parallel-hyperparameter-tuning/train.py at main · lillianpetersen/parallel-hyperparameter-tuning · GitHub

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
import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import Dataset
import torch.optim as optim

import base
import classes

############################################
# train model, given hyperparameters
############################################

def train_model(model_class, X_train, X_val, y_train, y_val, device, lr, size, batch_size, modeldir, save_descript, hidden_dim=None, dropout=None):
	'''
	Inputs:
		model_class: either classes.Linear or classes.OneHidden
		X_train: numpy array (N x size)
		X_val: numpy array (N x size)
		y_train: numpy array (N) (this code is made for regression)
		y_val: numpy array (N)
		device: str, eg 'cuda:0'
		lr: learning rate
		size: number of features in X
		batch_size: int, batch size
		model_dir: str, directory to save best model to
		save_descript: str, name to save best model under
		hidden_dim: int, hidden dimension of OneHidden
		dropout: float between 0 and 1. dropout percentage of OneHidden
	No output, just saves best model
	'''

	# data
	train_data = classes.RegularDataset(X_train, y_train)
	val_data = classes.RegularDataset(X_val, y_val)
	train_dataloader = DataLoader(train_data, batch_size=batch_size, shuffle=True)
	val_dataloader = DataLoader(val_data, batch_size=batch_size, shuffle=True)

	if model_class==classes.Linear:
		model = model_class(size).to(device)
	elif model_class==classes.OneHidden:
		model = model_class(size, hidden_dim, dropout).to(device)

	optimizer = optim.Adam(model.parameters(), lr=lr)

	val_loss_curve = []
	train_loss_curve = []

	best_loss = 3
	best_epoch = 0
	epoch = -1
	while(True):
		epoch += 1
		# Train model on training data
		epoch_loss = base.train(model, train_dataloader, optimizer, device=device)

		# Validate on validation data
		val_loss = base.validate(model, val_dataloader, device=device)

		# Record train and loss performance
		train_loss_curve.append(epoch_loss)
		val_loss_curve.append(val_loss)

		if val_loss < best_loss:
			best_loss = val_loss
			torch.save(model.state_dict(), modeldir+'/best_'+save_descript+'.pt')
			best_epoch = epoch
			print(best_epoch, best_loss)

		if epoch > best_epoch+50: break

############################################
# apply trained model to val set
############################################

def apply_model(model_class, X_test, y_test, device, size, batch_size, modeldir, figdir, save_descript, title, hidden_dim=None, dropout=None):
	'''
	Inputs:
		model_class: either classes.Linear or classes.OneHidden
		X_test: numpy array (N x size)
		y_train: numpy array (N) (this code is made for regression)
		device: str, eg 'cuda:0'
		size: number of features in X
		batch_size: int, batch size
		model_dir: str, directory to save best model to
		save_descript: str, name to save best model under
		title: title of plot
		hidden_dim: int, hidden dimension of OneHidden
		dropout: float between 0 and 1. dropout percentage of OneHidden
	No output, just saves best model
	'''

	# data
	test_data = classes.RegularDataset(X_test, y_test)
	test_dataloader = DataLoader(test_data, batch_size=batch_size, shuffle=False)

	if model_class==classes.Linear:
		model = model_class(size).to(device)
	elif model_class==classes.OneHidden:
		model = model_class(size, hidden_dim, dropout).to(device)
	else:
		raise Exception("Unsupported model type")

	model.load_state_dict(torch.load(modeldir+'/best_'+save_descript+'.pt'))
	model.eval()

	model = model.to(device)
	preds = []
	y_true = []
	with torch.no_grad():
		for batch in test_dataloader:
			X_batch, y_batch = batch
			X_batch = X_batch.to(device)
			if GRU:
				h = model.init_hidden(X_batch.shape[0])
				h = h.data
				y_pred, h = model(X_batch, h)
			else:
				y_pred = model(X_batch)
			preds.extend(y_pred.tolist())
			y_true.extend(y_batch.tolist())

	rmse_test, corr_test = benchmark.benchmark(np.array(y_true), np.array(preds), title, save_descript, figdir, test_within_group=test_within_group, groups=groups)

	return rmse_test, corr_test