def servant_scroll(self, line):
if not self.parse_scene_condition(["$all", [["滚动条-上"], ["滚动条-下"]]]):
self.notice("Can't Scroll")
return
if abs(line) < 3:
mid_x = config.getint("Device", "MainWidth") / 2
mid_y = config.getint("Device", "MainHeight") / 2
rand_drag(
self.hwnd,
rand_point([mid_x, mid_y], [50, 10]),
rand_point([mid_x, mid_y - mid_y * line * 0.5], [50, 10]),
30,
)
return
_, top_xy = self.search_resource("滚动条-上")
_, bot_xy = self.search_resource("滚动条-下")
width = self.resources["滚动条-上"]["Size"][0]
top = top_xy[1]
bottom = bot_xy[1]
x = top_xy[0] + width / 2
middle = (top + bottom) // 2
cross = bottom - top
dy = line * cross * 0.37
rand_drag(
self.hwnd,
rand_point([x, middle], [width / 6, cross / 6]),
rand_point([x, middle + dy], [50, 10]),
30,
)
def update_resource(resource, section, name=None):
"""根据当前配置更新资源属性, 加载图片并进行缩放"""
dw = config.getint("Device", "MainWidth")
dh = config.getint("Device", "MainHeight")
sdw, sdh = resource["MainSize"]
if dw != sdw or dh != sdh:
for key in resource:
if re.search("Offset|Position|Size", key):
resource[key] = rescale_item(resource[key], dw / sdw, dh / sdh)
if (
resource.get("Image", None) is not None
or resource.get("ImageData", None) is not None
):
load_image(resource, section)
if name is not None and "Name" not in resource:
resource["Name"] = name
def get_window_shot(hwnd):
# 对后台应用程序截图，程序窗口可以被覆盖，但如果最小化后只能截取到标题栏、菜单栏等。
# 使用自定义的窗口边缘和大小设置
dx = config.getint("Device", "EdgeOffsetX")
dy = config.getint("Device", "EdgeOffsetY")
w = config.getint("Device", "MainWidth")
h = config.getint("Device", "MainHeight")
window_w, window_h = detect_window_size(hwnd)
if dx + w > window_w or dy + h > window_h:
raise ValueError("截图区域超出窗口! 请检查配置文件")
# logger.debug("截图: %dx%d at %dx%d", w, h, dx, dy)
# 返回句柄窗口的设备环境、覆盖整个窗口，包括非客户区，标题栏，菜单，边框
hwndDC = win32gui.GetWindowDC(hwnd)
# 创建设备描述表
mfcDC = win32ui.CreateDCFromHandle(hwndDC)
# 创建内存设备描述表
saveDC = mfcDC.CreateCompatibleDC()
# 创建位图对象
saveBitMap = win32ui.CreateBitmap()
saveBitMap.CreateCompatibleBitmap(mfcDC, w, h)
saveDC.SelectObject(saveBitMap)
# 截图至内存设备描述表
saveDC.BitBlt((0, 0), (w, h), mfcDC, (dx, dy), win32con.SRCCOPY)
# 获取位图信息
bmpinfo = saveBitMap.GetInfo()
bmpdata = saveBitMap.GetBitmapBits(True)
# 生成图像
image_data = np.frombuffer(bmpdata, "uint8")
image_data = image_data.reshape(
(bmpinfo["bmHeight"], bmpinfo["bmWidth"], 4))
image_data = cv.cvtColor(image_data, cv.COLOR_BGRA2BGR)
# 内存释放
win32gui.DeleteObject(saveBitMap.GetHandle())
saveDC.DeleteDC()
mfcDC.DeleteDC()
win32gui.ReleaseDC(hwnd, hwndDC)
return image_data
def __init__(
self,
train: bool = True,
val: bool = False,
transform: Optional[object] = None,
target_transform: Optional[object] = None,
debug_with_few_samples: bool = False,
create_standardization_transform: bool = False,
) -> None:
if train and val:
print('"train" and "val" cannot be true at the same time!')
raise ValueError
min_year = config.getint('DataOptions', 'min_year')
max_year = config.getint('DataOptions', 'max_year')
max_train_year = config.getint('DataOptions', 'max_train_year')
max_val_year = config.getint('DataOptions', 'max_val_year')
remo_input_dir = config.get('Paths', 'remo_input')
remo_target_dir = config.get('Paths', 'remo_target')
elev_file = config.get('Paths', 'elevation')
input_var = config.get('DataOptions', 'input_variable')
target_var = config.get('DataOptions', 'target_variable')
aux_base_path = config.get('Paths', 'aux_base_path')
# The filter removes empty strings in the resulting list, which occur when there are no aux_variables specified
aux_vars = list(
filter(None,
config.get('DataOptions', 'aux_variables').split(',')))
aux_vars = [join(aux_base_path, p) for p in aux_vars]
if train:
self.years = list(range(min_year, max_train_year + 1))
print('Train years', self.years)
self.mode = 'train'
elif val:
self.years = list(range(max_train_year + 1, max_val_year + 1))
print('Validation years', self.years)
self.mode = 'val'
else:
self.years = list(range(max_val_year + 1, max_year + 1))
print('Test years', self.years)
self.mode = 'test'
self.transform = transform
self.target_transform = target_transform
# For testing purposes
# self.years = [2000]
self.dataset = RemoSuperRes(
remo_input_dir,
remo_target_dir,
self.years,
elev_file,
input_var=input_var,
target_var=target_var,
aux_features=aux_vars,
)
# This name is misleading - it does not necessarily create test data
(
self.X,
self.aux,
self.elev_arr,
self.Y,
self.lats,
self.lons,
self.times,
) = self.dataset.make_test()
if debug_with_few_samples:
num_debug_samples = 16
print(f'DEBUG: Using only {num_debug_samples} samples')
self.X = self.X[:num_debug_samples]
if self.aux is not None:
self.aux = self.aux[:num_debug_samples]
self.Y = self.Y[:num_debug_samples]
self.times = self.times[:num_debug_samples]
# Convert E-OBS temperature from Celsius to Kelvin
if target_var == 'tg':
self.Y += 272.15
if create_standardization_transform:
self.standardize_transform = self.calculate_standardization_transform(
)
if self.transform is not None:
self.transform = Compose(
[self.transform, self.standardize_transform])
else:
self.transform = self.standardize_transform
def rescale_point(hwnd, point):
dpi = ctypes.windll.user32.GetDpiForWindow(hwnd)
x = point[0] + config.getint("Device", "EdgeOffsetX")
y = point[1] + config.getint("Device", "EdgeOffsetY")
# logger.debug("Rescale: %s -> %s", point, (x, y))
return int(np.round(x * 96 / dpi)), int(np.round(y * 96 / dpi))
def export(lr: float, model_name: str, save_dir: str):
models_available = {
'globalnet': GlobalNet,
'localnet': LocalNet,
'convmos': ConvMOS,
}
model = models_available[model_name]()
device = 'cpu'
if torch.cuda.is_available():
device = 'cuda'
model = model.to(device=device)
# E-OBS only provides observational data for land so we need to use a mask to avoid fitting on the sea
land_mask_np = np.load('remo_eobs_land_mask.npy')
# Convert booleans to 1 and 0, and convert numpy array to torch Tensor
land_mask = torch.from_numpy(1 * land_mask_np).to(device)
loss_fn = partial(masked_mse_loss, mask=land_mask)
optimizer = Adam(model.parameters(), lr=lr)
trainer = create_supervised_trainer(model,
optimizer,
loss_fn,
device=device)
to_save = {'model': model, 'optimizer': optimizer, 'trainer': trainer}
checkpoint_files = glob(join(save_dir, 'best_checkpoint_*.pt'))
if len(checkpoint_files) > 0:
# Parse something like:
# /scratch/scratch_remo/APRL-rr-11-11-sdnext-loglonet-prec-ger11-maskedloss-7/best_checkpoint_194_val_loss=-11.8250.pt
# Sorry
epoch_to_score = {
int(c.split(sep)[-1].split('_')[2]):
float(c.split(sep)[-1].split('=')[-1][:-3])
for c in checkpoint_files
}
print(epoch_to_score)
best_epoch = max(epoch_to_score, key=epoch_to_score.get)
best_checkpoint_file = next(
cf for cf in checkpoint_files
if int(cf.split(sep)[-1].split('_')[2]) == best_epoch)
print('Loading best checkpoint', best_checkpoint_file)
checkpoint = torch.load(best_checkpoint_file, map_location=device)
Checkpoint.load_objects(to_load=to_save, checkpoint=checkpoint)
else:
print(
'ERROR: cannot find any files matching',
join(save_dir, 'best_checkpoint_*.pt'),
)
return
# This uses all aux variables, the temperature/precipitation, and elevation
input_depth = (len(
list(
filter(None,
config.get('DataOptions', 'aux_variables').split(',')))) +
2)
input_width = config.getint('NN', 'input_width')
input_height = config.getint('NN', 'input_height')
dummy_input = torch.randn(1,
input_depth,
input_width,
input_height,
device=device)
torch.onnx.export(
model,
dummy_input,
join(save_dir, 'convmos.onnx'),
verbose=True,
input_names=['input'],
output_names=['output'],
)
# directory already exists
pass
val_preds.to_netcdf(join(save_dir, f'val_predictions.nc'))
test_preds.to_netcdf(join(save_dir, f'test_predictions.nc'))
if __name__ == "__main__":
parser = ArgumentParser()
parser.add_argument('--num_jobs',
type=int,
default=1,
help='Number of jobs (default: 1)')
parser.add_argument(
'--non_local_dist',
type=int,
default=config.getint('NN', 'non_local_dist'),
help='Max cell distance to consider (Only relevant in non_local mode)',
)
parser.add_argument(
'--n_components',
type=float,
default=config.getfloat('NN', 'n_components'),
help=
'Number of components to retain for PCA (Only relevant in non_local mode)',
)
parser.add_argument(
'--n_param_sets',
type=int,
default=config.getint('NN', 'n_param_sets'),
help='Number of parameter sets tried per cell in the hp search',
)
{f'val_torch_{k}': v
for k, v in val_evaluator.state.metrics.items()})
# ... the validation metrics that I calculate,
results.update({f'val_{k}': v for k, v in val_res.items()})
# ... asnd the test metrics that I calculate
results.update({f'test_{k}': v for k, v in test_res.items()})
write_results_file(join('results', 'results.json'),
pd.json_normalize(results))
if __name__ == "__main__":
parser = ArgumentParser()
parser.add_argument(
'--batch_size',
type=int,
default=config.getint('NN', 'batch_size'),
help='input batch size for training (default: 64)',
)
parser.add_argument(
'--val_batch_size',
type=int,
default=config.getint('NN', 'batch_size'),
help='input batch size for validation (default: 1000)',
)
parser.add_argument(
'--epochs',
type=int,
default=config.getint('NN', 'training_epochs'),
help='number of epochs to train (default: 1000)',
)
parser.add_argument(