import pandas as pd
import numpy as np
import seaborn as sns; sns.set()
import matplotlib.pyplot as plt
import matplotlib.dates as mdates
from sklearn.linear_model import LinearRegression
from scipy import stats
import statsmodels.api as sm
import pylab
# from google.colab import files
# from io import StringIO
# uploaded = files.upload()
url = 'https://raw.githubusercontent.com/assemzh/ProbProg-COVID-19/master/full_grouped.csv'
data = pd.read_csv(url)
data.Date = pd.to_datetime(data.Date)
# for fancy python printing
from IPython.display import Markdown, display
def printmd(string):
display(Markdown(string))
import warnings
warnings.filterwarnings('ignore')
import matplotlib as mpl
mpl.rcParams['figure.dpi'] = 250
/usr/local/lib/python3.7/dist-packages/statsmodels/tools/_testing.py:19: FutureWarning: pandas.util.testing is deprecated. Use the functions in the public API at pandas.testing instead. import pandas.util.testing as tm
# function to make the time series of confirmed and daily confirmed cases for a specific country
def create_country (country, end_date, state = False) :
if state :
df = data.loc[data["Province/State"] == country, ["Province/State", "Date", "Confirmed", "Deaths", "Recovered"]]
else :
df = data.loc[data["Country/Region"] == country, ["Country/Region", "Date", "Confirmed", "Deaths", "Recovered"]]
df.columns = ["country", "date", "confirmed", "deaths", "recovered"]
# group by country and date, sum(confirmed, deaths, recovered). do this because countries have multiple cities
df = df.groupby(['country','date'])['confirmed', 'deaths', 'recovered'].sum().reset_index()
# convert date string to datetime
std_dateparser = lambda x: str(x)[5:10]
df.date = pd.to_datetime(df.date)
df['date_only'] = df.date.apply(std_dateparser)
df = df.sort_values(by = "date")
df = df[df.date <= end_date]
# make new confirmed cases every day:
cases_shifted = np.array([0] + list(df.confirmed[:-1]))
daily_confirmed = np.array(df.confirmed) - cases_shifted
df["daily_confirmed"] = daily_confirmed
fig, ax = plt.subplots(nrows=1, ncols=1, figsize=(7, 6))
ax = [ax]
sns.lineplot(x = df.date,
y = df.daily_confirmed,
ax = ax[0])
ax[0].set(ylabel='Daily Confirmed Cases')
ax[0].axvline(pd.to_datetime('2020-01-22'),
linestyle = '--', linewidth = 1.5,
label = "Date of Lockdown: Jan 22, 2020" ,
color = "red")
# ax[0].set_ylim([100,150000])
ax[0].xaxis.get_label().set_fontsize(16)
ax[0].yaxis.get_label().set_fontsize(16)
ax[0].title.set_fontsize(20)
ax[0].tick_params(labelsize=16)
myFmt = mdates.DateFormatter('%b %-d')
ax[0].xaxis.set_major_formatter(myFmt)
ax[0].set(ylabel='Daily Confirmed Cases', xlabel='');
ax[0].legend(loc = "bottom right", fontsize=12.8)
sns.set_style("ticks")
plt.tight_layout()
sns.despine()
plt.savefig('/content/sample_data/china_daily.pdf')
print(df.tail())
return df
def summary(samples):
site_stats = {}
for k, v in samples.items():
site_stats[k] = {
"mean": torch.mean(v, 0),
"std": torch.std(v, 0),
"5%": v.kthvalue(int(len(v) * 0.05), dim=0)[0],
"95%": v.kthvalue(int(len(v) * 0.95), dim=0)[0],
}
return site_stats
cad = create_country("China", end_date = "2020-04-01")
country date confirmed ... date_only daily_confirmed moving_avg 66 China 2020-03-28 81999 ... 03-28 102 102.00 67 China 2020-03-29 82122 ... 03-29 123 115.25 68 China 2020-03-30 82198 ... 03-30 76 104.00 69 China 2020-03-31 82279 ... 03-31 81 95.50 70 China 2020-04-01 82361 ... 04-01 82 90.50 [5 rows x 8 columns]
cad_start = "2020-01-22" # 13 confirmed cases
cad = cad[cad.date >= cad_start].reset_index(drop = True)
cad["days_since_start"] = np.arange(cad.shape[0]) + 1
cad.shape
cad_tmp = cad[cad.date < "2020-03-16"]
cad_tmp.shape
(54, 9)
# variable for data to easily swap it out:
country_ = "China"
reg_data = cad_tmp.copy()
reg_data.head()
country | date | confirmed | deaths | recovered | date_only | daily_confirmed | moving_avg | days_since_start | |
---|---|---|---|---|---|---|---|---|---|
0 | China | 2020-01-22 | 548 | 17 | 28 | 01-22 | 548 | NaN | 1 |
1 | China | 2020-01-23 | 643 | 18 | 30 | 01-23 | 95 | NaN | 2 |
2 | China | 2020-01-24 | 920 | 26 | 36 | 01-24 | 277 | NaN | 3 |
3 | China | 2020-01-25 | 1406 | 42 | 39 | 01-25 | 486 | 351.50 | 4 |
4 | China | 2020-01-26 | 2075 | 56 | 49 | 01-26 | 669 | 381.75 | 5 |
!pip install pyro-ppl
!pip install numpyro
Requirement already satisfied: pyro-ppl in /usr/local/lib/python3.7/dist-packages (1.6.0) Requirement already satisfied: tqdm>=4.36 in /usr/local/lib/python3.7/dist-packages (from pyro-ppl) (4.41.1) Requirement already satisfied: torch>=1.8.0 in /usr/local/lib/python3.7/dist-packages (from pyro-ppl) (1.8.1+cu101) Requirement already satisfied: opt-einsum>=2.3.2 in /usr/local/lib/python3.7/dist-packages (from pyro-ppl) (3.3.0) Requirement already satisfied: pyro-api>=0.1.1 in /usr/local/lib/python3.7/dist-packages (from pyro-ppl) (0.1.2) Requirement already satisfied: numpy>=1.7 in /usr/local/lib/python3.7/dist-packages (from pyro-ppl) (1.19.5) Requirement already satisfied: typing-extensions in /usr/local/lib/python3.7/dist-packages (from torch>=1.8.0->pyro-ppl) (3.7.4.3) Requirement already satisfied: numpyro in /usr/local/lib/python3.7/dist-packages (0.6.0) Requirement already satisfied: tqdm in /usr/local/lib/python3.7/dist-packages (from numpyro) (4.41.1) Requirement already satisfied: jaxlib==0.1.62 in /usr/local/lib/python3.7/dist-packages (from numpyro) (0.1.62) Requirement already satisfied: jax==0.2.10 in /usr/local/lib/python3.7/dist-packages (from numpyro) (0.2.10) Requirement already satisfied: flatbuffers in /usr/local/lib/python3.7/dist-packages (from jaxlib==0.1.62->numpyro) (1.12) Requirement already satisfied: numpy>=1.16 in /usr/local/lib/python3.7/dist-packages (from jaxlib==0.1.62->numpyro) (1.19.5) Requirement already satisfied: absl-py in /usr/local/lib/python3.7/dist-packages (from jaxlib==0.1.62->numpyro) (0.12.0) Requirement already satisfied: scipy in /usr/local/lib/python3.7/dist-packages (from jaxlib==0.1.62->numpyro) (1.4.1) Requirement already satisfied: opt-einsum in /usr/local/lib/python3.7/dist-packages (from jax==0.2.10->numpyro) (3.3.0) Requirement already satisfied: six in /usr/local/lib/python3.7/dist-packages (from absl-py->jaxlib==0.1.62->numpyro) (1.15.0)
import torch
import pyro
import pyro.distributions as dist
from torch import nn
from pyro.nn import PyroModule, PyroSample
from pyro.infer import MCMC, NUTS, HMC
from pyro.infer.autoguide import AutoGuide, AutoDiagonalNormal
from pyro.infer import SVI, Trace_ELBO
from pyro.infer import Predictive
# we should be able to have an empirical estimate for the mean of the prior for the 2nd regression bias term
# this will be something like b = log(max(daily_confirmed))
# might be able to have 1 regression model but change the data so that we have new terms for (tau < t)
# like an interaction term
class COVID_change(PyroModule):
def __init__(self, in_features, out_features, b1_mu, b2_mu):
super().__init__()
self.linear1 = PyroModule[nn.Linear](in_features, out_features, bias = False)
self.linear1.weight = PyroSample(dist.Normal(0.5, 0.25).expand([1, 1]).to_event(1))
self.linear1.bias = PyroSample(dist.Normal(b1_mu, 1.))
# could possibly have stronger priors for the 2nd regression line, because we wont have as much data
self.linear2 = PyroModule[nn.Linear](in_features, out_features, bias = False)
self.linear2.weight = PyroSample(dist.Normal(0., 0.25).expand([1, 1])) #.to_event(1))
self.linear2.bias = PyroSample(dist.Normal(b2_mu, b2_mu/4))
def forward(self, x, y=None):
tau = pyro.sample("tau", dist.Beta(4, 3))
sigma = pyro.sample("sigma", dist.Uniform(0., 3.))
# fit lm's to data based on tau
sep = int(np.ceil(tau.detach().numpy() * len(x)))
mean1 = self.linear1(x[:sep]).squeeze(-1)
mean2 = self.linear2(x[sep:]).squeeze(-1)
mean = torch.cat((mean1, mean2))
obs = pyro.sample("obs", dist.StudentT(2, mean, sigma), obs=y)
return mean
tensor_data = torch.tensor(reg_data[["confirmed", "days_since_start", "daily_confirmed"]].values, dtype=torch.float)
x_data = tensor_data[:, 1].unsqueeze_(1)
y_data = np.log(tensor_data[:, 0])
y_data_daily = np.log(tensor_data[:, 2])
# prior hyper params
# take log of the average of the 1st quartile to get the prior mean for the bias of the 2nd regression line
q1 = np.quantile(y_data, q = 0.25)
bias_1_mean = np.mean(y_data.numpy()[y_data <= q1])
print("Prior mean for Bias 1: ", bias_1_mean)
# take log of the average of the 4th quartile to get the prior mean for the bias of the 2nd regression line
q4 = np.quantile(y_data, q = 0.75)
bias_2_mean = np.mean(y_data.numpy()[y_data >= q4])
print("Prior mean for Bias 2: ", bias_2_mean)
Prior mean for Bias 1: 8.359699 Prior mean for Bias 2: 11.29878
model = COVID_change(1, 1,
b1_mu = bias_1_mean,
b2_mu = bias_2_mean)
# need more than 400 samples/chain if we want to use a flat prior on b_2 and w_2
num_samples = 400
# mcmc
nuts_kernel = NUTS(model)
mcmc = MCMC(nuts_kernel,
num_samples=num_samples,
warmup_steps = 200,
num_chains = 1)
mcmc.run(x_data, y_data)
# Save the model:
import dill
# with open('china.pkl', 'wb') as f:
# dill.dump(mcmc, f)
with open('china.pkl', 'rb') as f:
mcmc = dill.load(f)
samples = mcmc.get_samples()
# extract individual posteriors
weight_1_post = samples["linear1.weight"].detach().numpy()
weight_2_post = samples["linear2.weight"].detach().numpy()
bias_1_post = samples["linear1.bias"].detach().numpy()
bias_2_post = samples["linear2.bias"].detach().numpy()
tau_post = samples["tau"].detach().numpy()
sigma_post = samples["sigma"].detach().numpy()
# build likelihood distribution:
tau_days = list(map(int, np.ceil(tau_post * len(x_data))))
mean_ = torch.zeros(len(tau_days), len(x_data))
obs_ = torch.zeros(len(tau_days), len(x_data))
for i in range(len(tau_days)) :
mean_[i, :] = torch.cat((x_data[:tau_days[i]] * weight_1_post[i] + bias_1_post[i],
x_data[tau_days[i]:] * weight_2_post[i] + bias_2_post[i])).reshape(len(x_data))
obs_[i, :] = dist.Normal(mean_[i, :], sigma_post[i]).sample()
samples["_RETURN"] = mean_
samples["obs"] = obs_
pred_summary = summary(samples)
mu = pred_summary["_RETURN"] # mean
y = pred_summary["obs"] # samples from likelihood: mu + sigma
y_shift = np.exp(y["mean"]) - np.exp(torch.cat((y["mean"][0:1], y["mean"][:-1])))
print(y_shift)
predictions = pd.DataFrame({
"days_since_start": x_data[:, 0],
"mu_mean": mu["mean"], # mean of likelihood
"mu_perc_5": mu["5%"],
"mu_perc_95": mu["95%"],
"y_mean": y["mean"], # mean of likelihood + noise
"y_perc_5": y["5%"],
"y_perc_95": y["95%"],
"true_confirmed": y_data,
"true_daily_confirmed": y_data_daily,
"y_daily_mean": y_shift
})
w1_ = pred_summary["linear1.weight"]
w2_ = pred_summary["linear2.weight"]
b1_ = pred_summary["linear1.bias"]
b2_ = pred_summary["linear2.bias"]
tau_ = pred_summary["tau"]
sigma_ = pred_summary["sigma"]
ind = int(np.ceil(tau_["mean"] * len(x_data)))
tensor([ 0.0000e+00, 2.2725e+02, 3.0094e+02, 4.0804e+02, 5.1011e+02, 6.8209e+02, 9.4119e+02, 1.2054e+03, 1.5588e+03, 2.1987e+03, 2.8797e+03, 3.6862e+03, 4.8820e+03, 6.5520e+03, 9.3445e+03, 1.1462e+04, 1.3558e+04, 5.6779e+03, 1.5542e+03, 1.3815e+02, 9.5227e+02, -2.4232e+02, 2.4934e+02, 7.3088e+02, 5.2862e+02, 7.5919e+02, -1.0573e+02, 3.9531e+00, 8.0497e+02, 9.6861e+02, -5.0664e+01, 5.5251e+02, 9.4084e+02, -2.7266e+02, 1.3263e+03, 3.4009e+02, 1.9668e+02, 1.1166e+03, -3.3041e+02, 7.6644e+02, 5.8625e+01, 1.3439e+03, -1.9008e+01, 7.6425e+02, 6.2384e+02, 1.4739e+02, 1.3451e+02, 9.4297e+02, 7.5945e+02, 1.1752e+03, -9.4339e+02, 1.7658e+03, 2.9969e+01, 4.2532e+02])
mcmc.summary()
diag = mcmc.diagnostics()
mean std median 5.0% 95.0% n_eff r_hat tau 0.31 0.03 0.31 0.26 0.35 15.20 1.21 sigma 0.08 0.02 0.08 0.05 0.11 55.90 1.01 linear1.weight[0,0] 0.28 0.03 0.29 0.21 0.31 8.85 1.21 linear1.bias 6.27 0.33 6.15 5.98 7.08 8.38 1.20 linear2.weight[0,0] 0.01 0.00 0.01 0.00 0.01 22.25 1.10 linear2.bias 11.00 0.07 11.00 10.90 11.10 23.06 1.10 Number of divergences: 0
print(ind)
print(reg_data.date[ind])
sns.distplot(weight_1_post,
kde_kws = {"label": "Weight posterior before CP"},
color = "red",
norm_hist = True,
kde = True)
plt.axvline(x = w1_["mean"], linestyle = '--',label = "Mean weight before CP" ,
color = "red")
sns.distplot(weight_2_post,
kde_kws = {"label": "Weight posterior after CP"},
color = "teal",
norm_hist = True,
kde = True)
plt.axvline(x = w2_["mean"], linestyle = '--',label = "Mean weight after CP" ,
color = "teal")
legend = plt.legend(loc='upper right')
legend.get_frame().set_alpha(1)
sns.set_style("ticks")
plt.tight_layout()
sns.despine()
plt.savefig('/content/sample_data/china_weights.pdf')
17 2020-02-08 00:00:00
predictions['date'] = pd.to_datetime(reg_data.date)
start_date_ = str(reg_data.date[0]).split(' ')[0]
change_date_ = str(reg_data.date[ind]).split(' ')[0]
print("Date of change for {}: {}".format(country_, change_date_))
import seaborn as sns
# plot data:
fig, ax = plt.subplots(nrows=1, ncols=1, figsize=(7, 5))
ax = [ax]
# log regression model
ax[0].scatter(y = np.exp(y_data[:ind]), x = x_data[:ind], s = 15);
ax[0].scatter(y = np.exp(y_data[ind:]), x = x_data[ind:], s = 15, color = "red");
ax[0].plot(predictions["days_since_start"],
np.exp(predictions["y_mean"]),
color = "green",
label = "Fitted line by MCMC-NUTS model")
ax[0].axvline(1,
linestyle = '--', linewidth = 1.5,
label = "Date of Lockdown: Jan 22, 2020" ,
color = "red")
ax[0].axvline(ind,
linestyle = '--', linewidth = 1.5,
label = "Date of Change: Feb 8, 2020",
color = "black")
ax[0].fill_between(predictions["days_since_start"],
np.exp(predictions["y_perc_5"]),
np.exp(predictions["y_perc_95"]),
alpha = 0.25,
label = "90% CI of predictions",
color = "teal");
ax[0].fill_betweenx([0, 1],
tau_["5%"] * len(x_data),
tau_["95%"] * len(x_data),
alpha = 0.25,
label = "90% CI of changing point",
color = "lightcoral",
transform=ax[0].get_xaxis_transform());
ax[0].set(ylabel = "Total Cases",)
# xlabel = "Days since %s" % start_date_,
# title = "Confirmed Cases in China") /
ax[0].legend(loc = "lower right", fontsize=12.8)
ax[0].set_ylim([100,150000])
ax[0].xaxis.get_label().set_fontsize(16)
ax[0].yaxis.get_label().set_fontsize(16)
ax[0].title.set_fontsize(20)
ax[0].tick_params(labelsize=16)
plt.xticks(ticks=[1,17,34,51], labels=["Jan 22",
"Feb 8",
"Feb 25",
"Mar 13"], fontsize=15)
ax[0].set_yscale('log')
plt.setp(ax[0].get_xticklabels(), rotation=0, horizontalalignment='center')
print(reg_data.columns)
myFmt = mdates.DateFormatter('%m-%d')
sns.set_style("ticks")
sns.despine()
ax[0].figure.savefig('/content/sample_data/china_cp.pdf')
Date of change for China (Before March 16th): 2020-02-08 Index(['country', 'date', 'confirmed', 'deaths', 'recovered', 'date_only', 'daily_confirmed', 'moving_avg', 'days_since_start'], dtype='object')