Created
June 29, 2019 20:33
-
-
Save etienne87/4904b3207d31cdb7ff80e581807ff610 to your computer and use it in GitHub Desktop.
trying to draw mickey using binary cross entropy
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import torch | |
import torch.nn as nn | |
import torch.optim as optim | |
import torch.nn.functional as F | |
import matplotlib.pyplot as plt | |
import numpy as np | |
import random | |
def bce_loss_with_logits(x, y): | |
z = y.float() | |
x = x.squeeze() | |
losses = F.relu(x) - x * z + torch.log(1 + torch.exp(-torch.abs(x))) | |
return losses.mean() | |
def bce_focal_loss(x, y): | |
alpha = 0.25 | |
gamma = 2 | |
z = y.float() | |
x = x.squeeze() | |
p = x.sigmoid() | |
pt = torch.where(z>0, p, 1-p) | |
weights = (1-pt).pow(gamma) | |
#weights = torch.where(z > 0, alpha * weights, alpha * weights) | |
losses = F.relu(x) - x * z + torch.log(1 + torch.exp(-torch.abs(x))) | |
#losses = F.binary_cross_entropy_with_logits(x, z) | |
loss = losses * weights | |
return loss.mean() | |
def softmax_focal_loss(sx, y): | |
gamma = 2 | |
r = torch.arange(x.size(0)) | |
pt = F.softmax(x, dim=1)[r,y] | |
weights = (1-pt).pow(gamma) #should normalize? | |
ce = -F.log_softmax(x, dim=1)[r,y] | |
loss = weights * ce | |
return loss.mean() | |
def make_pic_distribution(): | |
import cv2 | |
img = cv2.imread('mickey.jpg', cv2.IMREAD_GRAYSCALE) | |
x1, x2 = np.where(img > -1) | |
x1 = img.shape[0]-x1 | |
x = np.concatenate([x2[:,None], x1[:,None]], axis=1) | |
y = (img>3).reshape(-1) | |
return x, y | |
def make_net(cin=2, hidden=64, cout=1): | |
return nn.Sequential(nn.BatchNorm1d(cin), | |
nn.Linear(cin, hidden), nn.ELU(), | |
nn.Linear(hidden, hidden), nn.ELU(), | |
nn.Linear(hidden, cout)) | |
class ResNet(nn.Module): | |
def __init__(self, cin=2, hidden=64, num_layers=5): | |
super(ResNet, self).__init__() | |
self.prepare = nn.Sequential(nn.BatchNorm1d(cin), | |
nn.Linear(cin, hidden), | |
nn.ReLU()) | |
self.residuals = nn.ModuleList() | |
for _ in range(num_layers): | |
self.residuals.append(nn.Sequential(nn.Linear(hidden, hidden), nn.BatchNorm1d(hidden), nn.ReLU())) | |
self.out = nn.Linear(hidden, 1) | |
def forward(self, x): | |
x = self.prepare(x) | |
for res in self.residuals: | |
x = x + res(x) | |
return self.out(x) | |
#N = 50000 | |
#Ntr = N*50/100 | |
#x = torch.randn(N, 4) | |
# xtr = x[:Ntr] | |
# ytr = xtr.norm(dim=1) | |
# ytr = ytr < (ytr.mean()-2*ytr.std()) | |
# xval = x[Ntr:] | |
# yval = xval.norm(dim=1) | |
# yval = yval < (yval.mean()-2*yval.std()) | |
x, y = make_pic_distribution() | |
N = len(x) | |
Ntr = N*100/100 | |
x = torch.from_numpy(x).float() | |
#x = (x-x.mean(dim=0))/(x.std(dim=0) + 1e7) | |
y = torch.from_numpy(y.astype(np.uint8)) | |
idx = range(N) | |
random.shuffle(idx) | |
x = x[idx] | |
y = y[idx] | |
xtr = x[:Ntr] | |
ytr = y[:Ntr] | |
xval = x[Ntr:] | |
yval = y[Ntr:] | |
ytrnp = ytr.numpy().astype(np.int32) | |
cuda = 1 | |
hidden = 256 | |
net1 = ResNet(num_layers=10) | |
net2 = ResNet(num_layers=10) | |
if cuda: | |
xtr = xtr.cuda() | |
ytr = ytr.cuda() | |
net1.cuda() | |
net2.cuda() | |
p = np.array([0.8, 0.2], dtype=np.float32) | |
p = p[ytrnp] | |
p = p/p.sum() | |
batchsize = 1024*5 | |
net1, net2 = net1.train(), net2.train() | |
opt1 = optim.Adam(net1.parameters(), lr=0.1, betas=(0.9, 0.99), weight_decay=1e-5) | |
opt2 = optim.Adam(net2.parameters(), lr=0.1, betas=(0.9, 0.99), weight_decay=1e-5) | |
for i in range(1000): | |
idx = np.random.choice(np.arange(0, len(ytr)), size=batchsize, p=p) | |
bx = xtr[idx] | |
by = ytr[idx] | |
opt1.zero_grad() | |
out = net1(bx) | |
loss1 = bce_loss_with_logits(out, by) | |
loss1.backward() | |
opt1.step() | |
opt2.zero_grad() | |
out = net2(bx) | |
loss2 = bce_focal_loss(out, by) | |
loss2.backward() | |
opt2.step() | |
if i%100 == 0: | |
print('loss1: ', loss1.item(), ' loss2: ', loss2.item()) | |
net1.cpu() | |
net2.cpu() | |
net1.eval() | |
net2.eval() | |
xval = xtr.cpu() | |
yval = ytr.cpu() | |
y_hat = (net1(xval)>=0).squeeze() | |
Error = (y_hat != yval).float().mean() | |
y_hat2 = (net2(xval)>=0).squeeze() | |
Error2 = (y_hat2 != yval).float().mean() | |
print('BCE Error: ', Error, ' FocalBCE Error2: ', Error2) | |
plt.subplot(311) | |
plt.scatter(xval[yval, 0], xval[yval, 1], marker='.', color='b', s=1) | |
plt.scatter(xval[~yval, 0], xval[~yval, 1], marker='.', color='r', s=1) | |
plt.subplot(312) | |
plt.scatter(xval[y_hat, 0], xval[y_hat, 1], marker='.', color='b', s=1) | |
plt.scatter(xval[~y_hat, 0], xval[~y_hat, 1], marker='.', color='r', s=1) | |
plt.subplot(313) | |
plt.scatter(xval[y_hat2, 0], xval[y_hat2, 1], marker='.', color='b', s=1) | |
plt.scatter(xval[~y_hat2, 0], xval[~y_hat2, 1], marker='.', color='r', s=1) | |
# xerr = xval[y_hat != yval] | |
# yerr = yval[y_hat != yval] | |
# | |
# plt.subplot(312) | |
# plt.scatter(xerr[yerr, 0], xerr[yerr, 1], marker='.', color='b', s=1) | |
# plt.scatter(xerr[~yerr, 0], xerr[~yerr, 1], marker='.', color='r', s=1) | |
# | |
# xerr = xval[y_hat2 != yval] | |
# yerr = yval[y_hat2 != yval] | |
# | |
# plt.subplot(313) | |
# plt.scatter(xerr[yerr, 0], xerr[yerr, 1], marker='.', color='b', s=1) | |
# plt.scatter(xerr[~yerr, 0], xerr[~yerr, 1], marker='.', color='r', s=1) | |
plt.show() |
Author
etienne87
commented
Jun 29, 2019
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment