def test_inference_train():
from glob import glob
from model import load_model, find_best, delete_model
import keras.backend as K
from utils import get_session
from transfer import inference_train
name = 'test/transfer'
delete_model(name)
K.set_session(get_session())
imgs_path = glob('data/DeepQ-Vivepaper/frame/**/*.png',
recursive=True)[:100]
valid_csv = 'data/DeepQ-Vivepaper/data/annotations.csv'
class_map = 'data/class-map.csv'
thresh = 0.98
model = load_model(find_best()[0], compile=True)
buff = inference_train(model,
imgs_path,
valid_csv,
class_map,
thresh,
batch_size=4,
image_min_side=460,
image_max_side=1024,
name=name)
buff.seek(0)
print(buff.read())
def test_main():
from inference import main
from model import find_best
weights = find_best()[0]
print(weights)
args = ('--weights %s --batch-size 1' % weights).split()
score = main(args)
assert score > 0
def main():
""" 命令行参数及其整理 """
parser = utils.MyArguments()
parser.add_argument("--De", default=5.2, type=float, help="平均潜伏期")
parser.add_argument("--Dq", default=14, type=float, help="平均隔离期")
parser.add_argument("--c", default=13.0046, type=float, help="初始平均接触率")
parser.add_argument("--q", default=0.0, type=float, help="初始隔离率")
parser.add_argument("--beta", default=2.03e-9, type=float, help="基础传染概率")
parser.add_argument("--theta",
default=1.6003,
type=float,
help="无症状感染者传染概率系数")
parser.add_argument("--nu", default=1.5008, type=float, help="潜伏期传染概率系数")
parser.add_argument("--phi", default=0.9, type=float, help="有症状感染者收治率")
parser.add_argument("--gammaI",
default=0.1029,
type=float,
help="有症状感染者自愈速率")
parser.add_argument("--gammaA",
default=0.2978,
type=float,
help="无症状感染者自愈速度")
parser.add_argument("--gammaH",
default=1 / 10.5,
type=float,
help="医院治愈速率")
parser.add_argument("--deltaI",
default=1 / 3.5,
type=float,
help="出现症状患者被收治的速率")
parser.add_argument("--deltaQ",
default=0.1259,
type=float,
help="隔离的潜伏者出现症状（及时被收治）的速率")
parser.add_argument("--rho",
default=0.6834,
type=float,
help="有症状感染者占所有感染者的比例")
parser.add_argument("--use_19", action="store_true")
parser.add_argument("--zero_spring", action="store_true")
parser.add_argument("-pil",
"--piecewise_interval_length",
default=3,
type=int)
args = parser.parse_args() # 对于一些通用的参数，这里已经进行整理了
""" 读取准备好的数据 """
dat_file = "./DATA/Provinces.pkl"
dataset = utils.Dataset(dat_file, args.t0, args.tm, args.fit_time_start)
""" 构建、或读取、或训练模型 """
# 根据不同的情况来得到合适的模型
if args.model is not None and args.model != "fit":
models = NetSEAIRQ_piecewise.load(args.model)
else: # 不然就进行训练
# 设置我们拟合模型需要的数据
if args.use_whhb:
mask = None
else:
mask = np.full(dataset.num_regions, True, dtype=np.bool)
mask[0] = False
fit_start_index = (dataset.fit_start_t.ord - dataset.epi_t0.ord)
fit_start_index = int(fit_start_index)
fit_data_all = dataset.epi_times.delta[fit_start_index:]
# 根据分段的宽度，设置多个模型，并将其训练用参数也
pil = args.piecewise_interval_length
n_models = int(ceil(fit_data_all.shape[0] / pil))
models = []
score_kwargs = []
for i in range(n_models):
model = NetSEAIRQ_piecewise(
populations=dataset.populations,
y0_hb=np.array([0, 0, 0, 0, args.y0, 0, 0]),
score_type=args.fit_score,
gamma_func_kwargs={
"gammas": (dataset.out19_dict
if args.use_19 else dataset.out20_dict),
"zero_period":
(dataset.zero_period.delta if args.zero_spring else None)
},
Pmn_func_kwargs={"pmn": dataset.pmn_matrix_relative},
De=args.De,
Dq=args.Dq,
c=args.c,
q=args.q,
beta=args.beta,
rho=args.rho,
deltaI=args.deltaI,
deltaQ=args.deltaQ,
gammaI=args.gammaI,
gammaA=args.gammaH,
gammaH=args.gammaH,
theta=args.theta,
nu=args.nu,
phi=args.phi,
)
use_dat_start = i * pil
use_dat_end = min((i + 1) * pil, fit_data_all.shape[0])
score_kwarg = {
"times": dataset.epi_times.delta[use_dat_start:use_dat_end],
"mask": mask,
"trueH": dataset.trueH[use_dat_start:use_dat_end],
# "trueR": (dataset.trueD + dataset.trueR)[
# use_dat_start:use_dat_end]
}
models.append(model)
score_kwargs.append(score_kwarg)
# 搜索最优参数
if args.fit_method == "annealing":
fit_kwargs = {"callback": utils.callback, "method": "annealing"}
else:
fit_kwargs = {
"method": args.fit_method,
"fig_dir": args.save_dir + "/",
"njobs": -1,
"NIND": args.geatpy_nind,
"MAXGEN": args.geatpy_maxgen,
"n_populations": args.geatpy_npop
}
last_y0 = None
predHs = []
all_opts = []
for i, (model, score_kwarg) in enumerate(zip(models, score_kwargs)):
# 被这次训练的时间整理出来
start_time = utils.CustomDate.from_delta(
score_kwarg["times"].min(), dataset.t0.str)
end_time = utils.CustomDate.from_delta(score_kwarg["times"].max(),
dataset.t0.str)
print("开始训练 %s<->%s" % (start_time.str, end_time.str))
# 第一次训练的模型和后面训练的模型使用不同的参数
# 之后训练的模型要使用前面模型的最后一天输出作为y0
if i == 0:
model.no_fit_params([])
else:
model.no_fit_params(["y0_hb[2:4]"])
model.set_y0(last_y0)
# 得到训练参数，进行训练
dim, lb, ub = model.fit_params_range()
opt_res = find_best(lambda x: score_func(x, model, score_kwarg),
dim, lb, ub, **fit_kwargs)
all_opts.append(opt_res)
# 将得到的最优参数设置到模型中
model.set_params(opt_res["BestParam"])
# 预测结果
preds = model.predict(score_kwarg["times"])
predHs.append(preds[0])
# 预测结果中最后一天作为新的y0
last_y0 = np.concatenate(preds, axis=1)[-1, :]
predHs = np.concatenate(predHs, axis=0)
utils.save(all_opts, os.path.join(args.save_dir, "opt_res.pkl"))
utils.save([m.kwargs for m in models],
os.path.join(args.save_dir, "models.pkl"))
# model.save(os.path.join(args.save_dir, "model.pkl"))
# utils.save(opt_res, os.path.join(args.save_dir, "opt_res.pkl"))
""" 计算相关指标以及绘制图像 """
# 预测R0
pass
# 计算每个地区的曲线下面积以及面积差,并保存
# auc = under_area(
# dataset.epi_times.delta,
# dataset.trueH,
# dataset.pred_times.delta,
# nopr_preds[0],
# )
# auc_df = pd.DataFrame(
# auc.T, columns=["true_area", "pred_area", "diff_area"],
# index=dataset.regions
# )
# auc_df["population"] = dataset.populations
# auc_df["diff_norm"] = auc_df.diff_area / auc_df.population
# auc_df.sort_values("diff_norm", inplace=True)
# 为每个地区绘制曲线图
plt.rcParams["font.sans-serif"] = ["SimHei"]
img_dir = os.path.join(args.save_dir, "imgs")
if not os.path.exists(img_dir):
os.mkdir(img_dir)
for i, reg in enumerate(dataset.regions):
"""
y0 = [H R E A I Sq Eq] + [S]
"""
plot_one_regions(
reg,
[
("trueH", dataset.epi_times.ord.astype("int"),
dataset.trueH[:, i], "ro"),
# ("trueR", dataset.epi_times.ord.astype("int"),
# dataset.trueR[:, i]+dataset.trueD[:, i], "bo"),
# ("predH", dataset.pred_times.ord.astype("int"),
# predHs[:, i], "r"),
("predH", dataset.epi_times.ord.astype("int"), predHs[:,
i], "r"),
# ("predR", dataset.pred_times.ord.astype("int"),
# prot_preds[1][:, i], "b"),
# ("predE", dataset.pred_times.ord.astype("int"),
# prot_preds[3][:, i], "y"),
# ("predA", dataset.pred_times.ord.astype("int"),
# prot_preds[4][:, i], "g"),
# ("predI", dataset.pred_times.ord.astype("int"),
# prot_preds[4][:, i], "c"),
],
[
("trueH", dataset.epi_times.ord.astype("int"),
dataset.trueH[:, i], "ro"),
# ("trueR", dataset.epi_times.ord.astype("int"),
# dataset.trueR[:, i]+dataset.trueD[:, i], "bo"),
# ("predH", dataset.pred_times.ord.astype("int"),
# predHs[:, i], "r"),
("predH", dataset.epi_times.ord.astype("int"), predHs[:,
i], "r"),
# ("predR", dataset.pred_times.ord.astype("int"),
# nopr_preds[1][:, i], "b"),
# ("predE", dataset.pred_times.ord.astype("int"),
# nopr_preds[3][:, i], "y"),
# ("predA", dataset.pred_times.ord.astype("int"),
# nopr_preds[4][:, i], "g"),
# ("predI", dataset.pred_times.ord.astype("int"),
# nopr_preds[4][:, i], "c"),
],
save_dir=img_dir)
# # 保存结果
# for i, name in enumerate([
# "predH", "predR", "predE", "predA", "predI"
# ]):
# pd.DataFrame(
# prot_preds[i],
# columns=dataset.regions,
# index=dataset.pred_times.str
# ).to_csv(
# os.path.join(args.save_dir, "protect_%s.csv" % name)
# )
# pd.DataFrame(
# nopr_preds[i],
# columns=dataset.regions,
# index=dataset.pred_times.str
# ).to_csv(
# os.path.join(args.save_dir, "noprotect_%s.csv" % name)
# )
# auc_df.to_csv(os.path.join(args.save_dir, "auc.csv"))
# # 这里保存的是原始数据
# for i, attr_name in enumerate(["trueD", "trueH", "trueR"]):
# save_arr = getattr(dataset, attr_name)
# pd.DataFrame(
# save_arr,
# columns=dataset.regions,
# index=dataset.epi_times.str
# ).to_csv(os.path.join(args.save_dir, "%s.csv" % attr_name))
# 保存args到路径中（所有事情都完成再保存数据，安全）
save_args = deepcopy(args.__dict__)
save_args["model_type"] = "NetSEAIRQ-piecewise"
utils.save(save_args, os.path.join(args.save_dir, "args.json"), "json")
def main():
""" 命令行参数及其整理 """
parser = utils.MyArguments()
parser.add_argument("--De", default=5.2, type=float, help="平均潜伏期")
parser.add_argument("--Dq", default=14, type=float, help="平均隔离期")
parser.add_argument("--c", default=13.0046, type=float, help="初始平均接触率")
parser.add_argument("--q", default=0.0, type=float, help="初始隔离率")
parser.add_argument(
"--beta", default=2.03e-9, type=float, help="基础传染概率"
)
parser.add_argument(
"--theta", default=1.6003, type=float, help="无症状感染者传染概率系数"
)
parser.add_argument(
"--nu", default=1.5008, type=float, help="潜伏期传染概率系数"
)
parser.add_argument(
"--phi", default=0.9, type=float, help="有症状感染者收治率"
)
parser.add_argument(
"--gammaI", default=0.1029, type=float, help="有症状感染者自愈速率"
)
parser.add_argument(
"--gammaA", default=0.2978, type=float, help="无症状感染者自愈速度"
)
parser.add_argument(
"--gammaH", default=1/10.5, type=float, help="医院治愈速率"
)
parser.add_argument(
"--deltaI", default=1/3.5, type=float, help="出现症状患者被收治的速率"
)
parser.add_argument(
"--deltaQ", default=0.1259, type=float,
help="隔离的潜伏者出现症状（及时被收治）的速率"
)
parser.add_argument(
"--rho", default=0.6834, type=float, help="有症状感染者占所有感染者的比例"
)
parser.add_argument("--protect_ck", default=0.0, type=float)
parser.add_argument("--protect_qk", default=0.0, type=float)
parser.add_argument("--use_19", action="store_true")
parser.add_argument("--zero_spring", action="store_true")
args = parser.parse_args() # 对于一些通用的参数，这里已经进行整理了
""" 读取准备好的数据 """
dat_file = "./DATA/Provinces.pkl"
dataset = utils.Dataset(dat_file, args.t0, args.tm, args.fit_time_start)
""" 构建、或读取、或训练模型 """
# 根据不同的情况来得到合适的模型
if args.model is not None and args.model != "fit":
model = NetSEAIRQ.load(args.model)
else:
model = NetSEAIRQ(
populations=dataset.populations,
y0for1=np.array([0, 0, 0, 0, args.y0, 0, 0]),
protect=True, score_type=args.fit_score,
protect_args={
"t0": dataset.protect_t0.delta,
"c_k": args.protect_ck,
"q_k": args.protect_qk
},
gamma_func_kwargs={
"gammas": (dataset.out19_dict if args.use_19
else dataset.out20_dict),
"zero_period": (dataset.zero_period.delta
if args.zero_spring else None)
},
Pmn_func_kwargs={"pmn": dataset.pmn_matrix_relative},
De=args.De, Dq=args.Dq, c=args.c, q=args.q, beta=args.beta,
rho=args.rho, deltaI=args.deltaI, deltaQ=args.deltaQ,
gammaI=args.gammaI, gammaA=args.gammaH, gammaH=args.gammaH,
theta=args.theta, nu=args.nu, phi=args.phi,
)
if args.model == "fit":
# 设置我们拟合模型需要的数据
if args.use_whhb:
mask = None
else:
mask = np.full(dataset.num_regions, True, dtype=np.bool)
mask[0] = False
fit_start_index = (dataset.fit_start_t.ord - dataset.epi_t0.ord)
fit_start_index = int(fit_start_index)
score_kwargs = {
"times": dataset.epi_times.delta[fit_start_index:],
"mask": mask,
}
score_kwargs["trueH"] = dataset.trueH
# score_kwargs["trueR"] = dataset.trueD + dataset.trueR
# 搜索
if args.fit_method == "annealing":
fit_kwargs = {
"callback": utils.callback, "method": "annealing"
}
else:
fit_kwargs = {
"method": args.fit_method,
"fig_dir": args.save_dir+"/",
"njobs": -1,
"NIND": args.geatpy_nind,
"MAXGEN": args.geatpy_maxgen,
"n_populations": args.geatpy_npop
}
dim, lb, ub = model.fit_params_range()
opt_res = find_best(
lambda x: score_func(x, model, score_kwargs),
dim, lb, ub, **fit_kwargs
)
# 把拟合得到的参数整理成dataframe，然后保存
temp_d, temp_i = {}, 0
for i, (k, vs) in enumerate(model.fit_params_info.items()):
params_k = opt_res["BestParam"][temp_i:(temp_i+vs[0])]
for j, v in enumerate(params_k):
temp_d[k+str(j)] = v
temp_i += vs[0]
pd.Series(temp_d).to_csv(
os.path.join(args.save_dir, "params.csv")
)
# 将得到的最优参数设置到模型中，并保存
model.set_params(opt_res["BestParam"])
model.save(os.path.join(args.save_dir, "model.pkl"))
utils.save(opt_res, os.path.join(args.save_dir, "opt_res.pkl"))
# 预测结果
prot_preds = model.predict(dataset.pred_times.delta)
model.protect = False
nopr_preds = model.predict(dataset.pred_times.delta)
""" 计算相关指标以及绘制图像 """
# 预测R0
pass
# 计算每个地区的曲线下面积以及面积差,并保存
auc = under_area(
dataset.epi_times.delta,
dataset.trueH,
dataset.pred_times.delta,
nopr_preds[0],
)
auc_df = pd.DataFrame(
auc.T, columns=["true_area", "pred_area", "diff_area"],
index=dataset.regions
)
auc_df["population"] = dataset.populations
auc_df["diff_norm"] = auc_df.diff_area / auc_df.population
auc_df.sort_values("diff_norm", inplace=True)
# 为每个地区绘制曲线图
plt.rcParams["font.sans-serif"] = ["SimHei"]
img_dir = os.path.join(args.save_dir, "imgs")
if not os.path.exists(img_dir):
os.mkdir(img_dir)
for i, reg in enumerate(dataset.regions):
"""
y0 = [H R E A I Sq Eq] + [S]
"""
plot_one_regions(
reg, [
("trueH", dataset.epi_times.ord.astype("int"),
dataset.trueH[:, i], "ro"),
# ("trueR", dataset.epi_times.ord.astype("int"),
# dataset.trueR[:, i]+dataset.trueD[:, i], "bo"),
("predH", dataset.pred_times.ord.astype("int"),
prot_preds[0][:, i], "r"),
# ("predR", dataset.pred_times.ord.astype("int"),
# prot_preds[1][:, i], "b"),
# ("predE", dataset.pred_times.ord.astype("int"),
# prot_preds[3][:, i], "y"),
# ("predA", dataset.pred_times.ord.astype("int"),
# prot_preds[4][:, i], "g"),
# ("predI", dataset.pred_times.ord.astype("int"),
# prot_preds[4][:, i], "c"),
],
[
("trueH", dataset.epi_times.ord.astype("int"),
dataset.trueH[:, i], "ro"),
# ("trueR", dataset.epi_times.ord.astype("int"),
# dataset.trueR[:, i]+dataset.trueD[:, i], "bo"),
("predH", dataset.pred_times.ord.astype("int"),
nopr_preds[0][:, i], "r"),
# ("predR", dataset.pred_times.ord.astype("int"),
# nopr_preds[1][:, i], "b"),
# ("predE", dataset.pred_times.ord.astype("int"),
# nopr_preds[3][:, i], "y"),
# ("predA", dataset.pred_times.ord.astype("int"),
# nopr_preds[4][:, i], "g"),
# ("predI", dataset.pred_times.ord.astype("int"),
# nopr_preds[4][:, i], "c"),
],
save_dir=img_dir
)
# 保存结果
for i, name in enumerate([
"predH", "predR", "predE", "predA", "predI"
]):
pd.DataFrame(
prot_preds[i],
columns=dataset.regions,
index=dataset.pred_times.str
).to_csv(
os.path.join(args.save_dir, "protect_%s.csv" % name)
)
pd.DataFrame(
nopr_preds[i],
columns=dataset.regions,
index=dataset.pred_times.str
).to_csv(
os.path.join(args.save_dir, "noprotect_%s.csv" % name)
)
auc_df.to_csv(os.path.join(args.save_dir, "auc.csv"))
# 这里保存的是原始数据
for i, attr_name in enumerate(["trueD", "trueH", "trueR"]):
save_arr = getattr(dataset, attr_name)
pd.DataFrame(
save_arr,
columns=dataset.regions,
index=dataset.epi_times.str
).to_csv(os.path.join(args.save_dir, "%s.csv" % attr_name))
# 保存args到路径中（所有事情都完成再保存数据，安全）
save_args = deepcopy(args.__dict__)
save_args["model_type"] = "NetSEAIRQ"
utils.save(save_args, os.path.join(args.save_dir, "args.json"), "json")
def main():
""" 命令行参数及其整理 """
parser = utils.MyArguments()
parser.add_argument("--De", default=5.2, type=float)
parser.add_argument("--Di", default=11.5, type=float)
parser.add_argument("--alpha_E", default=0.0, type=float)
parser.add_argument("--alpha_I", default=0.4, type=float)
parser.add_argument("--protect_k", default=0.0, type=float)
parser.add_argument("--use_19", action="store_true")
parser.add_argument("--zero_spring", action="store_true")
args = parser.parse_args() # 对于一些通用的参数，这里已经进行整理了
""" 读取准备好的数据 """
# if args.region_type == "city":
# dat_file = "./DATA/City.pkl"
# else:
dat_file = "./DATA/Provinces.pkl"
dataset = utils.Dataset(dat_file, args.t0, args.tm, args.fit_time_start)
""" 构建、或读取、或训练模型 """
# 根据不同的情况来得到合适的模型
if args.model is not None:
model = NetSEIR.load(args.model)
else:
model = NetSEIR(
De=args.De,
Di=args.Di,
populations=dataset.populations,
y0for1=args.y0,
alpha_I=args.alpha_I,
alpha_E=args.alpha_E,
protect=True,
score_type=args.fit_score,
protect_args={
"t0": dataset.protect_t0.delta,
"k": args.protect_k
},
gamma_func_kwargs={
"gammas":
(dataset.out19_dict if args.use_19 else dataset.out20_dict),
"zero_period":
(dataset.zero_period.delta if args.zero_spring else None)
},
Pmn_func_kwargs={"pmn": dataset.pmn_matrix_relative})
if args.fit:
# 设置我们拟合模型需要的数据
if args.use_whhb:
mask = None
else:
mask = np.full(dataset.num_regions, True, dtype=np.bool)
mask[0] = False
fit_start_index = (dataset.fit_start_t.ord - dataset.epi_t0.ord)
fit_start_index = int(fit_start_index)
score_kwargs = {
"times": dataset.epi_times.delta[fit_start_index:],
"true_infects": dataset.trueH[fit_start_index:, :],
"mask": mask,
}
# 搜索
if args.fit_method == "annealing":
fit_kwargs = {"callback": utils.callback, "method": "annealing"}
else:
fit_kwargs = {
"method": "SEGA",
"fig_dir": args.save_dir + "/",
"njobs": -1,
"NIND": args.geatpy_nind,
"MAXGEN": args.geatpy_maxgen,
"n_populations": args.geatpy_npop
}
dim, lb, ub = model.fit_params_range()
opt_res = find_best(lambda x: score_func(x, model, score_kwargs), dim,
lb, ub, **fit_kwargs)
# 把拟合得到的参数整理成dataframe，然后保存
temp_d, temp_i = {}, 0
for i, (k, vs) in enumerate(model.fit_params_info.items()):
params_k = opt_res["BestParam"][temp_i:(temp_i + vs[0])]
for j, v in enumerate(params_k):
temp_d[k + str(j)] = v
temp_i += vs[0]
pd.Series(temp_d).to_csv(os.path.join(args.save_dir, "params.csv"))
# 将得到的最优参数设置到模型中，并保存
model.set_params(opt_res["BestParam"])
model.save(os.path.join(args.save_dir, "model.pkl"))
utils.save(opt_res, os.path.join(args.save_dir, "opt_res.pkl"))
# 预测结果
prot_preds = model.predict(dataset.pred_times.delta)
model.protect = False
nopr_preds = model.predict(dataset.pred_times.delta)
""" 计算相关指标以及绘制图像 """
# 预测R0
pass
# 计算每个地区的曲线下面积以及面积差,并保存
auc = under_area(
dataset.epi_times.delta,
dataset.trueH,
dataset.pred_times.delta,
nopr_preds[2],
)
auc_df = pd.DataFrame(auc.T,
columns=["true_area", "pred_area", "diff_area"],
index=dataset.regions)
auc_df["population"] = dataset.populations
auc_df["diff_norm"] = auc_df.diff_area / auc_df.population
auc_df.sort_values("diff_norm", inplace=True)
# utils.save(auc, os.path.join(args.save_dir, "auc.pkl"))
# 为每个地区绘制曲线图
plt.rcParams["font.sans-serif"] = ["SimHei"]
img_dir = os.path.join(args.save_dir, "imgs")
if not os.path.exists(img_dir):
os.mkdir(img_dir)
for i, reg in enumerate(dataset.regions):
plot_one_regions(reg, [("true", dataset.epi_times.ord.astype("int"),
dataset.trueH[:, i], "ro"),
("predI", dataset.pred_times.ord.astype("int"),
prot_preds[2][:, i], "r"),
("predE", dataset.pred_times.ord.astype("int"),
prot_preds[1][:, i], "y"),
("predR", dataset.pred_times.ord.astype("int"),
prot_preds[3][:, i], "b")],
[("true", dataset.epi_times.ord.astype("int"),
dataset.trueH[:, i], "ro"),
("predI", dataset.pred_times.ord.astype("int"),
nopr_preds[2][:, i], "r"),
("predE", dataset.pred_times.ord.astype("int"),
nopr_preds[1][:, i], "y"),
("predR", dataset.pred_times.ord.astype("int"),
nopr_preds[3][:, i], "b")],
save_dir=img_dir)
# 保存结果
for i, name in enumerate(["predS", "predE", "predI", "predR"]):
pd.DataFrame(prot_preds[i],
columns=dataset.regions,
index=dataset.pred_times.str).to_csv(
os.path.join(args.save_dir, "protect_%s.csv" % name))
pd.DataFrame(nopr_preds[i],
columns=dataset.regions,
index=dataset.pred_times.str).to_csv(
os.path.join(args.save_dir,
"noprotect_%s.csv" % name))
auc_df.to_csv(os.path.join(args.save_dir, "auc.csv"))
# 这里保存的是原始数据
for i, attr_name in enumerate(["trueD", "trueH", "trueR"]):
save_arr = getattr(dataset, attr_name)
pd.DataFrame(save_arr,
columns=dataset.regions,
index=dataset.epi_times.str).to_csv(
os.path.join(args.save_dir, "%s.csv" % attr_name))
# 保存args到路径中（所有事情都完成再保存数据，安全）
save_args = deepcopy(args.__dict__)
save_args["model_type"] = "NetSEIR"
utils.save(save_args, os.path.join(args.save_dir, "args.json"), "json")