-
Notifications
You must be signed in to change notification settings - Fork 0
/
nn.py
182 lines (151 loc) · 7.3 KB
/
nn.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
import numpy as np
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.model_selection import LeaveOneGroupOut
import torch
from torch.utils.data import TensorDataset, DataLoader
import torch.nn as nn
import torch.optim as optim
from matplotlib import pyplot as plt
from config import columns_config, device
from trainer import Trainer, CV
class Net(nn.Module):
def __init__(self, d_in=10, k=2, n_hidden=1, batch_norm=False, dropout=False):
super(Net, self).__init__()
d_cur = d_in
self.layers = []
for i in range(n_hidden):
self.layers.append(nn.Linear(d_cur, d_cur // k))
if batch_norm:
self.layers.append(nn.BatchNorm1d(d_cur // k))
self.layers.append(nn.ReLU())
if dropout:
self.layers.append(nn.Dropout())
d_cur //= k
self.layers.append(nn.Linear(d_cur, 1))
self.layers = nn.ModuleList(self.layers)
def forward(self, x):
for layer in self.layers:
x = layer(x)
return x
class Scaler(CV):
def __init__(self, preprocessor, batch_size=1024):
super(Scaler, self).__init__()
self.df = preprocessor.df
self.train_idx = preprocessor.train_idx
self.batch_size = batch_size
self.scaler_labels = StandardScaler()
self.scaler_features = StandardScaler()
self.encoder = OneHotEncoder(handle_unknown='ignore', sparse=True)
self.create_scalers()
self.d_in = len(columns_config['numerical']) + sum(len(c) for c in self.encoder.categories_)
def create_scalers(self):
self.scaler_features.fit(self.df[columns_config['numerical']])
self.scaler_labels.fit(self.df.loc[self.train_idx, 'meter_reading'].values.reshape(-1, 1).astype(np.float))
self.encoder.fit(self.df.loc[self.train_idx, columns_config['categorical']])
def transform(self, data):
num_features = self.scaler_features.transform(data[columns_config['numerical']])
labels = self.scaler_labels.transform(data.loc[:, 'meter_reading'].values.reshape(-1, 1).astype(np.float))
cat_features = self.encoder.transform(data[columns_config['categorical']])
print('NaNs in scaled arrays:', np.isnan(labels).sum(), np.isnan(num_features).sum(),
np.isnan(cat_features.todense()).sum())
return cat_features, num_features, labels
@staticmethod
def create_tensors(cat, num, labels):
return torch.Tensor(np.concatenate([cat.todense(), num], 1)), torch.Tensor(labels)
def create_dataloader(self, X, y, shuffle=True):
dataset = TensorDataset(X, y)
dataloader = DataLoader(dataset, batch_size=self.batch_size, shuffle=shuffle)
return dataloader
def iter_cv(self):
cat, num, labels = self.transform(self.df.loc[self.train_idx])
X, y = self.create_tensors(cat, num, labels)
logo = LeaveOneGroupOut()
for train_idx, test_idx in logo.split(X, y, self.df.loc[self.train_idx, 'cv_group']):
yield self.create_dataloader(X[train_idx], y[train_idx]), self.create_dataloader(X[test_idx], y[test_idx])
class NetTrainer(Trainer):
def __init__(self, trainloader, testloader, scaler=None, net_config=None, lr=0):
super(NetTrainer, self).__init__()
self.scaler = scaler
self.trainloader = trainloader
self.testloader = testloader
self.optimizer = None
self.criterion = None
self.net = None
net_config['d_in'] = scaler.d_in
self.create_models(net_config, lr)
self.train_losses = []
self.test_losses = []
self.metrics = []
def create_models(self, net_config, lr):
self.net = Net(**net_config).to(device)
print('Net architecture:')
print(self.net)
self.criterion = nn.MSELoss()
self.optimizer = optim.Adam(self.net.parameters(), lr=lr)
def metric(self, pred, labels):
pred_raw = self.scaler.scaler_labels.inverse_transform(pred.detach().cpu().numpy())
labels_raw = self.scaler.scaler_labels.inverse_transform(labels.detach().cpu().numpy())
loss = np.mean((pred_raw - labels_raw) ** 2) ** 0.5
return loss
def train(self, n_epochs=0, verbose=True, do_val=True):
for epoch in range(n_epochs):
self.net.train()
losses = []
for i, data in enumerate(self.trainloader, 0):
inputs, labels = data
if inputs.size(0) <= 1:
continue
inputs = inputs.to(device)
labels = labels.to(device)
self.optimizer.zero_grad()
outputs = self.net(inputs)
loss = self.criterion(outputs, labels)
loss.backward()
self.optimizer.step()
losses.append(loss.item())
if verbose:
print('[%d] Train loss: %.3f' % (epoch + 1, np.mean(losses)))
self.train_losses.append(np.mean(losses))
if do_val:
self.net.eval()
losses = []
metrics = []
for i, data in enumerate(self.testloader, 0):
inputs, labels = data
inputs = inputs.to(device)
labels = labels.to(device)
with torch.no_grad():
outputs = self.net(inputs)
loss = self.criterion(outputs, labels)
losses.append(loss.item())
metrics.append(self.metric(outputs, labels))
if verbose:
print('[%d] Test loss: %.3f' % (epoch + 1, np.mean(losses)))
print('[%d] Test metric: %.3f' % (epoch + 1, np.mean(metrics)))
self.test_losses.append(np.mean(losses))
self.metrics.append(np.mean(metrics))
def predict(self, test_df, submission, batch_size=100000):
self.net.eval()
for i in range(0, test_df.shape[0], batch_size):
cat_test, num_test, labels_test = self.scaler.transform(test_df[i: min(i + batch_size, test_df.shape[0])])
inputs, _ = self.scaler.create_tensors(cat_test, num_test, labels_test)
inputs = inputs.to(device)
row_ids = test_df.row_id[i: min(i + batch_size, test_df.shape[0])]
with torch.no_grad():
outputs = self.net(inputs)
pred_raw = self.scaler.scaler_labels.inverse_transform(outputs.detach().cpu().numpy())
pred_raw = np.exp(pred_raw) - 1
submission = np.concatenate([submission,
np.concatenate([row_ids.values.reshape(-1, 1), pred_raw], axis=1)
], axis=0)
return submission
def plot(self, pic_name):
f, (ax1, ax2) = plt.subplots(2, 1, figsize=(20, 10))
ax1.plot(self.train_losses, color='b')
ax1.plot(self.test_losses, color='y')
ax2.plot(self.metrics, color='y')
f.savefig('plots/%s.png' % pic_name)
def save_model(self, path):
torch.save(self.net, path)
def load_model(self, path):
self.net = torch.load(path)