|
| 1 | +# Copyright (c) 2021, salesforce.com, inc. |
| 2 | +# All rights reserved. |
| 3 | +# SPDX-License-Identifier: BSD-3-Clause |
| 4 | +# For full license text, see the LICENSE file in the repo root |
| 5 | +# or https://opensource.org/licenses/BSD-3-Clause |
| 6 | + |
| 7 | +import unittest |
| 8 | + |
| 9 | +import numpy as np |
| 10 | +import torch |
| 11 | + |
| 12 | +from warp_drive.managers.numba_managers.numba_data_manager import NumbaDataManager |
| 13 | +from warp_drive.managers.numba_managers.numba_function_manager import ( |
| 14 | + NumbaEnvironmentReset, |
| 15 | + NumbaFunctionManager, |
| 16 | +) |
| 17 | +from warp_drive.utils.common import get_project_root |
| 18 | +from warp_drive.utils.data_feed import DataFeed |
| 19 | + |
| 20 | +_NUMBA_FILEPATH = f"warp_drive.numba_includes" |
| 21 | + |
| 22 | + |
| 23 | +class TestEnvironmentReset(unittest.TestCase): |
| 24 | + """ |
| 25 | + Unit tests for the CUDA environment resetter |
| 26 | + """ |
| 27 | + |
| 28 | + def __init__(self, *args, **kwargs): |
| 29 | + super().__init__(*args, **kwargs) |
| 30 | + self.dm = NumbaDataManager(num_agents=5, num_envs=2, episode_length=2) |
| 31 | + self.fm = NumbaFunctionManager( |
| 32 | + num_agents=int(self.dm.meta_info("n_agents")), |
| 33 | + num_envs=int(self.dm.meta_info("n_envs")), |
| 34 | + ) |
| 35 | + self.fm.import_numba_from_source_code(f"{_NUMBA_FILEPATH}.test_build") |
| 36 | + self.resetter = NumbaEnvironmentReset(function_manager=self.fm) |
| 37 | + |
| 38 | + def test_reset_for_different_dim(self): |
| 39 | + |
| 40 | + self.dm.data_on_device_via_torch("_done_")[:] = torch.from_numpy( |
| 41 | + np.array([1, 0]) |
| 42 | + ).cuda() |
| 43 | + |
| 44 | + done = self.dm.pull_data_from_device("_done_") |
| 45 | + self.assertSequenceEqual(list(done), [1, 0]) |
| 46 | + |
| 47 | + # expected mean would be around 0.5 * (1+2+3+15) / 4 = 2.625 |
| 48 | + a_reset_pool = np.random.rand(4, 10, 10) |
| 49 | + a_reset_pool[1] *= 2 |
| 50 | + a_reset_pool[2] *= 3 |
| 51 | + a_reset_pool[3] *= 15 |
| 52 | + |
| 53 | + b_reset_pool = np.random.rand(4, 100) |
| 54 | + b_reset_pool[1] *= 2 |
| 55 | + b_reset_pool[2] *= 3 |
| 56 | + b_reset_pool[3] *= 15 |
| 57 | + |
| 58 | + c_reset_pool = np.random.rand(100) |
| 59 | + |
| 60 | + data_feed = DataFeed() |
| 61 | + data_feed.add_data( |
| 62 | + name="a", data=np.random.randn(2, 10, 10), save_copy_and_apply_at_reset=False |
| 63 | + ) |
| 64 | + data_feed.add_pool_for_reset(name="a_reset_pool", data=a_reset_pool, reset_target="a") |
| 65 | + data_feed.add_data( |
| 66 | + name="b", data=np.random.randn(2, 100), save_copy_and_apply_at_reset=False |
| 67 | + ) |
| 68 | + data_feed.add_pool_for_reset(name="b_reset_pool", data=b_reset_pool, reset_target="b") |
| 69 | + data_feed.add_data( |
| 70 | + name="c", data=np.random.randn(2), save_copy_and_apply_at_reset=False |
| 71 | + ) |
| 72 | + data_feed.add_pool_for_reset(name="c_reset_pool", data=c_reset_pool, reset_target="c") |
| 73 | + |
| 74 | + self.dm.push_data_to_device(data_feed) |
| 75 | + |
| 76 | + self.resetter.init_reset_pool(self.dm) |
| 77 | + |
| 78 | + a = self.dm.pull_data_from_device("a") |
| 79 | + b = self.dm.pull_data_from_device("b") |
| 80 | + c = self.dm.pull_data_from_device("c") |
| 81 | + |
| 82 | + # soft reset |
| 83 | + a_after_reset_0_mean = [] |
| 84 | + a_after_reset_1_mean = [] |
| 85 | + b_after_reset_0_mean = [] |
| 86 | + b_after_reset_1_mean = [] |
| 87 | + c_after_reset_0_mean = [] |
| 88 | + c_after_reset_1_mean = [] |
| 89 | + |
| 90 | + for _ in range(2000): |
| 91 | + self.resetter.reset_when_done(self.dm, mode="if_done", undo_done_after_reset=False) |
| 92 | + a_after_reset = self.dm.pull_data_from_device("a") |
| 93 | + a_after_reset_0_mean.append(a_after_reset[0].mean()) |
| 94 | + a_after_reset_1_mean.append(a_after_reset[1].mean()) |
| 95 | + b_after_reset = self.dm.pull_data_from_device("b") |
| 96 | + b_after_reset_0_mean.append(b_after_reset[0].mean()) |
| 97 | + b_after_reset_1_mean.append(b_after_reset[1].mean()) |
| 98 | + c_after_reset = self.dm.pull_data_from_device("c") |
| 99 | + c_after_reset_0_mean.append(c_after_reset[0].mean()) |
| 100 | + c_after_reset_1_mean.append(c_after_reset[1].mean()) |
| 101 | + # env 0 should have 1000 times random reset from the pool, so it should close to a_reset_pool.mean() |
| 102 | + print(a_reset_pool.mean()) |
| 103 | + print(np.array(a_after_reset_0_mean).mean()) |
| 104 | + self.assertTrue(np.absolute(a_reset_pool.mean() - np.array(a_after_reset_0_mean).mean()) < 5e-1) |
| 105 | + print(b_reset_pool.mean()) |
| 106 | + print(np.array(b_after_reset_0_mean).mean()) |
| 107 | + self.assertTrue(np.absolute(b_reset_pool.mean() - np.array(b_after_reset_0_mean).mean()) < 5e-1) |
| 108 | + print(c_reset_pool.mean()) |
| 109 | + print(np.array(c_after_reset_0_mean).mean()) |
| 110 | + self.assertTrue(np.absolute(c_reset_pool.mean() - np.array(c_after_reset_0_mean).mean()) < 5e-1) |
| 111 | + # env 1 has no reset at all, so it should be exactly the same as the original one |
| 112 | + self.assertTrue(np.absolute(a[1].mean() - np.array(a_after_reset_1_mean).mean()) < 1e-5) |
| 113 | + self.assertTrue(np.absolute(b[1].mean() - np.array(b_after_reset_1_mean).mean()) < 1e-5) |
| 114 | + self.assertTrue(np.absolute(c[1].mean() - np.array(c_after_reset_1_mean).mean()) < 1e-5) |
| 115 | + |
| 116 | + # hard reset |
| 117 | + a_after_reset_0_mean = [] |
| 118 | + a_after_reset_1_mean = [] |
| 119 | + b_after_reset_0_mean = [] |
| 120 | + b_after_reset_1_mean = [] |
| 121 | + c_after_reset_0_mean = [] |
| 122 | + c_after_reset_1_mean = [] |
| 123 | + for _ in range(2000): |
| 124 | + self.resetter.reset_when_done(self.dm, mode="force_reset", undo_done_after_reset=False) |
| 125 | + a_after_reset = self.dm.pull_data_from_device("a") |
| 126 | + a_after_reset_0_mean.append(a_after_reset[0].mean()) |
| 127 | + a_after_reset_1_mean.append(a_after_reset[1].mean()) |
| 128 | + b_after_reset = self.dm.pull_data_from_device("b") |
| 129 | + b_after_reset_0_mean.append(b_after_reset[0].mean()) |
| 130 | + b_after_reset_1_mean.append(b_after_reset[1].mean()) |
| 131 | + c_after_reset = self.dm.pull_data_from_device("c") |
| 132 | + c_after_reset_0_mean.append(c_after_reset[0].mean()) |
| 133 | + c_after_reset_1_mean.append(c_after_reset[1].mean()) |
| 134 | + # env 0 should have 1000 times random reset from the pool, so it should close to a_reset_pool.mean() |
| 135 | + self.assertTrue(np.absolute(a_reset_pool.mean() - np.array(a_after_reset_0_mean).mean()) < 5e-1) |
| 136 | + self.assertTrue(np.absolute(b_reset_pool.mean() - np.array(b_after_reset_0_mean).mean()) < 5e-1) |
| 137 | + self.assertTrue(np.absolute(c_reset_pool.mean() - np.array(c_after_reset_0_mean).mean()) < 5e-1) |
| 138 | + # env 1 should have 1000 times random reset from the pool, so it should close to a_reset_pool.mean() |
| 139 | + self.assertTrue(np.absolute(a_reset_pool.mean() - np.array(a_after_reset_1_mean).mean()) < 5e-1) |
| 140 | + self.assertTrue(np.absolute(b_reset_pool.mean() - np.array(b_after_reset_1_mean).mean()) < 5e-1) |
| 141 | + self.assertTrue(np.absolute(c_reset_pool.mean() - np.array(c_after_reset_1_mean).mean()) < 5e-1) |
| 142 | + |
| 143 | + |
| 144 | + |
| 145 | + |
0 commit comments