liuyu/qlib
1
0
Fork 0
mirror of https://github.com/microsoft/qlib.git synced 2026-07-12 23:36:54 +08:00
Code Activity

Migrate NeuTrader to Qlib RL (#1169)

Browse Source 
* Refine previous version RL codes

* Polish utils/__init__.py

* Draft

* Use | instead of Union

* Simulator & action interpreter

* Test passed

* Migrate to SAOEState & new qlib interpreter

* Black format

* . Revert file_storage change

* Refactor file structure & renaming functions

* Enrich test cases

* Add QlibIntradayBacktestData

* Test interpreter

* Black format

* .

.

.

* Rename receive_execute_result()

* Use indicator to simplify state update

* Format code

* Modify data path

* Adjust file structure

* Minor change

* Add copyright message

* Format code

* Rename util functions

* Add CI

* Pylint issue

* Remove useless code to pass pylint

* Pass mypy

* Mypy issue

* mypy issue

* mypy issue

* Revert "mypy issue"

This reverts commit 8eb1b0174e.

* mypy issue

* mypy issue

* Fix the numpy version incompatible bug

* Fix a minor typing issue

* Try to skip python 3.7 test for qlib simulator

* Resolve PR comments by Yuge; solve several CI issues.

* Black issue

* Fix a low-level type error

* Change data name

* Resolve PR comments. Leave TODOs in the code base.

Co-authored-by: Young <afe.young@gmail.com>
This commit is contained in:
Huoran Li
2022-08-01 09:56:07 +08:00
committed by GitHub
parent 687edd79d0
commit 2752bdc92c
35 changed files with 1305 additions and 257 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
qlib/rl/order_execution/network.py
View File

@@ -3,13 +3,14 @@
from __future__ import annotations
from typing import cast
from typing import List, Tuple, cast
import torch
import torch.nn as nn
from tianshou.data import Batch
from qlib.typehint import Literal
from .interpreter import FullHistoryObs
__all__ = ["Recurrent"]
@@ -18,7 +19,7 @@ __all__ = ["Recurrent"]
class Recurrent(nn.Module):
"""The network architecture proposed in `OPD <https://seqml.github.io/opd/opd_aaai21_supplement.pdf>`_.
At every timestep the input of policy network is divided into two parts,
At every time step the input of policy network is divided into two parts,
the public variables and the private variables. which are handled by ``raw_rnn``
and ``pri_rnn`` in this network, respectively.
@@ -33,7 +34,7 @@ class Recurrent(nn.Module):
output_dim: int = 32,
rnn_type: Literal["rnn", "lstm", "gru"] = "gru",
rnn_num_layers: int = 1,
):
) -> None:
super().__init__()
self.hidden_dim = hidden_dim
@@ -62,10 +63,10 @@ class Recurrent(nn.Module):
nn.ReLU(),
)
def _init_extra_branches(self):
def _init_extra_branches(self) -> None:
pass
def _source_features(self, obs: FullHistoryObs, device: torch.device) -> tuple[list[torch.Tensor], torch.Tensor]:
def _source_features(self, obs: FullHistoryObs, device: torch.device) -> Tuple[List[torch.Tensor], torch.Tensor]:
bs, _, data_dim = obs["data_processed"].size()
data = torch.cat((torch.zeros(bs, 1, data_dim, device=device), obs["data_processed"]), 1)
cur_step = obs["cur_step"].long()