def run_eval_mend():
img = cv2.imread('road-car.png')[np.newaxis, :, :, :]
img = np.pad(img, ((0, 0), (32, 32), (32, 32), (0, 0)), 'reflect')
# mask = cv2.imread('road-label.png')[np.newaxis, :, :, :]
mask = cv2.imread('road-cloud0.png')[np.newaxis, :, :, :]
mask = np.pad(mask, ((0, 0), (32, 32), (32, 32), (0, 0)), 'reflect')[:, :, :, 0:1]
threshold = 244
mask[mask < threshold] = 0
mask[mask >= threshold] = 255
# cv2.imshow('', mask[0])
# cv2.waitKey(5432)
eval_list = [img, mask, img, mask]
from mod_mend_dila import init_train
C = Config('mod_mend_dila')
from mod_mend_nres import init_train
C = Config('mod_mend_nres')
inp_ground, inp_mask01, inp_grdbuf, inp_mskbuf, fetch, eval_fetch = init_train()
C.size = img.shape[1]
sess = mod_util.get_sess(C)
mod_util.get_saver_logger(C, sess)
print("||Training Check")
eval_feed_dict = {inp_ground: eval_list[0],
inp_mask01: eval_list[1],
inp_grdbuf: eval_list[2],
inp_mskbuf: eval_list[3], }
img_util.get_eval_img(mat_list=sess.run(eval_fetch, eval_feed_dict), channel=3,
img_path="%s/eval-%08d.jpg" % ('temp', 0))
def run_eval_haze():
img = cv2.imread('road-thin.png')[np.newaxis, :, :, :]
img = np.pad(img, ((0, 0), (32, 32), (32, 32), (0, 0)), 'reflect')
eval_list = [img, np.zeros_like(img[:, :, :, 0:1])]
from mod_haze_unet import init_train
inp_ground, inp_mask01, train_fetch, eval_fetch = init_train()
C = Config('mod_haze_unet')
C.size = img.shape[1]
sess = mod_util.get_sess(C)
mod_util.get_saver_logger(C, sess)
print("||Training Check")
eval_feed_dict = {inp_ground: eval_list[0],
inp_mask01: eval_list[1], }
img_util.get_eval_img(mat_list=sess.run(eval_fetch, eval_feed_dict), channel=3,
img_path="%s/eval-%08d.jpg" % ('temp', 0))
def configure(self, main_config):
self.main_config = Config.read_config(main_config)
scripts_config = Convpy.pathify(True,
self.main_config['scripts']['path'],
self.main_config['scripts']['config'])
self.scripts = Config.read_config(scripts_config)
engines_config = Convpy.pathify(True,
self.main_config['engines']['path'],
self.main_config['engines']['config'])
self.engines = Config.read_config(engines_config)
self.tmp_dir = os.path.dirname(Convpy.pathify(True, Convpy.tmp_file))
output_params = self.main_config['output']
self.output_dir = output_params['dir']
def cloud_detect(aerials):
import tensorflow as tf
from configure import Config
from utils import mod_util
from mod_cloud_detect import unet
C = Config('mod_cloud_detect')
size = aerials.shape[1]
unet_name, unet_dim = 'unet', 24
inp_aerial = tf.placeholder(tf.uint8, [None, size, size, 3])
ten_aerial = tf.to_float(inp_aerial) / 255
eva_grdcld = unet(ten_aerial, unet_dim, 4, unet_name, reuse=False, training=False)
eva_ground = eva_grdcld[:, :, :, 0:3]
eva_mask01 = eva_grdcld[:, :, :, 3:4]
sess = mod_util.get_sess(C)
mod_util.get_saver_logger(C, sess)
print("||Training Check")
# aerials_shape = list(aerials.shape[-1:])
# aerials_shape = [-1, 16] + aerials_shape
# aerials = aerials.reshape(aerials_shape)
grounds = list()
mask01s = list()
for i, aerial in enumerate(aerials):
eval_feed_dict = {inp_aerial: aerial[np.newaxis, :, :, :]}
# eval_fetch = [ten_aerial, eva_ground, eva_mask01]
eval_fetch = [eva_ground, eva_mask01]
mat_list = sess.run(eval_fetch, eval_feed_dict)
grounds.append(np.clip(mat_list[0] * 255, 0, 255).astype(np.uint8))
mask01s.append(np.clip(mat_list[1] * 255, 0, 255).astype(np.uint8))
# img_util.get_eval_img(mat_list=mat_list,channel=3, img_write=False
# img_path="%s/eval-%08d.jpg" % ('temp', 0),)
if rd.rand() < 0.01:
print('Eval:', i)
# def mats_list2jpg(mats_list, save_name):
# mats = np.concatenate(mats_list, axis=0)
# img = img_grid_reverse(mats)
# cv2.imwrite(save_name, img)
#
# mats_list2jpg(grounds, 'su_zhou/ground.jpg')
# mats_list2jpg(mask01s, 'su_zhou/mask01.jpg')
grounds = np.concatenate(grounds, axis=0)
mask01s = np.concatenate(mask01s, axis=0)
return grounds, mask01s
def cloud_removal(aerials, label1s):
import tensorflow as tf
from configure import Config
from utils import mod_util
from mod_cloud_remove_rec import auto_encoder
C = Config('mod_cloud_remove_rec')
size = aerials.shape[1]
gene_name, gene_dim = 'gene', 32
inp_ground = tf.placeholder(tf.uint8, [None, size, size, 3])
ten_ground = tf.to_float(inp_ground) / 255
inp_mask01 = tf.placeholder(tf.uint8, [None, size, size, 1])
ten_mask01 = tf.to_float(inp_mask01) / 255
ten_mask10 = (1.0 - ten_mask01)
ten_ragged = ten_ground * ten_mask10
ten_patch3 = auto_encoder(ten_ragged - ten_mask01,
gene_dim, 3, gene_name,
reuse=False, training=False)
out_ground = ten_ragged + ten_patch3 * ten_mask01
sess = mod_util.get_sess(C)
mod_util.get_saver_logger(C, sess)
print("||Training Check")
patch3s = list()
grounds = list()
for i, (aerial, label1) in enumerate(zip(aerials, label1s)):
aerial = aerial[np.newaxis, :, :, :]
label1 = label1[np.newaxis, :, :, 0:1]
eval_feed_dict = {inp_ground: aerial,
inp_mask01: label1, }
eval_fetch = [ten_patch3, out_ground]
mat_list = sess.run(eval_fetch, eval_feed_dict)
patch3s.append(np.clip(mat_list[0] * 255, 0, 255).astype(np.uint8))
grounds.append(np.clip(mat_list[1] * 255, 0, 255).astype(np.uint8))
if i % 64 == 0:
print('Eval:', i)
grounds = np.concatenate(grounds, axis=0)
patch3s = np.concatenate(patch3s, axis=0)
return grounds, patch3s,
def tf_test():
"""
https://github.com/chiralsoftware/tensorflow/blob/master/convolve-blur.py
"""
import cv2
from util.img_util import Tools
from configure import Config
C = Config()
T = Tools()
cv2.namedWindow('beta', cv2.WINDOW_KEEPRATIO)
# img = cv2.imread(os.path.join(C.aerial_dir, 'bellingham1.tif')) # test
img = cv2.imread(os.path.join(C.aerial_dir, 'austin1.tif')) # train
img = img[np.newaxis, :C.size, :C.size, :3]
blur_size = 3
channel = 3
kernel_ary = np.zeros((blur_size, blur_size, channel, channel), np.float32)
kernel_ary[:, :, 0, 0] = 1.0 / (blur_size**2)
kernel_ary[:, :, 1, 1] = 1.0 / (blur_size**2)
kernel_ary[:, :, 2, 2] = 1.0 / (blur_size**2)
inp = tf.placeholder(tf.float32, [None, C.size, C.size, 3])
ten = tf.nn.conv2d(
tf.pad(inp, ((0, 0), (2, 2), (2, 2), (0, 0)), 'REFLECT'),
tf.constant(kernel_ary), (1, 1, 1, 1), 'VALID')
tf_config = tf.ConfigProto(device_count={'GPU': 0})
sess = tf.Session(config=tf_config)
img = sess.run(ten, {inp: img})
sess.close()
img = np.array(img[0], np.uint8)
T.img_check(img)
cv2.imshow('beta', img)
cv2.waitKey(3456)
import re
import time
from datetime import datetime, timedelta
import os
import json
import requests
from configure import Config
since = datetime.strptime(Config.date_conf()['since'], "%Y%m%d")
if Config.date_conf()['until']:
until = since + timedelta(days=Config.date_conf()['until'])
else:
until = since + timedelta(days=1)
article = {'title': '', 'content': []}
my_voices_dir = Config().out_path().get('my', 'output_voice')
host_voices_dir = Config().out_path().get('host', 'output_voice')
output_notes = Config().out_path().get('notes', 'output_notes')
# 確保檔案目錄已經存在
# for path in [my_voices_dir, host_voices_dir, output_notes]:
# if not os.path.exists(os.path.abspath(path)):
# os.mkdir(os.path.abspath(path))
# 登入
login = '******'
headers = {
'User-Agent':
2018-10-10 Modify: Yonv1943
2018-10-11 save eval jpg
2018-10-12 'TF_CPP_MIN_LOG_LEVEL' tf.Session()
2018-10-12 origin, tensorflow.contrib.layers --> tf.layers
2018-10-12 elegant coding, stable
2018-10-13 C.size 28 --> 32, deeper, dcgan
2018-10-15 cloud removal
2018-10-21 'class Tools' move from mod_*.py to util.img_util.py
2018-10-22 change mask from 'middle square' to 'spot'
2018-10-23 spot --> polygon
2018-10-23 for discriminator, tf.concat([tenx, mask], axis=0)
'''
C = Config('mod_GAN_circle')
T = img_util.Tools()
rd.seed(1943)
def model_save_npy(sess, print_info):
tf_vars = tf.global_variables()
'''save as singal npy'''
npy_dict = dict()
for var in tf_vars:
npy_dict[var.name] = var.eval(session=sess)
print("| FETCH: %s" % var.name) if print_info else None
np.savez(C.model_npz, npy_dict)
with open(C.model_npz + '.txt', 'w') as f:
f.writelines(["%s\n" % key for key in npy_dict.keys()])
device = cfg.DEVICE
model = cfg.MODEL
model.eval()
predictions = []
with torch.no_grad():
for data, label in loader:
data = data.to(device)
label = label.to(device)
pred = model(data)['clipwise_output']
predictions.append(pred.cpu().detach().numpy())
pred = np.concatenate(predictions)
label = dataset.label
thresholds = np.linspace(0, 1, 101)
for threshold in thresholds:
score = mean_f1_score(pred, label, threshold)
print(threshold, score)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--config',
type=str,
default='default',
help="YAML file name under configs/")
args = parser.parse_args()
cfg = Config(args.config, train=False)
test(cfg)
2018-10-12 'TF_CPP_MIN_LOG_LEVEL' tf.Session()
2018-10-12 origin, tensorflow.contrib.layers --> tf.layers
2018-10-12 elegant coding, stable
2018-10-13 C.size 28 --> 32, deeper, dcgan
2018-10-15 cloud removal
2018-10-21 'class Tools' move from mod_*.py to util.img_util.py
2018-10-22 change mask from 'middle square' to 'spot'
2018-10-23 spot --> polygon
2018-10-23 for discriminator, tf.concat([tenx, mask], axis=0)
2018-11-05 for generator
2018-11-06 for process_data, tiny, U-net for generator
2018-11-06 buffer mask01 for discriminator, inp_dim_4
2018-11-06 inp, buf, mask
'''
C = Config('mod_mend_mask')
T = img_util.Tools()
rd.seed(1943)
def model_save_npy(sess, print_info):
tf_vars = tf.global_variables()
'''save as singal npy'''
npy_dict = dict()
for var in tf_vars:
npy_dict[var.name] = var.eval(session=sess)
print("| FETCH: %s" % var.name) if print_info else None
np.savez(C.model_npz, npy_dict)
with open(C.model_npz + '.txt', 'w') as f:
f.writelines(["%s\n" % key for key in npy_dict.keys()])
import tensorflow as tf
import tensorflow.layers as tl
from configure import Config
from util import img_util
from util import mod_util
'''
2018-10-10 Yonv1943
2018-11-14 kernel2 better than kernel4, but low solution
2018-11-15 add conv_tp_conv to decoder
2018-11-15 Debug: np.savez(C.model_npz, **npy_dict) FOR mod_save_npy()
2018-11-15 Debug: load from mod.npz FOR mod_sess_saver_logger()
'''
C = Config('mod_coderGAN_buff')
T = img_util.Tools()
def leru_batch_norm(ten):
ten = tf.layers.batch_normalization(ten, training=True)
ten = tf.nn.leaky_relu(ten)
return ten
def conv(ten, dim, idx):
filters = (2**(idx - 1)) * dim
return tl.conv2d(ten, filters, 3, 2, 'same', activation=tf.nn.leaky_relu)
def conv_tp(ten, dim, idx):
2018-10-10 Modify: Yonv1943
2018-10-11 save eval jpg
2018-10-12 'TF_CPP_MIN_LOG_LEVEL' tf.Session()
2018-10-12 origin, tensorflow.contrib.layers --> tf.layers
2018-10-12 elegant coding, stable
2018-10-13 C.size 28 --> 32, deeper, dcgan
2018-10-15 cloud removal
2018-10-21 'class Tools' move from mod_*.py to util.img_util.py
2018-10-22 change mask from 'middle square' to 'spot'
2018-10-23 spot --> polygon
2018-10-23 for discriminator, tf.concat([tenx, mask], axis=0)
'''
C = Config('mod_defog')
T = img_util.Tools()
rd.seed(1943)
def model_save_npy(sess, print_info):
tf_vars = tf.global_variables()
'''save as singal npy'''
npy_dict = dict()
for var in tf_vars:
npy_dict[var.name] = var.eval(session=sess)
print("| FETCH: %s" % var.name) if print_info else None
np.savez(C.model_npz, npy_dict)
with open(C.model_npz + '.txt', 'w') as f:
f.writelines(["%s\n" % key for key in npy_dict.keys()])
'''save as several npy'''
import os
import glob
import cv2
import numpy as np
import numpy.random as rd
from configure import Config
C = Config()
"""
2018-09-18 23:23:23 fix bug: img_grid()
2018-09-19 upgrade: img_grid(), random cut
2018-10-21 'class Tools' move from mod_*.py to util.img_util.py
2018-10-21 poly, blur debug
2018-10-24 stagger plot
2018-11-06 get_data__circle, circle_radius
2018-11-14 image_we_have
"""
class Tools(object):
def img_check(self, img):
print("| min,max %6.2f %6.2f |%s" %
(np.min(img), np.max(img), img.shape))
def ary_check(self, ary):
print("| min,max %6.2f %6.2f |ave,std %6.2f %6.2f |%s" % (
np.min(ary),
np.max(ary),
np.average(ary),
float(np.std(ary)),
from torch.utils.data import Dataset
from torchvision import models, transforms
from configure import Config
from utlity import *
from AudiotoImg import PrecomputedAudio
from torch.optim import lr_scheduler
import torch.optim as optim
import torch.nn as nn
if __name__ == '__main__':
# Configuration file
cfg = Config()
# Prepare data - you need to uncomment the following line to enable data preprocessing
# prepare_data(cfg)
# create new dataset and normalized images - train
pre_train = PrecomputedAudio(cfg.train_path,
img_transforms=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]))
# create new dataset and normalized images -val
pre_val = PrecomputedAudio(cfg.valid_path,
img_transforms=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(
mean=[0.485, 0.456, 0.406],
grey_dir = os.path.join(data_dir, 'CloudGreyDataset_%dx%d' % (size, size))
def __init__(self, model_dir='mod'):
self.model_dir = model_dir
self.model_name = 'mod'
self.model_path = os.path.join(self.model_dir, self.model_name)
self.model_npz = os.path.join(self.model_dir, self.model_name + '.npz')
self.model_log = os.path.join(self.model_dir, 'training_npy.txt')
if __name__ != '__main__':
from configure import Config # for test
print("|| TEST")
C = Config('mod_cloud_remove')
tf.set_random_seed(time.time() * 1943 % 178320049)
def auto_encoder(inp0, dim, out_dim, name, reuse, training=True):
padding1 = tf.constant(((0, 0), (1, 1), (1, 1), (0, 0)))
def leru_batch_norm(ten):
ten = tl.batch_normalization(ten, training=training)
ten = tf.nn.leaky_relu(ten)
return ten
def conv_tp(ten, idx):
filters = (2**idx) * dim
ten = tl.conv2d_transpose(ten,
filters,
from tensorboardX import SummaryWriter
from configure import Config
from model import WNet
from torch.utils.data import DataLoader
from Ncuts import NCutsLoss
#from AttenUnet import Attention_block
import time
import os
import argparse
import sys
from saveimg_helper import makegridimg3D
from AbdomenDS import AbdomenDS
if __name__ == '__main__':
config = Config()
parser = argparse.ArgumentParser(description='3D Attention WNet')
parser.add_argument(
"-d",
"--dataset",
required=False,
help="Dataset Type: 1 = InPhase, 2 = OutPhase, 3 = T2, 4 = In+OutPhase"
)
parser.add_argument("-l",
"--losscombine",
required=False,
help="Loss Combine: 1 = True, 0 = False")
parser.add_argument(
"-c",
"--cuda",
required=False,
Reference: https://github.com/cameronfabbri/Improved-Wasserstein-GAN
Reference: https://github.com/znxlwm/tensorflow-MNIST-GAN-DCGAN
2018-10-10 Modify: Yonv1943
2018-10-11 save eval jpg
2018-10-12 'TF_CPP_MIN_LOG_LEVEL' tf.Session()
2018-10-12 origin, tensorflow.contrib.layers --> tf.layers
2018-10-12 elegant coding, stable
2018-10-13 C.size 28 --> 32, deeper, dcgan
2018-10-15 cloud removal
2018-10-21 'class Tools' move from mod_*.py to util.img_util.py
2018-10-22 change mask from 'middle square' to 'spot'
'''
C = Config('mod_GAN_spot')
T = img_util.Tools()
rd.seed(1943)
def model_save_npy(sess, print_info):
tf_vars = tf.global_variables()
'''save as singal npy'''
npy_dict = dict()
for var in tf_vars:
npy_dict[var.name] = var.eval(session=sess)
print("| FETCH: %s" % var.name) if print_info else None
np.savez(C.model_npz, npy_dict)
with open(C.model_npz + '.txt', 'w') as f:
f.writelines(["%s\n" % key for key in npy_dict.keys()])
grey_dir = os.path.join(data_dir, 'CloudGreyDataset')
def __init__(self, model_dir='mod'):
self.model_dir = model_dir
self.model_name = 'mod'
self.model_path = os.path.join(self.model_dir, self.model_name)
self.model_npz = os.path.join(self.model_dir, self.model_name + '.npz')
self.model_log = os.path.join(self.model_dir, 'training_npy.txt')
if __name__ != '__main__':
from configure import Config # for test
print("||TEST")
C = Config('mod_cloud_detect')
tf.set_random_seed(time.time() * 1943 % 178320049)
def unet(inp0, dim, out_dim, name, reuse, training=True):
def leru_batch_norm(ten):
ten = tf.layers.batch_normalization(ten, training=training)
ten = tf.nn.leaky_relu(ten)
return ten
paddings = tf.constant(((0, 0), (1, 1), (1, 1), (0, 0)))
def conv_pad(ten, idx, step=2):
filters = (2**idx) * dim
ten = tf.pad(ten, paddings, 'REFLECT')
ten = tl.conv2d(ten,
from logging import getLogger
from experiment import Experiment
from configure import Config
from utils.seeder import seed_everything
from utils.mylog import timer, create_logger
@timer
def main(config):
experiment = Experiment(config)
experiment.run()
if __name__ == "__main__":
c = Config()
c.set_parameter(config_dir=pathlib.Path('src/config'), use_option=True)
gc.enable()
seed_everything(c.runtime.RANDOM_SEED)
create_logger('main', c.log)
create_logger('train', c.log)
logger = getLogger('main')
logger.info(f':thinking_face: ============ {datetime.now():%Y-%m-%d %H:%M:%S} ============ :thinking_face:')
try:
main(c)
logger.info(f':sunglasses: ============ {datetime.now():%Y-%m-%d %H:%M:%S} ============ :sunglasses:')
except Exception:
logger.critical(f':smiling_imp: Exception occured \n {traceback.format_exc()}')
logger.critical(f':skull: ============ {datetime.now():%Y-%m-%d %H:%M:%S} ============ :skull:')
def run_eval():
mat_list = list()
for name in ['ground', 'out_aer', 'out_cld']:
img = cv2.imread('eval_replace_%s.jpg' % name)
mat_list.append(img[np.newaxis, :, :, :])
import tensorflow as tf
import mod_mend_buff as mod
import os
class Config(object):
train_epoch = 2 ** 14
train_size = int(2 ** 17 * 1.9)
eval_size = 2 ** 3
batch_size = 2 ** 4
batch_epoch = train_size // batch_size
size = int(2 ** 9) # size = int(2 ** 7)
replace_num = int(0.25 * batch_size)
learning_rate = 1e-5 # 1e-4
show_gap = 2 ** 5 # time
eval_gap = 2 ** 9 # time
gpu_limit = 0.48 # 0.0 ~ 1.0
gpu_id = 1
data_dir = '/mnt/sdb1/data_sets'
aerial_dir = os.path.join(data_dir, 'AerialImageDataset/train')
cloud_dir = os.path.join(data_dir, 'ftp.nnvl.noaa.gov_color_IR_2018')
grey_dir = os.path.join(data_dir, 'CloudGreyDataset')
def __init__(self, model_dir='mod'):
self.model_dir = model_dir
self.model_name = 'mod'
self.model_path = os.path.join(self.model_dir, self.model_name)
self.model_npz = os.path.join(self.model_dir, self.model_name + '.npz')
self.model_log = os.path.join(self.model_dir, 'training_npy.txt')
C = Config('mod_mend_GAN_buff')
gene_name = 'gene'
inp_ground = tf.placeholder(tf.uint8, [None, C.size, C.size, 3])
inp_cloud1 = tf.placeholder(tf.uint8, [None, C.size, C.size, 1])
flt_ground = tf.to_float(inp_ground) / 255.0
flt_cloud1 = tf.to_float(inp_cloud1) / 255.0
ten_repeat = tf.ones([1, 1, 1, 3])
ten_ground = flt_ground[:C.batch_size]
# buf_ground = flt_ground[C.batch_size:]
ten_cloud1 = flt_cloud1[:C.batch_size]
ten_cloud3 = ten_cloud1 * ten_repeat
ten_mask10 = (1.0 - ten_cloud3)
ten_ragged = ten_ground * ten_mask10
ten_patch3 = mod.generator(tf.concat((ten_ragged, ten_cloud3), axis=3),
32, 3, gene_name, reuse=False)
out_ground = ten_ragged + ten_patch3 * ten_cloud3
from utils import mod_util
sess = mod_util.get_sess(C)
saver, logger, pre_epoch = mod_util.get_saver_logger(C, sess)
print("||Training Check")
# eval_fetch = [ten_ground, out_ground, ten_patch3, ten_cloud3]
eval_fetch = [out_ground, ten_patch3]
eval_feed_dict = {inp_ground: mat_list[0],
inp_cloud1: mat_list[2][:, :, :, 0:1]}
mat_list = sess.run(eval_fetch, eval_feed_dict)
for img, name in zip(mat_list, ['out_ground', 'ten_patch3']):
img = (img[0] * 255).astype(np.uint8)
cv2.imshow('beta', img)
cv2.waitKey(4321)
print(img.shape, np.max(img))
cv2.imwrite('eval_gan_%s.jpg' % name, img)
print(end=" EVAL")
import os
import glob
import cv2
import numpy as np
import numpy.random as rd
from configure import Config
G = Config()
"""
2018-09-18 23:23:23 fix bug: img_grid()
2018-09-19 15:12:12 upgrade: img_grid(), random cut
"""
class Cloud2Grey(object):
def __init__(self):
self.map_pts = np.load(
os.path.join(os.path.dirname(os.path.realpath(__file__)),
'white_line_eliminate_map.npy'))
def run(self, img):
for i, j, x, y in self.map_pts: # eliminate white line
img[i, j] = img[x, y]
switch = np.array(img[:, :, 2], dtype=np.int) - img[:, :, 0]
switch = np.clip(
switch - 4, 0,
1) # The part with thick cloud, could not see the ground
out = np.clip(img[:, :, 1], 60, 195)
grey_dir = os.path.join(data_dir, 'CloudGreyDataset_%dx%d' % (size, size))
def __init__(self, model_dir='mod'):
self.model_dir = model_dir
self.model_name = 'mod'
self.model_path = os.path.join(self.model_dir, self.model_name)
self.model_npz = os.path.join(self.model_dir, self.model_name + '.npz')
self.model_log = os.path.join(self.model_dir, 'training_npy.txt')
if __name__ != '__main__':
from configure import Config # for test
print("|| TEST")
C = Config('mod_cloud_remove_rec')
tf.set_random_seed(time.time() * 1943 % 178320049)
def auto_encoder(inp0, dim, out_dim, name, reuse, training=True):
padding1 = tf.constant(((0, 0), (1, 1), (1, 1), (0, 0)))
def leru_batch_norm(ten):
ten = tl.batch_normalization(ten, training=training)
ten = tf.nn.leaky_relu(ten)
return ten
def conv_tp(ten, idx):
filters = (2**idx) * dim
ten = tl.conv2d_transpose(ten,
filters,
import numpy.random as rd
import tensorflow as tf
import tensorflow.layers as tl
from configure import Config
from util import img_util
from util import mod_util
'''
2018-10-10 Yonv1943
2018-11-14 kernel2 better than kernel4, but low solution
2018-11-15 add conv_tp_conv to decoder
2018-11-15 Debug: np.savez(C.model_npz, **npy_dict) FOR mod_save_npy()
'''
C = Config('mod_AutoEncoderGAN')
T = img_util.Tools()
rd.seed(1943)
def leru_batch_norm(ten):
ten = tf.layers.batch_normalization(ten, training=True)
ten = tf.nn.leaky_relu(ten)
return ten
def conv(ten, dim, idx):
filters = (2**(idx - 1)) * dim
return tl.conv2d(ten, filters, 3, 2, 'same', activation=tf.nn.leaky_relu)
import os
import time
import shutil
import cv2
import numpy as np
import numpy.random as rd
import tensorflow as tf
import tensorflow.layers as tl
from configure import Config
from util import img_util
C = Config('mod_mend_Unet')
T = img_util.Tools()
rd.seed(1943)
def run():
cv2.namedWindow('beta', cv2.WINDOW_KEEPRATIO)
img = cv2.imread(os.path.join(C.aerial_dir, 'bellingham1.tif'))
# cv2.imshow('beta', img)
# cv2.waitKey(3456)
channel = 3
xlen, ylen = img.shape[0:2]
xmod = xlen % C.size
ymod = ylen % C.size
xrnd = int(rd.rand() * xmod)
import os, shutil
from os import path
from os.path import basename
from configure import Config
config = Config()
folders = []
# Quick N Dirty - non-configurable
folders.append("/content/bach-thes/notebooks/UNet/Train_images/train")
folders.append("/content/bach-thes/notebooks/UNet/Train_images/train_masks")
folders.append(
"/content/bach-thes/notebooks/UNet/Train_images/resampled_train")
folders.append(
"/content/bach-thes/notebooks/UNet/Train_images/resampled_train_masks")
folders.append(
"/content/bach-thes/notebooks/UNet/Train_images/extracted_train")
folders.append(
"/content/bach-thes/notebooks/UNet/Train_images/extracted_train_masks")
folders.append("/content/bach-thes/notebooks/UNet/Train_images/train2")
folders.append("/content/bach-thes/notebooks/UNet/Train_images/train2_masks")
folders.append(
"/content/bach-thes/notebooks/UNet/Train_images/resampled_train2")
folders.append(
"/content/bach-thes/notebooks/UNet/Train_images/resampled_train2_masks")
folders.append(
"/content/bach-thes/notebooks/UNet/Train_images/extracted_train2")
folders.append(
"/content/bach-thes/notebooks/UNet/Train_images/extracted_train2_masks")
Reference: https://github.com/znxlwm/tensorflow-MNIST-GAN-DCGAN
2018-10-10 Modify: Yonv1943
2018-10-11 save eval jpg
2018-10-12 'TF_CPP_MIN_LOG_LEVEL' tf.Session()
2018-10-12 origin, tensorflow.contrib.layers --> tf.layers
2018-10-12 elegant coding, stable
2018-10-13 C.size 28 --> 32, deeper, dcgan
2018-10-15 cloud removal
2018-10-21 'class Tools' move from mod_*.py to util.img_util.py
2018-10-22 change mask from 'middle square' to 'spot'
2018-10-23 spot --> polygon
'''
C = Config('mod_GAN_poly')
T = img_util.Tools()
rd.seed(1943)
def model_save_npy(sess, print_info):
tf_vars = tf.global_variables()
'''save as singal npy'''
npy_dict = dict()
for var in tf_vars:
npy_dict[var.name] = var.eval(session=sess)
print("| FETCH: %s" % var.name) if print_info else None
np.savez(C.model_npz, npy_dict)
with open(C.model_npz + '.txt', 'w') as f:
f.writelines(["%s\n" % key for key in npy_dict.keys()])
'''save as several npy'''
'val': val_score
}, epoch)
scheduler.step()
# save model
if best_score <= val_score and not cfg.SANITY_CHECK:
best_score = val_score
checkpoint = {
'config': cfg.cfg,
'epoch': epoch,
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
}
torch.save(checkpoint, cfg.CKPT_PATH)
writer.close()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--config',
type=str,
default='default',
help="YAML file name under configs/")
args = parser.parse_args()
cfg = Config(args.config, train=True)
train(cfg)
# from aiogram.contrib.fsm_storage.redis import RedisStorage2
from aiogram.dispatcher import FSMContext
from aiogram.dispatcher.filters.state import State, StatesGroup
from aiogram.types import ParseMode
from aiogram.utils.executor import start_webhook
from aiogram.utils.markdown import bold, code, italic, pre, text
from loguru import logger
from configure import Config
from middlewares import AccessMiddleware
from parser import SounCloudDownloader
scp = SounCloudDownloader('https://www.forhub.io/download.php')
config_file_path = join(dirname(abspath('config.ini')), 'config.ini')
config = Config(config_file_path)
loop = asyncio.get_event_loop()
bot = Bot(config.API_TOKEN,
proxy=config.PROXY_URL,
parse_mode=ParseMode.MARKDOWN_V2)
# storage = RedisStorage2(config.REDIS_HOST)
dp = Dispatcher(bot, loop=loop)
@dp.message_handler(state='*', commands=['start'])
async def send_welcome(message: types.Message, state: FSMContext):
await message.answer(text='Кидай ссылку на тречик в soundcloud')
@dp.message_handler(state='*', content_types=['text'])