[Feature] Support LAMB optimizer. (#591)

* impl lamb * Add unit tests * Fix unit test * Fix unit tests * Use list instead of tuple in `__all__` according to PEP8 Co-authored-by: mzr1996 <mzr1996@163.com>
2021-12-13 17:24:44 +08:00 · 2021-12-13 17:24:44 +08:00 · 851b438574
parent 188aa6ed5d
commit 851b438574
4 changed files with 545 additions and 0 deletions
--- a/mmcls/core/init.py
+++ b/mmcls/core/init.py
@ -1,4 +1,5 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 from .evaluation import *  # noqa: F401, F403
 from .fp16 import *  # noqa: F401, F403
+from .optimizers import *  # noqa: F401, F403
 from .utils import *  # noqa: F401, F403
--- a/mmcls/core/optimizers/init.py
+++ b/mmcls/core/optimizers/init.py
@ -0,0 +1,5 @@
+from .lamb import Lamb
+
+__all__ = [
+    'Lamb',
+]
--- a/mmcls/core/optimizers/lamb.py
+++ b/mmcls/core/optimizers/lamb.py
@ -0,0 +1,231 @@
+"""PyTorch Lamb optimizer w/ behaviour similar to NVIDIA FusedLamb.
+
+This optimizer code was adapted from the following (starting with latest)
+* https://github.com/HabanaAI/Model-References/blob/
+2b435114fe8e31f159b1d3063b8280ae37af7423/PyTorch/nlp/bert/pretraining/lamb.py
+* https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/
+LanguageModeling/Transformer-XL/pytorch/lamb.py
+* https://github.com/cybertronai/pytorch-lamb
+
+Use FusedLamb if you can (GPU). The reason for including this variant of Lamb
+is to have a version that is
+similar in behaviour to APEX FusedLamb if you aren't using NVIDIA GPUs or
+cannot install/use APEX.
+
+In addition to some cleanup, this Lamb impl has been modified to support
+PyTorch XLA and has been tested on TPU.
+
+Original copyrights for above sources are below.
+
+Modifications Copyright 2021 Ross Wightman
+"""
+# Copyright (c) 2021, Habana Labs Ltd.  All rights reserved.
+
+# Copyright (c) 2019-2020, NVIDIA CORPORATION. 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.
+
+# MIT License
+#
+# Copyright (c) 2019 cybertronai
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in
+# all copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+import math
+
+import torch
+from mmcv.runner import OPTIMIZERS
+from torch.optim import Optimizer
+
+
+@OPTIMIZERS.register_module()
+class Lamb(Optimizer):
+    """Implements a pure pytorch variant of FuseLAMB (NvLamb variant) optimizer
+    from apex.optimizers.FusedLAMB
+    reference: https://github.com/NVIDIA/DeepLearningExamples/blob/master/
+    PyTorch/LanguageModeling/Transformer-XL/pytorch/lamb.py
+
+    LAMB was proposed in `Large Batch Optimization for Deep Learning: Training
+    BERT in 76 minutes`_.
+
+    Arguments:
+        params (iterable): iterable of parameters to optimize or dicts defining
+        parameter groups.
+        lr (float, optional): learning rate. (default: 1e-3)
+        betas (Tuple[float, float], optional): coefficients used for computing
+            running averages of gradient and its norm. (default: (0.9, 0.999))
+        eps (float, optional): term added to the denominator to improve
+            numerical stability. (default: 1e-8)
+        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
+        grad_averaging (bool, optional): whether apply (1-beta2) to grad when
+            calculating running averages of gradient. (default: True)
+        max_grad_norm (float, optional): value used to clip global grad norm
+            (default: 1.0)
+        trust_clip (bool): enable LAMBC trust ratio clipping (default: False)
+        always_adapt (boolean, optional): Apply adaptive learning rate to 0.0
+            weight decay parameter (default: False)
+
+    .. _Large Batch Optimization for Deep Learning - Training BERT in 76
+        minutes:
+        https://arxiv.org/abs/1904.00962
+    .. _On the Convergence of Adam and Beyond:
+        https://openreview.net/forum?id=ryQu7f-RZ
+    """
+
+    def __init__(self,
+                 params,
+                 lr=1e-3,
+                 bias_correction=True,
+                 betas=(0.9, 0.999),
+                 eps=1e-6,
+                 weight_decay=0.01,
+                 grad_averaging=True,
+                 max_grad_norm=1.0,
+                 trust_clip=False,
+                 always_adapt=False):
+        defaults = dict(
+            lr=lr,
+            bias_correction=bias_correction,
+            betas=betas,
+            eps=eps,
+            weight_decay=weight_decay,
+            grad_averaging=grad_averaging,
+            max_grad_norm=max_grad_norm,
+            trust_clip=trust_clip,
+            always_adapt=always_adapt)
+        super().__init__(params, defaults)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        device = self.param_groups[0]['params'][0].device
+        one_tensor = torch.tensor(
+            1.0, device=device
+        )  # because torch.where doesn't handle scalars correctly
+        global_grad_norm = torch.zeros(1, device=device)
+        for group in self.param_groups:
+            for p in group['params']:
+                if p.grad is None:
+                    continue
+                grad = p.grad
+                if grad.is_sparse:
+                    raise RuntimeError(
+                        'Lamb does not support sparse gradients, consider '
+                        'SparseAdam instead.')
+                global_grad_norm.add_(grad.pow(2).sum())
+
+        global_grad_norm = torch.sqrt(global_grad_norm)
+        # FIXME it'd be nice to remove explicit tensor conversion of scalars
+        #  when torch.where promotes
+        # scalar types properly https://github.com/pytorch/pytorch/issues/9190
+        max_grad_norm = torch.tensor(
+            self.defaults['max_grad_norm'], device=device)
+        clip_global_grad_norm = torch.where(global_grad_norm > max_grad_norm,
+                                            global_grad_norm / max_grad_norm,
+                                            one_tensor)
+
+        for group in self.param_groups:
+            bias_correction = 1 if group['bias_correction'] else 0
+            beta1, beta2 = group['betas']
+            grad_averaging = 1 if group['grad_averaging'] else 0
+            beta3 = 1 - beta1 if grad_averaging else 1.0
+
+            # assume same step across group now to simplify things
+            # per parameter step can be easily support by making it tensor, or
+            # pass list into kernel
+            if 'step' in group:
+                group['step'] += 1
+            else:
+                group['step'] = 1
+
+            if bias_correction:
+                bias_correction1 = 1 - beta1**group['step']
+                bias_correction2 = 1 - beta2**group['step']
+            else:
+                bias_correction1, bias_correction2 = 1.0, 1.0
+
+            for p in group['params']:
+                if p.grad is None:
+                    continue
+                grad = p.grad.div_(clip_global_grad_norm)
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    # Exponential moving average of gradient valuesa
+                    state['exp_avg'] = torch.zeros_like(p)
+                    # Exponential moving average of squared gradient values
+                    state['exp_avg_sq'] = torch.zeros_like(p)
+
+                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
+
+                # Decay the first and second moment running average coefficient
+                exp_avg.mul_(beta1).add_(grad, alpha=beta3)  # m_t
+                exp_avg_sq.mul_(beta2).addcmul_(
+                    grad, grad, value=1 - beta2)  # v_t
+
+                denom = (exp_avg_sq.sqrt() / math.sqrt(bias_correction2)).add_(
+                    group['eps'])
+                update = (exp_avg / bias_correction1).div_(denom)
+
+                weight_decay = group['weight_decay']
+                if weight_decay != 0:
+                    update.add_(p, alpha=weight_decay)
+
+                if weight_decay != 0 or group['always_adapt']:
+                    # Layer-wise LR adaptation. By default, skip adaptation on
+                    # parameters that are
+                    # excluded from weight decay, unless always_adapt == True,
+                    # then always enabled.
+                    w_norm = p.norm(2.0)
+                    g_norm = update.norm(2.0)
+                    # FIXME nested where required since logical and/or not
+                    #  working in PT XLA
+                    trust_ratio = torch.where(
+                        w_norm > 0,
+                        torch.where(g_norm > 0, w_norm / g_norm, one_tensor),
+                        one_tensor,
+                    )
+                    if group['trust_clip']:
+                        # LAMBC trust clipping, upper bound fixed at one
+                        trust_ratio = torch.minimum(trust_ratio, one_tensor)
+                    update.mul_(trust_ratio)
+
+                p.add_(update, alpha=-group['lr'])
+
+        return loss
--- a/tests/test_runtime/test_optimizer.py
+++ b/tests/test_runtime/test_optimizer.py
@ -0,0 +1,308 @@
+import functools
+from collections import OrderedDict
+from copy import deepcopy
+from typing import Iterable
+
+import torch
+import torch.nn as nn
+from mmcv.runner import build_optimizer
+from mmcv.runner.optimizer.builder import OPTIMIZERS
+from mmcv.utils.registry import build_from_cfg
+from torch.autograd import Variable
+from torch.optim.optimizer import Optimizer
+
+import mmcls.core  # noqa: F401
+
+base_lr = 0.01
+base_wd = 0.0001
+
+
+def assert_equal(x, y):
+    if isinstance(x, torch.Tensor) and isinstance(y, torch.Tensor):
+        torch.testing.assert_allclose(x, y.to(x.device))
+    elif isinstance(x, OrderedDict) and isinstance(y, OrderedDict):
+        for x_value, y_value in zip(x.values(), y.values()):
+            assert_equal(x_value, y_value)
+    elif isinstance(x, dict) and isinstance(y, dict):
+        assert x.keys() == y.keys()
+        for key in x.keys():
+            assert_equal(x[key], y[key])
+    elif isinstance(x, str) and isinstance(y, str):
+        assert x == y
+    elif isinstance(x, Iterable) and isinstance(y, Iterable):
+        assert len(x) == len(y)
+        for x_item, y_item in zip(x, y):
+            assert_equal(x_item, y_item)
+    else:
+        assert x == y
+
+
+class SubModel(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+        self.conv1 = nn.Conv2d(2, 2, kernel_size=1, groups=2)
+        self.gn = nn.GroupNorm(2, 2)
+        self.fc = nn.Linear(2, 2)
+        self.param1 = nn.Parameter(torch.ones(1))
+
+    def forward(self, x):
+        return x
+
+
+class ExampleModel(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+        self.param1 = nn.Parameter(torch.ones(1))
+        self.conv1 = nn.Conv2d(3, 4, kernel_size=1, bias=False)
+        self.conv2 = nn.Conv2d(4, 2, kernel_size=1)
+        self.bn = nn.BatchNorm2d(2)
+        self.sub = SubModel()
+        self.fc = nn.Linear(2, 1)
+
+    def forward(self, x):
+        return x
+
+
+def check_lamb_optimizer(optimizer,
+                         model,
+                         bias_lr_mult=1,
+                         bias_decay_mult=1,
+                         norm_decay_mult=1,
+                         dwconv_decay_mult=1):
+    param_groups = optimizer.param_groups
+    assert isinstance(optimizer, Optimizer)
+    assert optimizer.defaults['lr'] == base_lr
+    assert optimizer.defaults['weight_decay'] == base_wd
+    model_parameters = list(model.parameters())
+    assert len(param_groups) == len(model_parameters)
+    for i, param in enumerate(model_parameters):
+        param_group = param_groups[i]
+        assert torch.equal(param_group['params'][0], param)
+    # param1
+    param1 = param_groups[0]
+    assert param1['lr'] == base_lr
+    assert param1['weight_decay'] == base_wd
+    # conv1.weight
+    conv1_weight = param_groups[1]
+    assert conv1_weight['lr'] == base_lr
+    assert conv1_weight['weight_decay'] == base_wd
+    # conv2.weight
+    conv2_weight = param_groups[2]
+    assert conv2_weight['lr'] == base_lr
+    assert conv2_weight['weight_decay'] == base_wd
+    # conv2.bias
+    conv2_bias = param_groups[3]
+    assert conv2_bias['lr'] == base_lr * bias_lr_mult
+    assert conv2_bias['weight_decay'] == base_wd * bias_decay_mult
+    # bn.weight
+    bn_weight = param_groups[4]
+    assert bn_weight['lr'] == base_lr
+    assert bn_weight['weight_decay'] == base_wd * norm_decay_mult
+    # bn.bias
+    bn_bias = param_groups[5]
+    assert bn_bias['lr'] == base_lr
+    assert bn_bias['weight_decay'] == base_wd * norm_decay_mult
+    # sub.param1
+    sub_param1 = param_groups[6]
+    assert sub_param1['lr'] == base_lr
+    assert sub_param1['weight_decay'] == base_wd
+    # sub.conv1.weight
+    sub_conv1_weight = param_groups[7]
+    assert sub_conv1_weight['lr'] == base_lr
+    assert sub_conv1_weight['weight_decay'] == base_wd * dwconv_decay_mult
+    # sub.conv1.bias
+    sub_conv1_bias = param_groups[8]
+    assert sub_conv1_bias['lr'] == base_lr * bias_lr_mult
+    assert sub_conv1_bias['weight_decay'] == base_wd * dwconv_decay_mult
+    # sub.gn.weight
+    sub_gn_weight = param_groups[9]
+    assert sub_gn_weight['lr'] == base_lr
+    assert sub_gn_weight['weight_decay'] == base_wd * norm_decay_mult
+    # sub.gn.bias
+    sub_gn_bias = param_groups[10]
+    assert sub_gn_bias['lr'] == base_lr
+    assert sub_gn_bias['weight_decay'] == base_wd * norm_decay_mult
+    # sub.fc1.weight
+    sub_fc_weight = param_groups[11]
+    assert sub_fc_weight['lr'] == base_lr
+    assert sub_fc_weight['weight_decay'] == base_wd
+    # sub.fc1.bias
+    sub_fc_bias = param_groups[12]
+    assert sub_fc_bias['lr'] == base_lr * bias_lr_mult
+    assert sub_fc_bias['weight_decay'] == base_wd * bias_decay_mult
+    # fc1.weight
+    fc_weight = param_groups[13]
+    assert fc_weight['lr'] == base_lr
+    assert fc_weight['weight_decay'] == base_wd
+    # fc1.bias
+    fc_bias = param_groups[14]
+    assert fc_bias['lr'] == base_lr * bias_lr_mult
+    assert fc_bias['weight_decay'] == base_wd * bias_decay_mult
+
+
+def _test_state_dict(weight, bias, input, constructor):
+    weight = Variable(weight, requires_grad=True)
+    bias = Variable(bias, requires_grad=True)
+    inputs = Variable(input)
+
+    def fn_base(optimizer, weight, bias):
+        optimizer.zero_grad()
+        i = input_cuda if weight.is_cuda else inputs
+        loss = (weight.mv(i) + bias).pow(2).sum()
+        loss.backward()
+        return loss
+
+    optimizer = constructor(weight, bias)
+    fn = functools.partial(fn_base, optimizer, weight, bias)
+
+    # Prime the optimizer
+    for _ in range(20):
+        optimizer.step(fn)
+    # Clone the weights and construct new optimizer for them
+    weight_c = Variable(weight.data.clone(), requires_grad=True)
+    bias_c = Variable(bias.data.clone(), requires_grad=True)
+    optimizer_c = constructor(weight_c, bias_c)
+    fn_c = functools.partial(fn_base, optimizer_c, weight_c, bias_c)
+    # Load state dict
+    state_dict = deepcopy(optimizer.state_dict())
+    state_dict_c = deepcopy(optimizer.state_dict())
+    optimizer_c.load_state_dict(state_dict_c)
+    # Run both optimizations in parallel
+    for _ in range(20):
+        optimizer.step(fn)
+        optimizer_c.step(fn_c)
+        assert_equal(weight, weight_c)
+        assert_equal(bias, bias_c)
+    # Make sure state dict wasn't modified
+    assert_equal(state_dict, state_dict_c)
+    # Make sure state dict is deterministic with equal
+    # but not identical parameters
+    # NOTE: The state_dict of optimizers in PyTorch 1.5 have random keys,
+    state_dict = deepcopy(optimizer.state_dict())
+    state_dict_c = deepcopy(optimizer_c.state_dict())
+    keys = state_dict['param_groups'][-1]['params']
+    keys_c = state_dict_c['param_groups'][-1]['params']
+    for key, key_c in zip(keys, keys_c):
+        assert_equal(optimizer.state_dict()['state'][key],
+                     optimizer_c.state_dict()['state'][key_c])
+    # Make sure repeated parameters have identical representation in state dict
+    optimizer_c.param_groups.extend(optimizer_c.param_groups)
+    assert_equal(optimizer_c.state_dict()['param_groups'][0],
+                 optimizer_c.state_dict()['param_groups'][1])
+
+    # Check that state dict can be loaded even when we cast parameters
+    # to a different type and move to a different device.
+    if not torch.cuda.is_available():
+        return
+
+    input_cuda = Variable(inputs.data.float().cuda())
+    weight_cuda = Variable(weight.data.float().cuda(), requires_grad=True)
+    bias_cuda = Variable(bias.data.float().cuda(), requires_grad=True)
+    optimizer_cuda = constructor(weight_cuda, bias_cuda)
+    fn_cuda = functools.partial(fn_base, optimizer_cuda, weight_cuda,
+                                bias_cuda)
+
+    state_dict = deepcopy(optimizer.state_dict())
+    state_dict_c = deepcopy(optimizer.state_dict())
+    optimizer_cuda.load_state_dict(state_dict_c)
+
+    # Make sure state dict wasn't modified
+    assert_equal(state_dict, state_dict_c)
+
+    for _ in range(20):
+        optimizer.step(fn)
+        optimizer_cuda.step(fn_cuda)
+        assert_equal(weight, weight_cuda)
+        assert_equal(bias, bias_cuda)
+
+    # validate deepcopy() copies all public attributes
+    def getPublicAttr(obj):
+        return set(k for k in obj.__dict__ if not k.startswith('_'))
+
+    assert_equal(getPublicAttr(optimizer), getPublicAttr(deepcopy(optimizer)))
+
+
+def _test_basic_cases_template(weight, bias, inputs, constructor,
+                               scheduler_constructors):
+    """Copied from PyTorch."""
+    weight = Variable(weight, requires_grad=True)
+    bias = Variable(bias, requires_grad=True)
+    inputs = Variable(inputs)
+    optimizer = constructor(weight, bias)
+    schedulers = []
+    for scheduler_constructor in scheduler_constructors:
+        schedulers.append(scheduler_constructor(optimizer))
+
+    # to check if the optimizer can be printed as a string
+    optimizer.__repr__()
+
+    def fn():
+        optimizer.zero_grad()
+        y = weight.mv(inputs)
+        if y.is_cuda and bias.is_cuda and y.get_device() != bias.get_device():
+            y = y.cuda(bias.get_device())
+        loss = (y + bias).pow(2).sum()
+        loss.backward()
+        return loss
+
+    initial_value = fn().item()
+    for _ in range(200):
+        for scheduler in schedulers:
+            scheduler.step()
+        optimizer.step(fn)
+
+    assert fn().item() < initial_value
+
+
+def _test_basic_cases(constructor,
+                      scheduler_constructors=None,
+                      ignore_multidevice=False):
+    """Copied from PyTorch."""
+    if scheduler_constructors is None:
+        scheduler_constructors = []
+    _test_state_dict(
+        torch.randn(10, 5), torch.randn(10), torch.randn(5), constructor)
+    _test_basic_cases_template(
+        torch.randn(10, 5), torch.randn(10), torch.randn(5), constructor,
+        scheduler_constructors)
+    # non-contiguous parameters
+    _test_basic_cases_template(
+        torch.randn(10, 5, 2)[..., 0],
+        torch.randn(10, 2)[..., 0], torch.randn(5), constructor,
+        scheduler_constructors)
+    # CUDA
+    if not torch.cuda.is_available():
+        return
+    _test_basic_cases_template(
+        torch.randn(10, 5).cuda(),
+        torch.randn(10).cuda(),
+        torch.randn(5).cuda(), constructor, scheduler_constructors)
+    # Multi-GPU
+    if not torch.cuda.device_count() > 1 or ignore_multidevice:
+        return
+    _test_basic_cases_template(
+        torch.randn(10, 5).cuda(0),
+        torch.randn(10).cuda(1),
+        torch.randn(5).cuda(0), constructor, scheduler_constructors)
+
+
+def test_lamb_optimizer():
+    model = ExampleModel()
+    optimizer_cfg = dict(
+        type='Lamb',
+        lr=base_lr,
+        betas=(0.9, 0.999),
+        eps=1e-8,
+        weight_decay=base_wd,
+        paramwise_cfg=dict(
+            bias_lr_mult=2,
+            bias_decay_mult=0.5,
+            norm_decay_mult=0,
+            dwconv_decay_mult=0.1))
+    optimizer = build_optimizer(model, optimizer_cfg)
+    check_lamb_optimizer(optimizer, model, **optimizer_cfg['paramwise_cfg'])
+
+    _test_basic_cases(lambda weight, bias: build_from_cfg(
+        dict(type='Lamb', params=[weight, bias], lr=base_lr), OPTIMIZERS))