-
Notifications
You must be signed in to change notification settings - Fork 23
/
ml_builder.py
197 lines (165 loc) · 8.52 KB
/
ml_builder.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
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
import numpy as np
import random as rd
import xarray as xr
import pandas as pd
import matplotlib.pyplot as plt
import time as tm
import os
from model.stconvs2s import STConvS2S_R, STConvS2S_C
from model.baselines import *
from model.ablation import *
from tool.train_evaluate import Trainer, Evaluator
from tool.dataset import NetCDFDataset
from tool.loss import RMSELoss
from tool.utils import Util
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from torch import optim
class MLBuilder:
def __init__(self, config, device):
self.config = config
self.device = device
self.dataset_type = 'small-dataset' if (self.config.small_dataset) else 'full-dataset'
self.step = str(config.step)
self.dataset_name, self.dataset_file = self.__get_dataset_file()
self.dropout_rate = self.__get_dropout_rate()
self.filename_prefix = self.dataset_name + '_step' + self.step
def run_model(self, number):
self.__define_seed(number)
validation_split = 0.2
test_split = 0.2
# Loading the dataset
ds = xr.open_mfdataset(self.dataset_file)
if (self.config.small_dataset):
ds = ds[dict(sample=slice(0,500))]
train_dataset = NetCDFDataset(ds, test_split=test_split,
validation_split=validation_split)
val_dataset = NetCDFDataset(ds, test_split=test_split,
validation_split=validation_split, is_validation=True)
test_dataset = NetCDFDataset(ds, test_split=test_split,
validation_split=validation_split, is_test=True)
if (self.config.verbose):
print('[X_train] Shape:', train_dataset.X.shape)
print('[y_train] Shape:', train_dataset.y.shape)
print('[X_val] Shape:', val_dataset.X.shape)
print('[y_val] Shape:', val_dataset.y.shape)
print('[X_test] Shape:', test_dataset.X.shape)
print('[y_test] Shape:', test_dataset.y.shape)
print(f'Train on {len(train_dataset)} samples, validate on {len(val_dataset)} samples')
params = {'batch_size': self.config.batch,
'num_workers': self.config.workers,
'worker_init_fn': self.__init_seed}
train_loader = DataLoader(dataset=train_dataset, shuffle=True,**params)
val_loader = DataLoader(dataset=val_dataset, shuffle=False,**params)
test_loader = DataLoader(dataset=test_dataset, shuffle=False, **params)
models = {
'stconvs2s-r': STConvS2S_R,
'stconvs2s-c': STConvS2S_C,
'convlstm': STConvLSTM,
'predrnn': PredRNN,
'mim': MIM,
'conv2plus1d': Conv2Plus1D,
'conv3d': Conv3D,
'enc-dec3d': Endocer_Decoder3D,
'ablation-stconvs2s-nocausalconstraint': AblationSTConvS2S_NoCausalConstraint,
'ablation-stconvs2s-notemporal': AblationSTConvS2S_NoTemporal,
'ablation-stconvs2s-r-nochannelincrease': AblationSTConvS2S_R_NoChannelIncrease,
'ablation-stconvs2s-c-nochannelincrease': AblationSTConvS2S_C_NoChannelIncrease,
'ablation-stconvs2s-r-inverted': AblationSTConvS2S_R_Inverted,
'ablation-stconvs2s-c-inverted': AblationSTConvS2S_C_Inverted,
'ablation-stconvs2s-r-notfactorized': AblationSTConvS2S_R_NotFactorized,
'ablation-stconvs2s-c-notfactorized': AblationSTConvS2S_C_NotFactorized
}
if not(self.config.model in models):
raise ValueError(f'{self.config.model} is not a valid model name. Choose between: {models.keys()}')
quit()
# Creating the model
model_bulder = models[self.config.model]
model = model_bulder(train_dataset.X.shape, self.config.num_layers, self.config.hidden_dim,
self.config.kernel_size, self.device, self.dropout_rate, int(self.step))
model.to(self.device)
criterion = RMSELoss()
opt_params = {'lr': 0.001,
'alpha': 0.9,
'eps': 1e-6}
optimizer = torch.optim.RMSprop(model.parameters(), **opt_params)
util = Util(self.config.model, self.dataset_type, self.config.version, self.filename_prefix)
train_info = {'train_time': 0}
if self.config.pre_trained is None:
train_info = self.__execute_learning(model, criterion, optimizer, train_loader, val_loader, util)
eval_info = self.__load_and_evaluate(model, criterion, optimizer, test_loader,
train_info['train_time'], util)
if (torch.cuda.is_available()):
torch.cuda.empty_cache()
return {**train_info, **eval_info}
def __execute_learning(self, model, criterion, optimizer, train_loader, val_loader, util):
checkpoint_filename = util.get_checkpoint_filename()
trainer = Trainer(model, criterion, optimizer, train_loader, val_loader, self.config.epoch,
self.device, util, self.config.verbose, self.config.patience, self.config.no_stop)
start_timestamp = tm.time()
# Training the model
train_losses, val_losses = trainer.fit(checkpoint_filename, is_chirps=self.config.chirps)
end_timestamp = tm.time()
# Learning curve
util.save_loss(train_losses, val_losses)
util.plot([train_losses, val_losses], ['Training', 'Validation'], 'Epochs', 'Loss',
'Learning curve - ' + self.config.model.upper(), self.config.plot)
train_time = end_timestamp - start_timestamp
print(f'\nTraining time: {util.to_readable_time(train_time)} [{train_time}]')
return {'dataset': self.dataset_name,
'dropout_rate': self.dropout_rate,
'train_time': train_time
}
def __load_and_evaluate(self, model, criterion, optimizer, test_loader, train_time, util):
evaluator = Evaluator(model, criterion, optimizer, test_loader, self.device, util, self.step)
if self.config.pre_trained is not None:
# Load pre-trained model
best_epoch, val_loss = evaluator.load_checkpoint(self.config.pre_trained, self.dataset_type, self.config.model)
else:
# Load model with minimal loss after training phase
checkpoint_filename = util.get_checkpoint_filename()
best_epoch, val_loss = evaluator.load_checkpoint(checkpoint_filename)
time_per_epochs = 0
if not(self.config.no_stop): # Earling stopping during training
time_per_epochs = train_time / (best_epoch + self.config.patience)
print(f'Training time/epochs: {util.to_readable_time(time_per_epochs)} [{time_per_epochs}]')
test_rmse, test_mae = evaluator.eval(is_chirps=self.config.chirps)
print(f'Test RMSE: {test_rmse:.4f}\nTest MAE: {test_mae:.4f}')
return {'best_epoch': best_epoch,
'val_rmse': val_loss,
'test_rmse': test_rmse,
'test_mae': test_mae,
'train_time_epochs': time_per_epochs
}
def __define_seed(self, number):
if (~self.config.no_seed):
# define a different seed in every iteration
seed = (number * 10) + 1000
np.random.seed(seed)
rd.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic=True
def __init_seed(self, number):
seed = (number * 10) + 1000
np.random.seed(seed)
def __get_dataset_file(self):
dataset_file, dataset_name = None, None
if (self.config.chirps):
dataset_file = 'data/dataset-chirps-1981-2019-seq5-ystep' + self.step + '.nc'
dataset_name = 'chirps'
else:
dataset_file = 'data/dataset-ucar-1979-2015-seq5-ystep' + self.step + '.nc'
dataset_name = 'cfsr'
return dataset_name, dataset_file
def __get_dropout_rate(self):
dropout_rates = {
'predrnn': 0.5,
'mim': 0.5
}
if self.config.model in dropout_rates:
dropout_rate = dropout_rates[self.config.model]
else:
dropout_rate = 0.
return dropout_rate