-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathtrain.py
145 lines (129 loc) · 5.92 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
133
134
135
136
137
138
139
140
141
142
143
144
145
import random
import numpy as np
import os
import torch as torch
from load_data import load_relation_data, load_EOD_data
from evaluator import evaluate
from model import get_loss, RelationLSTM
np.random.seed(123456789)
torch.random.manual_seed(12345678)
device = torch.device("cuda") if torch.cuda.is_available() else 'cpu'
data_path = 'data/2013-01-01'
market_name = 'NASDAQ'
relation_name = 'wikidata' # or sector_industry
parameters = {'seq': 16, 'unit': 64, 'alpha': 0.1}
epochs = 50
valid_index = 756
test_index = 1008
fea_dim = 5
steps = 1
tickers_fname = market_name + '_tickers_qualify_dr-0.98_min-5_smooth.csv'
tickers = np.genfromtxt(os.path.join(data_path, '..', tickers_fname), dtype=str, delimiter='\t', skip_header=False)
batch_size = len(tickers)
print('#tickers selected:', len(tickers))
eod_data, mask_data, gt_data, price_data = load_EOD_data(data_path, market_name, tickers, steps)
trade_dates = mask_data.shape[1]
# relation data
rname_tail = {'sector_industry': '_industry_relation.npy', 'wikidata': '_wiki_relation.npy'}
rel_encoding, rel_mask = load_relation_data(
os.path.join(data_path, '..', 'relation', relation_name, market_name + rname_tail[relation_name])
)
print('relation encoding shape:', rel_encoding.shape)
print('relation mask shape:', rel_mask.shape)
model = RelationLSTM(
batch_size=batch_size,
rel_encoding=rel_encoding,
rel_mask=rel_mask
).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
best_valid_loss = np.inf
best_valid_perf = None
best_test_perf = None
batch_offsets = np.arange(start=0, stop=valid_index, dtype=int)
def validate(start_index, end_index):
"""
get loss on validate/test set
"""
with torch.no_grad():
cur_valid_pred = np.zeros([len(tickers), end_index - start_index], dtype=float)
cur_valid_gt = np.zeros([len(tickers), end_index - start_index], dtype=float)
cur_valid_mask = np.zeros([len(tickers), end_index - start_index], dtype=float)
loss = 0.
reg_loss = 0.
rank_loss = 0.
for cur_offset in range(start_index - parameters['seq'] - steps + 1, end_index - parameters['seq'] - steps + 1):
data_batch, mask_batch, price_batch, gt_batch = map(
lambda x: torch.Tensor(x).to(device),
get_batch(cur_offset)
)
prediction = model(data_batch)
cur_loss, cur_reg_loss, cur_rank_loss, cur_rr = get_loss(prediction, gt_batch, price_batch, mask_batch,
batch_size, parameters['alpha'])
loss += cur_loss.item()
reg_loss += cur_reg_loss.item()
rank_loss += cur_rank_loss.item()
cur_valid_pred[:, cur_offset - (start_index - parameters['seq'] - steps + 1)] = cur_rr[:, 0].cpu()
cur_valid_gt[:, cur_offset - (start_index - parameters['seq'] - steps + 1)] = gt_batch[:, 0].cpu()
cur_valid_mask[:, cur_offset - (start_index - parameters['seq'] - steps + 1)] = mask_batch[:, 0].cpu()
loss = loss / (end_index - start_index)
reg_loss = reg_loss / (end_index - start_index)
rank_loss = rank_loss / (end_index - start_index)
cur_valid_perf = evaluate(cur_valid_pred, cur_valid_gt, cur_valid_mask)
return loss, reg_loss, rank_loss, cur_valid_perf
def get_batch(offset=None):
if offset is None:
offset = random.randrange(0, valid_index)
seq_len = parameters['seq']
mask_batch = mask_data[:, offset: offset + seq_len + steps]
mask_batch = np.min(mask_batch, axis=1)
return (
eod_data[:, offset:offset + seq_len, :],
np.expand_dims(mask_batch, axis=1),
np.expand_dims(price_data[:, offset + seq_len - 1], axis=1),
np.expand_dims(gt_data[:, offset + seq_len + steps - 1], axis=1))
# train loop
for epoch in range(epochs):
np.random.shuffle(batch_offsets)
tra_loss = 0.0
tra_reg_loss = 0.0
tra_rank_loss = 0.0
# steps
for j in range(valid_index - parameters['seq'] - steps + 1):
data_batch, mask_batch, price_batch, gt_batch = map(
lambda x: torch.Tensor(x).to(device),
get_batch(batch_offsets[j])
)
optimizer.zero_grad()
prediction = model(data_batch)
cur_loss, cur_reg_loss, cur_rank_loss, _ = get_loss(prediction, gt_batch, price_batch, mask_batch,
batch_size, parameters['alpha'])
# update model
cur_loss.backward()
optimizer.step()
tra_loss += cur_loss.item()
tra_reg_loss += cur_reg_loss.item()
tra_rank_loss += cur_rank_loss.item()
# train loss
# loss = reg_loss(mse) + alpha*rank_loss
tra_loss = tra_loss / (valid_index - parameters['seq'] - steps + 1)
tra_reg_loss = tra_reg_loss / (valid_index - parameters['seq'] - steps + 1)
tra_rank_loss = tra_rank_loss / (valid_index - parameters['seq'] - steps + 1)
print('\n\nTrain : loss:{} reg_loss:{} rank_loss:{}'.format(tra_loss, tra_reg_loss, tra_rank_loss))
# show performance on valid set
val_loss, val_reg_loss, val_rank_loss, val_perf = validate(valid_index, test_index)
print('Valid : loss:{} reg_loss:{} rank_loss:{}'.format(val_loss, val_reg_loss, val_rank_loss))
print('\t Valid performance:', val_perf)
# show performance on valid set
test_loss, test_reg_loss, test_rank_loss, test_perf = validate(test_index, trade_dates)
print('Test: loss:{} reg_loss:{} rank_loss:{}'.format(test_loss, test_reg_loss, test_rank_loss))
print('\t Test performance:', test_perf)
# best result
if val_loss < best_valid_loss:
best_valid_loss = val_loss
# In this place, remove some var that wouldn't be printed
# without copy.copy()
best_valid_perf = val_perf
best_test_perf = test_perf
print('Better valid loss:', best_valid_loss)
print('\nBest Valid performance:', best_valid_perf)
print('Best Test performance:', best_test_perf)