BMN模型和AttentionLSTM模型量化压缩和蒸馏方案

[AvaMins]

1、BMN模型和AttentionLSTM模型是否有量化压缩和蒸馏的方案？训练完成的AttentionLSTM模型的模型参数和计算量如何计算？
2、模型使用自动压缩，参数文件大小没有变化，如何处理这种情况？

[wanghaoshuang]

BMN模型和AttentionLSTM模型是否有量化压缩和蒸馏的方案？

飞桨没有提供这两个模型的量化压缩何蒸馏方案，可以参考其它模型来做，基础原理是一样的。

训练完成的AttentionLSTM模型的模型参数和计算量如何计算？

可以使用飞桨接口 paddle.flops 计算flops(计算量)；
paddle.nn.Layer 提供 parameters function，用于获取模型的所有参数。相关文档：https://www.paddlepaddle.org.cn/documentation/docs/zh/api/paddle/nn/Layer_cn.html#parameters-include-sublayers-true

模型使用自动压缩，参数文件大小没有变化，如何处理这种情况？

辛苦上传下日志，我们确认下具体触发了什么压缩算法？

[AvaMins]

BMN模型和AttentionLSTM模型是否有量化压缩和蒸馏的方案？

飞桨没有提供这两个模型的量化压缩何蒸馏方案，可以参考其它模型来做，基础原理是一样的。

训练完成的AttentionLSTM模型的模型参数和计算量如何计算？

可以使用飞桨接口 paddle.flops 计算flops(计算量)； paddle.nn.Layer 提供 parameters function，用于获取模型的所有参数。相关文档：https://www.paddlepaddle.org.cn/documentation/docs/zh/api/paddle/nn/Layer_cn.html#parameters-include-sublayers-true

模型使用自动压缩，参数文件大小没有变化，如何处理这种情况？

辛苦上传下日志，我们确认下具体触发了什么压缩算法？

---

压缩脚本使用的是：PaddleVideo/deploy/slim/quant_post_static.py

---

压缩配置文件内容如下：
DATASET:
batch_size: 1
batch_nums: 10
num_workers: 0
quant:
format: "FrameDataset_Sport"
data_prefix: ""
file_path: "/home/jetion/Action/apps/cpr/datasets/CPR/input_for_pptsm/val.list"
suffix: 'img_{:05}.jpg'

PIPELINE:
quant:
decode:
name: "FrameDecoder"
sample:
name: "SamplerPkl"
num_seg: 8
seg_len: 1
valid_mode: True
transform:
- Scale:
short_size: 256
- CenterCrop:
target_size: 224
- Image2Array:
- Normalization:
mean: [0.537, 0.541, 0.543]
std: [0.230, 0.221, 0.225]

inference_model_dir: "/home/jetion/Action/apps/cpr/export_models/ppTSM_server"
quant_output_dir: "/home/jetion/PaddleVideo2.2/deploy/slim/quant_cpu_models/ppTSM"

model_name: "ppTSM"
log_level: "INFO" # Optional, the logger level. default: "INFO"

---

静态离线量化日志如下：
[05-29 17:25:06 MainThread @logger.py:242] Argv: quant_post_static.py -c ./pptsm_frames_uniform_quantization.yaml --use_gpu=True
/home/jetion/anaconda3/envs/pdslim/lib/python3.8/site-packages/setuptools/sandbox.py:13: DeprecationWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html
import pkg_resources
/home/jetion/anaconda3/envs/pdslim/lib/python3.8/site-packages/pkg_resources/init.py:2871: DeprecationWarning: Deprecated call to pkg_resources.declare_namespace('mpl_toolkits').
Implementing implicit namespace packages (as specified in PEP 420) is preferred to pkg_resources.declare_namespace. See https://setuptools.pypa.io/en/latest/references/keywords.html#keyword-namespace-packages
declare_namespace(pkg)
/home/jetion/anaconda3/envs/pdslim/lib/python3.8/site-packages/pkg_resources/init.py:2871: DeprecationWarning: Deprecated call to pkg_resources.declare_namespace('google').
Implementing implicit namespace packages (as specified in PEP 420) is preferred to pkg_resources.declare_namespace. See https://setuptools.pypa.io/en/latest/references/keywords.html#keyword-namespace-packages
declare_namespace(pkg)
WARNING:root:cannot import name 'layers' from 'parl' (/home/jetion/anaconda3/envs/pdslim/lib/python3.8/site-packages/parl/init.py)
No module named 'av', [av] package and it's dependencies is required for TimeSformer and other models.
No module named 'skimage', [scikit-image] package and it's dependencies is required for ADDS.
No module named 'paddlenlp', [paddlenlp] package and it's dependencies is required for ActBERT.
No module named 'SimpleITK', [SimpleITK] package and it's dependencies is required for PP-Care.
No module named 'SimpleITK', [SimpleITK] package and it's dependencies is required for PP-Care.
No module named 'lmdb', [lmdb] package and it's dependencies is required for ActBERT.
No module named 'paddlenlp', [paddlenlp] package and it's dependencies is required for ActBERT.
No module named 'lmdb', [lmdb] package and it's dependencies is required for ActBERT.
No module named 'paddlenlp', [paddlenlp] package and it's dependencies is required for ActBERT.
[05/29 17:25:06] DATASET :
[05/29 17:25:06] batch_nums : 10
[05/29 17:25:06] batch_size : 1
[05/29 17:25:06] num_workers : 0
[05/29 17:25:06] quant :
[05/29 17:25:06] data_prefix :
[05/29 17:25:06] file_path : /home/jetion/Action/apps/cpr/datasets/CPR/input_for_pptsm/val.list
[05/29 17:25:06] format : FrameDataset_Sport
[05/29 17:25:06] suffix : img_{:05}.jpg
[05/29 17:25:06] ------------------------------------------------------------
[05/29 17:25:06] PIPELINE :
[05/29 17:25:06] quant :
[05/29 17:25:06] decode :
[05/29 17:25:06] name : FrameDecoder
[05/29 17:25:06] sample :
[05/29 17:25:06] name : SamplerPkl
[05/29 17:25:06] num_seg : 8
[05/29 17:25:06] seg_len : 1
[05/29 17:25:06] valid_mode : True
[05/29 17:25:06] transform :
[05/29 17:25:06] Scale :
[05/29 17:25:06] short_size : 256
[05/29 17:25:06] CenterCrop :
[05/29 17:25:06] target_size : 224
[05/29 17:25:06] Image2Array : None
[05/29 17:25:06] Normalization :
[05/29 17:25:06] mean : [0.537, 0.541, 0.543]
[05/29 17:25:06] std : [0.23, 0.221, 0.225]
[05/29 17:25:06] ------------------------------------------------------------
[05/29 17:25:06] inference_model_dir : /home/jetion/Action/apps/cpr/export_models/ppTSM_server
[05/29 17:25:06] log_level : INFO
[05/29 17:25:06] model_name : ppTSM
[05/29 17:25:06] quant_output_dir : /home/jetion/PaddleVideo2.2/deploy/slim/quant_cpu_models/ppTSM
inference_file_name: ppTSM
[05/29 17:25:06] Build slim_loader finished
[05/29 17:25:06] Staring Post-Training Quantization...
/home/jetion/anaconda3/envs/pdslim/lib/python3.8/site-packages/_distutils_hack/init.py:33: UserWarning: Setuptools is replacing distutils.
warnings.warn("Setuptools is replacing distutils.")
Wed May 29 17:25:06-INFO: Load model and set data loader ...
W0529 17:25:07.024253 16356 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 8.6, Driver API Version: 12.1, Runtime API Version: 11.2
W0529 17:25:07.026293 16356 gpu_resources.cc:91] device: 0, cuDNN Version: 8.2.
Wed May 29 17:25:07-INFO: Collect quantized variable names ...

Preparation stage, Run batch:| | 0/10
Preparation stage, Run batch:|████▍ | 1/10
Preparation stage, Run batch:|████████▊ | 2/10
Preparation stage, Run batch:|█████████████▏ | 3/10
Preparation stage, Run batch:|█████████████████▌ | 4/10
Preparation stage, Run batch:|██████████████████████ | 5/10
Preparation stage, Run batch:|██████████████████████████▍ | 6/10
Preparation stage, Run batch:|██████████████████████████████▊ | 7/10
Preparation stage, Run batch:|███████████████████████████████████▏ | 8/10
Preparation stage, Run batch:|███████████████████████████████████████▌ | 9/10
Preparation stage, Run batch:|███████████████████████████████████████████| 10/10
Preparation stage, Run batch:|███████████████████████████████████████████| 10/10

Sampling stage, Run batch:| | 0/10
Sampling stage, Run batch:|████▋ | 1/10
Sampling stage, Run batch:|█████████▍ | 2/10
Sampling stage, Run batch:|██████████████ | 3/10
Sampling stage, Run batch:|██████████████████▊ | 4/10
Sampling stage, Run batch:|███████████████████████▌ | 5/10
Sampling stage, Run batch:|████████████████████████████▏ | 6/10
Sampling stage, Run batch:|████████████████████████████████▉ | 7/10
Sampling stage, Run batch:|█████████████████████████████████████▌ | 8/10
Sampling stage, Run batch:|██████████████████████████████████████████▎ | 9/10
Sampling stage, Run batch:|██████████████████████████████████████████████| 10/10
Sampling stage, Run batch:|██████████████████████████████████████████████| 10/10
Wed May 29 17:28:20-INFO: Calculate KL threshold ...
Wed May 29 17:29:31-INFO: Update the program ...

Adding quant op with weight:| | 0/202
Adding quant op with weight:|█████████▎ | 44/202
Adding quant op with weight:|██████████████████▋ | 88/202
Adding quant op with weight:|███████████████████████████ | 130/202
Adding quant op with weight:|██████████████████████████████████████▉ | 187/202
Adding quant op with weight:|██████████████████████████████████████████| 202/202

Adding quant activation op:| | 0/422
Adding quant activation op:| | 1/422
Wed May 29 17:29:32-INFO: The quantized model is saved in /home/jetion/PaddleVideo2.2/deploy/slim/quant_cpu_models/ppTSM
[05/29 17:29:32] Post-Training Quantization finished...

模型压缩后，参数文件大小如下：

第一个文件是静态量化压缩前的模型参数文件，第二个是压缩后的文件。可以看到，压缩后文件大小没有变小，反而增大了。

麻烦帮忙看一下是什么问题，谢谢！