test_lambv2_op.py 10.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import unittest
16

17
import numpy as np
18

19
import paddle
20 21
import paddle.fluid as fluid
import paddle.fluid.layers as layers
22 23
from paddle.fluid import core
from paddle.fluid.dygraph.base import switch_to_static_graph
24 25 26 27 28 29 30


class LAMBOptimizer(paddle.optimizer.Lamb):
    def _append_optimize_op(self, block, param_and_grad):
        assert isinstance(block, fluid.framework.Block)
        block.program._use_lamb = True

31 32 33
        m = moment1 = self._get_accumulator(
            self._moment1_acc_str, param_and_grad[0]
        )
34
        v = self._get_accumulator(self._moment2_acc_str, param_and_grad[0])
35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
        beta_1_pow_acc = self._get_accumulator(
            self._beta1_pow_acc_str, param_and_grad[0]
        )
        beta_2_pow_acc = self._get_accumulator(
            self._beta2_pow_acc_str, param_and_grad[0]
        )

        beta_1 = layers.fill_constant(
            dtype='float32', shape=[1], value=self._beta1, name='lamb_beta_1'
        )
        beta_2 = layers.fill_constant(
            dtype='float32', shape=[1], value=self._beta2, name='lamb_beta_2'
        )
        epsilon = layers.fill_constant(
            dtype='float32', shape=[1], value=self._epsilon, name='epsilon'
        )
51 52 53 54

        one = paddle.ones(shape=[1]).astype('float32')
        zero = paddle.zeros(shape=[1]).astype('float32')

55
        next_m = paddle.multiply(m, beta_1) + paddle.multiply(
56 57
            param_and_grad[1], one - beta_1
        )
58
        next_v = paddle.multiply(v, beta_2) + paddle.multiply(
59 60
            paddle.pow(param_and_grad[1], 2), one - beta_2
        )
61 62 63 64 65 66 67 68 69

        beta1_correction = one - beta_1_pow_acc
        beta2_correction = one - beta_2_pow_acc

        next_m_unbiased = next_m / beta1_correction
        next_v_unbiased = next_v / beta2_correction

        update = next_m_unbiased / (paddle.sqrt(next_v_unbiased) + epsilon)

70 71 72 73
        if (
            self._exclude_from_weight_decay_fn is not None
            and self._exclude_from_weight_decay_fn(param_and_grad[0])
        ):
74 75 76 77 78 79 80 81 82 83
            self._lamb_weight_decay = 0.0
        update += self._lamb_weight_decay * param_and_grad[0]

        w_norm = paddle.norm(param_and_grad[0], p=2)
        g_norm = paddle.norm(update, p=2)

        learning_rate = self._create_param_lr(param_and_grad)

        ratio = paddle.where(
            paddle.greater_than(w_norm, zero),
84 85 86 87 88
            paddle.where(
                paddle.greater_than(g_norm, zero), (w_norm / g_norm), one
            ),
            one,
        )
89 90 91 92 93 94 95 96 97 98 99
        update_with_lr = ratio * learning_rate * update
        next_param = param_and_grad[0] - update_with_lr

        beta_1_pow_acc *= beta_1
        beta_2_pow_acc *= beta_2

        paddle.assign(next_m, m)
        paddle.assign(next_v, v)
        paddle.assign(next_param, param_and_grad[0])

        return None
100 101 102 103


class TestLambOpV2(unittest.TestCase):
    def test_lamb_op(self):
104 105 106 107 108
        shape = [2, 4, 8, 8]
        data = paddle.to_tensor(np.random.random(size=shape).astype("float32"))
        conv = paddle.nn.Conv2D(4, 6, (3, 3))
        data = conv(data)
        loss = paddle.mean(data)
109 110 111
        opt = paddle.optimizer.Lamb(
            learning_rate=1e-5, epsilon=1e-8, parameters=conv.parameters()
        )
112 113 114 115 116 117 118 119
        loss.backward()
        opt.minimize(loss)

        assert loss.numpy() is not None


class TestLambOpWithCombinedOp(unittest.TestCase):
    def test_lamb_op_with_multi_steps(self):
120
        paddle.enable_static()
121 122 123 124 125

        def _build_static_model(main, startup, seed=100):
            with fluid.program_guard(main, startup):
                main.random_seed = seed
                startup.random_seed = seed
G
GGBond8488 已提交
126 127 128 129
                x = paddle.static.data(
                    name='X', shape=[-1, 13], dtype='float32'
                )
                y = paddle.static.data(name='Y', shape=[-1, 1], dtype='float32')
C
Charles-hit 已提交
130
                prediction = paddle.static.nn.fc(x, size=1, activation=None)
131 132 133
                loss = paddle.nn.functional.square_error_cost(
                    input=prediction, label=y
                )
134
                avg_loss = paddle.mean(loss)
135 136
            return avg_loss

137
        place = fluid.CPUPlace()
138 139 140 141 142 143 144 145 146 147 148 149 150 151 152
        num_steps = 10

        for i in range(num_steps):
            feed_x = np.random.random(size=(10, 13)).astype('float32')
            feed_y = np.random.random(size=(10, 1)).astype('float32')

            main_program = fluid.Program()
            startup_program = fluid.Program()
            with fluid.program_guard(main_program, startup_program):
                avg_loss = _build_static_model(main_program, startup_program)
                lamb_kernel = paddle.optimizer.Lamb(learning_rate=0.2)
                lamb_kernel.minimize(avg_loss)

            executor = fluid.Executor(place)
            executor.run(startup_program)
153 154 155 156 157
            output = executor.run(
                program=main_program,
                feed={'X': feed_x, 'Y': feed_y},
                fetch_list=[avg_loss.name],
            )
158 159 160 161 162 163 164 165 166 167

            main = fluid.Program()
            startup = fluid.Program()
            with fluid.program_guard(main, startup):
                loss = _build_static_model(main, startup)
                lamb = LAMBOptimizer(learning_rate=0.2)
                lamb.minimize(loss)

            exe = fluid.Executor(place)
            exe.run(startup)
168 169 170 171 172
            out = exe.run(
                program=main,
                feed={'X': feed_x, 'Y': feed_y},
                fetch_list=[loss.name],
            )
173

174
            np.testing.assert_allclose(out, output, rtol=1e-05)
175 176


177 178 179 180 181 182 183 184
class TestLambOpV2Group(TestLambOpV2):
    def test_lamb_op(self):
        paddle.disable_static()
        value = np.arange(26).reshape(2, 13).astype("float32")
        a = paddle.to_tensor(value)
        linear_1 = paddle.nn.Linear(13, 5)
        linear_2 = paddle.nn.Linear(5, 3)
        # This can be any optimizer supported by dygraph.
185 186 187 188 189 190 191 192 193 194 195 196 197
        adam = paddle.optimizer.Lamb(
            learning_rate=0.01,
            parameters=[
                {'params': linear_1.parameters()},
                {
                    'params': linear_2.parameters(),
                    'lamb_weight_decay': 0.001,
                    'beta1': 0.9,
                    'beta2': 0.99,
                },
            ],
            lamb_weight_decay=0.01,
        )
198 199 200 201 202 203 204
        out = linear_1(a)
        out = linear_2(out)
        out.backward()
        adam.step()
        adam.clear_gradients()


205 206 207 208 209 210 211
class TestLambOpMultiPrecision(unittest.TestCase):
    def check_main(self, x_np, place, multi_precision=False, seed=10, n=10):
        main_prog = paddle.static.Program()
        startup_prog = paddle.static.Program()
        with paddle.static.program_guard(main_prog, startup_prog):
            paddle.seed(seed)
            with paddle.static.amp.fp16_guard():
212 213 214
                x = paddle.static.data(
                    name='x', shape=[None, 10], dtype='float32'
                )
215 216 217 218
                linear = paddle.nn.Linear(10, 2)
                hidden = linear(x)
                loss = paddle.mean(hidden)

219 220
            original_optimizer = paddle.optimizer.Lamb(learning_rate=1e-3)
            original_optimizer._multi_precision = multi_precision
221
            if multi_precision:
222 223 224
                optimizer = paddle.static.amp.decorate(
                    original_optimizer, use_pure_fp16=True, use_fp16_guard=True
                )
225 226
            else:
                optimizer = original_optimizer
227 228 229 230 231 232 233 234 235
            optimizer.minimize(loss)

        weight, bias = linear.weight, linear.bias
        exe = paddle.static.Executor(place)
        scope = paddle.static.Scope()
        x = main_prog.global_block().var(x.name)
        if x.dtype == core.VarDesc.VarType.FP16:
            x_np = x_np.astype(np.float16)

236 237 238 239 240 241 242 243 244
        def get_parameter(var):
            name = var if isinstance(var, (str, bytes)) else var.name
            params = original_optimizer._get_parameter(name, scope)
            assert isinstance(params, (list, tuple))
            params = list(params)
            assert len(params) == 2
            if multi_precision:
                params[0] = np.array(params[0])
                params[1] = np.array(params[1])
245 246 247
                np.testing.assert_array_equal(
                    params[0], params[1].astype(np.float16)
                )
248 249
                return params[0].astype(np.float32)
            else:
250 251
                self.assertIsNotNone(params[0])
                self.assertIsNone(params[1])
252 253 254
                params[0] = np.array(params[0])
                return params[0]

255 256 257 258
        with paddle.static.scope_guard(scope):
            exe.run(startup_prog)
            if multi_precision:
                optimizer.amp_init(place)
259

260 261 262
            weight_np, bias_np = None, None
            for i in range(n):
                feed_dict = {x.name: x_np}
263 264 265
                weight_np, bias_np = exe.run(
                    main_prog, feed=feed_dict, fetch_list=[weight, bias]
                )
266 267
                weight_np = weight_np.astype('float32')
                bias_np = bias_np.astype('float32')
268 269
                np.testing.assert_array_equal(weight_np, get_parameter(weight))
                np.testing.assert_array_equal(bias_np, get_parameter(bias))
270
            return weight_np, bias_np
271 272 273 274 275 276 277 278 279 280 281 282 283 284

    @switch_to_static_graph
    def test_main(self):
        if not paddle.is_compiled_with_cuda():
            return

        place = paddle.CUDAPlace(0)
        x_np = np.random.random(size=[5, 10]).astype('float32')
        weight_1, bias_1 = self.check_main(x_np, place, multi_precision=False)
        weight_2, bias_2 = self.check_main(x_np, place, multi_precision=True)
        self.assertTrue(np.all(np.abs(weight_1 - weight_2) < 1e-3))
        self.assertTrue(np.all(np.abs(bias_1 - bias_2) < 1e-7))


285 286
if __name__ == "__main__":
    unittest.main()