Unstructured prune (#4653) (#4675)

* unstructured prune for picodet

configs/picodet/README.md

@@ -267,6 +267,15 @@ python tools/post_quant.py -c configs/picodet/picodet_s_320_coco.yml \
 
 </details>
 
+## Unstructured Pruning
+
+<details open>
+<summary>Toturial:</summary>
+
+Please refer this [documentation](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/README_PRUNER.md) for details such as requirements, training and deployment.
+
+</details>
+
 ## Application
 
 - **Pedestrian detection:** model zoo of `PicoDet-S-Pedestrian` please refer to [PP-TinyPose](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.3/configs/keypoint/tiny_pose#%E8%A1%8C%E4%BA%BA%E6%A3%80%E6%B5%8B%E6%A8%A1%E5%9E%8B)

configs/picodet/README_PRUNER.md

+# 非结构化稀疏在 PicoDet 上的应用教程
+
+## 1. 介绍
+在模型压缩中，常见的稀疏方式为结构化稀疏和非结构化稀疏，前者在某个特定维度（特征通道、卷积核等等）上对卷积、矩阵乘法进行剪枝操作，然后生成一个更小的模型结构，这样可以复用已有的卷积、矩阵乘计算，无需特殊实现推理算子；后者以每一个参数为单元进行稀疏化，然而并不会改变参数矩阵的形状，所以更依赖于推理库、硬件对于稀疏后矩阵运算的加速能力。我们在 PP-PicoDet （以下简称PicoDet） 模型上运用了非结构化稀疏技术，在精度损失较小时，获得了在 ARM CPU 端推理的显著性能提升。本文档会介绍如何非结构化稀疏训练 PicoDet，关于非结构化稀疏的更多介绍请参照[这里](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/demo/dygraph/unstructured_pruning)。
+
+## 2. 版本要求
+```bash
+PaddlePaddle >= 2.1.2
+PaddleSlim develop分支 （pip install paddleslim -i https://pypi.tuna.tsinghua.edu.cn/simple）
+```
+
+## 3. 数据准备
+同 PicoDet
+
+## 4. 预训练模型
+在非结构化稀疏训练中，我们规定预训练模型是已经收敛完成的模型参数，所以需要额外在相关配置文件中声明。
+
+声明预训练模型地址的配置文件：./configs/picodet/pruner/picodet_m_320_coco_pruner.yml
+预训练模型地址请参照 PicoDet 文档：./configs/picodet/README.md
+
+## 5. 自定义稀疏化的作用范围
+为达到最佳推理加速效果，我们建议只对 1x1 卷积层进行稀疏化，其他层参数保持稠密。另外，有些层对于精度影响较大（例如head的最后几层，se-block的若干层），我们同样不建议对他们进行稀疏化，我们支持开发者通过传入自定义函数的形式，方便的指定哪些层不参与稀疏。例如，基于picodet_m_320这个模型，我们稀疏时跳过了后4层卷积以及6层se-block中的卷积，自定义函数如下：
+
+```python
+NORMS_ALL = [ 'BatchNorm', 'GroupNorm', 'LayerNorm', 'SpectralNorm', 'BatchNorm1D',
+    'BatchNorm2D', 'BatchNorm3D', 'InstanceNorm1D', 'InstanceNorm2D',
+    'InstanceNorm3D', 'SyncBatchNorm', 'LocalResponseNorm' ]
+
+def skip_params_self(model):
+    skip_params = set()
+    for _, sub_layer in model.named_sublayers():
+        if type(sub_layer).__name__.split('.')[-1] in NORMS_ALL:
+            skip_params.add(sub_layer.full_name())
+        for param in sub_layer.parameters(include_sublayers=False):
+            cond_is_conv1x1 = len(param.shape) == 4 and param.shape[2] == 1 and param.shape[3] == 1
+            cond_is_head_m = cond_is_conv1x1 and param.shape[0] == 112 and param.shape[1] == 128
+            cond_is_se_block_m = param.name.split('.')[0] in ['conv2d_17', 'conv2d_18', 'conv2d_56', 'conv2d_57', 'conv2d_75', 'conv2d_76']
+            if not cond_is_conv1x1 or cond_is_head_m or cond_is_se_block_m:
+                skip_params.add(param.name)
+    return skip_params
+```
+
+## 6. 训练
+我们已经将非结构化稀疏的核心功能通过 API 调用的方式嵌入到了训练中，所以如果您没有更细节的需求，直接运行 6.1 的命令启动训练即可。同时，为帮助您根据自己的需求更改、适配代码，我们也提供了更为详细的使用介绍，请参照 6.2。
+
+### 6.1 直接使用
+```bash
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+python3.7 -m paddle.distributed.launch --log_dir=log_test --gpus 0,1,2,3 tools/train.py -c configs/picodet/pruner/picodet_m_320_coco_pruner.yml --slim_config configs/slim/prune/picodet_m_unstructured_prune_75.yml --eval
+```
+
+### 6.2 详细介绍
+- 自定义稀疏化的作用范围：可以参照本教程的第 5 节
+- 如何添加稀疏化训练所需的 4 行代码
+
+```python
+# after constructing model and before training
+
+# Pruner Step1: configs
+configs = {
+    'pruning_strategy': 'gmp',
+    'stable_iterations': self.stable_epochs * steps_per_epoch,
+    'pruning_iterations': self.pruning_epochs * steps_per_epoch,
+    'tunning_iterations': self.tunning_epochs * steps_per_epoch,
+    'resume_iteration': 0,
+    'pruning_steps': self.pruning_steps,
+    'initial_ratio': self.initial_ratio,
+}
+
+# Pruner Step2: construct a pruner object
+self.pruner = GMPUnstructuredPruner(
+    model,
+    ratio=self.cfg.ratio,
+    skip_params_func=skip_params_self, # Only pass in this value when you design your own skip_params function. And the following argument (skip_params_type) will be ignored.
+    skip_params_type=self.cfg.skip_params_type,
+    local_sparsity=True,
+    configs=configs)
+
+# training
+for epoch_id in range(self.start_epoch, self.cfg.epoch):
+    model.train()
+    for step_id, data in enumerate(self.loader):
+        # model forward
+        outputs = model(data)
+        loss = outputs['loss']
+        # model backward
+        loss.backward()
+        self.optimizer.step()
+
+        # Pruner Step3: step during training
+        self.pruner.step()
+
+    # Pruner Step4: save the sparse model
+    self.pruner.update_params()
+    # model-saving API
+```
+
+## 7. 模型评估与推理部署
+这部分与 PicoDet 文档中基本一致，只是在转换到 PaddleLite 模型时，需要添加一个输入参数（sparse_model）：
+
+```bash
+paddle_lite_opt --model_dir=inference_model/picodet_m_320_coco --valid_targets=arm --optimize_out=picodet_m_320_coco_fp32_sparse --sparse_model=True
+```
+
+**注意：** 目前稀疏化推理适用于 PaddleLite的 FP32 和 INT8 模型，所以执行上述命令时，请不要打开 FP16 开关。
+
+## 8. 稀疏化结果
+我们在75%和85%稀疏度下，训练得到了 FP32 PicoDet-m模型，并在 SnapDragon-835设备上实测推理速度，效果如下表。其中：
+- 对于 m 模型，mAP损失1.5，获得了 34\%-58\% 的加速性能
+- 同样对于 m 模型，除4线程推理速度基本持平外，单线程推理速度、mAP、模型体积均优于 s 模型。
+
+
+| Model     | Input size | Sparsity | mAP<sup>val<br>0.5:0.95 | Size<br><sup>(MB) | Latency single-thread<sup><small>[Lite](#latency)</small><sup><br><sup>(ms) |  speed-up single-thread |  Latency 4-thread<sup><small>[Lite](#latency)</small><sup><br><sup>(ms) |  speed-up 4-thread |  Download  | SlimConfig |
+| :-------- | :--------: |:--------: | :---------------------: | :----------------: | :----------------: |:----------------: | :---------------: | :-----------------------------: | :-----------------------------: | :----------------------------------------: |
+| PicoDet-m-1.0 |  320*320   |   0      |          30.9         | 8.9 |  127     | 0    |  43     |    0       | [model](https://paddledet.bj.bcebos.com/models/picodet_m_320_coco.pdparams)&#124; [log](https://paddledet.bj.bcebos.com/logs/train_picodet_m_320_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.3/configs/picodet/picodet_m_320_coco.yml)|
+| PicoDet-m-1.0 |  320*320   |   75%    |          29.4         | 5.6 |  **80**  | 58%  | **32**  |   34%      | [model](https://paddledet.bj.bcebos.com/models/slim/picodet_m_320__coco_sparse_75.pdparams)&#124; [log](https://paddledet.bj.bcebos.com/logs/train_picodet_m_320__coco_sparse_75.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/slim/prune/picodet_m_unstructured_prune_75.yml)|
+| PicoDet-s-1.0 |  320*320   |   0      |          27.1         | 4.6 |    68    |  0   |    26   |    0       | [model](https://paddledet.bj.bcebos.com/models/picodet_s_320_coco.pdparams) &#124; [log](https://paddledet.bj.bcebos.com/logs/train_picodet_s_320_coco.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.3/configs/picodet/picodet_s_320_coco.yml)|
+| PicoDet-m-1.0 |  320*320   |   85%    |          27.6         | 4.1 |  **65**  | 96%  |  **27** |   59%      | [model](https://paddledet.bj.bcebos.com/models/slim/picodet_m_320__coco_sparse_85.pdparams) &#124; [log](https://paddledet.bj.bcebos.com/logs/train_picodet_m_320__coco_sparse_85.log) | [config](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/slim/prune/picodet_m_unstructured_prune_85.yml)|
+
+**注意：**
+- 上述模型体积是**部署模型体积**，即 PaddleLite 转换得到的 *.nb 文件的体积。
+- 加速一栏我们按照 FPS 增加百分比计算，即：$(dense\_latency - sparse\_latency) / sparse\_latency$
+- 上述稀疏化训练时，我们额外添加了一种数据增强方式到 _base_/picodet_320_reader.yml，代码如下。但是不添加的话，预期mAP也不会有明显下降（<0.1），且对速度和模型体积没有影响。
+```yaml
+worker_num: 6
+TrainReader:
+  sample_transforms:
+  - Decode: {}
+  - RandomCrop: {}
+  - RandomFlip: {prob: 0.5}
+  - RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
+  - RandomDistort: {}
+  batch_transforms:
+etc.
+```

configs/picodet/pruner/optimizer_300e_pruner.yml

+epoch: 300
+
+LearningRate:
+  base_lr: 0.15
+  schedulers:
+  - !CosineDecay
+    max_epochs: 300
+  - !LinearWarmup
+    start_factor: 1.0
+    steps: 34350
+
+OptimizerBuilder:
+  optimizer:
+    momentum: 0.9
+    type: Momentum
+  regularizer:
+    factor: 0.00004
+    type: L2

configs/picodet/pruner/picodet_m_320_coco_pruner.yml

+_BASE_: [
+  '../../datasets/coco_detection.yml',
+  '../../runtime.yml',
+  '../_base_/picodet_esnet.yml',
+  './optimizer_300e_pruner.yml',
+  '../_base_/picodet_320_reader.yml',
+]
+
+weights: output/picodet_m_320_coco/model_final
+find_unused_parameters: True
+use_ema: true
+cycle_epoch: 40
+snapshot_epoch: 10

configs/slim/prune/picodet_m_unstructured_prune_75.yml

+pretrain_weights: https://paddledet.bj.bcebos.com/models/picodet_m_320_coco.pdparams
+slim: UnstructuredPruner
+
+UnstructuredPruner:
+    stable_epochs: 0
+    pruning_epochs: 150
+    tunning_epochs: 150
+    pruning_steps: 300
+    ratio: 0.75
+    initial_ratio: 0.15
+    prune_params_type: conv1x1_only

configs/slim/prune/picodet_m_unstructured_prune_85.yml

+pretrain_weights: https://paddledet.bj.bcebos.com/models/picodet_m_320_coco.pdparams
+slim: UnstructuredPruner
+
+UnstructuredPruner:
+    stable_epochs: 0
+    pruning_epochs: 150
+    tunning_epochs: 150
+    pruning_steps: 300
+    ratio: 0.85
+    initial_ratio: 0.20
+    prune_params_type: conv1x1_only

ppdet/engine/trainer.py

@@ -81,7 +81,8 @@ class Trainer(object):
             # JDE only support single class MOT now.
 
         if cfg.architecture == 'FairMOT' and self.mode == 'train':
-            cfg['FairMOTEmbeddingHead']['num_identities_dict'] = self.dataset.num_identities_dict
+            cfg['FairMOTEmbeddingHead'][
+                'num_identities_dict'] = self.dataset.num_identities_dict
             # FairMOT support single class and multi-class MOT now.
 
         # build model
@@ -119,6 +120,10 @@ class Trainer(object):
             self.lr = create('LearningRate')(steps_per_epoch)
             self.optimizer = create('OptimizerBuilder')(self.lr, self.model)
 
+        if self.cfg.get('unstructured_prune'):
+            self.pruner = create('UnstructuredPruner')(self.model,
+                                                       steps_per_epoch)
+
         self._nranks = dist.get_world_size()
         self._local_rank = dist.get_rank()
 
@@ -395,9 +400,10 @@ class Trainer(object):
                     # model backward
                     loss.backward()
                     self.optimizer.step()
-
                 curr_lr = self.optimizer.get_lr()
                 self.lr.step()
+                if self.cfg.get('unstructured_prune'):
+                    self.pruner.step()
                 self.optimizer.clear_grad()
                 self.status['learning_rate'] = curr_lr
 
@@ -414,6 +420,8 @@ class Trainer(object):
             if self.use_ema:
                 weight = copy.deepcopy(self.model.state_dict())
                 self.model.set_dict(self.ema.apply())
+            if self.cfg.get('unstructured_prune'):
+                self.pruner.update_params()
 
             self._compose_callback.on_epoch_end(self.status)
 

ppdet/slim/__init__.py

@@ -15,10 +15,12 @@
 from . import prune
 from . import quant
 from . import distill
+from . import unstructured_prune
 
 from .prune import *
 from .quant import *
 from .distill import *
+from .unstructured_prune import *
 
 import yaml
 from ppdet.core.workspace import load_config
@@ -56,6 +58,12 @@ def build_slim_model(cfg, slim_cfg, mode='train'):
         cfg['slim_type'] = cfg.slim
         cfg['model'] = slim(model)
         cfg['slim'] = slim
+    elif slim_load_cfg['slim'] == 'UnstructuredPruner':
+        load_config(slim_cfg)
+        slim = create(cfg.slim)
+        cfg['slim_type'] = cfg.slim
+        cfg['slim'] = slim
+        cfg['unstructured_prune'] = True
     else:
         load_config(slim_cfg)
         model = create(cfg.architecture)

ppdet/slim/unstructured_prune.py

+# Copyright (c) 2021 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.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from paddle.utils import try_import
+
+from ppdet.core.workspace import register, serializable
+from ppdet.utils.logger import setup_logger
+logger = setup_logger(__name__)
+
+
+@register
+@serializable
+class UnstructuredPruner(object):
+    def __init__(self,
+                 stable_epochs,
+                 pruning_epochs,
+                 tunning_epochs,
+                 pruning_steps,
+                 ratio,
+                 initial_ratio,
+                 prune_params_type=None):
+        self.stable_epochs = stable_epochs
+        self.pruning_epochs = pruning_epochs
+        self.tunning_epochs = tunning_epochs
+        self.ratio = ratio
+        self.prune_params_type = prune_params_type
+        self.initial_ratio = initial_ratio
+        self.pruning_steps = pruning_steps
+
+    def __call__(self, model, steps_per_epoch, skip_params_func=None):
+        paddleslim = try_import('paddleslim')
+        from paddleslim import GMPUnstructuredPruner
+        configs = {
+            'pruning_strategy': 'gmp',
+            'stable_iterations': self.stable_epochs * steps_per_epoch,
+            'pruning_iterations': self.pruning_epochs * steps_per_epoch,
+            'tunning_iterations': self.tunning_epochs * steps_per_epoch,
+            'resume_iteration': 0,
+            'pruning_steps': self.pruning_steps,
+            'initial_ratio': self.initial_ratio,
+        }
+
+        pruner = GMPUnstructuredPruner(
+            model,
+            ratio=self.ratio,
+            skip_params_func=skip_params_func,
+            prune_params_type=self.prune_params_type,
+            local_sparsity=True,
+            configs=configs)
+
+        return pruner