocessor:
SEP_TOKEN = '[SEP]'
@morrisalp
morrisalp / html2text.py
Created April 7, 2020 11:45
sane text extraction given html string, using BeautifulSoup
from bs4 import BeautifulSoup as bs
def html2text(html):
soup = bs(html, features='lxml')
for script in soup(["script", "style"]):
script.decompose()
for br in soup.find_all("br"):
br.replace_with("\n")
return soup.get_text(separator=' ').strip()