File size: 8,442 Bytes
079c32c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
import shutil
from time import sleep
import pytest
import numpy as np
import tempfile

import torch
from ding.data.model_loader import FileModelLoader
from ding.data.storage_loader import FileStorageLoader
from ding.framework import task
from ding.framework.context import OnlineRLContext
from ding.framework.middleware.distributer import ContextExchanger, ModelExchanger, PeriodicalModelExchanger
from ding.framework.parallel import Parallel
from ding.utils.default_helper import set_pkg_seed
from os import path


def context_exchanger_main():
    with task.start(ctx=OnlineRLContext()):
        if task.router.node_id == 0:
            task.add_role(task.role.LEARNER)
        elif task.router.node_id == 1:
            task.add_role(task.role.COLLECTOR)

        task.use(ContextExchanger(skip_n_iter=1))

        if task.has_role(task.role.LEARNER):

            def learner_context(ctx: OnlineRLContext):
                assert len(ctx.trajectories) == 2
                assert len(ctx.trajectory_end_idx) == 4
                assert len(ctx.episodes) == 8
                assert ctx.env_step > 0
                assert ctx.env_episode > 0
                yield
                ctx.train_iter += 1

            task.use(learner_context)
        elif task.has_role(task.role.COLLECTOR):

            def collector_context(ctx: OnlineRLContext):
                if ctx.total_step > 0:
                    assert ctx.train_iter > 0
                yield
                ctx.trajectories = [np.random.rand(10, 10) for _ in range(2)]
                ctx.trajectory_end_idx = [1 for _ in range(4)]
                ctx.episodes = [np.random.rand(10, 10) for _ in range(8)]
                ctx.env_step += 1
                ctx.env_episode += 1

            task.use(collector_context)

        task.run(max_step=3)


@pytest.mark.tmp
def test_context_exchanger():
    Parallel.runner(n_parallel_workers=2)(context_exchanger_main)


def context_exchanger_with_storage_loader_main():
    with task.start(ctx=OnlineRLContext()):
        if task.router.node_id == 0:
            task.add_role(task.role.LEARNER)
        elif task.router.node_id == 1:
            task.add_role(task.role.COLLECTOR)

        tempdir = path.join(tempfile.gettempdir(), "test_storage_loader")
        storage_loader = FileStorageLoader(dirname=tempdir)
        try:
            task.use(ContextExchanger(skip_n_iter=1, storage_loader=storage_loader))

            if task.has_role(task.role.LEARNER):

                def learner_context(ctx: OnlineRLContext):
                    assert len(ctx.trajectories) == 2
                    assert len(ctx.trajectory_end_idx) == 4
                    assert len(ctx.episodes) == 8
                    assert ctx.env_step > 0
                    assert ctx.env_episode > 0
                    yield
                    ctx.train_iter += 1

                task.use(learner_context)
            elif task.has_role(task.role.COLLECTOR):

                def collector_context(ctx: OnlineRLContext):
                    if ctx.total_step > 0:
                        assert ctx.train_iter > 0
                    yield
                    ctx.trajectories = [np.random.rand(10, 10) for _ in range(2)]
                    ctx.trajectory_end_idx = [1 for _ in range(4)]
                    ctx.episodes = [np.random.rand(10, 10) for _ in range(8)]
                    ctx.env_step += 1
                    ctx.env_episode += 1

                task.use(collector_context)

            task.run(max_step=3)
        finally:
            storage_loader.shutdown()
            sleep(1)
            if path.exists(tempdir):
                shutil.rmtree(tempdir)


@pytest.mark.tmp
def test_context_exchanger_with_storage_loader():
    Parallel.runner(n_parallel_workers=2)(context_exchanger_with_storage_loader_main)


class MockPolicy:

    def __init__(self) -> None:
        self._model = self._get_model(10, 10)

    def _get_model(self, X_shape, y_shape) -> torch.nn.Module:
        return torch.nn.Sequential(
            torch.nn.Linear(X_shape, 24), torch.nn.ReLU(), torch.nn.Linear(24, 24), torch.nn.ReLU(),
            torch.nn.Linear(24, y_shape)
        )

    def train(self, X, y):
        loss_fn = torch.nn.MSELoss(reduction="mean")
        optimizer = torch.optim.Adam(self._model.parameters(), lr=0.01)
        y_pred = self._model(X)
        loss = loss_fn(y_pred, y)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

    def predict(self, X):
        with torch.no_grad():
            return self._model(X)


def model_exchanger_main():
    with task.start(ctx=OnlineRLContext()):
        set_pkg_seed(0, use_cuda=False)
        policy = MockPolicy()
        X = torch.rand(10)
        y = torch.rand(10)

        if task.router.node_id == 0:
            task.add_role(task.role.LEARNER)
        else:
            task.add_role(task.role.COLLECTOR)

        task.use(ModelExchanger(policy._model))

        if task.has_role(task.role.LEARNER):

            def train(ctx):
                policy.train(X, y)
                sleep(0.3)

            task.use(train)
        else:
            y_pred1 = policy.predict(X)

            def pred(ctx):
                if ctx.total_step > 0:
                    y_pred2 = policy.predict(X)
                    # Ensure model is upgraded
                    assert any(y_pred1 != y_pred2)
                sleep(0.3)

            task.use(pred)

        task.run(2)


@pytest.mark.tmp
def test_model_exchanger():
    Parallel.runner(n_parallel_workers=2, startup_interval=0)(model_exchanger_main)


def model_exchanger_main_with_model_loader():
    with task.start(ctx=OnlineRLContext()):
        set_pkg_seed(0, use_cuda=False)
        policy = MockPolicy()
        X = torch.rand(10)
        y = torch.rand(10)

        if task.router.node_id == 0:
            task.add_role(task.role.LEARNER)
        else:
            task.add_role(task.role.COLLECTOR)

        tempdir = path.join(tempfile.gettempdir(), "test_model_loader")
        model_loader = FileModelLoader(policy._model, dirname=tempdir)
        task.use(ModelExchanger(policy._model, model_loader=model_loader))

        try:
            if task.has_role(task.role.LEARNER):

                def train(ctx):
                    policy.train(X, y)
                    sleep(0.3)

                task.use(train)
            else:
                y_pred1 = policy.predict(X)

                def pred(ctx):
                    if ctx.total_step > 0:
                        y_pred2 = policy.predict(X)
                        # Ensure model is upgraded
                        assert any(y_pred1 != y_pred2)
                    sleep(0.3)

                task.use(pred)
            task.run(2)
        finally:
            model_loader.shutdown()
            sleep(0.3)
            if path.exists(tempdir):
                shutil.rmtree(tempdir)


@pytest.mark.tmp
def test_model_exchanger_with_model_loader():
    Parallel.runner(n_parallel_workers=2, startup_interval=0)(model_exchanger_main_with_model_loader)


def periodical_model_exchanger_main():
    with task.start(ctx=OnlineRLContext()):
        set_pkg_seed(0, use_cuda=False)
        policy = MockPolicy()
        X = torch.rand(10)
        y = torch.rand(10)

        if task.router.node_id == 0:
            task.add_role(task.role.LEARNER)
            task.use(PeriodicalModelExchanger(policy._model, mode="send", period=3))
        else:
            task.add_role(task.role.COLLECTOR)
            task.use(PeriodicalModelExchanger(policy._model, mode="receive", period=1, stale_toleration=3))

        if task.has_role(task.role.LEARNER):

            def train(ctx):
                policy.train(X, y)
                sleep(0.3)

            task.use(train)
        else:
            y_pred1 = policy.predict(X)
            print("y_pred1: ", y_pred1)
            stale = 1

            def pred(ctx):
                nonlocal stale
                y_pred2 = policy.predict(X)
                print("y_pred2: ", y_pred2)
                stale += 1
                assert stale <= 3 or all(y_pred1 == y_pred2)
                if any(y_pred1 != y_pred2):
                    stale = 1

                sleep(0.3)

            task.use(pred)
        task.run(8)


@pytest.mark.tmp
def test_periodical_model_exchanger():
    Parallel.runner(n_parallel_workers=2, startup_interval=0)(periodical_model_exchanger_main)