1
0
mirror of https://github.com/microsoft/qlib.git synced 2026-07-05 12:00:58 +08:00
Files
qlib/qlib/contrib/model/pytorch_gats.py
2020-11-27 10:09:17 +08:00

388 lines
12 KiB
Python
Executable File

# Copyright (c) Microsoft Corporation.
# Licensed under the MIT License.
from __future__ import division
from __future__ import print_function
import os
import numpy as np
import pandas as pd
import copy
from ...utils import create_save_path
from ...log import get_module_logger
import torch
import torch.nn as nn
import torch.optim as optim
from ...model.base import Model
from ...data.dataset import DatasetH
from ...data.dataset.handler import DataHandlerLP
class GAT(Model):
"""GAT Model
Parameters
----------
lr : float
learning rate
d_feat : int
input dimensions for each time step
metric : str
the evaluate metric used in early stop
optimizer : str
optimizer name
GPU : str
the GPU ID(s) used for training
"""
def __init__(
self,
d_feat=6,
hidden_size=64,
num_layers=2,
dropout=0.0,
n_epochs=200,
lr=0.001,
metric="",
early_stop=20,
loss="mse",
base_model="GRU",
with_pretrain=True,
optimizer="adam",
GPU="0",
seed=0,
**kwargs
):
# Set logger.
self.logger = get_module_logger("GAT")
self.logger.info("GAT pytorch version...")
# set hyper-parameters.
self.d_feat = d_feat
self.hidden_size = hidden_size
self.num_layers = num_layers
self.dropout = dropout
self.n_epochs = n_epochs
self.lr = lr
self.metric = metric
self.early_stop = early_stop
self.optimizer = optimizer.lower()
self.loss = loss
self.base_model = base_model
self.with_pretrain = with_pretrain
self.visible_GPU = GPU
self.use_gpu = torch.cuda.is_available()
self.seed = seed
self.logger.info(
"GAT parameters setting:"
"\nd_feat : {}"
"\nhidden_size : {}"
"\nnum_layers : {}"
"\ndropout : {}"
"\nn_epochs : {}"
"\nlr : {}"
"\nmetric : {}"
"\nearly_stop : {}"
"\noptimizer : {}"
"\nloss_type : {}"
"\nbase_model : {}"
"\nwith_pretrain : {}"
"\nvisible_GPU : {}"
"\nuse_GPU : {}"
"\nseed : {}".format(
d_feat,
hidden_size,
num_layers,
dropout,
n_epochs,
lr,
metric,
early_stop,
optimizer.lower(),
loss,
base_model,
with_pretrain,
GPU,
self.use_gpu,
seed,
)
)
self.GAT_model = GATModel(
d_feat=self.d_feat,
hidden_size=self.hidden_size,
num_layers=self.num_layers,
dropout=self.dropout,
base_model=self.base_model,
)
if optimizer.lower() == "adam":
self.train_optimizer = optim.Adam(self.GAT_model.parameters(), lr=self.lr)
elif optimizer.lower() == "gd":
self.train_optimizer = optim.SGD(self.GAT_model.parameters(), lr=self.lr)
else:
raise NotImplementedError("optimizer {} is not supported!".format(optimizer))
self._fitted = False
if self.use_gpu:
self.GAT_model.cuda()
# set the visible GPU
if self.visible_GPU:
os.environ["CUDA_VISIBLE_DEVICES"] = self.visible_GPU
def mse(self, pred, label):
loss = (pred - label) ** 2
return torch.mean(loss)
def loss_fn(self, pred, label):
mask = ~torch.isnan(label)
if self.loss == "mse":
return self.mse(pred[mask], label[mask])
raise ValueError("unknown loss `%s`" % self.loss)
def metric_fn(self, pred, label):
mask = torch.isfinite(label)
if self.metric == "" or self.metric == "loss": # use loss
return -self.loss_fn(pred[mask], label[mask])
raise ValueError("unknown metric `%s`" % self.metric)
def get_daily_inter(self, df, shuffle=False):
# organize the train data into daily inter as daily batches
daily_count = df.groupby(level=0).size().values
daily_index = np.roll(np.cumsum(daily_count), 1)
daily_index[0] = 0
if shuffle:
# shuffle the daily inter data
daily_shuffle = list(zip(daily_index, daily_count))
np.random.shuffle(daily_shuffle)
daily_index, daily_count = zip(*daily_shuffle)
return daily_index, daily_count
def train_epoch(self, x_train, y_train):
x_train_values = x_train.values
y_train_values = np.squeeze(y_train.values)
self.GAT_model.train()
# organize the train data into daily inter as daily batches
daily_index, daily_count = self.get_daily_inter(x_train, shuffle=True)
for idx, count in zip(daily_index, daily_count):
batch = slice(idx, idx + count)
feature = torch.from_numpy(x_train_values[batch]).float()
label = torch.from_numpy(y_train_values[batch]).float()
if self.use_gpu:
feature = feature.cuda()
label = label.cuda()
pred = self.GAT_model(feature)
loss = self.loss_fn(pred, label)
self.train_optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_value_(self.GAT_model.parameters(), 3.0)
self.train_optimizer.step()
def test_epoch(self, data_x, data_y):
# prepare training data
x_values = data_x.values
y_values = np.squeeze(data_y.values)
self.GAT_model.eval()
scores = []
losses = []
# organize the test data into daily inter as daily batches
daily_index, daily_count = self.get_daily_inter(data_x, shuffle=False)
for idx, count in zip(daily_index, daily_count):
batch = slice(idx, idx + count)
feature = torch.from_numpy(x_values[batch]).float()
label = torch.from_numpy(y_values[batch]).float()
if self.use_gpu:
feature = feature.cuda()
label = label.cuda()
pred = self.GAT_model(feature)
loss = self.loss_fn(pred, label)
losses.append(loss.item())
score = self.metric_fn(pred, label)
scores.append(score.item())
return np.mean(losses), np.mean(scores)
def fit(
self,
dataset: DatasetH,
evals_result=dict(),
verbose=True,
save_path=None,
):
df_train, df_valid, df_test = dataset.prepare(
["train", "valid", "test"], col_set=["feature", "label"], data_key=DataHandlerLP.DK_L
)
x_train, y_train = df_train["feature"], df_train["label"]
x_valid, y_valid = df_valid["feature"], df_valid["label"]
if save_path == None:
save_path = create_save_path(save_path)
stop_steps = 0
best_score = -np.inf
best_epoch = 0
evals_result["train"] = []
evals_result["valid"] = []
# load pretrained base_model
if self.with_pretrain:
self.logger.info("Loading pretrained model...")
if self.base_model == "LSTM":
from ...contrib.model.pytorch_lstm import LSTMModel
pretrained_model = LSTMModel()
pretrained_model.load_state_dict(torch.load("benchmarks/LSTM/model_lstm_csi300.pkl"))
elif self.base_model == "GRU":
from ...contrib.model.pytorch_gru import GRUModel
pretrained_model = GRUModel()
pretrained_model.load_state_dict(torch.load("benchmarks/GRU/model_gru_csi300.pkl"))
model_dict = self.GAT_model.state_dict()
pretrained_dict = {k: v for k, v in pretrained_model.state_dict().items() if k in model_dict}
model_dict.update(pretrained_dict)
self.GAT_model.load_state_dict(model_dict)
self.logger.info("Loading pretrained model Done...")
# train
self.logger.info("training...")
self._fitted = True
# return
for step in range(self.n_epochs):
self.logger.info("Epoch%d:", step)
self.logger.info("training...")
self.train_epoch(x_train, y_train)
self.logger.info("evaluating...")
train_loss, train_score = self.test_epoch(x_train, y_train)
val_loss, val_score = self.test_epoch(x_valid, y_valid)
self.logger.info("train %.6f, valid %.6f" % (train_score, val_score))
evals_result["train"].append(train_score)
evals_result["valid"].append(val_score)
if val_score > best_score:
best_score = val_score
stop_steps = 0
best_epoch = step
best_param = copy.deepcopy(self.GAT_model.state_dict())
else:
stop_steps += 1
if stop_steps >= self.early_stop:
self.logger.info("early stop")
break
self.logger.info("best score: %.6lf @ %d" % (best_score, best_epoch))
self.GAT_model.load_state_dict(best_param)
torch.save(best_param, save_path)
if self.use_gpu:
torch.cuda.empty_cache()
def predict(self, dataset):
if not self._fitted:
raise ValueError("model is not fitted yet!")
x_test = dataset.prepare("test", col_set="feature")
index = x_test.index
self.GAT_model.eval()
x_values = x_test.values
preds = []
# organize the data into daily inter as daily batches
daily_index, daily_count = self.get_daily_inter(x_test, shuffle=False)
for idx, count in zip(daily_index, daily_count):
batch = slice(idx, idx + count)
x_batch = torch.from_numpy(x_values[batch]).float()
if self.use_gpu:
x_batch = x_batch.cuda()
with torch.no_grad():
if self.use_gpu:
pred = self.GAT_model(x_batch).detach().cpu().numpy()
else:
pred = self.GAT_model(x_batch).detach().numpy()
preds.append(pred)
return pd.Series(np.concatenate(preds), index=index)
class GATModel(nn.Module):
def __init__(self, d_feat=6, hidden_size=64, num_layers=2, dropout=0.0, base_model="GRU"):
super().__init__()
if base_model == "GRU":
self.rnn = nn.GRU(
input_size=d_feat,
hidden_size=hidden_size,
num_layers=num_layers,
batch_first=True,
dropout=dropout,
)
elif base_model == "LSTM":
self.rnn = nn.LSTM(
input_size=d_feat,
hidden_size=hidden_size,
num_layers=num_layers,
batch_first=True,
dropout=dropout,
)
else:
raise ValueError("unknown base model name `%s`" % base_model)
self.hidden_size = hidden_size
self.bn1 = nn.BatchNorm1d(num_features=hidden_size, track_running_stats=False)
self.fc = nn.Linear(hidden_size, hidden_size)
self.bn2 = nn.BatchNorm1d(num_features=hidden_size, track_running_stats=False)
self.fc_out = nn.Linear(hidden_size, 1)
self.leaky_relu = nn.LeakyReLU()
self.softmax = nn.Softmax(dim=1)
self.d_feat = d_feat
def cal_convariance(self, x, y): # the 2nd dimension of x and y are the same
e_x = torch.mean(x, dim=1).reshape(-1, 1)
e_y = torch.mean(y, dim=1).reshape(-1, 1)
e_x_e_y = e_x.mm(torch.t(e_y))
x_extend = x.reshape(x.shape[0], 1, x.shape[1]).repeat(1, y.shape[0], 1)
y_extend = y.reshape(1, y.shape[0], y.shape[1]).repeat(x.shape[0], 1, 1)
e_xy = torch.mean(x_extend * y_extend, dim=2)
return e_xy - e_x_e_y
def forward(self, x):
# x: [N, F*T]
x = x.reshape(len(x), self.d_feat, -1) # [N, F, T]
x = x.permute(0, 2, 1) # [N, T, F]
out, _ = self.rnn(x)
hidden = out[:, -1, :]
hidden = self.bn1(hidden)
gamma = self.cal_convariance(hidden, hidden)
output = gamma.mm(hidden)
output = self.fc(output)
output = self.bn2(output)
output = self.leaky_relu(output)
return self.fc_out(output).squeeze()