def plot_evolvement():
import sklearn.datasets
import torch
import sys
sys.path.append('../')
from models import csnn_learnable_r
epochs = [0, 5, 10, 50, 100, 500]
outputDir = '../results/evolvement/'
# load models
models = []
for epoch in epochs:
model = csnn_learnable_r(2, 64, bias=False)
PATH = outputDir + 'csnn_epoch{}.pth'.format(epoch)
model.load_state_dict(torch.load(PATH))
models.append(model)
print('load models successfully')
# load confidences
confidences = []
x_lin = None
y_lin = None
for epoch in epochs:
dir = outputDir + 'moons_confidence_alpha1_epoch{}.npz'.format(epoch)
f = np.load(dir)
confidences.append(f['c'])
if x_lin is None:
x_lin = f['a']
y_lin = f['b']
print('load confidences successfully')
# Moons
noise = 0.1
# sklearn has no random seed, it depends on numpy to get random numbers
np.random.seed(0)
x_train, y_train = sklearn.datasets.make_moons(n_samples=1500, noise=noise)
x_train0 = x_train
mask = y_train.astype(np.bool)
mean = np.mean(x_train, axis=0)
std = np.std(x_train, axis=0)
print('mean, std', mean, std)
x_train = (x_train - mean) / std / np.sqrt(2)
# support circles
fig, axs = plt.subplots(2, 3, figsize=(10, 6.67), dpi=300)
alpha = 1
for i in range(len(epochs)):
row = i // 3
col = i % 3
w = models[i].fc1.weight.data.numpy()
r = models[i].r.detach().numpy()
center = w / alpha
radius2 = r * r + np.sum(w * w, axis=1) * (1 / alpha / alpha - 1)
radius = np.sqrt(radius2)
# cycle = plt.rcParams['axes.prop_cycle'].by_key()['color']
N = w.shape[0]
color = 1. / N * np.arange(0, N)
for j in range(w.shape[0]):
circle = plt.Circle(center[j],
radius[j],
color=cm.jet(color[j]),
fill=False)
axs[row, col].add_artist(circle)
# axs[0].axis('equal')
axs[row, col].scatter(x_train[mask, 0], x_train[mask, 1], s=1, c='r')
axs[row, col].scatter(x_train[~mask, 0], x_train[~mask, 1], s=1)
axs[row, col].set(xlim=(-2.4, 2.4), ylim=(-2.4, 2.4))
axs[row, col].set_aspect('equal')
if row == 0: axs[row, col].set_xticklabels([])
if col != 0: axs[row, col].set_yticklabels([])
axs[row, col].set_title('epoch={}'.format(epochs[i]))
plt.savefig(outputDir + 'evolvement_circles.png', dpi=300)
# confidence
fig, axs = plt.subplots(2, 3, figsize=(10, 5), dpi=300)
for i in range(len(epochs)):
row = i // 3
col = i % 3
level = np.linspace(0.5, 1., 21)
contour = axs[row, col].contourf(x_lin,
y_lin,
confidences[i],
cmap=plt.get_cmap('inferno'),
levels=level) # , extend='both')
fig.colorbar(contour, ax=axs[row, col], format='%.2f')
axs[row, col].scatter(x_train[mask, 0], x_train[mask, 1], s=1, c='r')
axs[row, col].scatter(x_train[~mask, 0], x_train[~mask, 1], s=1)
axs[row, col].set(xlim=(-2.4, 2.4), ylim=(-2.4, 2.4))
axs[row, col].set_aspect('equal')
if row == 0: axs[row, col].set_xticklabels([])
if col != 0: axs[row, col].set_yticklabels([])
axs[row, col].set_title('epoch={}'.format(epochs[i]))
plt.savefig(outputDir + 'evolvement_conf.png', dpi=300)
losses = []
accs_validate = []
losses_validate = []
# pre_train
for run in range(runs):
np.random.seed(seeds[run])
torch.manual_seed(seeds[run])
dl_train = torch.utils.data.DataLoader(ds_train,
batch_size=batchSize,
shuffle=True,
drop_last=False)
dl_test = torch.utils.data.DataLoader(ds_test,
batch_size=x_validate.shape[0],
shuffle=False)
model = csnn_learnable_r(2, features, bias=BIAS)
if learnable_r:
model.set_lambda(LAMBDA)
model.set_miu(MIU)
optimizer = torch.optim.Adam(model.parameters(),
lr=learningRate,
weight_decay=l2Penalty)
accuracy, loss, accuracy_validate, loss_validate = pre_train(model,
optimizer,
dl_train,
dl_test,
x_train,
y_train,
x_validate,
y_validate,
run,
def plot_radius_penalty_effect():
import sklearn.datasets
import torch
import sys
sys.path.append('../')
from models import csnn_learnable_r
# lambdas = [0.00, 0.01, 0.02, 0.04, 0.16, 0.64]
lambdas = [0.00, 0.02, 0.04, 0.64]
outputDir = '../results/radius_penalty_impact/'
# load models
models = []
for l in lambdas:
model = csnn_learnable_r(2, 64, bias=False)
PATH = outputDir + 'lambda_{:.2f}/csnn_run0_epoch500.pth'.format(l)
model.load_state_dict(torch.load(PATH))
models.append(model)
print('load models successfully')
# load rs
rs = []
for l in lambdas:
f = np.load(outputDir +
'lambda_{:.2f}/mean_std_accs_aucs_net4.npz'.format(l))
rs.append(f['l'])
print('load radius successfully')
# load confidences
confidences = []
x_lin = None
y_lin = None
for l in lambdas:
f = np.load(outputDir +
'lambda_{:.2f}/moons_confidence_alpha1.npz'.format(l))
confidences.append(f['c'])
if x_lin is None:
x_lin = f['a']
y_lin = f['b']
print('load confidences successfully')
# Moons
noise = 0.1
# sklearn has no random seed, it depends on numpy to get random numbers
np.random.seed(0)
x_train, y_train = sklearn.datasets.make_moons(n_samples=1500, noise=noise)
x_train0 = x_train
mask = y_train.astype(np.bool)
mean = np.mean(x_train, axis=0)
std = np.std(x_train, axis=0)
print('mean, std', mean, std)
x_train = (x_train - mean) / std / np.sqrt(2)
fig, axs = plt.subplots(len(lambdas), 3, figsize=(10, 10.6), dpi=300)
alpha = 1
for i in range(len(lambdas)):
# support circles
w = models[i].fc1.weight.data.numpy()
r = models[i].r.detach().numpy()
center = w / alpha
radius2 = r * r + np.sum(w * w, axis=1) * (1 / alpha / alpha - 1)
radius = np.sqrt(radius2)
N = w.shape[0]
col = 1. / N * np.arange(0, N)
for j in range(w.shape[0]):
circle = plt.Circle(center[j],
radius[j],
color=cm.jet(col[j]),
fill=False)
axs[i, 0].add_artist(circle)
axs[i, 0].scatter(x_train[mask, 0], x_train[mask, 1], s=1, c='r')
axs[i, 0].scatter(x_train[~mask, 0], x_train[~mask, 1], s=1)
axs[i, 0].set(xlim=(-2.4, 2.4), ylim=(-2.4, 2.4))
axs[i, 0].set_aspect('equal')
if i != len(lambdas) - 1: axs[i, 0].set_xticklabels([])
# confidence
level = np.linspace(0.5, 1., 21)
contour = axs[i, 1].contourf(x_lin,
y_lin,
confidences[i],
cmap=plt.get_cmap('inferno'),
levels=level) # , extend='both')
fig.colorbar(contour, ax=axs[i, 1], format='%.2f')
axs[i, 1].scatter(x_train[mask, 0], x_train[mask, 1], s=1, c='r')
axs[i, 1].scatter(x_train[~mask, 0], x_train[~mask, 1], s=1)
axs[i, 1].set(xlim=(-2.4, 2.4), ylim=(-2.4, 2.4))
axs[i, 1].set_aspect('equal')
axs[i, 1].set_yticklabels([])
if i != len(lambdas) - 1: axs[i, 1].set_xticklabels([])
# radius
epochs = np.arange(101) * 5
axs[i, 2].plot(epochs, rs[i][:, 0], lw=2, label='$||r||_\infty$')
axs[i, 2].plot(epochs,
rs[i][:, 1] / 8.,
lw=2,
label='$||r||_2/\sqrt{n_{hidden}}$')
axs[i, 2].fill_between(epochs,
rs[i][:, 1] / 8. - rs[i][:, 2],
rs[i][:, 1] / 8. + rs[i][:, 2],
facecolor='orange',
alpha=0.2)
axs[i, 2].axis([0, 500, 0., 1.8])
axs[i, 2].set_ylabel('r norm')
if i == len(lambdas) - 1: axs[i, 2].set_xlabel('epochs')
axs[i, 2].legend()
axs[i, 2].yaxis.tick_right()
axs[i, 2].yaxis.set_label_position("right")
if i != len(lambdas) - 1: axs[i, 2].set_xticklabels([])
plt.savefig(outputDir + 'evolvement.png', dpi=300)
outputDir = '/home/hh/data/ngsim/combined_dataset/'
seed = 0
np.random.seed(seed)
torch.manual_seed(seed)
for LAMBDA in LAMBDAS:
print('lambda {:.2f}'.format(LAMBDA))
for run in range(runs):
dl_train = torch.utils.data.DataLoader(ds_train,
batch_size=batchSize,
shuffle=True,
drop_last=False)
dl_test = torch.utils.data.DataLoader(ds_test,
batch_size=x_validate.shape[0],
shuffle=False)
model = csnn_learnable_r(inputs, hiddenUnits, bias=BIAS)
# model = csnn_learnable_r_3layers(inputs, hiddenUnits, bias=BIAS)
# model = Net2(inputs, hiddenUnits, bias=BIAS)
if learnable_r:
model.set_lambda(LAMBDA)
model.set_miu(MIU)
optimizer = torch.optim.Adam(model.parameters(),
lr=learningRate,
weight_decay=l2Penalty)
pre_train(model,
optimizer,
dl_train,
dl_test,
x_train,
y_train,
x_validate,