hushell

import torch

F = torch.rand(5, 10, 10) # C, H, W

M = torch.randint(8, (10, 10)) # 8 superpixels
M1hot = torch.nn.functional.one_hot(M, 8) # H, W, K

counts = M1hot.sum(dim=(0, 1)) # counts of pixels per superpixel

Favg = M1hot.view(-1, 8).T.float() @ F0.view(5, -1).T # K, C

hushell

import torch 
from torch.distributions.utils import clamp_probs
from torch.distributions import RelaxedOneHotCategorical

class QuantizeCategorical(torch.autograd.Function):
    @staticmethod
    def forward(ctx, soft_value):
        argmax = soft_value.max(-1)[1]
        hard_value = torch.zeros_like(soft_value)
        hard_value._unquantize = soft_value

hushell

hushell

hushell

hushell

wget https://repo.continuum.io/archive/Anaconda3-5.1.0-Linux-x86_64.sh
bash Anaconda3-5.1.0-Linux-x86_64.sh
conda install pytorch torchvision -c pytorch
conda install -c menpo opencv3
pip install dominate tqdm
pip install git+https://github.com/pytorch/tnt.git@master

hushell

from joblib import Parallel, delayed
import Queue
import os

# Define number of GPUs available
N_GPU = 4

# Put indices in queue
q = Queue.Queue(maxsize=N_GPU)
for i in range(N_GPU):

hushell

hushell

hushell

function ww = approx_int_func_solve(xy_pairs, xmin, xmax, intval, w0, f, lbda)
% example: 
% xy_pairs = [0, 1.32; 33.9, .55];
% xmin = 0;
% xmax = 33.9;
% intval = 38.985;
% w0 = [0.1,-2,2]
% ww = approx_int_func_solve(xy_pairs, xmin, xmax, intval, w0);

if nargin < 6