Skip to content

Instantly share code, notes, and snippets.

View janhuenermann's full-sized avatar
🤖
Working on robots

Jan Hünermann janhuenermann

🤖
Working on robots
View GitHub Profile
@janhuenermann
janhuenermann / poker-head.swift
Last active April 15, 2017 16:46
A little program that calculates poker probabilities.
//
// main.swift
// poker-head
//
// Created by Jan on 14.04.17.
// Copyright © 2017 Jan Hünermann. All rights reserved.
//
import Foundation
@janhuenermann
janhuenermann / mnist.py
Last active May 9, 2018 08:00
TensorFlow using MNIST. Example of how to implement SGD and fully-connected layers.
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
def fullyConnected(x, name, input_size, output_size=64):
with tf.variable_scope(name):
kernel = tf.get_variable(name="kernel",
initializer=tf.random_normal(shape=(input_size, output_size), stddev=0.1))
@janhuenermann
janhuenermann / mnist2.py
Created May 9, 2018 08:13
TensorFlow using MNIST. Using standard TF APIs, in 30 lines of code.
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
session = tf.Session()
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
input = tf.placeholder(tf.float32, shape=(None, 784))
x = tf.layers.dense(input, 128, activation=tf.nn.relu) # layer 1
x = tf.layers.dense(x, 128, activation=tf.nn.relu) # layer 2
x = tf.layers.dense(x, 10, activation=None) # layer 3
@janhuenermann
janhuenermann / sinkhorn.py
Last active September 13, 2023 11:49
Sinkhorn Optimal Transport Algorithm in PyTorch
import torch
@torch.jit.script
def log_optimal_transport(Z, iters: int):
m, n = Z.shape
log_mu = -torch.tensor(m).to(Z).log().expand(Z.shape[:-2] + [m])
log_nu = -torch.tensor(n).to(Z).log().expand(Z.shape[:-2] + [n])
u, v = torch.zeros_like(log_mu), torch.zeros_like(log_nu)
for _ in range(iters):
v = log_nu - torch.logsumexp(Z + u.unsqueeze(-1), dim=-2)
@janhuenermann
janhuenermann / giou.py
Last active May 21, 2021 00:22
Efficient implementation of Generalized Intersection over Union (GIoU) in PyTorch, as in https://giou.stanford.edu/
import torch
@torch.jit.script
def generalized_box_iou(boxes1, boxes2, strict: bool = False):
"""
Generalized IoU from https://giou.stanford.edu/
The boxes should be in [x0, y0, x1, y1] format
Returns a [*, N, M] pairwise matrix,
where N = boxes1.size(-2) and M = boxes2.size(-2)
"""
from torch import nn, chunk, einsum
from torch.nn.functional import softmax
from math import sqrt
class TinyAttention(nn.Module):
def __init__(self, d_attn: int, d_ffn: int):
super().__init__()
self.proj_qkv = nn.Linear(d_ffn, 3 * d_attn)
self.proj_ffn = nn.Linear(d_attn, d_ffn)
@janhuenermann
janhuenermann / min-rotated-rect.py
Last active May 21, 2021 13:00
Find the smallest rotated rectangle that covers a given polygon
import numpy as np
# Gift wrapping algorithm, from https://en.wikipedia.org/wiki/Gift_wrapping_algorithm
def find_convex_hull(poly):
"""
Returns the convex hull of the given polygon
Can be replaced using OpenCV:
```
hull = cv2.convexHull(poly, clockwise=True, returnPoints=False)
@janhuenermann
janhuenermann / polygon-intersection.py
Last active November 4, 2021 12:44
Find the intersection of two polygons in NumPy using the Sutherland–Hodgman algorithm
# Line intersection utility
# Works with broadcasting
def intersect_lines(line1, line2):
"""Computes the intersection of line1 and line2,
both described by tuple of origin and direction vector"""
p, r = line1
q, s = line2
rs = np.cross(r, s)[..., None]
d = (q - p) / (rs + 1e-24)
t = np.cross(d, s)[..., None]
@janhuenermann
janhuenermann / shoelace.py
Last active May 21, 2021 00:20
Compute the area of a polygon
def shoelace(poly):
"""Returns the area of the polygon using the shoelace formula"""
s1 = np.sum(poly[:,0] * np.roll(poly[:,1], -1))
s2 = np.sum(poly[:,1] * np.roll(poly[:,0], -1))
return np.absolute(s1 - s2) / 2.
@janhuenermann
janhuenermann / positional_encoding_2d.py
Last active May 22, 2021 21:52
Two-dimensional positional encoding in PyTorch (inspired by https://arxiv.org/abs/1706.03762)
import torch
from typing import Tuple, Optional
@torch.jit.script
def positional_encoding_2d(shape: Tuple[int, int, int], temperature: float = 1e4, scale: float = 2*math.pi,
dtype: Optional[torch.dtype] = None, device: Optional[torch.device] = None):
"""Returns the two-dimensional positional encoding as shape [d_model, h, w]"""
d_model, h, w = shape[-3:]
i = torch.arange(d_model // 4, dtype=dtype, device=device)