-
Notifications
You must be signed in to change notification settings - Fork 1
/
mtask_pytorch.py
130 lines (111 loc) · 5.22 KB
/
mtask_pytorch.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
"""
SCRIPT TO IMPLEMENT DIFFERENT MULTITASK ARCHITECTURES
"""
import numpy as np
import pandas as pd
import os
from support.support_funs import stopifnot, makeifnot, find_dir_nsqip
from sklearn import metrics
from time import time
import torch
from support.mtask_network import mtask_nn
from support.fpc_lasso import FPC
###############################
# ---- STEP 1: LOAD DATA ---- #
# Set directories
dir_NSQIP = find_dir_nsqip()
dir_output = os.path.join(dir_NSQIP, 'output')
assert os.path.exists(dir_output)
dir_figures = os.path.join(dir_NSQIP, 'figures')
makeifnot(dir_figures)
dir_weights = os.path.join(dir_output, 'weights')
makeifnot(dir_weights)
fn_X = 'X_imputed.csv'
fn_Y = 'y_agg.csv'
dat_X = pd.read_csv(os.path.join(dir_output,fn_X))
dat_Y = pd.read_csv(os.path.join(dir_output,fn_Y))
print(dat_X.shape); print(dat_Y.shape)
stopifnot(all(dat_X.caseid == dat_Y.caseid))
u_years = dat_X.operyr.unique()
# !! ENCODE CPT AS CATEGORICAL !! #
dat_X['cpt'] = 'c'+dat_X.cpt.astype(str)
cn_X = list(dat_X.columns[2:])
# Split Y into the agg vs not
dat_agg = dat_Y.loc[:,dat_Y.columns.str.contains('^agg|caseid|operyr')]
dat_Y = dat_Y.loc[:,~dat_Y.columns.str.contains('^agg')]
cn_Y = list(dat_Y.columns[2:])
cn_agg = list(dat_agg.columns[2:])
# # If we use 2012/13 as baseline years, what is the y-prop?
# prop_Y = dat_Y.groupby('operyr')[cn_Y].apply(lambda x: x[~(x==-1)].mean()).reset_index()
# prop_Y = prop_Y.melt('operyr',var_name='outcome')
# tmp = dat_Y.groupby('operyr')[cn_Y].apply(lambda x: (x==-1).sum()).reset_index().melt('operyr',
# value_name='n',var_name='outcome')
# prop_Y = prop_Y.merge(tmp[tmp.n > 0],how='left',on=['operyr','outcome'])
# prop_Y = prop_Y[prop_Y.n.isnull()].reset_index(drop=True).drop(columns='n')
# prop_Y['l10'] = -np.log10(prop_Y.value)
# g = sns.FacetGrid(data=prop_Y,col='outcome',col_wrap=5,sharey=True,sharex=True)
# g.map(sns.scatterplot,'operyr','l10')
# g.savefig(os.path.join(dir_figures,'outcome_prop.png'))
#####################################
# ---- STEP 2: TRAIN THE MODEL ---- #
seed = 1234
np.random.seed(seed)
torch.manual_seed(seed)
train_years = [2012, 2013]
test_years = np.setdiff1d(u_years, train_years)
yy=test_years[0]
for yy in test_years:
print('Training years: %s, test year: %i' % (', '.join([str(x) for x in train_years]),yy))
idx_train = dat_X.operyr.isin(train_years)
idx_test = (dat_X.operyr == yy)
Xtrain, Xtest = dat_X.loc[idx_train, cn_X].reset_index(drop=True), \
dat_X.loc[idx_test, cn_X].reset_index(drop=True)
Ytrain, Ytest = dat_Y.loc[idx_train, cn_Y].reset_index(drop=True), \
dat_Y.loc[idx_test, cn_Y].reset_index(drop=True)
YAggtrain, YAggtest = dat_agg.loc[idx_train,cn_agg].reset_index(drop=True), \
dat_agg.loc[idx_test,cn_agg].reset_index(drop=True)
cpt_train , cpt_test = dat_X.loc[idx_train,'cpt'].reset_index(drop=True), \
dat_X.loc[idx_test, 'cpt'].reset_index(drop=True)
caseid_train, caseid_test = dat_X.loc[idx_train,'caseid'].values, \
dat_X.loc[idx_test,'caseid']
# Initialize NN model
mdl = mtask_nn()
# Fit model
stime = time()
mdl.fit(data=Xtrain,lbls=Ytrain,nepochs=2000,mbatch=1000,val_prop=0.1,lr=0.001)
rtime = (time() - stime)/60
print('----- took %0.1f minutes to train model -----' % (rtime))
# fn_weights = pd.Series(os.listdir(dir_weights))
# fn_weights = fn_weights[fn_weights.str.contains(str(yy)+'.pt$')].to_list()
# if len(fn_weights)==1:
# mdl.load_state_dict(torch.load(os.path.join(dir_weights, fn_weights[0])))
# Save network weights
torch.save(mdl.nnet.state_dict(),
os.path.join(dir_weights, 'mtask5_' + str(yy) + '.pt'))
# Train sparse model on top for aggregated outcomes
phat_train = mdl.predict(data=Xtrain, mbatch=10000)
mdl_FPC = dict(zip(cn_agg,[FPC(standardize=True) for ii in range(len(cn_agg))]))
for ii, cc in enumerate(cn_agg):
print('Aggregated column: %s (%i of %i)' % (cc, ii+1, len(cn_agg)))
y_cc = YAggtrain[cc].values
idx_cc = np.where((y_cc == 0) | (y_cc == 1))[0]
mdl_FPC[cc].fit(phat_train[idx_cc],y_cc[idx_cc],2)
# Get test probabilities
phat_test = mdl.predict(data=Xtest, mbatch=10000)
fpc_test = np.vstack([mdl_FPC[cc].predict(phat_test) for cc in cn_agg]).T
y_test = pd.concat([Ytest, YAggtest], axis=1)
df_test = pd.DataFrame(np.c_[phat_test, fpc_test], columns=cn_Y + cn_agg)
stopifnot(all(df_test.columns == y_test.columns))
holder = []
for cc in df_test.columns:
df_cc = pd.DataFrame({'lbl':cc,'y':y_test[cc],'phat':df_test[cc],
'cpt':cpt_test,'caseid':caseid_test.values})
holder.append(df_cc)
df_test = pd.concat(holder)
df_test.insert(0,'operyr',yy)
df_test.to_csv(os.path.join(dir_weights,'df_test_' + str(yy) + '.csv'),index=True)
# Print the average test performance
print(df_test[~(df_test.y == -1)].groupby('lbl').apply(lambda x: pd.Series({'auc': metrics.roc_auc_score(x['y'], x['phat']),
'pr':metrics.average_precision_score(x['y'], x['phat'])})))
# Update the training years
train_years.append(yy)