def plot_W_from_pixel(model, i):
H = int(np.sqrt(model.W.shape[0]))
W = H
toplot = model.W[:, i]
utils.data_plot(toplot[None, :])
climabs = np.abs(toplot).max()
plt.clim(-climabs, climabs)
plt.set_cmap('gray')
import lane_spline_area as lsa
from utils.data_plot import *
import json
if __name__ == '__main__':
offline_annotation = True # 离线标注数据
normalize = False # 拉框/标点数据归一化处理
offline_annotation_path = '/home/huhaoyu/Downloads/3d标点离线标注/annotation' # 标注结果存放文件夹
# offline_annotation_path = '/home/huhaoyu/Downloads/离线标注/annotation'
# offline_annotation_path = '/home/huhaoyu/Downloads/车道画线_FLV_V10/车道画线_FLV_V10'
# 选择评测类型
task_dict = {
1: 'box',
2: 'point',
3: 'line'
}
task = task_dict[2]
# todo 标注结果可视化
if offline_annotation:
old_cam_ann, new_cam_ann, ann_num = convert_annotation_offline(offline_annotation_path) # 离线数据格式转换
output_dict = eval_offline(task, old_cam_ann, new_cam_ann, ann_num, normalize) # 评测
data_plot(output_dict, task) # 可视化
else:
# 预留的线上数据接口
test_img = get_test_image(task) # 获取数据列表
# output_list = eval(task, test_img) # 获取评估结果
# output_list = eval_offline(task, old_cam_ann, new_cam_ann, ann_num)
# data_plot(output_list) # 可视化
print(
tabulate([[
'PcDGAN',
str(PcDGAN_MAE_overall) + '+/-' + str(PcDGAN_MAE_std_overall),
str(PcDGAN_KDE_overall) + '+/-' + str(PcDGAN_KDE_std_overall),
str(PcDGAN_diver_overall) + '+/-' +
str(PcDGAN_diver_std_overall)
]],
headers=[
'Model', 'Label Score', 'Probability Density',
'Diversity'
]))
ind = np.random.choice(X.shape[0], replace=False, size=1000)
data_plot(
X[ind], equation,
'./' + folder + '/Evaluation/PcDGAN/Data_' + str(args.id) + '.png')
dist_anim(
Xs, conds, equation, './' + folder +
'/Evaluation/PcDGAN/out_put_samples_' + str(args.id) + '.mp4')
plt.figure(figsize=(18, 12))
plt.rc('font', size=45)
plt.plot(conds, PcDGAN_diver, color='#003F5C')
plt.fill_between(conds,
PcDGAN_diver - PcDGAN_diver_std,
PcDGAN_diver + PcDGAN_diver_std,
facecolor='#003F5C',
edgecolor="#003F5C",
alpha=0.3)
# load nmist data
(X_dtr, y_dtr), (X_dvl, y_dvl), (X_dts, y_dts) = utils.load_data()
# one hot encoding of labels
ohe = preprocess.OneHotEncoder(sparse=False)
ohe.fit(y_dtr[:, None])
Y_dtr = ohe.transform(y_dtr[:, None])
Y_dvl = ohe.transform(y_dvl[:, None])
Y_dts = ohe.transform(y_dts[:, None])
# plot example data
importlib.reload(utils)
h_fig, h_ax = plt.subplots(nrows=4, ncols=5)
for ax in h_ax.ravel():
plt.axes(ax)
utils.data_plot(X_dtr, y_dtr)
# get augmented data: horizontal flip
X_dtr_flip = utils.data_ravel(utils.data_unravel(X_dtr)[:, :, ::-1])
importlib.reload(utils)
h_fig, h_ax = plt.subplots(nrows=4, ncols=5)
for ax in h_ax.ravel():
plt.axes(ax)
utils.data_plot(X_dtr_flip, y_dtr)
X_dtr_all = np.concatenate((X_dtr, X_dtr_flip), axis=0)
Y_dtr_all = np.concatenate((Y_dtr, Y_dtr), axis=0)
indx_reorder = np.random.permutation(X_dtr_all.shape[0])
X_dtr_all = X_dtr_all[indx_reorder]
# X_dvl = utils.data_unravel(X_dvl)[:, :, :, None]
# X_dts = utils.data_unravel(X_dts)[:, :, :, None]
# one hot encoding of labels
ohe = preprocess.OneHotEncoder(sparse=False)
ohe.fit(y_dtr[:, None])
Y_dtr = ohe.transform(y_dtr[:, None])
Y_dvl = ohe.transform(y_dvl[:, None])
Y_dts = ohe.transform(y_dts[:, None])
# plot example data
importlib.reload(utils)
h_fig, h_ax = plt.subplots(nrows=4, ncols=5)
for ax in h_ax.ravel():
plt.axes(ax)
utils.data_plot(X_dtr, y_dtr)
M0 = X_dtr.shape[1]
M1 = 512
batch_size = 32
total_steps = 1000000
learning_rate = 0.03
wd_l2 = 0.001 # weight decay
""" define a tf graph for computation """
graph = tf.Graph()
with graph.as_default():
""" constant, variable and placeholder """
# place holder
X0_in = tf.placeholder(dtype=tf.float32, shape=[batch_size, M0])
with tf.summary.FileWriter('./model_log') as writer:
writer.add_graph(session.graph)
tf.global_variables_initializer().run()
if yn_load_file:
model.load_parameters(filedir='./model_save', filename='RBM_tf')
model.params_dict_to_tensor()
gibss_result = session.run(gibbs_outcome, feed_dict={x0_in: x_batch})
##
utils.data_plot(gibss_result[1], n=10)
utils.data_plot(model.dict_params['w'][:, :10].transpose(), n=10)
##
with tf.Session(graph=model.graph) as session:
tf.global_variables_initializer().run()
temp0 = session.run(model.cal_energy(x0=x_in))
print(temp0)
session.run(
model.load_parameters(filedir='./model_save', filename='RBM_tf'))
temp1 = session.run(model.cal_energy(x0=x_in))
print(temp1)
##
X_dvl = utils.data_unravel(X_dvl)[:, :, :, None]
X_dts = utils.data_unravel(X_dts)[:, :, :, None]
# one hot encoding of labels
ohe = preprocess.OneHotEncoder(sparse=False)
ohe.fit(y_dtr[:, None])
Y_dtr = ohe.transform(y_dtr[:, None])
Y_dvl = ohe.transform(y_dvl[:, None])
Y_dts = ohe.transform(y_dts[:, None])
# plot example data
importlib.reload(utils)
h_fig, h_ax = plt.subplots(nrows=4, ncols=5)
for ax in h_ax.ravel():
plt.axes(ax)
utils.data_plot(X_dtr, y_dtr)
##
X_dvl_flp = X_dvl[:, :, ::-1]
X_dvl_rot = (np.rot90(X_dvl, k=1, axes=[1, 2]))
X_tr = np.reshape(X_dtr, [X_dtr.shape[0], -1])
X_vl = np.reshape(X_dvl, [X_dvl.shape[0], -1])
X_vl_rot = np.reshape(X_dvl_rot, [X_dvl_rot.shape[0], -1])
X_tr = utils.data_binarize(X_tr, threshold=0.5, states='0,1')
X_vl = utils.data_binarize(X_vl, threshold=0.5, states='0,1')
X_vl_rot = utils.data_binarize(X_vl_rot, threshold=0.5, states='0,1')
utils.data_plot(X_vl, n=100, yn_random=False)
plt.suptitle('validation, original')
# one hot encoding of labels
ohe = preprocess.OneHotEncoder(sparse=False)
ohe.fit(y_dtr[:, None])
Y_dtr = ohe.transform(y_dtr[:, None])
Y_dvl = ohe.transform(y_dvl[:, None])
Y_dts = ohe.transform(y_dts[:, None])
# plot example data
importlib.reload(utils)
h_fig, h_ax = plt.subplots(nrows=4, ncols=5)
for ax in h_ax.ravel():
plt.axes(ax)
utils.data_plot(X_dtr, y_dtr)
# get augmented data: horizontal flip
X_dtr_flip = X_dtr[:, :, ::-1]
h_fig, h_ax = plt.subplots(nrows=4, ncols=5)
for ax in h_ax.ravel():
plt.axes(ax)
utils.data_plot(X_dtr_flip, y_dtr)
X_dtr_all = np.concatenate((X_dtr, X_dtr_flip), axis=0)
Y_dtr_all = np.concatenate((Y_dtr, Y_dtr), axis=0)
indx_reorder = np.random.permutation(X_dtr_all.shape[0])
X_dtr_all = X_dtr_all[indx_reorder]
Y_dtr_all = Y_dtr_all[indx_reorder]