def load_image(imgdir,
size,
norm=True,
median=True,
median_size=3,
mean=True,
consider=True):
"""
input: image directory
output: numpy array
"""
images = []
imglist = os.listdir(imgdir)
imglist.sort()
for filename in imglist:
if not filename.startswith('.'):
if len(filename) > 10:
img = Image.open(os.path.join(imgdir, filename))
if img is not None:
img = img.resize(size)
img = np.array(img, dtype=np.float32)
if consider is True:
img = make_considered_picture(img=img, level=4)
if norm is True:
img = img / 255.
if median is True:
img = ndimage.median_filter(img, median_size)
images.append(img)
img_array = np.array(images, dtype=np.float32)
print("---------- Image Load Done")
return img_array
def main():
# 画像の読み込み
if args.mnist:
(x_train, y_train), (x_test, y_test) = mnist.load_data()
else:
x_data = []
x_list = glob(os.path.join(data_path, "*.png"))
for path in tqdm(ls):
img = Image.open(path)
img = img.resize((100, 100))
img = np.array(img, dtype=np.float32)
x_data.append(img)
x_data = np.array(x_data)
img_ori = np.zeros((224, 224, 3))
level = 4
img_fish, pos = fish.make_considered_picture(img=img_ori, level=level, return_array=1)
pos_x = rotation_xyz(pos, alpha=np.pi/2, mode="x")
print("level: ", level)
print(img_fish.shape)
# from keras.utils import np_utils
from tqdm import tqdm
from PIL import Image
(x_train, t_train), (x_test, t_test) = cifar100.load_data()
# load sphere array
# x_train, t_train = x_train[:50], t_train[:50000]
# x_test, t_test = x_test[:10], t_test[:10000]
level = 6
hemisphere = 60
r = 1
Q = 2**level
img_ori = np.zeros((224, 224, 3))
img_fish, pos = fish.make_considered_picture(img=img_ori,
level=level,
return_array=1)
# process equi image
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train, x_train_sp = it.make_sphere_batch(train_generator=x_train,
pos=pos,
start='random',
hemisphere=hemisphere,
r=r,
return_equ=False)
x_test, x_test_sp = it.make_sphere_batch(train_generator=x_test,
pos=pos,
start='random',
hemisphere=hemisphere,
def main(args):
data_dir, save_dir, mode = args.data_dir, args.save_dir, args.mode
# pos
img_ori = np.zeros((224, 224, 3))
level = 4
img_fish, pos = fish.make_considered_picture(img=img_ori,
level=level,
return_array=1)
print("level: ", level)
print(img_fish.shape)
# data load
ndar = np.load(data_dir)
train = ndar['train']
test = ndar['test']
x_train, y_train = make_data(train)
x_test, y_test = make_data(test)
x_train = get_sphere_array(x_train, pos)
x_test = get_sphere_array(x_test, pos)
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
if x_train.shape[-1] == 3:
print('change channel')
x_train = x_train[:, :, :, 0]
x_test = x_test[:, :, :, 0]
# if len(x_test) > 1000:
# tmp = 1000
# x_train = np.concatenate((x_train, x_test[tmp:]), axis=0)
# x_test = x_test[:tmp]
# y_train = np.concatenate((y_train, y_test[tmp:]), axis=0)
# y_test = y_test[:tmp]
x_test = x_test[:1000]
y_test = y_test[:1000]
# パラメータの設定
hemisphere = 50
Q = 2**level
title = 'SphereConv'
batch_size = 100
max_epochs = 10
out_num = 10
print('conducting in SphereConv')
# パラメータの吐き出し
with open(save_dir + 'setting.txt', 'w') as f:
f.write('save_dir: ' + save_dir + '\n')
f.write('hemisphere: ' + str(hemisphere) + '\n')
f.write('level: ' + str(level) + '\n')
f.write('Q: ' + str(Q) + '\n')
f.write('batch_size: ' + str(batch_size) + '\n')
f.write('max_epochs: ' + str(max_epochs) + '\n')
f.write('out_num :' + str(out_num) + '\n')
f.write('train : 10000' + '\n')
f.write('test 1000 :' + '\n')
f.write('COMMENT: SGD, SphereConv\n')
print('train size;', x_train.shape)
network = SphereConv(input_dim=(3, 6, Q + 2, 2 * Q + 2),
conv_param={
'filter_num': 32,
'filter_size': 7,
'pad': 0,
'stride': 1
},
hidden_size=128,
output_size=out_num,
weight_init_std='he',
level=level)
trainer = Trainer(network,
x_train,
y_train,
x_test[:100],
y_test[:100],
epochs=max_epochs,
mini_batch_size=batch_size,
optimizer='SGD',
optimizer_param={'lr': 0.001},
evaluate_sample_num_per_epoch=100)
trainer.train()
# パラメータの保存
network.save_params(save_dir + "params.pkl")
print("Saved Network Parameters!")
try:
print('save acc')
a = np.array(trainer.train_acc_list)
b = np.array(train_acc_list.test_acc_list)
np.savez('acc.npz', train=a, test=b)
except:
print('error in saving acc')
# グラフの描画
markers = {'train': 'o', 'test': 's'}
x = np.arange(max_epochs)
plt.plot(x, trainer.train_acc_list, marker='o', label='train', markevery=2)
plt.plot(x, trainer.test_acc_list, marker='s', label='test', markevery=2)
plt.xlabel("epochs")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
plt.legend(loc='lower right')
plt.savefig(save_dir + "test.png")
def main(args):
data_dir, save_dir = args.data_dir, args.save_dir
(x_train, y_train), (x_test, y_test) = mnist.load_data()
# y_test = np_utils.to_categorical(y_test)
# y_train = np_utils.to_categorical(y_train)
img_ori = np.zeros((224, 224, 3))
level = 4
img_fish, pos = fish.make_considered_picture(img=img_ori, level=level, return_array=1)
pos_x = rotation_xyz(pos, alpha=np.pi/2, mode="x")
print("level: ", level)
print(img_fish.shape)
print('making train samples')
# x_train, y_train = make_spherical_mnist(x=x_train[:10000], y=y_train[:10000])
# output_train = make_spherical_mnist(x=x_train[:30000], y=y_train[:30000])
output_train = make_spherical_mnist(x=x_train[:10000], y=y_train[:10000], pos=pos_x)
print("arraying")
output_train = np.array(output_train)
print('making test samples')
# x_test, y_test = make_spherical_mnist(x=x_test[:1000], y=y_test[:1000])
output_test = make_spherical_mnist(x=x_test, y=y_test, pos=pos_x)
output_test = np.array(output_test)
title = 'sphere_mnist_pica_level4_cluster7x_10000.npz'
print(title)
path_sp = os.path.join(save_dir, title)
# np.save(path_sp, output)
np.savez(path_sp,
train=output_train,
test=output_test,
)
pos_xz = rotation_xyz(pos, alpha=np.pi/2, mode="xz")
print('making train samples')
output_train = make_spherical_mnist(x=x_train[:10000], y=y_train[:10000], pos=pos_xz)
print("arraying")
output_train = np.array(output_train)
print('making test samples')
output_test = make_spherical_mnist(x=x_test, y=y_test, pos=pos_xz)
output_test = np.array(output_test)
title = 'sphere_mnist_pica_level4_cluster7xz_10000.npz'
# title = 'sphere_mnist_unique_v1_cluster5_10000.npz'
print(title)
path_sp = os.path.join(save_dir, title)
# np.save(path_sp, output)
np.savez(path_sp,
train=output_train,
test=output_test,
)
print('making train samples')
output_train = make_spherical_mnist(x=x_train[:10000], y=y_train[:10000])
print("arraying")
output_train = np.array(output_train)
print('making test samples')
output_test = make_spherical_mnist(x=x_test, y=y_test)
output_test = np.array(output_test)
title = 'sphere_mnist_unique_v1_cluster7_10000.npz'
print(title)
path_sp = os.path.join(save_dir, title)
# np.save(path_sp, output)
np.savez(path_sp,
train=output_train,
test=output_test,
)
def main(args):
#######################
# ディレクトリパスの読み込み
#######################
# 今日の日付
today_time = dt.datetime.today().strftime("%Y_%m_%d")
# 各ディレクトリの設定
DATA_DIR, SAVE_DIR = \
args.data_dir, args.save_dir
# 'SAVE_DIR'に今日の日付を足す
do_time = os.datetime.today().strftime("%H_%M_%S")
SAVE_DIR = os.path.join(SAVE_DIR, today_time, do_time)
if not os.path.isdir(SAVE_DIR):
os.makedirs(SAVE_DIR)
#######################
# Hyper Parameter
#######################
param_dic = {}
config = configparser.ConfigParser()
config.read('config.ini')
param_sphere = config['sphere']
level = param_dic['level'] = int(param_sphere['level'])
hemisphere = param_dic['hemisphere'] int(param_sphere['hemisphere'])
Q = param_dic['Q'] = 2**level
title = param_dic['title'] = param_sphere['title']
param = config['parameter']
batch_size = param_dic['batch_size'] = param['batch_size']
max_epochs = param_dic['max_epochs'] = param['max_epochs']
evaluate_num = param_dic['evaluate_sample_num_per_epoch'] = param['evaluate_num']
out_num = param_dic['out_num'] = param['out_num']
filter_num = param_dic['filter_num'] = param['filter_num']
filter_size = param_dic['filter_size'] = param['filter_size']
padding = param_dic['padding'] = param['padding']
stride = param_dic['stride'] = param['stride']
hidden_size = param_dic['hidden_size'] = param['hidden_size']
optimizer = param_dic['optimizer'] = param['optimizer']
lr = param_dic['lr'] = param['optimizer']
weight_init_std = param_dic['weight_init_std'] = param['weight_init_std']
# 保存
with open(SAVE_DIR + 'setting.txt', 'a') as f:
for key, value in param_dic.items():
f.write(str(key) + ':' + str(value))
f.write('\n')
#######################
# 球面座標の獲得
#######################
img_ori = np.zeros((224, 224, 3))
img_fish, pos = fish.make_considered_picture(img=img_ori, level=level, return_array=1)
# print("level: ", level)
# print(img_fish.shape)
#######################
# Load Data
#######################
ndar = np.load(data_dir)
train = ndar['train']
test = ndar['test']
#######################
# Make data
# データをネットワークに適した形に変形する
#######################
x_train, y_train = make_data(train)
x_test, y_test = make_data(test)
#######################
# 球面画像の獲得
#######################
x_train = get_sphere_array(x_train, pos)
x_test = get_sphere_array(x_test, pos)
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
if x_train.shape[-1] == 3:
print('change channel')
x_train = x_train[:, :, :, 0]
x_test = x_test[:, :, :, 0]
# if len(x_test) > 1000:
# tmp = 1000
# x_train = np.concatenate((x_train, x_test[tmp:]), axis=0)
# x_test = x_test[:tmp]
# y_train = np.concatenate((y_train, y_test[tmp:]), axis=0)
# y_test = y_test[:tmp]
x_test = x_test[:1000]
y_test = y_test[:1000]
#######################
# Train
#######################
network = SphereConv(input_dim=(3, 6, Q+2, 2*Q+2),
conv_param={'filter_num': filter_num,
'filter_size': filter_size,
'pad': padding, 'stride': stride},
hidden_size=hidden_size, output_size=out_num,
weight_init_std=weight_init_std, level=level)
trainer = Trainer(network, x_train, y_train, x_test[:100], y_test[:100],
epochs=max_epochs, mini_batch_size=batch_size,
optimizer=optimizer, optimizer_param={'lr': lr},
evaluate_sample_num_per_epoch=evaluate_num)
trainer.train()
#######################
# パラメータの保存
#######################
network.save_params(save_dir+"params.pkl")
print("Saved Network Parameters!")
try:
print('save acc')
a = np.array(trainer.train_acc_list)
b = np.array(train_acc_list.test_acc_list)
np.savez('acc.npz', train=a, test=b)
except:
print('error in saving acc')
# グラフの描画
markers = {'train': 'o', 'test': 's'}
x = np.arange(max_epochs)
plt.plot(x, trainer.train_acc_list, marker='o', label='train', markevery=2)
plt.plot(x, trainer.test_acc_list, marker='s', label='test', markevery=2)
plt.xlabel("epochs")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
plt.legend(loc='lower right')
plt.savefig(save_dir+"test.png")