-
Notifications
You must be signed in to change notification settings - Fork 10
/
train.py
132 lines (104 loc) · 5.4 KB
/
train.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
import numpy as np
import matplotlib.pyplot as plt
from sklearn.cross_validation import StratifiedShuffleSplit
from sklearn.metrics import accuracy_score
from tqdm import tqdm as tqdm
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader
from malenov.dataset import MalenovDataset
from malenov.options import prepare_output_directory, get_args
from malenov.model import MalenovNet
from malenov.utils import save_checkpoints
def get_valid_indices_and_labels(index_mins, index_maxs, indices, labels, cube_half_size):
valid_indices = []
valid_labels = []
for index, label in zip(indices, labels):
if [(index_mins[i]+cube_half_size <= index[i] < index_maxs[i]-cube_half_size) for i in range(1,3)] == [True]*2:
valid_indices.append(index)
valid_labels.append(label)
valid_indices = np.array(valid_indices)
valid_labels = np.array(valid_labels)
return valid_indices, valid_labels
def main():
args = get_args()
# CUDA setting
if not torch.cuda.is_available():
raise ValueError("Doesn't make much sense without a GPU. Expect long training times.")
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
device = torch.device('cuda')
torch.set_default_tensor_type('torch.cuda.FloatTensor')
torch.backends.cudnn.benchmark = False #To combat randomness
args, train_writer, val_writer, time_str = prepare_output_directory(args)
labels = np.load(args.data_root+"labels.npy")
indices = np.load(args.data_root+"indices.npy")
index_mins = indices.min(0)
index_maxs = indices.max(0)
cube_half_size = args.cube_size
if args.use_stratified_kfold:
valid_indices, valid_labels = get_valid_indices_and_labels(index_mins, index_maxs, indices, labels, cube_half_size)
splitter = StratifiedShuffleSplit(n_iter=1, random_state=args.seed, y=valid_labels, train_size=args.train_size)
train_index, test_index = [*splitter][0]
print("Labels in TRAIN:", len(train_index), "Labels in TEST:", len(test_index))
X_train, X_val = valid_indices[train_index], valid_indices[test_index]
y_train, y_val = valid_labels[train_index], valid_labels[test_index]
else:
train_indices, train_labels = np.load("./split/train_split.npy")
val_indices, val_labels = np.load("./split/val_split.npy")
X_train, y_train = get_valid_indices_and_labels(index_mins, index_maxs, train_indices, train_labels, cube_half_size)
X_val, y_val = get_valid_indices_and_labels(index_mins, index_maxs, val_indices, val_labels, cube_half_size)
seismic = torch.FloatTensor(np.load(args.data_root+"seismic_cube.npy")[:, :, :, 0]).unsqueeze(0)
y_train = torch.LongTensor(y_train)
y_val = torch.LongTensor(y_val)
train_dset = MalenovDataset(seismic, X_train, y_train, args.cube_size)
val_dset = MalenovDataset(seismic, X_val, y_val, args.cube_size)
train_loader = DataLoader(train_dset, shuffle=True, batch_size=args.batch_size, num_workers=args.num_workers)
val_loader = DataLoader(val_dset, shuffle=False, batch_size=args.batch_size, num_workers=args.num_workers)
model = MalenovNet()
model.to(device)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, betas=(0.5, 0.9))
for epoch in range(args.epochs):
model.train()
avg_acc = 0.
avg_loss = 0.
examples = 0
for i, (X, y) in enumerate(tqdm(train_loader, total=np.ceil(args.num_examples/args.batch_size))):
optimizer.zero_grad()
y_pred = model(X.to(device))
loss = criterion(y_pred, y.to(device))
loss.backward()
avg_loss += loss.detach().item()*X.size(0)
optimizer.step()
pred = torch.argmax(F.softmax(y_pred.detach(), 1), 1)
avg_acc += accuracy_score(y_true=y.cpu().numpy(), y_pred=pred.cpu().numpy())*X.size(0)
examples += X.size(0)
if examples >= args.num_examples:
break
if epoch % args.log_interval == args.log_interval-1:
print("Train: Average Loss: ", avg_loss/examples)
print("Train: Average Accuracy: ", avg_acc/examples)
train_writer.add_scalar('loss', avg_loss/examples, epoch)
train_writer.add_scalar('accuracy', avg_acc/examples, epoch)
model.eval()
avg_acc = 0.
avg_loss = 0.
with torch.set_grad_enabled(False):
for i, (X, y) in enumerate(tqdm(val_loader)):
optimizer.zero_grad()
y_pred = model(X.to(device))
loss = criterion(y_pred, y.to(device))
avg_loss += loss.detach().item()*X.size(0)
pred = torch.argmax(F.softmax(y_pred.detach(), 1), 1)
avg_acc += accuracy_score(y_true=y.cpu().numpy(), y_pred=pred.cpu().numpy())*X.size(0)
if epoch % args.log_interval == args.log_interval-1:
print("Val: Average Loss: ", avg_loss/len(val_dset))
print("Val: Average Accuracy: ", avg_acc/len(val_dset))
val_writer.add_scalar('loss', avg_loss/len(val_dset), epoch)
val_writer.add_scalar('accuracy', avg_acc/len(val_dset), epoch)
if epoch % args.checkpoint_interval == args.checkpoint_interval-1:
save_checkpoints(args, epoch, model, optimizer, time_str)
if __name__ == '__main__':
main()