-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy path06a_CPM_deep.py
165 lines (135 loc) · 6.79 KB
/
06a_CPM_deep.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
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
#### Master Script 6a: Train deep learning concise-predictor-based models (CPM) ####
#
# Shubhayu Bhattacharyay
# University of Cambridge
# email address: sb2406@cam.ac.uk
#
### Contents:
# I. Initialisation
# II. Create grid of training combinations
# III. Train CPM_deep model based on provided hyperparameter row index
### I. Initialisation
# Fundamental libraries
import os
import re
import sys
import time
import glob
import random
import datetime
import warnings
import itertools
import numpy as np
import pandas as pd
import pickle as cp
import seaborn as sns
import multiprocessing
from scipy import stats
from pathlib import Path
from ast import literal_eval
import matplotlib.pyplot as plt
from collections import Counter
from argparse import ArgumentParser
os.environ['CUDA_LAUNCH_BLOCKING'] = "1"
warnings.filterwarnings(action="ignore")
# PyTorch, PyTorch.Text, and Lightning-PyTorch methods
import torch
from torch import nn, optim, Tensor
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
from torchtext.vocab import Vocab
import pytorch_lightning as pl
from pytorch_lightning.callbacks import ModelCheckpoint
from pytorch_lightning.callbacks.early_stopping import EarlyStopping
# SciKit-Learn methods
from sklearn.metrics import confusion_matrix, accuracy_score, roc_auc_score
from sklearn.preprocessing import LabelEncoder, KBinsDiscretizer, OneHotEncoder, StandardScaler
from sklearn.model_selection import StratifiedShuffleSplit
from sklearn.utils import resample
from sklearn.utils.class_weight import compute_class_weight
# TQDM for progress tracking
from tqdm import tqdm
# Custom methods
from classes.datasets import CONCISE_PREDICTOR_SET
from models.CPM import CPM_deep
from functions.model_building import train_CPM_deep
# Set version code
VERSION = 'DEEP_v1-0'
# Initialise model output directory based on version code
model_dir = '/home/sb2406/rds/hpc-work/CPM_outputs/'+VERSION
os.makedirs(model_dir,exist_ok=True)
# Load cross-validation split information
cv_splits = pd.read_csv('../cross_validation_splits.csv')
# Load the tuning grid
tuning_grid = pd.read_csv(os.path.join(model_dir,'CPM_deep_tuning_grid.csv'))
tuning_grid['TUNE_IDX'] = tuning_grid['TUNE_IDX'].astype(str).str.zfill(4)
tuning_grid['NEURONS'] = tuning_grid['NEURONS'].apply(eval)
### II. Create grid of training combinations
# Identify unique token index-CV combinations
uniq_splits = cv_splits[['repeat','fold']].drop_duplicates().reset_index(drop=True)
uniq_splits['key'] = 1
uniq_tune_idx = tuning_grid
uniq_tune_idx['key'] = 1
uniq_tune_idx = tuning_grid[['key','TUNE_IDX']]
cv_ti_combos = pd.merge(uniq_splits,uniq_tune_idx,how='outer',on='key').drop(columns='key')
# Define which repeat partitions to train in current run
REPEAT = [1]
cv_ti_combos = cv_ti_combos[cv_ti_combos.repeat.isin(REPEAT)].reset_index(drop=True)
### III. Train CPM_deep model based on provided hyperparameter row index
def main(array_task_id):
# Extract current repeat, fold, and tuning index information
curr_repeat = cv_ti_combos.repeat[array_task_id]
curr_fold = cv_ti_combos.fold[array_task_id]
curr_TUNE_IDX = cv_ti_combos.TUNE_IDX[array_task_id]
# Create a directory for the current repeat
repeat_dir = os.path.join(model_dir,'repeat'+str(curr_repeat).zfill(int(np.log10(cv_splits.repeat.max()))+1))
os.makedirs(repeat_dir,exist_ok=True)
# Create a directory for the current fold
fold_dir = os.path.join(repeat_dir,'fold'+str(curr_fold).zfill(int(np.log10(cv_splits.fold.max()))+1))
os.makedirs(fold_dir,exist_ok=True)
# Create a directory for the current tune index
tune_dir = os.path.join(fold_dir,'tune_'+str(curr_TUNE_IDX).zfill(int(np.log10(cv_splits.fold.max()))+1))
os.makedirs(tune_dir,exist_ok=True)
# Load current imputed training and testing sets
training_set = pd.read_csv('/home/sb2406/rds/hpc-work/imputed_CPM_sets/repeat'+str(curr_repeat).zfill(2)+'/fold'+str(curr_fold)+'/training_set.csv')
testing_set = pd.read_csv('/home/sb2406/rds/hpc-work/imputed_CPM_sets/repeat'+str(curr_repeat).zfill(2)+'/fold'+str(curr_fold)+'/testing_set.csv')
# One-hot encode categorical predictors
cat_encoder = OneHotEncoder(drop = 'first',categories=[[1,2,3,4,5,6],[1,2,3,4,5,6],[0,1,2]])
cat_column_names = ['GCSm_'+str(i+1) for i in range(1,6)] + ['marshall_'+str(i+1) for i in range(1,6)] + ['unreactive_pupils_'+str(i+1) for i in range(2)]
training_categorical = pd.DataFrame(cat_encoder.fit_transform(training_set[['GCSm','marshall','unreactive_pupils']]).toarray(),
columns=cat_column_names)
training_set = pd.concat([training_set.drop(columns=['GCSm','marshall','unreactive_pupils']),training_categorical],axis=1)
testing_categorical = pd.DataFrame(cat_encoder.transform(testing_set[['GCSm','marshall','unreactive_pupils']]).toarray(),
columns=cat_column_names)
testing_set = pd.concat([testing_set.drop(columns=['GCSm','marshall','unreactive_pupils']),testing_categorical],axis=1)
cp.dump(cat_encoder, open(os.path.join(fold_dir,'one_hot_encoder.pkl'), "wb"))
# Standardize numerical predictors
num_scaler = StandardScaler()
training_set[['age','Hb','glu']] = num_scaler.fit_transform(training_set[['age','Hb','glu']])
testing_set[['age','Hb','glu']] = num_scaler.transform(testing_set[['age','Hb','glu']])
cp.dump(num_scaler, open(os.path.join(fold_dir,'standardizer.pkl'), "wb"))
# Set aside 15% of the training set for validation, independent of the final testing set
sss = StratifiedShuffleSplit(n_splits=1, test_size=0.15)
for train_index, val_index in sss.split(training_set.drop(columns='GOSE'),training_set.GOSE):
training_set, val_set = training_set.loc[train_index].reset_index(drop=True), training_set.loc[val_index].reset_index(drop=True)
cp.dump(sss, open(os.path.join(tune_dir,'val_set_splitter.pkl'), "wb"))
curr_tune_configs = tuning_grid[tuning_grid.TUNE_IDX == curr_TUNE_IDX].reset_index(drop=True)
train_CPM_deep(training_set,
val_set,
testing_set,
curr_TUNE_IDX,
curr_repeat,
curr_fold,
fold_dir,
curr_tune_configs.BATCH_SIZE.values[0],
curr_tune_configs.LEARNING_RATE.values[0],
curr_tune_configs.LAYERS.values[0],
curr_tune_configs.NEURONS.values[0],
curr_tune_configs.DROPOUT.values[0],
curr_tune_configs.ES_PATIENCE.values[0],
curr_tune_configs.EPOCHS.values[0],
curr_tune_configs.CLASS_WEIGHTS.values[0],
curr_tune_configs.OUTPUT_ACTIVATION.values[0])
if __name__ == '__main__':
array_task_id = int(sys.argv[1])
main(array_task_id)