Merge pull request #227 from Arthur-Null/high-freq-execution

High freq execution

examples/trade/README.md

@@ -0,0 +1,16 @@
+# Universal Trading for Order Execution with Oracle Policy Distillation
+This is the experiment code for our AAAI 2021 paper "[Universal Trading for Order Execution with Oracle Policy Distillation](https://seqml.github.io/opd/opd_aaai21.pdf)", including the implementations of all the compared methods in the paper and a general reinforcement learning framework for order execution in quantitative finance. 
+
+## Abstract
+As a fundamental problem in algorithmic trading, order execution aims at fulfilling a specific trading order, either liquidation or acquirement, for a given instrument. Towards effective execution strategy, recent years have witnessed the shift from the analytical view with model-based market assumptions to model-free perspective, i.e., reinforcement learning, due to its nature of sequential decision optimization. However, the noisy and yet imperfect market information that can be leveraged by the policy has made it quite challenging to build up sample efficient reinforcement learning methods to achieve effective order execution. In this paper, we propose a novel universal trading policy optimization framework to bridge the gap between the noisy yet imperfect market states and the optimal action sequences for order execution. Particularly, this framework leverages a policy distillation method that can better guide the learning of the common policy towards practically optimal execution by an oracle teacher with perfect information to approximate the optimal trading strategy. The extensive experiments have shown significant improvements of our method over various strong baselines, with reasonable trading actions.
+
+### Citation
+You are more than welcome to cite our paper:
+```
+@inproceedings{fang2021universal,
+  title={Universal Trading for Order Execution with Oracle Policy Distillation},
+  author={Fang, Yuchen and Ren, Kan and Liu, Weiqing and Zhou, Dong and Zhang, Weinan and Bian, Jiang and Yu, Yong and Liu, Tie-Yan},
+  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
+  year={2021}
+}
+```

examples/trade/agent/basic.py

@@ -1,5 +1,3 @@
-from joblib import Parallel, delayed
-from numba import njit, prange
 from tianshou.policy import BasePolicy
 from tianshou.data import Batch
 import numpy as np

examples/trade/executor.py

@@ -5,7 +5,7 @@ import logger
 import json
 import os
 import agent
-import model
+import network
 import policy
 import random
 import tianshou as ts
@@ -48,7 +48,15 @@ def setup_seed(seed):
 
 class BaseExecutor(object):
     def __init__(
-        self, log_dir, resources, env_conf, optim=None, policy_conf=None, network=None, policy_path=None, seed=None,
+        self,
+        log_dir,
+        resources,
+        env_conf,
+        optim=None,
+        policy_conf=None,
+        network_conf=None,
+        policy_path=None,
+        seed=None,
     ):
         """A base class for executor
 
@@ -62,8 +70,8 @@ class BaseExecutor(object):
         :type optim: dict, optional
         :param policy_conf: Configurations for the RL algorithm, defaults to None
         :type policy_conf: dict, optional
-        :param network: Configurations for policy network, defaults to None
-        :type network: dict, optional
+        :param network_conf: Configurations for policy network_conf, defaults to None
+        :type network_conf: dict, optional
         :param policy_path: If is not None, would load the policy from this path, defaults to None
         :type policy_path: string, optional
         :param seed: Random seed, defaults to None
@@ -90,17 +98,23 @@ class BaseExecutor(object):
             self.policy = getattr(agent, policy_conf["name"])(policy_conf["config"])
             # print(self.policy)
         else:
-            assert not network is None
-            if "extractor" in network.keys():
-                net = getattr(model, network["extractor"]["name"] + "_Extractor")(
-                    device=self.device, **network["config"]
+            assert not network_conf is None
+            if "extractor" in network_conf.keys():
+                net = getattr(network, network_conf["extractor"]["name"] + "_Extractor")(
+                    device=self.device, **network_conf["config"]
                 )
             else:
-                net = getattr(model, network["name"] + "_Extractor")(device=self.device, **network["config"])
+                net = getattr(network, network_conf["name"] + "_Extractor")(
+                    device=self.device, **network_conf["config"]
+                )
             net.to(self.device)
-            actor = getattr(model, network["name"] + "_Actor")(extractor=net, device=self.device, **network["config"])
+            actor = getattr(network, network_conf["name"] + "_Actor")(
+                extractor=net, device=self.device, **network_conf["config"]
+            )
             actor.to(self.device)
-            critic = getattr(model, network["name"] + "_Critic")(extractor=net, device=self.device, **network["config"])
+            critic = getattr(network, network_conf["name"] + "_Critic")(
+                extractor=net, device=self.device, **network_conf["config"]
+            )
             critic.to(self.device)
             self.optim = torch.optim.Adam(
                 list(actor.parameters()) + list(critic.parameters()),
@@ -180,7 +194,7 @@ class Executor(BaseExecutor):
         io_conf,
         optim=None,
         policy_conf=None,
-        network=None,
+        network_conf=None,
         policy_path=None,
         seed=None,
         share_memory=False,
@@ -210,7 +224,7 @@ class Executor(BaseExecutor):
         :param buffer_size: The size of replay buffer, defaults to 200000
         :type buffer_size: int, optional
         """
-        super().__init__(log_dir, resources, env_conf, optim, policy_conf, network, policy_path, seed)
+        super().__init__(log_dir, resources, env_conf, optim, policy_conf, network_conf, policy_path, seed)
         single_env = getattr(env, env_conf["name"])
         env_conf = merge_dicts(env_conf, train_paths)
         env_conf["log"] = True

examples/trade/network/__init__.py

@@ -0,0 +1,5 @@
+from .ppo import *
+from .qmodel import *
+from .teacher import *
+from .util import *
+from .opd import *

examples/trade/network/opd.py

@@ -0,0 +1,74 @@
+import torch
+import numpy as np
+from torch import nn
+import torch.nn.functional as F
+from copy import deepcopy
+import sys
+
+from tianshou.data import to_torch
+
+
+class OPD_Extractor(nn.Module):
+    def __init__(self, device="cpu", **kargs):
+        super().__init__()
+        self.device = device
+        hidden_size = kargs["hidden_size"]
+        fc_size = kargs["fc_size"]
+        self.cnn_shape = kargs["cnn_shape"]
+
+        self.rnn = nn.GRU(64, hidden_size, batch_first=True)
+        self.rnn2 = nn.GRU(64, hidden_size, batch_first=True)
+        self.dnn = nn.Sequential(nn.Linear(2, 64), nn.ReLU(),)
+        self.cnn = nn.Sequential(nn.Conv1d(self.cnn_shape[1], 3, 3), nn.ReLU(),)
+        self.raw_fc = nn.Sequential(nn.Linear((self.cnn_shape[0] - 2) * 3, 64), nn.ReLU(),)
+
+        self.fc = nn.Sequential(
+            nn.Linear(hidden_size * 2, hidden_size), nn.ReLU(), nn.Linear(hidden_size, 32), nn.ReLU(),
+        )
+
+    def forward(self, inp):
+        inp = to_torch(inp, dtype=torch.float32, device=self.device)
+        teacher_action = inp[:, 0]
+        inp = inp[:, 1:]
+        seq_len = inp[:, -1].to(torch.long)
+        batch_size = inp.shape[0]
+        raw_in = inp[:, : 6 * 240]
+        raw_in = torch.cat((torch.zeros_like(inp[:, : 6 * 30]), raw_in), dim=-1)
+        raw_in = raw_in.reshape(-1, 30, 6).transpose(1, 2)
+        dnn_in = inp[:, 6 * 240 : -1].reshape(batch_size, -1, 2)
+        cnn_out = self.cnn(raw_in).view(batch_size, 9, -1)
+        rnn_in = self.raw_fc(cnn_out)
+        rnn2_in = self.dnn(dnn_in)
+        rnn2_out = self.rnn2(rnn2_in)[0]
+        rnn_out = self.rnn(rnn_in)[0]
+        rnn_out = rnn_out[torch.arange(rnn_out.size(0)), seq_len]
+        rnn2_out = rnn2_out[torch.arange(rnn2_out.size(0)), seq_len]
+        # dnn_out = self.dnn(dnn_in)
+        fc_in = torch.cat((rnn_out, rnn2_out), dim=-1)
+        feature = self.fc(fc_in)
+        return feature, teacher_action / 2
+
+
+class OPD_Actor(nn.Module):
+    def __init__(self, extractor, out_shape, device=torch.device("cpu"), **kargs):
+        super().__init__()
+        self.extractor = extractor
+        self.layer_out = nn.Sequential(nn.Linear(32, out_shape), nn.Softmax(dim=-1))
+        self.device = device
+
+    def forward(self, obs, state=None, info={}):
+        feature, self.teacher_action = self.extractor(obs)
+        out = self.layer_out(feature)
+        return out, state
+
+
+class OPD_Critic(nn.Module):
+    def __init__(self, extractor, out_shape, device=torch.device("cpu"), **kargs):
+        super().__init__()
+        self.extractor = extractor
+        self.value_out = nn.Linear(32, 1)
+        self.device = device
+
+    def forward(self, obs, state=None, info={}):
+        feature, self.teacher_action = self.extractor(obs)
+        return self.value_out(feature).squeeze(dim=-1)

examples/trade/network/ppo.py

@@ -0,0 +1,79 @@
+import torch
+import numpy as np
+from torch import nn
+import torch.nn.functional as F
+from copy import deepcopy
+import sys
+
+from tianshou.data import to_torch
+
+
+class PPO_Extractor(nn.Module):
+    def __init__(self, device="cpu", **kargs):
+        super().__init__()
+        self.device = device
+        hidden_size = kargs["hidden_size"]
+        fc_size = kargs["fc_size"]
+        self.cnn_shape = kargs["cnn_shape"]
+
+        self.rnn = nn.GRU(64, hidden_size, batch_first=True)
+        self.rnn2 = nn.GRU(64, hidden_size, batch_first=True)
+        self.dnn = nn.Sequential(nn.Linear(2, 64), nn.ReLU(),)
+        self.cnn = nn.Sequential(nn.Conv1d(self.cnn_shape[1], 3, 3), nn.ReLU(),)
+        self.raw_fc = nn.Sequential(nn.Linear((self.cnn_shape[0] - 2) * 3, 64), nn.ReLU(),)
+
+        self.fc = nn.Sequential(
+            nn.Linear(hidden_size * 2, hidden_size), nn.ReLU(), nn.Linear(hidden_size, 32), nn.ReLU(),
+        )
+
+    def forward(self, inp):
+        inp = to_torch(inp, dtype=torch.float32, device=self.device)
+        # inp = torch.from_numpy(inp).to(torch.device('cpu'))
+        seq_len = inp[:, -1].to(torch.long)
+        batch_size = inp.shape[0]
+        raw_in = inp[:, : 6 * 240]
+        raw_in = torch.cat((torch.zeros_like(inp[:, : 6 * 30]), raw_in), dim=-1)
+        raw_in = raw_in.reshape(-1, 30, 6).transpose(1, 2)
+        dnn_in = inp[:, -19:-1].reshape(batch_size, -1, 2)
+        cnn_out = self.cnn(raw_in).view(batch_size, 9, -1)
+        assert not torch.isnan(cnn_out).any()
+        rnn_in = self.raw_fc(cnn_out)
+        assert not torch.isnan(rnn_in).any()
+        rnn2_in = self.dnn(dnn_in)
+        assert not torch.isnan(rnn2_in).any()
+        rnn2_out = self.rnn2(rnn2_in)[0]
+        assert not torch.isnan(rnn2_out).any()
+        rnn_out = self.rnn(rnn_in)[0]
+        assert not torch.isnan(rnn_out).any()
+        rnn_out = rnn_out[torch.arange(rnn_out.size(0)), seq_len]
+        rnn2_out = rnn2_out[torch.arange(rnn2_out.size(0)), seq_len]
+        # dnn_out = self.dnn(dnn_in)
+        fc_in = torch.cat((rnn_out, rnn2_out), dim=-1)
+        self.feature = self.fc(fc_in)
+        return self.feature
+
+
+class PPO_Actor(nn.Module):
+    def __init__(self, extractor, out_shape, device=torch.device("cpu"), **kargs):
+        super().__init__()
+        self.extractor = extractor
+        self.layer_out = nn.Sequential(nn.Linear(32, out_shape), nn.Softmax(dim=-1))
+        self.device = device
+
+    def forward(self, obs, state=None, info={}):
+        self.feature = self.extractor(obs)
+        assert not (torch.isnan(self.feature).any() | torch.isinf(self.feature).any()), f"{self.feature}"
+        out = self.layer_out(self.feature)
+        return out, state
+
+
+class PPO_Critic(nn.Module):
+    def __init__(self, extractor, out_shape, device=torch.device("cpu"), **kargs):
+        super().__init__()
+        self.extractor = extractor
+        self.value_out = nn.Linear(32, 1)
+        self.device = device
+
+    def forward(self, obs, state=None, info={}):
+        self.feature = self.extractor(obs)
+        return self.value_out(self.feature).squeeze(dim=-1)

examples/trade/network/qmodel.py

@@ -0,0 +1,52 @@
+import torch
+import numpy as np
+from torch import nn
+import torch.nn.functional as F
+from copy import deepcopy
+import sys
+
+from tianshou.data import to_torch
+
+
+class RNNQModel(nn.Module):
+    def __init__(self, device="cpu", out_shape=10, **kargs):
+        super().__init__()
+        self.device = device
+        hidden_size = kargs["hidden_size"]
+        fc_size = kargs["fc_size"]
+        self.cnn_shape = kargs["cnn_shape"]
+
+        self.rnn = nn.GRU(64, hidden_size, batch_first=True)
+        self.rnn2 = nn.GRU(64, hidden_size, batch_first=True)
+        self.dnn = nn.Sequential(nn.Linear(2, 64), nn.ReLU(),)
+        self.cnn = nn.Sequential(nn.Conv1d(self.cnn_shape[1], 3, 3), nn.ReLU(),)
+        self.raw_fc = nn.Sequential(nn.Linear((self.cnn_shape[0] - 2) * 3, 64), nn.ReLU(),)
+
+        self.fc = nn.Sequential(
+            nn.Linear(hidden_size * 2, hidden_size),
+            nn.ReLU(),
+            nn.Linear(hidden_size, 32),
+            nn.ReLU(),
+            nn.Linear(32, out_shape),
+        )
+
+    def forward(self, obs, state=None, info={}):
+        inp = to_torch(obs, dtype=torch.float32, device=self.device)
+        inp = inp[:, 182:]
+        seq_len = inp[:, -1].to(torch.long)
+        batch_size = inp.shape[0]
+        raw_in = inp[:, : 6 * 240]
+        raw_in = torch.cat((torch.zeros_like(inp[:, : 6 * 30]), raw_in), dim=-1)
+        raw_in = raw_in.reshape(-1, 30, 6).transpose(1, 2)
+        dnn_in = inp[:, 6 * 240 : -1].reshape(batch_size, -1, 2)
+        cnn_out = self.cnn(raw_in).view(batch_size, 9, -1)
+        rnn_in = self.raw_fc(cnn_out)
+        rnn2_in = self.dnn(dnn_in)
+        rnn2_out = self.rnn2(rnn2_in)[0]
+        rnn_out = self.rnn(rnn_in)[0]
+        rnn_out = rnn_out[torch.arange(rnn_out.size(0)), seq_len]
+        rnn2_out = rnn2_out[torch.arange(rnn2_out.size(0)), seq_len]
+        # dnn_out = self.dnn(dnn_in)
+        fc_in = torch.cat((rnn_out, rnn2_out), dim=-1)
+        out = self.fc(fc_in)
+        return out, state

examples/trade/network/teacher.py

@@ -0,0 +1,70 @@
+import torch
+import numpy as np
+from torch import nn
+import torch.nn.functional as F
+from copy import deepcopy
+import sys
+
+from tianshou.data import to_torch
+
+
+class Teacher_Extractor(nn.Module):
+    def __init__(self, device="cpu", feature_size=180, **kargs):
+        super().__init__()
+        self.device = device
+        hidden_size = kargs["hidden_size"]
+        fc_size = kargs["fc_size"]
+        self.cnn_shape = kargs["cnn_shape"]
+
+        self.rnn = nn.GRU(64, hidden_size, batch_first=True)
+        self.rnn2 = nn.GRU(64, hidden_size, batch_first=True)
+        self.dnn = nn.Sequential(nn.Linear(2, 64), nn.ReLU(),)
+        self.cnn = nn.Sequential(nn.Conv1d(self.cnn_shape[1], 3, 3), nn.ReLU(),)
+        self.raw_fc = nn.Sequential(nn.Linear((self.cnn_shape[0] - 2) * 3, 64), nn.ReLU(),)
+
+        self.fc = nn.Sequential(
+            nn.Linear(hidden_size * 2, hidden_size), nn.ReLU(), nn.Linear(hidden_size, 32), nn.ReLU(),
+        )
+
+    def forward(self, inp):
+        inp = to_torch(inp, dtype=torch.float32, device=self.device)
+        inp = inp[:, 182:]
+        seq_len = inp[:, -1].to(torch.long)
+        batch_size = inp.shape[0]
+        raw_in = inp[:, : 6 * 240].reshape(-1, 30, 6).transpose(1, 2)
+        dnn_in = inp[:, 6 * 240 : -1].reshape(batch_size, -1, 2)
+        cnn_out = self.cnn(raw_in).view(batch_size, 8, -1)
+        rnn_in = self.raw_fc(cnn_out)
+        rnn2_in = self.dnn(dnn_in)
+        rnn2_out = self.rnn2(rnn2_in)[0]
+        rnn_out = self.rnn(rnn_in)[0][:, -1, :]
+        rnn2_out = rnn2_out[torch.arange(rnn2_out.size(0)), seq_len]
+        # dnn_out = self.dnn(dnn_in)
+        fc_in = torch.cat((rnn_out, rnn2_out), dim=-1)
+        self.feature = self.fc(fc_in)
+        return self.feature
+
+
+class Teacher_Actor(nn.Module):
+    def __init__(self, extractor, out_shape, device=torch.device("cpu"), **kargs):
+        super().__init__()
+        self.extractor = extractor
+        self.layer_out = nn.Sequential(nn.Linear(32, out_shape), nn.Softmax(dim=-1))
+        self.device = device
+
+    def forward(self, obs, state=None, info={}):
+        self.feature = self.extractor(obs)
+        out = self.layer_out(self.feature)
+        return out, state
+
+
+class Teacher_Critic(nn.Module):
+    def __init__(self, extractor, out_shape, device=torch.device("cpu"), **kargs):
+        super().__init__()
+        self.extractor = extractor
+        self.value_out = nn.Linear(32, 1)
+        self.device = device
+
+    def forward(self, obs, state=None, info={}):
+        self.feature = self.extractor(obs)
+        return self.value_out(self.feature).squeeze(-1)

examples/trade/network/util.py

@@ -0,0 +1,191 @@
+import torch
+import numpy as np
+from torch import nn
+import torch.nn.functional as F
+from copy import deepcopy
+import sys
+
+from tianshou.data import to_torch
+
+
+class Attention(nn.Module):
+    def __init__(self, in_dim, out_dim):
+        super().__init__()
+        self.get_w = nn.Sequential(nn.Linear(in_dim * 2, in_dim), nn.ReLU(), nn.Linear(in_dim, 1))
+
+        self.fc = nn.Sequential(nn.Linear(in_dim, out_dim), nn.ReLU(),)
+
+    def forward(self, value, key):
+        key = key.unsqueeze(dim=1)
+        length = value.shape[1]
+        key = key.repeat([1, length, 1])
+        weight = self.get_w(torch.cat((key, value), dim=-1)).squeeze()  # B * l
+        weight = weight.softmax(dim=-1).unsqueeze(dim=-1)  # B * l * 1
+        out = (value * weight).sum(dim=1)
+        out = self.fc(out)
+        return out
+
+
+class MaskAttention(nn.Module):
+    def __init__(self, in_dim, out_dim):
+        super().__init__()
+        self.get_w = nn.Sequential(nn.Linear(in_dim * 2, in_dim), nn.ReLU(), nn.Linear(in_dim, 1))
+
+        self.fc = nn.Sequential(nn.Linear(in_dim, out_dim), nn.ReLU(),)
+
+    def forward(self, value, key, seq_len, maxlen=9):
+        # seq_len: (batch,)
+        device = value.device
+        key = key.unsqueeze(dim=1)
+        length = value.shape[1]
+        key = key.repeat([1, length, 1])  # (batch, 9, 64)
+        weight = self.get_w(torch.cat((key, value), dim=-1)).squeeze(-1)  # (batch, 9)
+        mask = sequence_mask(seq_len + 1, maxlen=maxlen, device=device)
+        weight[~mask] = float("-inf")
+        weight = weight.softmax(dim=-1).unsqueeze(dim=-1)
+        out = (value * weight).sum(dim=1)
+        out = self.fc(out)
+        return out
+
+
+class TFMaskAttention(nn.Module):
+    def __init__(self, in_dim, out_dim):
+        super().__init__()
+        self.get_w = nn.Sequential(nn.Linear(in_dim * 2, in_dim), nn.ReLU(), nn.Linear(in_dim, 1))
+
+        self.fc = nn.Sequential(nn.Linear(in_dim, out_dim), nn.ReLU(),)
+
+    def forward(self, value, key, seq_len, maxlen=9):
+        device = value.device
+        key = key.unsqueeze(dim=1)
+        length = value.shape[1]
+        key = key.repeat([1, length, 1])
+        weight = self.get_w(torch.cat((key, value), dim=-1)).squeeze(-1)
+        mask = sequence_mask(seq_len + 1, maxlen=maxlen, device=device)
+        mask = mask.repeat(1, 3)  # (batch, 9*3)
+        weight[~mask] = float("-inf")
+        weight = weight.softmax(dim=-1).unsqueeze(dim=-1)
+        out = (value * weight).sum(dim=1)
+        out = self.fc(out)
+        return out
+
+
+class NNAttention(nn.Module):
+    def __init__(self, in_dim, out_dim):
+        super().__init__()
+        self.q_net = nn.Linear(in_dim, out_dim)
+        self.k_net = nn.Linear(in_dim, out_dim)
+        self.v_net = nn.Linear(in_dim, out_dim)
+
+    def forward(self, Q, K, V):
+        q = self.q_net(Q)
+        k = self.k_net(K)
+        v = self.v_net(V)
+
+        attn = torch.einsum("ijk,ilk->ijl", q, k)
+        attn = attn.to(Q.device)
+        attn_prob = torch.softmax(attn, dim=-1)
+
+        attn_vec = torch.einsum("ijk,ikl->ijl", attn_prob, v)
+
+        return attn_vec
+
+
+class Reshape(nn.Module):
+    def __init__(self, *args):
+        super(Reshape, self).__init__()
+        self.shape = args
+
+    def forward(self, x):
+        return x.view(self.shape)
+
+
+class DARNN(nn.Module):
+    def __init__(self, device="cpu", **kargs):
+        super().__init__()
+        self.emb_dim = kargs["emb_dim"]
+        self.hidden_size = kargs["hidden_size"]
+        self.num_layers = kargs["num_layers"]
+        self.is_bidir = kargs["is_bidir"]
+        self.dropout = kargs["dropout"]
+        self.seq_len = kargs["seq_len"]
+        self.interval = kargs["interval"]
+        self.today_length = 238
+        self.prev_length = 240
+        self.input_length = 480
+        self.input_size = 6
+
+        self.rnn = nn.LSTM(
+            input_size=self.input_size + self.emb_dim,
+            hidden_size=self.hidden_size,
+            num_layers=self.num_layers,
+            batch_first=True,
+            bidirectional=self.is_bidir,
+            dropout=self.dropout,
+        )
+        self.prev_rnn = nn.LSTM(
+            input_size=self.input_size,
+            hidden_size=self.hidden_size,
+            num_layers=self.num_layers,
+            batch_first=True,
+            bidirectional=self.is_bidir,
+            dropout=self.dropout,
+        )
+        self.fc_out = nn.Linear(in_features=self.hidden_size * 2, out_features=1)
+        self.attention = NNAttention(self.hidden_size, self.hidden_size)
+        self.act_out = nn.Sigmoid()
+        if self.emb_dim != 0:
+            self.pos_emb = nn.Embedding(self.input_length, self.emb_dim)
+
+    def forward(self, inputs):
+        inputs = inputs.view(-1, self.input_length, self.input_size)  # [B, T, F]
+        today_input = inputs[:, : self.today_length, :]
+        today_input = torch.cat((torch.zeros_like(today_input[:, :1, :]), today_input), dim=1)
+        prev_input = inputs[:, 240 : 240 + self.prev_length, :]
+        if self.emb_dim != 0:
+            embedding = self.pos_emb(torch.arange(end=self.today_length + 1, device=inputs.device))
+            embedding = embedding.repeat([today_input.size()[0], 1, 1])
+            today_input = torch.cat((today_input, embedding), dim=-1)
+        prev_outs, _ = self.prev_rnn(prev_input)
+        today_outs, _ = self.rnn(today_input)
+
+        outs = self.attention(today_outs, prev_outs, prev_outs)
+        outs = torch.cat((today_outs, outs), dim=-1)
+        outs = outs[:, range(0, self.seq_len * self.interval, self.interval), :]
+        # outs = self.fc_out(outs).squeeze()
+        return self.act_out(self.fc_out(outs).squeeze(-1)), outs
+
+
+class Transpose(nn.Module):
+    def __init__(self, dim1=0, dim2=1):
+        super().__init__()
+        self.dim1 = dim1
+        self.dim2 = dim2
+
+    def forward(self, x):
+        return x.transpose(self.dim1, self.dim2)
+
+
+class SelfAttention(nn.Module):
+    def __init__(self, *args, **kargs):
+        super().__init__()
+        self.attention = nn.MultiheadAttention(*args, **kargs)
+
+    def forward(self, x):
+        return self.attention(x, x, x)[0]
+
+
+def onehot_enc(y, len):
+    y = y.unsqueeze(-1)
+    y_onehot = torch.zeros(y.shape[0], len)
+    # y_onehot.zero_()
+    y_onehot.scatter(1, y, 1)
+    return y_onehot
+
+
+def sequence_mask(lengths, maxlen=None, dtype=torch.bool, device=None):
+    if maxlen is None:
+        maxlen = lengths.max()
+    mask = ~(torch.ones((len(lengths), maxlen), device=device).cumsum(dim=1).t() > lengths).t()
+    mask.type(dtype)
+    return mask

examples/trade/requirements.txt

@@ -0,0 +1,10 @@
+gym==0.17.3
+torch==1.6.0
+numba==0.51.2
+numpy==1.19.1
+pandas==1.1.3
+tqdm==4.50.2
+tianshou==0.3.0.post1
+env==0.1.0
+PyYAML==5.4.1
+redis==3.5.3