[Feature] Add DAFL Distillation (#235)

* 1.Add DAFL, including config, DAFLLoss and readme. 2.Add DataFreeDistillationtillation. 3.Add Generator, including base_generator and dafl_generator. 4.Add get_module_device and set_requires_grad functions in utils. * 1.Amend the file that report error in mypy test under py37, including gather_tensors, datafree_distillation, base_generator. 2.Revise other linting error. * 1.Revise some docstrings. * 1.Add UT for datafreedistillation. 2.Add all typing.hints. * 1.Add UT for generators and gather_tensors. * 1.Add assert of batch_size in base_generator * 1.Isort Co-authored-by: zhangzhongyu.vendor < zhangzhongyu.vendor@sensetime.com>
open-mmlab · Aug 23, 2022 · 57aec1f · 57aec1f
1 parent 72c1175
commit 57aec1f
Show file tree

Hide file tree

Showing 23 changed files with 1,092 additions and 18 deletions.
diff --git a/configs/distill/mmcls/dafl/README.md b/configs/distill/mmcls/dafl/README.md
@@ -0,0 +1,42 @@
+# Data-Free Learning of Student Networks (DAFL)
+
+> [Data-Free Learning of Student Networks](https://doi.org/10.1109/ICCV.2019.00361)
+
+<!-- [ALGORITHM] -->
+
+## Abstract
+
+Learning portable neural networks is very essential for computer vision for the purpose that pre-trained heavy deep models can be well applied on edge devices such as mobile phones and micro sensors. Most existing deep neural network compression and speed-up methods are very effective for training compact deep models, when we can directly access the training dataset. However, training data for the given deep network are often unavailable due to some practice problems (e.g. privacy, legal issue, and transmission), and the architecture of the given network are also unknown except some interfaces. To this end, we propose a novel framework for training efficient deep neural networks by exploiting generative adversarial networks (GANs). To be specific, the pre-trained teacher networks are regarded as a fixed discriminator and the generator is utilized for deviating training samples which can obtain the maximum response on the discriminator. Then, an efficient network with smaller model size and computational complexity is trained using the generated data and the teacher network, simultaneously. Efficient student networks learned using the pro- posed Data-Free Learning (DAFL) method achieve 92.22% and 74.47% accuracies using ResNet-18 without any training data on the CIFAR-10 and CIFAR-100 datasets, respectively. Meanwhile, our student network obtains an 80.56% accuracy on the CelebA benchmark.
+
+![pipeline](/docs/en/imgs/model_zoo/dafl/pipeline.png)
+
+## Results and models
+
+### Classification
+
+|               Location               | Dataset |                                                     Teacher                                                     |                                                     Student                                                     |  Acc  | Acc(T) | Acc(S) |                      Config                       | Download                                                                                                                                     |
+| :----------------------------------: | :-----: | :-------------------------------------------------------------------------------------------------------------: | :-------------------------------------------------------------------------------------------------------------: | :---: | :----: | :----: | :-----------------------------------------------: | :------------------------------------------------------------------------------------------------------------------------------------------- |
+| backbone (pretrain) & logits (train) | Cifar10 | [resnet34](https://github.com/open-mmlab/mmclassification/blob/master/configs/resnet/resnet34_8xb16_cifar10.py) | [resnet18](https://github.com/open-mmlab/mmclassification/blob/master/configs/resnet/resnet18_8xb16_cifar10.py) | 93.11 | 95.34  | 94.82  | [config](./dafl_logits_r34_r18_8xb256_cifar10.py) | [teacher](https://download.openmmlab.com/mmclassification/v0/resnet/resnet34_b16x8_cifar10_20210528-a8aa36a6.pth) \|[model](<>) \| [log](<>) |
+
+## Citation
+
+```latex
+@inproceedings{DBLP:conf/iccv/ChenW0YLSXX019,
+  author    = {Hanting Chen, Yunhe Wang, Chang Xu, Zhaohui Yang, Chuanjian Liu,
+               Boxin Shi, Chunjing Xu, Chao Xu and Qi Tian},
+  title     = {Data-Free Learning of Student Networks},
+  booktitle = {2019 {IEEE/CVF} International Conference on Computer Vision, {ICCV}
+               2019, Seoul, Korea (South), October 27 - November 2, 2019},
+  pages     = {3513--3521},
+  publisher = {{IEEE}},
+  year      = {2019},
+  url       = {https://doi.org/10.1109/ICCV.2019.00361},
+  doi       = {10.1109/ICCV.2019.00361},
+  timestamp = {Mon, 17 May 2021 08:18:18 +0200},
+  biburl    = {https://dblp.org/rec/conf/iccv/ChenW0YLSXX019.bib},
+  bibsource = {dblp computer science bibliography, https://dblp.org}
+```
+
+## Acknowledgement
+
+Shout out to Davidgzx.
diff --git a/configs/distill/mmcls/dafl/dafl_logits_r34_r18_8xb256_cifar10.py b/configs/distill/mmcls/dafl/dafl_logits_r34_r18_8xb256_cifar10.py
@@ -0,0 +1,104 @@
+_base_ = [
+    'mmcls::_base_/datasets/cifar10_bs16.py',
+    'mmcls::_base_/schedules/cifar10_bs128.py',
+    'mmcls::_base_/default_runtime.py'
+]
+
+model = dict(
+    _scope_='mmrazor',
+    type='DAFLDataFreeDistillation',
+    data_preprocessor=dict(
+        type='ImgDataPreprocessor',
+        # RGB format normalization parameters
+        mean=[125.307, 122.961, 113.8575],
+        std=[51.5865, 50.847, 51.255],
+        # convert image from BGR to RGB
+        bgr_to_rgb=False),
+    architecture=dict(
+        cfg_path='mmcls::resnet/resnet18_8xb16_cifar10.py', pretrained=False),
+    teachers=dict(
+        res34=dict(
+            build_cfg=dict(
+                cfg_path='mmcls::resnet/resnet34_8xb16_cifar10.py',
+                pretrained=True),
+            ckpt_path='resnet34_b16x8_cifar10_20210528-a8aa36a6.pth')),
+    generator=dict(
+        type='DAFLGenerator',
+        img_size=32,
+        latent_dim=1000,
+        hidden_channels=128),
+    distiller=dict(
+        type='ConfigurableDistiller',
+        student_recorders=dict(
+            fc=dict(type='ModuleOutputs', source='head.fc')),
+        teacher_recorders=dict(
+            res34_fc=dict(type='ModuleOutputs', source='res34.head.fc')),
+        distill_losses=dict(
+            loss_kl=dict(type='KLDivergence', tau=6, loss_weight=1)),
+        loss_forward_mappings=dict(
+            loss_kl=dict(
+                preds_S=dict(from_student=True, recorder='fc'),
+                preds_T=dict(from_student=False, recorder='res34_fc')))),
+    generator_distiller=dict(
+        type='ConfigurableDistiller',
+        teacher_recorders=dict(
+            res34_neck_gap=dict(type='ModuleOutputs', source='res34.neck.gap'),
+            res34_fc=dict(type='ModuleOutputs', source='res34.head.fc')),
+        distill_losses=dict(
+            loss_res34_oh=dict(type='OnehotLikeLoss', loss_weight=0.05),
+            loss_res34_ie=dict(type='InformationEntropyLoss', loss_weight=5),
+            loss_res34_ac=dict(type='ActivationLoss', loss_weight=0.01)),
+        loss_forward_mappings=dict(
+            loss_res34_oh=dict(
+                preds_T=dict(from_student=False, recorder='res34_fc')),
+            loss_res34_ie=dict(
+                preds_T=dict(from_student=False, recorder='res34_fc')),
+            loss_res34_ac=dict(
+                feat_T=dict(from_student=False, recorder='res34_neck_gap')))))
+
+# model wrapper
+model_wrapper_cfg = dict(
+    type='mmengine.MMSeparateDistributedDataParallel',
+    broadcast_buffers=False,
+    find_unused_parameters=False)
+
+find_unused_parameters = True
+
+# optimizer wrapper
+optim_wrapper = dict(
+    _delete_=True,
+    constructor='mmrazor.SeparateOptimWrapperConstructor',
+    architecture=dict(optimizer=dict(type='AdamW', lr=1e-1)),
+    generator=dict(optimizer=dict(type='AdamW', lr=1e-3)))
+
+auto_scale_lr = dict(base_batch_size=256)
+
+param_scheduler = dict(
+    _delete_=True,
+    architecture=[
+        dict(type='LinearLR', end=500, by_epoch=False, start_factor=0.0001),
+        dict(
+            type='MultiStepLR',
+            begin=500,
+            milestones=[100 * 120, 200 * 120],
+            by_epoch=False)
+    ],
+    generator=dict(
+        type='LinearLR', end=500, by_epoch=False, start_factor=0.0001))
+
+train_cfg = dict(
+    _delete_=True, by_epoch=False, max_iters=250 * 120, val_interval=150)
+
+train_dataloader = dict(
+    batch_size=256, sampler=dict(type='InfiniteSampler', shuffle=True))
+val_dataloader = dict(batch_size=256)
+val_evaluator = dict(type='Accuracy', topk=(1, 5))
+
+default_hooks = dict(
+    logger=dict(type='LoggerHook', interval=75, log_metric_by_epoch=False),
+    checkpoint=dict(
+        type='CheckpointHook', by_epoch=False, interval=150, max_keep_ckpts=2))
+
+log_processor = dict(by_epoch=False)
+# Must set diff_rank_seed to True!
+randomness = dict(seed=None, diff_rank_seed=True)
diff --git a/docs/en/imgs/model_zoo/dafl/pipeline.png b/docs/en/imgs/model_zoo/dafl/pipeline.png
diff --git a/mmrazor/models/algorithms/__init__.py b/mmrazor/models/algorithms/__init__.py
@@ -1,11 +1,13 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 from .base import BaseAlgorithm
-from .distill import FpnTeacherDistill, SelfDistill, SingleTeacherDistill
+from .distill import (DAFLDataFreeDistillation, DataFreeDistillation,
+                      FpnTeacherDistill, SelfDistill, SingleTeacherDistill)
 from .nas import SPOS, AutoSlim, AutoSlimDDP, Darts, DartsDDP
 from .pruning import SlimmableNetwork, SlimmableNetworkDDP
 
 __all__ = [
     'SingleTeacherDistill', 'BaseAlgorithm', 'FpnTeacherDistill', 'SPOS',
     'SlimmableNetwork', 'SlimmableNetworkDDP', 'AutoSlim', 'AutoSlimDDP',
-    'Darts', 'DartsDDP', 'SelfDistill'
+    'Darts', 'DartsDDP', 'SelfDistill', 'DataFreeDistillation',
+    'DAFLDataFreeDistillation'
 ]
diff --git a/mmrazor/models/algorithms/distill/__init__.py b/mmrazor/models/algorithms/distill/__init__.py
@@ -1,4 +1,9 @@
 # Copyright (c) OpenMMLab. All rights reserved.
-from .configurable import FpnTeacherDistill, SelfDistill, SingleTeacherDistill
+from .configurable import (DAFLDataFreeDistillation, DataFreeDistillation,
+                           FpnTeacherDistill, SelfDistill,
+                           SingleTeacherDistill)
 
-__all__ = ['SingleTeacherDistill', 'FpnTeacherDistill', 'SelfDistill']
+__all__ = [
+    'SingleTeacherDistill', 'FpnTeacherDistill', 'SelfDistill',
+    'DataFreeDistillation', 'DAFLDataFreeDistillation'
+]
diff --git a/mmrazor/models/algorithms/distill/configurable/__init__.py b/mmrazor/models/algorithms/distill/configurable/__init__.py
@@ -1,6 +1,11 @@
 # Copyright (c) OpenMMLab. All rights reserved.
+from .datafree_distillation import (DAFLDataFreeDistillation,
+                                    DataFreeDistillation)
 from .fpn_teacher_distill import FpnTeacherDistill
 from .self_distill import SelfDistill
 from .single_teacher_distill import SingleTeacherDistill
 
-__all__ = ['SelfDistill', 'SingleTeacherDistill', 'FpnTeacherDistill']
+__all__ = [
+    'SelfDistill', 'SingleTeacherDistill', 'FpnTeacherDistill',
+    'DataFreeDistillation', 'DAFLDataFreeDistillation'
+]