[to #42322933] Add S4: child-tuning

1. add child-tuning optimizer and ut
2. fix a training bug which can cause interruption after cross-evaluation
3. move model.params from cfg to default args in build_optimizer to prevent the saving of params in save_pretrained
        Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/9891963

modelscope/trainers/optimizer/__init__.py

@@ -1,4 +1,5 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
 from .builder import OPTIMIZERS, build_optimizer
+from .child_tuning_adamw_optimizer import ChildTuningAdamW
 
-__all__ = ['OPTIMIZERS', 'build_optimizer']
+__all__ = ['OPTIMIZERS', 'build_optimizer', 'ChildTuningAdamW']

modelscope/trainers/optimizer/builder.py

@@ -20,7 +20,10 @@ def build_optimizer(model: torch.nn.Module,
     """
     if hasattr(model, 'module'):
         model = model.module
-    cfg.params = model.parameters()
+
+    if default_args is None:
+        default_args = {}
+    default_args['params'] = model.parameters()
 
     return build_from_cfg(
         cfg, OPTIMIZERS, group_key=default_group, default_args=default_args)

modelscope/trainers/optimizer/child_tuning_adamw_optimizer.py

@@ -0,0 +1,188 @@
+# Copyright 2021-2022 The Alibaba DAMO NLP Team 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 math
+import types
+from typing import Callable, Iterable, Tuple
+
+import numpy as np
+import torch
+from torch.distributions.bernoulli import Bernoulli
+from torch.optim import Optimizer
+
+from modelscope.utils.logger import get_logger
+from .builder import OPTIMIZERS, default_group
+
+logger = get_logger(__name__)
+
+__all__ = ['calculate_fisher', 'ChildTuningAdamW']
+
+
+def calculate_fisher(model: torch.nn.Module,
+                     data_loader,
+                     forward_step,
+                     reserve_p,
+                     grad_clip=None):
+
+    gradient_mask = dict()
+    model.train()
+    for name, params in model.named_parameters():
+        if 'layer' in name:
+            gradient_mask[params] = params.new_zeros(params.size())
+
+    iters = len(data_loader)
+    for inputs in data_loader:
+        loss = forward_step(model, inputs)
+        loss.backward()
+        for name, params in model.named_parameters():
+            if 'layer' in name:
+                if grad_clip is not None:
+                    torch.nn.utils.clip_grad_norm_(params, **grad_clip)
+                gradient_mask[params] += (params.grad**2) / iters
+        model.zero_grad()
+
+    logger.info('Calculate Fisher Information...')
+
+    # Numpy
+    r = None
+    for k, v in gradient_mask.items():
+        v = v.view(-1).cpu().numpy()
+        if r is None:
+            r = v
+        else:
+            r = np.append(r, v)
+    polar = np.percentile(r, (1 - reserve_p) * 100)
+    for k in gradient_mask:
+        gradient_mask[k] = gradient_mask[k] >= polar
+    print('Polar => {}'.format(polar))
+
+    # TODO: pytorch: torch.kthvalue
+
+    return gradient_mask
+
+
+@OPTIMIZERS.register_module(
+    group_key=default_group, module_name='ChildTuningAdamW')
+class ChildTuningAdamW(Optimizer):
+
+    def __init__(self,
+                 params: Iterable[torch.nn.parameter.Parameter],
+                 lr: float = 1e-3,
+                 betas: Tuple[float, float] = (0.9, 0.999),
+                 eps: float = 1e-6,
+                 weight_decay: float = 0.0,
+                 correct_bias: bool = True,
+                 reserve_p=1.0,
+                 mode=None):
+        if lr < 0.0:
+            raise ValueError(
+                'Invalid learning rate: {} - should be >= 0.0'.format(lr))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError(
+                'Invalid beta parameter: {} - should be in [0.0, 1.0['.format(
+                    betas[0]))
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError(
+                'Invalid beta parameter: {} - should be in [0.0, 1.0['.format(
+                    betas[1]))
+        if not 0.0 <= eps:
+            raise ValueError(
+                'Invalid epsilon value: {} - should be >= 0.0'.format(eps))
+        defaults = dict(
+            lr=lr,
+            betas=betas,
+            eps=eps,
+            weight_decay=weight_decay,
+            correct_bias=correct_bias)
+        super().__init__(params, defaults)
+
+        self.gradient_mask = None
+        self.reserve_p = reserve_p
+        self.mode = mode
+
+    def set_gradient_mask(self, gradient_mask):
+        self.gradient_mask = gradient_mask
+
+    def step(self, closure: Callable = None):
+        """
+        Performs a single optimization step.
+        Arguments:
+            closure (:obj:`Callable`, `optional`): A closure that reevaluates the model and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            loss = closure()
+        for group in self.param_groups:
+            for p in group['params']:
+                if p.grad is None:
+                    continue
+                grad = p.grad.data
+                if grad.is_sparse:
+                    raise RuntimeError(
+                        'Adam does not support sparse gradients, please consider SparseAdam instead'
+                    )
+
+                # ChildTuning code
+                if self.mode is not None:
+                    if self.mode == 'ChildTuning-D':
+                        if p in self.gradient_mask:
+                            grad *= self.gradient_mask[p]
+                    else:
+                        # ChildTuning-F
+                        grad_mask = Bernoulli(
+                            grad.new_full(
+                                size=grad.size(), fill_value=self.reserve_p))
+                        grad *= grad_mask.sample() / self.reserve_p
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state['step'] = 0
+                    # Exponential moving average of gradient values
+                    state['exp_avg'] = torch.zeros_like(p.data)
+                    # Exponential moving average of squared gradient values
+                    state['exp_avg_sq'] = torch.zeros_like(p.data)
+
+                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
+                beta1, beta2 = group['betas']
+
+                state['step'] += 1
+
+                # Decay the first and second moment running average coefficient
+                # In-place operations to update the averages at the same time
+                exp_avg.mul_(beta1).add_(grad, alpha=1.0 - beta1)
+                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1.0 - beta2)
+                denom = exp_avg_sq.sqrt().add_(group['eps'])
+
+                step_size = group['lr']
+                if group['correct_bias']:  # No bias correction for Bert
+                    bias_correction1 = 1.0 - beta1**state['step']
+                    bias_correction2 = 1.0 - beta2**state['step']
+                    step_size = step_size * math.sqrt(
+                        bias_correction2) / bias_correction1
+
+                p.data.addcdiv_(exp_avg, denom, value=-step_size)
+
+                # Just adding the square of the weights to the loss function is *not*
+                # the correct way of using L2 regularization/weight decay with Adam,
+                # since that will interact with the m and v parameters in strange ways.
+                #
+                # Instead we want to decay the weights in a manner that doesn't interact
+                # with the m/v parameters. This is equivalent to adding the square
+                # of the weights to the loss with plain (non-momentum) SGD.
+                # Add weight decay at the end (fixed version)
+                p.data.add_(p.data, alpha=-group['lr'] * group['weight_decay'])
+
+        return loss

modelscope/trainers/trainer.py

@@ -800,6 +800,7 @@ class EpochBasedTrainer(BaseTrainer):
                 self.invoke_hook(TrainerStages.after_train_iter)
                 del self.data_batch
                 self._iter += 1
+                self._mode = ModeKeys.TRAIN
 
                 if i + 1 >= self.iters_per_epoch:
                     break

tests/trainers/test_finetune_sequence_classification.py

@@ -6,9 +6,15 @@ import unittest
 
 from modelscope.metainfo import Preprocessors, Trainers
 from modelscope.models import Model
+from modelscope.msdatasets import MsDataset
 from modelscope.pipelines import pipeline
 from modelscope.trainers import build_trainer
+from modelscope.trainers.hooks import Hook
+from modelscope.trainers.nlp_trainer import NlpEpochBasedTrainer
+from modelscope.trainers.optimizer.child_tuning_adamw_optimizer import \
+    calculate_fisher
 from modelscope.utils.constant import ModelFile, Tasks
+from modelscope.utils.data_utils import to_device
 
 
 class TestFinetuneSequenceClassification(unittest.TestCase):
@@ -69,6 +75,10 @@ class TestFinetuneSequenceClassification(unittest.TestCase):
 
     @unittest.skip
     def test_finetune_afqmc(self):
+        """This unittest is used to reproduce the clue:afqmc dataset + structbert model training results.
+
+        User can train a custom dataset by modifying this piece of code and comment the @unittest.skip.
+        """
 
         def cfg_modify_fn(cfg):
             cfg.task = Tasks.sentence_similarity
@@ -114,7 +124,7 @@ class TestFinetuneSequenceClassification(unittest.TestCase):
         dc.local_files_only = True
         dataset = load_dataset('clue', 'afqmc', download_config=dc)
         self.finetune(
-            model_id='damo/nlp_structbert_backbone_tiny_std',
+            model_id='damo/nlp_structbert_backbone_base_std',
             train_dataset=dataset['train'],
             eval_dataset=dataset['validation'],
             cfg_modify_fn=cfg_modify_fn)
@@ -124,6 +134,10 @@ class TestFinetuneSequenceClassification(unittest.TestCase):
 
     @unittest.skip
     def test_finetune_tnews(self):
+        """This unittest is used to reproduce the clue:tnews dataset + structbert model training results.
+
+        User can train a custom dataset by modifying this piece of code and comment the @unittest.skip.
+        """
 
         def cfg_modify_fn(cfg):
             # TODO no proper task for tnews
@@ -175,13 +189,21 @@ class TestFinetuneSequenceClassification(unittest.TestCase):
         dataset = load_dataset('clue', 'tnews', download_config=dc)
 
         self.finetune(
-            model_id='damo/nlp_structbert_backbone_tiny_std',
+            model_id='damo/nlp_structbert_backbone_base_std',
             train_dataset=dataset['train'],
             eval_dataset=dataset['validation'],
             cfg_modify_fn=cfg_modify_fn)
 
     @unittest.skip
     def test_veco_xnli(self):
+        """This unittest is used to reproduce the xnli dataset + veco model training results.
+
+        Here we follow the training scenario listed in the Alicemind open source project:
+        https://github.com/alibaba/AliceMind/tree/main/VECO
+        by training the english language subset.
+        User can train a custom dataset by modifying this piece of code and comment the @unittest.skip.
+        """
+
         from datasets import load_dataset
         langs = ['en']
         langs_eval = ['en']
@@ -267,6 +289,112 @@ class TestFinetuneSequenceClassification(unittest.TestCase):
             name=Trainers.nlp_veco_trainer,
             cfg_modify_fn=cfg_modify_fn)
 
+    @unittest.skip
+    def test_finetune_cluewsc(self):
+        """This unittest is used to reproduce the clue:wsc dataset + structbert model training results.
+
+        A runnable sample of child-tuning is also showed here.
+
+        User can train a custom dataset by modifying this piece of code and comment the @unittest.skip.
+        """
+
+        child_tuning_type = 'ChildTuning-F'
+        mode = {}
+        if child_tuning_type is not None:
+            mode = {'mode': child_tuning_type, 'reserve_p': 0.2}
+
+        def cfg_modify_fn(cfg):
+            cfg.task = 'nli'
+            cfg['preprocessor'] = {'type': 'nli-tokenizer'}
+            cfg['dataset'] = {
+                'train': {
+                    'labels': ['0', '1'],
+                    'first_sequence': 'text',
+                    'second_sequence': 'text2',
+                    'label': 'label',
+                }
+            }
+            cfg.train.dataloader.batch_size_per_gpu = 16
+            cfg.train.max_epochs = 30
+            cfg.train.optimizer = {
+                'type':
+                'AdamW' if child_tuning_type is None else 'ChildTuningAdamW',
+                'lr': 1e-5,
+                'options': {},
+                **mode,
+            }
+            cfg.train.lr_scheduler = {
+                'type':
+                'LinearLR',
+                'start_factor':
+                1.0,
+                'end_factor':
+                0.0,
+                'total_iters':
+                int(
+                    len(dataset['train'])
+                    / cfg.train.dataloader.batch_size_per_gpu)
+                * cfg.train.max_epochs,
+                'options': {
+                    'by_epoch': False
+                }
+            }
+            cfg.train.hooks = [{
+                'type': 'CheckpointHook',
+                'interval': 1
+            }, {
+                'type': 'TextLoggerHook',
+                'interval': 1
+            }, {
+                'type': 'IterTimerHook'
+            }, {
+                'type': 'EvaluationHook',
+                'by_epoch': False,
+                'interval': 30
+            }]
+            return cfg
+
+        def add_sentence2(features):
+            return {
+                'text2':
+                features['target']['span2_text'] + '指代'
+                + features['target']['span1_text']
+            }
+
+        dataset = MsDataset.load('clue', subset_name='cluewsc2020')
+        dataset = {
+            k: v.to_hf_dataset().map(add_sentence2)
+            for k, v in dataset.items()
+        }
+
+        kwargs = dict(
+            model='damo/nlp_structbert_backbone_base_std',
+            train_dataset=dataset['train'],
+            eval_dataset=dataset['validation'],
+            work_dir=self.tmp_dir,
+            cfg_modify_fn=cfg_modify_fn)
+
+        os.environ['LOCAL_RANK'] = '0'
+        trainer: NlpEpochBasedTrainer = build_trainer(
+            name=Trainers.nlp_base_trainer, default_args=kwargs)
+
+        class CalculateFisherHook(Hook):
+
+            @staticmethod
+            def forward_step(model, inputs):
+                inputs = to_device(inputs, trainer.device)
+                trainer.train_step(model, inputs)
+                return trainer.train_outputs['loss']
+
+            def before_run(self, trainer: NlpEpochBasedTrainer):
+                v = calculate_fisher(trainer.model, trainer.train_dataloader,
+                                     self.forward_step, 0.2)
+                trainer.optimizer.set_gradient_mask(v)
+
+        if child_tuning_type == 'ChildTuning-D':
+            trainer.register_hook(CalculateFisherHook())
+        trainer.train()
+
 
 if __name__ == '__main__':
     unittest.main()

tests/trainers/test_finetune_token_classificatin.py

@@ -47,6 +47,11 @@ class TestFinetuneTokenClassification(unittest.TestCase):
 
     @unittest.skip
     def test_word_segmentation(self):
+        """This unittest is used to reproduce the icwb2:pku dataset + structbert model training results.
+
+        User can train a custom dataset by modifying this piece of code and comment the @unittest.skip.
+        """
+
         os.system(
             f'curl http://sighan.cs.uchicago.edu/bakeoff2005/data/icwb2-data.zip > {self.tmp_dir}/icwb2-data.zip'
         )
@@ -114,7 +119,7 @@ class TestFinetuneTokenClassification(unittest.TestCase):
             return cfg
 
         self.finetune(
-            'damo/nlp_structbert_backbone_tiny_std',
+            'damo/nlp_structbert_backbone_base_std',
             train_dataset,
             dev_dataset,
             cfg_modify_fn=cfg_modify_fn)