add ppdet auto compression demo (#1039)

demo/auto-compression/configs/PaddleDet/README.md

+# 使用预测模型进行量化训练示例
+
+预测模型保存接口：
+动态图使用``paddle.jit.save``保存；
+静态图使用``paddle.static.save_inference_model``保存。
+
+本示例将介绍如何使用PaddleDetection中预测模型进行蒸馏量化训练。
+
+## 模型量化蒸馏训练流程
+
+### 1. 准备COCO格式数据
+
+参考[COCO数据准备文档](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.4/docs/tutorials/PrepareDataSet.md#coco%E6%95%B0%E6%8D%AE)
+
+### 2. 准备需要量化的环境
+
+- PaddlePaddle >= 2.2
+- PaddleDet >= 2.3
+
+```shell
+pip install paddledet
+```
+
+#### 3 准备待量化模型
+- 下载代码
+```
+git clone https://github.com/PaddlePaddle/PaddleDetection.git
+```
+- 导出预测模型
+```shell
+python tools/export_model.py -c configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams
+```
+或直接下载：
+```shell
+wget https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_270e_coco.tar
+tar -xf yolov3_mobilenet_v1_270e_coco.tar
+```
+
+#### 2.4 测试模型精度
+拷贝``yolov3_mobilenet_v1_270e_coco``文件夹到``PaddleSlim/demo/auto-compression/``文件夹。
+```
+cd PaddleSlim/demo/auto-compression/
+```
+使用[demo_coco.py](../demo_coco.py)脚本得到模型的分类精度：
+```
+python3.7 ../demo_coco.py --model_dir=../yolov3_mobilenet_v1_270e_coco/ --model_filename=model.pdmodel --params_filename=model.pdiparams --eval=True
+```
+
+### 3. 进行多策略融合压缩
+
+每一个小章节代表一种多策略融合压缩，不代表需要串行执行。
+
+### 3.1 进行量化蒸馏压缩
+蒸馏量化训练示例脚本为[demo_coco.py](../demo_coco.py)，使用接口``paddleslim.auto_compression.AutoCompression``对模型进行量化训练。运行命令为：
+```
+python ../demo_coco.py \
+    --model_dir='infermodel_mobilenetv2' \
+    --model_filename='model.pdmodel' \
+    --params_filename='./model.pdiparams' \
+    --save_dir='./output/' \
+    --devices='gpu' \
+    --config_path='./yolov3_mbv1_qat_dis.yaml'
+```

demo/auto-compression/configs/PaddleDet/coco_dataset.yml

@@ -5,14 +5,13 @@ TrainDataset:
   !COCODataSet
     image_dir: train2017
     anno_path: annotations/instances_train2017.json
-    dataset_dir: dataset/coco
-    data_fields: ['image', 'gt_bbox', 'gt_class', 'is_crowd']
+    dataset_dir: dataset/coco/
 
 EvalDataset:
   !COCODataSet
     image_dir: val2017
     anno_path: annotations/instances_val2017.json
-    dataset_dir: dataset/coco
+    dataset_dir: dataset/coco/
 
 TestDataset:
   !ImageFolder

demo/auto-compression/configs/PaddleDet/ppyoloe_reader.yml

-_BASE_: [
-  './coco_dataset.yml',
-]
-
-worker_num: 8
-TrainReader:
-  sample_transforms:
-    - Decode: {}
-    - RandomDistort: {}
-    - RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
-    - RandomCrop: {}
-    - RandomFlip: {}
-  batch_transforms:
-    - BatchRandomResize: {target_size: [320, 352, 384, 416, 448, 480, 512, 544, 576, 608, 640, 672, 704, 736, 768], random_size: True, random_interp: True, keep_ratio: False}
-    - NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
-    - Permute: {}
-    - PadGT: {}
-  batch_size: 24
-  shuffle: true
-  drop_last: true
-  use_shared_memory: true
-  collate_batch: true
-
-EvalReader:
-  sample_transforms:
-    - Decode: {}
-    - Resize: {target_size: [640, 640], keep_ratio: False, interp: 2}
-    - NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
-    - Permute: {}
-  batch_size: 4
-
-TestReader:
-  inputs_def:
-    image_shape: [3, 640, 640]
-  sample_transforms:
-    - Decode: {}
-    - Resize: {target_size: [640, 640], keep_ratio: False, interp: 2}
-    - NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
-    - Permute: {}
-  batch_size: 1

demo/auto-compression/configs/PaddleDet/yolo_reader.yml

+_BASE_: [
+  './coco_dataset.yml',
+]
+
+worker_num: 8
+
+TestReader:
+  inputs_def:
+    image_shape: [3, 640, 640]
+  sample_transforms:
+    - Decode: {}
+    - Resize: {target_size: [640, 640], keep_ratio: False, interp: 2}
+    - NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
+    - Permute: {}
+  batch_size: 4

demo/auto-compression/configs/PaddleDet/yolov3_mbv1_qat_dis.yaml

+Distillation:
+  distill_lambda: 1.0
+  distill_loss: l2_loss
+  distill_node_pair:
+  - teacher_conv2d_84.tmp_0
+  - conv2d_84.tmp_0
+  - teacher_conv2d_85.tmp_0
+  - conv2d_85.tmp_0
+  - teacher_conv2d_86.tmp_0
+  - conv2d_86.tmp_0
+  merge_feed: true
+  teacher_model_dir: ./yolov3_mobilenet_v1_270e_coco/
+  teacher_model_filename: model.pdmodel
+  teacher_params_filename: model.pdiparams
+Quantization:
+  activation_bits: 8
+  is_full_quantize: false
+  not_quant_pattern:
+  - skip_quant
+  quantize_op_types:
+  - conv2d
+  - depthwise_conv2d
+  weight_bits: 8
+TrainConfig:
+  epochs: 1
+  eval_iter: 1000
+  learning_rate: 0.0001
+  optimizer: SGD
+  optim_args:
+    weight_decay: 4.0e-05
+  #origin_metric: 0.289
+

demo/auto-compression/demo_coco.py

@@ -2,15 +2,156 @@ import os
 import sys
 sys.path[0] = os.path.join(os.path.dirname("__file__"), os.path.pardir)
 import argparse
-
+import functools
+from functools import partial
+import numpy as np
+import paddle
 from ppdet.core.workspace import load_config, merge_config
 from ppdet.core.workspace import create
+from ppdet.metrics import COCOMetric
+from paddleslim.auto_compression.config_helpers import load_config as load_slim_config
+from paddleslim.auto_compression import AutoCompression
+
+paddle.enable_static()
+
+from utility import add_arguments, print_arguments
+
+parser = argparse.ArgumentParser(description=__doc__)
+add_arg = functools.partial(add_arguments, argparser=parser)
+
+# yapf: disable
+add_arg('model_dir',                   str,    None,         "inference model directory.")
+add_arg('model_filename',              str,    None,         "inference model filename.")
+add_arg('params_filename',             str,    None,         "inference params filename.")
+add_arg('save_dir',                    str,    'output',         "directory to save compressed model.")
+add_arg('devices',                     str,    'gpu',        "which device used to compress.")
+add_arg('batch_size',                  int,    1,            "train batch size.")
+add_arg('config_path',                 str,    None,         "path of compression strategy config.")
+add_arg('eval',                  bool,    False,            "whether to run evaluation.")
+# yapf: enable
+
+
+def reader_wrapper(reader):
+    def gen():
+        for data in reader:
+            yield {
+                "image": data['image'],
+                'im_shape': data['im_shape'],
+                'scale_factor': data['scale_factor']
+            }
+
+    return gen
+
+
+def eval():
+    dataset = reader_cfg['EvalDataset']
+    val_loader = create('TestReader')(dataset,
+                                      reader_cfg['worker_num'],
+                                      return_list=True)
+
+    place = paddle.CUDAPlace(0) if args.devices == 'gpu' else paddle.CPUPlace()
+    exe = paddle.static.Executor(place)
+
+    val_program, feed_target_names, fetch_targets = paddle.fluid.io.load_inference_model(
+        args.model_dir,
+        exe,
+        model_filename=args.model_filename,
+        params_filename=args.params_filename)
+    clsid2catid = {v: k for k, v in dataset.catid2clsid.items()}
+
+    anno_file = dataset.get_anno()
+    metric = COCOMetric(
+        anno_file=anno_file, clsid2catid=clsid2catid, bias=0, IouType='bbox')
+    for batch_id, data in enumerate(val_loader):
+        data_new = {k: np.array(v) for k, v in data.items()}
+        outs = exe.run(val_program,
+                       feed={
+                           'image': data['image'],
+                           'im_shape': data['im_shape'],
+                           'scale_factor': data['scale_factor']
+                       },
+                       fetch_list=fetch_targets,
+                       return_numpy=False)
+        res = {}
+        for out in outs:
+            v = np.array(out)
+            if len(v.shape) > 1:
+                res['bbox'] = v
+            else:
+                res['bbox_num'] = v
+
+        metric.update(data_new, res)
+        if batch_id % 100 == 0:
+            print('Eval iter:', batch_id)
+    metric.accumulate()
+    metric.log()
+    metric.reset()
+
+
+def eval_function(exe, compiled_test_program, test_feed_names, test_fetch_list):
+    clsid2catid = {v: k for k, v in dataset.catid2clsid.items()}
+
+    anno_file = dataset.get_anno()
+    metric = COCOMetric(
+        anno_file=anno_file, clsid2catid=clsid2catid, bias=1, IouType='bbox')
+    for batch_id, data in enumerate(val_loader):
+        data_new = {k: np.array(v) for k, v in data.items()}
+        outs = exe.run(compiled_test_program,
+                       feed={
+                           'image': data['image'],
+                           'im_shape': data['im_shape'],
+                           'scale_factor': data['scale_factor']
+                       },
+                       fetch_list=test_fetch_list,
+                       return_numpy=False)
+        res = {}
+        for out in outs:
+            v = np.array(out)
+            if len(v.shape) > 1:
+                res['bbox'] = v
+            else:
+                res['bbox_num'] = v
+
+        metric.update(data_new, res)
+        if batch_id % 100 == 0:
+            print('Eval iter:', batch_id)
+    metric.accumulate()
+    metric.log()
+    map_res = metric.get_results()
+    metric.reset()
+    return map_res['bbox'][0]
+
+
+if __name__ == '__main__':
+    args = parser.parse_args()
+    print_arguments(args)
+    paddle.enable_static()
+    reader_cfg = load_config('./configs/PaddleDet/yolo_reader.yml')
+    if args.eval:
+        eval()
+        sys.exit(0)
+
+    compress_config, train_config = load_slim_config(args.config_path)
 
-cfg = load_config('./configs/PaddleDet/ppyoloe_reader.yml')
+    train_loader = create('TestReader')(reader_cfg['TrainDataset'],
+                                        reader_cfg['worker_num'],
+                                        return_list=True)
+    dataset = reader_cfg['EvalDataset']
+    val_loader = create('TestReader')(reader_cfg['EvalDataset'],
+                                      reader_cfg['worker_num'],
+                                      return_list=True)
 
-print(cfg)
+    train_dataloader = reader_wrapper(train_loader)
 
-coco_loader = create('TestReader')(cfg['TrainDataset'], cfg['worker_num'])
+    ac = AutoCompression(
+        model_dir=args.model_dir,
+        model_filename=args.model_filename,
+        params_filename=args.params_filename,
+        save_dir=args.save_dir,
+        strategy_config=compress_config,
+        train_config=train_config,
+        train_dataloader=train_dataloader,
+        eval_callback=eval_function,
+        devices=args.devices)
 
-for data in coco_loader:
-    print(data.keys())
+    ac.compress()

paddleslim/auto_compression/compressor.py

@@ -315,7 +315,8 @@ class AutoCompression:
                     _logger.info("epoch: {}, batch: {}, loss: {}".format(
                         epoch_id, batch_id, np_probs_float))
 
-                if batch_id % int(self.train_config.eval_iter) == 0:
+                if batch_id % int(
+                        self.train_config.eval_iter) == 0 and batch_id != 0:
                     if self.eval_function is not None:
 
                         # GMP pruner step 3: update params before summrizing sparsity, saving model or evaluation. 

paddleslim/quant/quanter.py

@@ -198,7 +198,8 @@ def quant_aware(program,
                 optimizer_func=None,
                 executor=None,
                 onnx_format=False,
-                return_program=False):
+                return_program=False,
+                draw_graph=False):
     """Add quantization  and dequantization operators to "program" 
     for quantization training or testing.
 
@@ -241,6 +242,8 @@ def quant_aware(program,
                 initialization. Default is None.
         return_program(bool): If user want return value is a Program rather than Compiled Program, This argument should be set True.
                 Default is False.
+        draw_graph(bool): whether to draw graph when quantization is initialized. In order to prevent cycle,
+                the ERNIE model needs to be set to True. Default is False.
     Returns:
         paddle.static.CompiledProgram | paddle.static.Program: Program with quantization and dequantization ``operators``
     """
@@ -308,15 +311,10 @@ def quant_aware(program,
             VARS_MAPPING_TABLE))
         save_dict(main_graph.out_node_mapping_table)
 
-    main_graph.draw('./', 'graph.pdf')
-    #remove_ctr_vars = set()
-    #from paddle.fluid.framework import IrVarNode
-    #all_var_nodes = {IrVarNode(node) for node in main_graph.nodes() if node.is_var()}
-    #for node in all_var_nodes:
-    #    print("node: ", node)
-    #    if node.is_ctrl_var():
-    #        remove_ctr_vars.add(node)
-    #self.safe_remove_nodes(remove_ctr_vars)
+    # TDOD: remove it.
+    if draw_graph:
+        main_graph.draw('./', 'graph.pdf')
+
     if for_test or return_program:
         quant_program = main_graph.to_program()
     else: