mirrors
/
mmrazor
mirror of https://github.com/open-mmlab/mmrazor.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

Add get_prune_config and a demo config_pruning (#389) 

* update tools and test

* add demo

* disable test doc

* add switch for test tools and test_doc

* fix bug

* update doc

* update tools name

* mv get_channel_units

Co-authored-by: liukai <your_email@abc.example>
pull/398/head
LKJacky 2022-12-13 10:56:29 +08:00 committed by GitHub
parent c8e14e5489
commit f886821ba1
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
19 changed files with 991 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
configs/pruning/mmcls/dcff/README.md
View File

@ -47,10 +47,10 @@ The mainstream approach for filter pruning is usually either to force a hard-cod
### Generate channel_config file
Generate `resnet_cls.json` with `tools/get_channel_units.py`.
Generate `resnet_cls.json` with `tools/pruning/get_channel_units.py`.
```bash
python tools/get_channel_units.py
python tools/pruning/get_channel_units.py
configs/pruning/mmcls/dcff/dcff_resnet50_8xb32_in1k.py \
-c -i --output-path=configs/pruning/mmcls/dcff/resnet_cls.json
```

2
configs/pruning/mmcls/dcff/dcff_resnet_8xb32_in1k.py
View File

@ -10,7 +10,7 @@ stage_ratio_3 = 0.9
stage_ratio_4 = 0.7
# the config template of target_pruning_ratio can be got by
# python ./tools/get_channel_units.py {config_file} --choice
# python ./tools/pruning/get_channel_units.py {config_file} --choice
target_pruning_ratio = {
'backbone.layer1.0.conv1_(0, 64)_64': stage_ratio_1,
'backbone.layer1.0.conv2_(0, 64)_64': stage_ratio_2,

2
configs/pruning/mmcls/l1-norm/l1-norm_resnet34_8xb32_in1k.py
View File

@ -6,7 +6,7 @@ stage_ratio_3 = 0.7
stage_ratio_4 = 1.0
# the config template of target_pruning_ratio can be got by
# python ./tools/get_channel_units.py {config_file} --choice
# python ./tools/pruning/get_channel_units.py {config_file} --choice
target_pruning_ratio = {
'backbone.conv1_(0, 64)_64': stage_ratio_1,
'backbone.layer1.0.conv1_(0, 64)_64': stage_ratio_1,

4
configs/pruning/mmdet/dcff/README.md
View File

@ -47,10 +47,10 @@ The mainstream approach for filter pruning is usually either to force a hard-cod
### Generate channel_config file
Generate `resnet_det.json` with `tools/get_channel_units.py`.
Generate `resnet_det.json` with `tools/pruning/get_channel_units.py`.
```bash
python tools/get_channel_units.py
python tools/pruning/get_channel_units.py
configs/pruning/mmdet/dcff/dcff_faster_rcnn_resnet50_8xb4_coco.py \
-c -i --output-path=configs/pruning/mmcls/dcff/resnet_det.json
```

2
configs/pruning/mmdet/dcff/dcff_faster_rcnn_resnet50_8xb4_coco.py
View File

@ -11,7 +11,7 @@ stage_ratio_3 = 0.9
stage_ratio_4 = 0.7
# the config template of target_pruning_ratio can be got by
# python ./tools/get_channel_units.py {config_file} --choice
# python ./tools/pruning/get_channel_units.py {config_file} --choice
target_pruning_ratio = {
'backbone.layer1.0.conv1_(0, 64)_64': stage_ratio_1,
'backbone.layer1.0.conv2_(0, 64)_64': stage_ratio_2,

4
configs/pruning/mmpose/dcff/README.md
View File

@ -47,10 +47,10 @@ The mainstream approach for filter pruning is usually either to force a hard-cod
### Generate channel_config file
Generate `resnet_pose.json` with `tools/get_channel_units.py`.
Generate `resnet_pose.json` with `tools/pruning/get_channel_units.py`.
```bash
python tools/get_channel_units.py
python tools/pruning/get_channel_units.py
configs/pruning/mmpose/dcff/dcff_topdown_heatmap_resnet50.py \
-c -i --output-path=configs/pruning/mmpose/dcff/resnet_pose.json
```

2
configs/pruning/mmpose/dcff/dcff_topdown_heatmap_resnet50_coco.py
View File

@ -61,7 +61,7 @@ stage_ratio_3 = 0.9
stage_ratio_4 = 0.85
# the config template of target_pruning_ratio can be got by
# python ./tools/get_channel_units.py {config_file} --choice
# python ./tools/pruning/get_channel_units.py {config_file} --choice
target_pruning_ratio = {
'backbone.layer1.0.conv1_(0, 64)_64': stage_ratio_1,
'backbone.layer1.0.conv2_(0, 64)_64': stage_ratio_2,

4
configs/pruning/mmseg/dcff/README.md
View File

@ -47,10 +47,10 @@ The mainstream approach for filter pruning is usually either to force a hard-cod
### Generate channel_config file
Generate `resnet_seg.json` with `tools/get_channel_units.py`.
Generate `resnet_seg.json` with `tools/pruning/get_channel_units.py`.
```bash
python tools/get_channel_units.py
python tools/pruning/get_channel_units.py
configs/pruning/mmseg/dcff/dcff_pointrend_resnet50_8xb2_cityscapes.py \
-c -i --output-path=configs/pruning/mmseg/dcff/resnet_seg.json
```

2
configs/pruning/mmseg/dcff/dcff_pointrend_resnet50_8xb2_cityscapes.py
View File

@ -32,7 +32,7 @@ stage_ratio_3 = 0.9
stage_ratio_4 = 0.7
# the config template of target_pruning_ratio can be got by
# python ./tools/get_channel_units.py {config_file} --choice
# python ./tools/pruning/get_channel_units.py {config_file} --choice
target_pruning_ratio = {
'backbone.layer1.0.conv1_(0, 64)_64': stage_ratio_1,
'backbone.layer1.0.conv2_(0, 64)_64': stage_ratio_2,

689
demo/pruning/config_pruning.ipynb 100644
View File

@ -0,0 +1,689 @@
{
"cells": [
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [],
"source": [
"import os\n",
"\n",
"# please set cwd to the root of mmrazor repo."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# 使用MMRazor对ResNet34进行剪枝"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"本教程主要介绍如何手动配置剪枝config。"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 回顾MMCls"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"### 跨库调用resnet34配置文件\n",
"\n",
"首先我们先跨库调用resnet34的配置文件。通过跨库调用，我们可以继承原有配置文件的所有内容。"
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"outputs": [],
"source": [
"# prepare work_dir\n",
"work_dir = './demo/tmp/'\n",
"if not os.path.exists(work_dir):\n",
" os.mkdir(work_dir)"
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [],
"source": [
"from mmengine import Config\n",
"\n",
"\n",
"def write_config(config_str, filename):\n",
" with open(filename, 'w') as f:\n",
" f.write(config_str)\n"
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'type': 'ImageClassifier', 'backbone': {'type': 'ResNet', 'depth': 34, 'num_stages': 4, 'out_indices': (3,), 'style': 'pytorch'}, 'neck': {'type': 'GlobalAveragePooling'}, 'head': {'type': 'LinearClsHead', 'num_classes': 1000, 'in_channels': 512, 'loss': {'type': 'CrossEntropyLoss', 'loss_weight': 1.0}, 'topk': (1, 5)}, '_scope_': 'mmcls'}\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"/home/liukai/miniconda3/envs/lab2max/lib/python3.9/site-packages/mmengine/config/utils.py:51: UserWarning: There is not `Config` define in {'Name': 'convnext-base_3rdparty_in21k', 'Metadata': {'Training Data': 'ImageNet-21k', 'FLOPs': 15359124480, 'Parameters': 88591464}, 'In Collection': 'ConvNeXt', 'Results': None, 'Weights': 'https://download.openmmlab.com/mmclassification/v0/convnext/convnext-base_3rdparty_in21k_20220124-13b83eec.pth', 'Converted From': {'Weights': 'https://dl.fbaipublicfiles.com/convnext/convnext_base_22k_224.pth', 'Code': 'https://github.com/facebookresearch/ConvNeXt'}}\n",
" warnings.warn(f'There is not `Config` define in {model_cfg}')\n",
"/home/liukai/miniconda3/envs/lab2max/lib/python3.9/site-packages/mmengine/config/utils.py:51: UserWarning: There is not `Config` define in {'Name': 'convnext-large_3rdparty_in21k', 'Metadata': {'Training Data': 'ImageNet-21k', 'FLOPs': 34368026112, 'Parameters': 197767336}, 'In Collection': 'ConvNeXt', 'Results': None, 'Weights': 'https://download.openmmlab.com/mmclassification/v0/convnext/convnext-large_3rdparty_in21k_20220124-41b5a79f.pth', 'Converted From': {'Weights': 'https://dl.fbaipublicfiles.com/convnext/convnext_large_22k_224.pth', 'Code': 'https://github.com/facebookresearch/ConvNeXt'}}\n",
" warnings.warn(f'There is not `Config` define in {model_cfg}')\n",
"/home/liukai/miniconda3/envs/lab2max/lib/python3.9/site-packages/mmengine/config/utils.py:51: UserWarning: There is not `Config` define in {'Name': 'convnext-xlarge_3rdparty_in21k', 'Metadata': {'Training Data': 'ImageNet-21k', 'FLOPs': 60929820672, 'Parameters': 350196968}, 'In Collection': 'ConvNeXt', 'Results': None, 'Weights': 'https://download.openmmlab.com/mmclassification/v0/convnext/convnext-xlarge_3rdparty_in21k_20220124-f909bad7.pth', 'Converted From': {'Weights': 'https://dl.fbaipublicfiles.com/convnext/convnext_xlarge_22k_224.pth', 'Code': 'https://github.com/facebookresearch/ConvNeXt'}}\n",
" warnings.warn(f'There is not `Config` define in {model_cfg}')\n",
"/home/liukai/miniconda3/envs/lab2max/lib/python3.9/site-packages/mmengine/config/utils.py:51: UserWarning: There is not `Config` define in {'Name': 'swinv2-base-w12_3rdparty_in21k-192px', 'Metadata': {'Training Data': 'ImageNet-21k', 'FLOPs': 8510000000, 'Parameters': 87920000}, 'In Collection': 'Swin-Transformer V2', 'Results': None, 'Weights': 'https://download.openmmlab.com/mmclassification/v0/swin-v2/pretrain/swinv2-base-w12_3rdparty_in21k-192px_20220803-f7dc9763.pth', 'Converted From': {'Weights': 'https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_base_patch4_window12_192_22k.pth', 'Code': 'https://github.com/microsoft/Swin-Transformer'}}\n",
" warnings.warn(f'There is not `Config` define in {model_cfg}')\n",
"/home/liukai/miniconda3/envs/lab2max/lib/python3.9/site-packages/mmengine/config/utils.py:51: UserWarning: There is not `Config` define in {'Name': 'swinv2-large-w12_3rdparty_in21k-192px', 'Metadata': {'Training Data': 'ImageNet-21k', 'FLOPs': 19040000000, 'Parameters': 196740000}, 'In Collection': 'Swin-Transformer V2', 'Results': None, 'Weights': 'https://download.openmmlab.com/mmclassification/v0/swin-v2/pretrain/swinv2-large-w12_3rdparty_in21k-192px_20220803-d9073fee.pth', 'Converted From': {'Weights': 'https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_large_patch4_window12_192_22k.pth', 'Code': 'https://github.com/microsoft/Swin-Transformer'}}\n",
" warnings.warn(f'There is not `Config` define in {model_cfg}')\n"
]
}
],
"source": [
"# Prepare pretrain config\n",
"pretrain_config_path = f'{work_dir}/pretrain.py'\n",
"config_string = \"\"\"\n",
"_base_ = ['mmcls::resnet/resnet34_8xb32_in1k.py']\n",
"\"\"\"\n",
"write_config(config_string, pretrain_config_path)\n",
"print(Config.fromfile(pretrain_config_path)['model'])"
]
},
{
"cell_type": "code",
"execution_count": 5,
"metadata": {},
"outputs": [],
"source": [
"# Run config\n",
"! timeout 2 python ./tools/train.py $prune_config_path"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 准备剪枝config"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"1. 增加pretrained参数\n",
"2. 将resnet34模型装入剪枝算法wrapper中\n",
"3. 配置剪枝比例\n",
"4. 运行"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"### 1. 增加预训练参数\n",
"我们将原有的’model‘字段取出，命名为architecture，并且给archtecture增加init_cfg字段用来加载预训练模型参数。"
]
},
{
"cell_type": "code",
"execution_count": 6,
"metadata": {},
"outputs": [],
"source": [
"checkpoint_path = 'https://download.openmmlab.com/mmclassification/v0/resnet/resnet34_8xb32_in1k_20210831-f257d4e6.pth'\n",
"prune_config_path = work_dir + 'prune.py'\n",
"config_string += \"\"\"\\n\n",
"data_preprocessor = {'type': 'mmcls.ClsDataPreprocessor'}\n",
"architecture = _base_.model\n",
"architecture.update({\n",
" 'init_cfg': {\n",
" 'type':\n",
" 'Pretrained',\n",
" 'checkpoint':\n",
" 'https://download.openmmlab.com/mmclassification/v0/resnet/resnet34_8xb32_in1k_20210831-f257d4e6.pth' # noqa\n",
" }\n",
"})\n",
"\"\"\"\n",
"write_config(config_string, prune_config_path)"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"### 2. 将resnet34模型装入剪枝算法wrapper中\n",
"\n",
"我们将原有的model作为architecture放入到ItePruneAlgorithm算法中，并且将ItePruneAlgorithm作为新的model字段。"
]
},
{
"cell_type": "code",
"execution_count": 7,
"metadata": {},
"outputs": [],
"source": [
"config_string += \"\"\"\n",
"target_pruning_ratio={}\n",
"model = dict(\n",
" _delete_=True,\n",
" _scope_='mmrazor',\n",
" type='ItePruneAlgorithm',\n",
" architecture=architecture,\n",
" mutator_cfg=dict(\n",
" type='ChannelMutator',\n",
" channel_unit_cfg=dict(\n",
" type='L1MutableChannelUnit',\n",
" default_args=dict(choice_mode='ratio'))),\n",
" target_pruning_ratio=target_pruning_ratio,\n",
" step_freq=1,\n",
" prune_times=1,\n",
")\n",
"\"\"\"\n",
"write_config(config_string, prune_config_path)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"配置到这一步时，我们的config文件已经能够运行了。但是因为我们没有配置target_pruning_ratio，因此现在跑起来就和直接用原有config跑起来没有区别，接下来我们会介绍如何配置剪枝比例"
]
},
{
"cell_type": "code",
"execution_count": 8,
"metadata": {},
"outputs": [],
"source": [
"! timeout 2 python ./tools/train.py $prune_config_path"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### 4. 配置剪枝比例"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"我们的模型使用tracer解析模型，进而获得剪枝节点，为了方便用户配置剪枝节点比例，我们提供了一个获得剪枝节点剪枝比例配置的工具。通过该工具，我们可以方便地对剪枝比例进行配置。"
]
},
{
"cell_type": "code",
"execution_count": 9,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{\n",
" \"backbone.conv1_(0, 64)_64\":1.0,\n",
" \"backbone.layer1.0.conv1_(0, 64)_64\":1.0,\n",
" \"backbone.layer1.1.conv1_(0, 64)_64\":1.0,\n",
" \"backbone.layer1.2.conv1_(0, 64)_64\":1.0,\n",
" \"backbone.layer2.0.conv1_(0, 128)_128\":1.0,\n",
" \"backbone.layer2.0.conv2_(0, 128)_128\":1.0,\n",
" \"backbone.layer2.1.conv1_(0, 128)_128\":1.0,\n",
" \"backbone.layer2.2.conv1_(0, 128)_128\":1.0,\n",
" \"backbone.layer2.3.conv1_(0, 128)_128\":1.0,\n",
" \"backbone.layer3.0.conv1_(0, 256)_256\":1.0,\n",
" \"backbone.layer3.0.conv2_(0, 256)_256\":1.0,\n",
" \"backbone.layer3.1.conv1_(0, 256)_256\":1.0,\n",
" \"backbone.layer3.2.conv1_(0, 256)_256\":1.0,\n",
" \"backbone.layer3.3.conv1_(0, 256)_256\":1.0,\n",
" \"backbone.layer3.4.conv1_(0, 256)_256\":1.0,\n",
" \"backbone.layer3.5.conv1_(0, 256)_256\":1.0,\n",
" \"backbone.layer4.0.conv1_(0, 512)_512\":1.0,\n",
" \"backbone.layer4.0.conv2_(0, 512)_512\":1.0,\n",
" \"backbone.layer4.1.conv1_(0, 512)_512\":1.0,\n",
" \"backbone.layer4.2.conv1_(0, 512)_512\":1.0\n",
"}"
]
}
],
"source": [
"ratio_template_path=work_dir+'prune_ratio_template.json'\n",
"! python ./tools/pruning/get_channel_units.py $pretrain_config_path --choice -o $ratio_template_path &> /dev/null 2>&1\n",
"! cat $ratio_template_path\n",
"! rm $ratio_template_path"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"我们修改该配置模板如下，并且将替换到我们的剪枝配置文件中。\n",
"\n",
"（该配置来源于：Li, Hao, et al. \"Pruning filters for efficient convnets.\" arXiv preprint arXiv:1608.08710 (2016).）"
]
},
{
"cell_type": "code",
"execution_count": 10,
"metadata": {},
"outputs": [],
"source": [
"target_config = \"\"\"\n",
"un_prune = 1.0\n",
"stage_ratio_1 = 0.5\n",
"stage_ratio_2 = 0.4\n",
"stage_ratio_3 = 0.6\n",
"stage_ratio_4 = un_prune\n",
"\n",
"target_pruning_ratio = {\n",
" # stage 1\n",
" 'backbone.conv1_(0, 64)_64': un_prune, # short cut layers\n",
" 'backbone.layer1.0.conv1_(0, 64)_64': stage_ratio_1,\n",
" 'backbone.layer1.1.conv1_(0, 64)_64': stage_ratio_1,\n",
" 'backbone.layer1.2.conv1_(0, 64)_64': un_prune,\n",
" # stage 2\n",
" 'backbone.layer2.0.conv1_(0, 128)_128': un_prune,\n",
" 'backbone.layer2.0.conv2_(0, 128)_128': un_prune, # short cut layers\n",
" 'backbone.layer2.1.conv1_(0, 128)_128': stage_ratio_2,\n",
" 'backbone.layer2.2.conv1_(0, 128)_128': stage_ratio_2,\n",
" 'backbone.layer2.3.conv1_(0, 128)_128': un_prune,\n",
" # stage 3\n",
" 'backbone.layer3.0.conv1_(0, 256)_256': un_prune,\n",
" 'backbone.layer3.0.conv2_(0, 256)_256': un_prune, # short cut layers\n",
" 'backbone.layer3.1.conv1_(0, 256)_256': stage_ratio_3,\n",
" 'backbone.layer3.2.conv1_(0, 256)_256': stage_ratio_3,\n",
" 'backbone.layer3.3.conv1_(0, 256)_256': stage_ratio_3,\n",
" 'backbone.layer3.4.conv1_(0, 256)_256': stage_ratio_3,\n",
" 'backbone.layer3.5.conv1_(0, 256)_256': un_prune,\n",
" # stage 4\n",
" 'backbone.layer4.0.conv1_(0, 512)_512': stage_ratio_4,\n",
" 'backbone.layer4.0.conv2_(0, 512)_512': un_prune, # short cut layers\n",
" 'backbone.layer4.1.conv1_(0, 512)_512': stage_ratio_4,\n",
" 'backbone.layer4.2.conv1_(0, 512)_512': stage_ratio_4\n",
"}\n",
"\"\"\""
]
},
{
"cell_type": "code",
"execution_count": 11,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"_base_ = ['mmcls::resnet/resnet34_8xb32_in1k.py']\n",
"\n",
"\n",
"data_preprocessor = {'type': 'mmcls.ClsDataPreprocessor'}\n",
"architecture = _base_.model\n",
"architecture.update({\n",
" 'init_cfg': {\n",
" 'type':\n",
" 'Pretrained',\n",
" 'checkpoint':\n",
" 'https://download.openmmlab.com/mmclassification/v0/resnet/resnet34_8xb32_in1k_20210831-f257d4e6.pth' # noqa\n",
" }\n",
"})\n",
"\n",
"\n",
"un_prune = 1.0\n",
"stage_ratio_1 = 0.5\n",
"stage_ratio_2 = 0.4\n",
"stage_ratio_3 = 0.6\n",
"stage_ratio_4 = un_prune\n",
"\n",
"target_pruning_ratio = {\n",
" # stage 1\n",
" 'backbone.conv1_(0, 64)_64': un_prune, # short cut layers\n",
" 'backbone.layer1.0.conv1_(0, 64)_64': stage_ratio_1,\n",
" 'backbone.layer1.1.conv1_(0, 64)_64': stage_ratio_1,\n",
" 'backbone.layer1.2.conv1_(0, 64)_64': un_prune,\n",
" # stage 2\n",
" 'backbone.layer2.0.conv1_(0, 128)_128': un_prune,\n",
" 'backbone.layer2.0.conv2_(0, 128)_128': un_prune, # short cut layers\n",
" 'backbone.layer2.1.conv1_(0, 128)_128': stage_ratio_2,\n",
" 'backbone.layer2.2.conv1_(0, 128)_128': stage_ratio_2,\n",
" 'backbone.layer2.3.conv1_(0, 128)_128': un_prune,\n",
" # stage 3\n",
" 'backbone.layer3.0.conv1_(0, 256)_256': un_prune,\n",
" 'backbone.layer3.0.conv2_(0, 256)_256': un_prune, # short cut layers\n",
" 'backbone.layer3.1.conv1_(0, 256)_256': stage_ratio_3,\n",
" 'backbone.layer3.2.conv1_(0, 256)_256': stage_ratio_3,\n",
" 'backbone.layer3.3.conv1_(0, 256)_256': stage_ratio_3,\n",
" 'backbone.layer3.4.conv1_(0, 256)_256': stage_ratio_3,\n",
" 'backbone.layer3.5.conv1_(0, 256)_256': un_prune,\n",
" # stage 4\n",
" 'backbone.layer4.0.conv1_(0, 512)_512': stage_ratio_4,\n",
" 'backbone.layer4.0.conv2_(0, 512)_512': un_prune, # short cut layers\n",
" 'backbone.layer4.1.conv1_(0, 512)_512': stage_ratio_4,\n",
" 'backbone.layer4.2.conv1_(0, 512)_512': stage_ratio_4\n",
"}\n",
"\n",
"model = dict(\n",
" _delete_=True,\n",
" _scope_='mmrazor',\n",
" type='ItePruneAlgorithm',\n",
" architecture=architecture,\n",
" mutator_cfg=dict(\n",
" type='ChannelMutator',\n",
" channel_unit_cfg=dict(\n",
" type='L1MutableChannelUnit',\n",
" default_args=dict(choice_mode='ratio'))),\n",
" target_pruning_ratio=target_pruning_ratio,\n",
" step_freq=1,\n",
" prune_times=1,\n",
")\n"
]
}
],
"source": [
"config_string=config_string.replace('target_pruning_ratio={}',target_config)\n",
"write_config(config_string,prune_config_path)\n",
"! cat $prune_config_path"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### 5. 运行"
]
},
{
"cell_type": "code",
"execution_count": 12,
"metadata": {},
"outputs": [],
"source": [
"! timeout 2 python ./tools/train.py $prune_config_path"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"# 自动生成剪枝Config"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"我们提供了一键生成剪枝config的工具get_prune_config.py"
]
},
{
"cell_type": "code",
"execution_count": 13,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"usage: get_l1_prune_config.py [-h] [--checkpoint CHECKPOINT] [--subnet SUBNET]\n",
" [-o O]\n",
" config\n",
"\n",
"Get the config to prune a model.\n",
"\n",
"positional arguments:\n",
" config config of the model\n",
"\n",
"optional arguments:\n",
" -h, --help show this help message and exit\n",
" --checkpoint CHECKPOINT\n",
" checkpoint path of the model\n",
" --subnet SUBNET pruning structure for the model\n",
" -o O output path to store the pruning config.\n"
]
}
],
"source": [
"! python ./tools/pruning/get_l1_prune_config.py -h"
]
},
{
"cell_type": "code",
"execution_count": 14,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"model = dict(\n",
" _scope_='mmrazor',\n",
" type='ItePruneAlgorithm',\n",
" architecture=dict(\n",
" type='ImageClassifier',\n",
" backbone=dict(\n",
" type='ResNet',\n",
" depth=34,\n",
" num_stages=4,\n",
" out_indices=(3, ),\n",
" style='pytorch'),\n",
" neck=dict(type='GlobalAveragePooling'),\n",
" head=dict(\n",
" type='LinearClsHead',\n",
" num_classes=1000,\n",
" in_channels=512,\n",
" loss=dict(type='CrossEntropyLoss', loss_weight=1.0),\n",
" topk=(1, 5)),\n",
" _scope_='mmcls',\n",
" init_cfg=dict(\n",
" type='Pretrained',\n",
" checkpoint=\n",
" 'https://download.openmmlab.com/mmclassification/v0/resnet/resnet34_8xb32_in1k_20210831-f257d4e6.pth'\n",
" ),\n",
" data_preprocessor=dict(\n",
" mean=[123.675, 116.28, 103.53],\n",
" std=[58.395, 57.12, 57.375],\n",
" to_rgb=True)),\n",
" target_pruning_ratio=dict({\n",
" 'backbone.conv1_(0, 64)_64': 1.0,\n",
" 'backbone.layer1.0.conv1_(0, 64)_64': 1.0,\n",
" 'backbone.layer1.1.conv1_(0, 64)_64': 1.0,\n",
" 'backbone.layer1.2.conv1_(0, 64)_64': 1.0,\n",
" 'backbone.layer2.0.conv1_(0, 128)_128': 1.0,\n",
" 'backbone.layer2.0.conv2_(0, 128)_128': 1.0,\n",
" 'backbone.layer2.1.conv1_(0, 128)_128': 1.0,\n",
" 'backbone.layer2.2.conv1_(0, 128)_128': 1.0,\n",
" 'backbone.layer2.3.conv1_(0, 128)_128': 1.0,\n",
" 'backbone.layer3.0.conv1_(0, 256)_256': 1.0,\n",
" 'backbone.layer3.0.conv2_(0, 256)_256': 1.0,\n",
" 'backbone.layer3.1.conv1_(0, 256)_256': 1.0,\n",
" 'backbone.layer3.2.conv1_(0, 256)_256': 1.0,\n",
" 'backbone.layer3.3.conv1_(0, 256)_256': 1.0,\n",
" 'backbone.layer3.4.conv1_(0, 256)_256': 1.0,\n",
" 'backbone.layer3.5.conv1_(0, 256)_256': 1.0,\n",
" 'backbone.layer4.0.conv1_(0, 512)_512': 1.0,\n",
" 'backbone.layer4.0.conv2_(0, 512)_512': 1.0,\n",
" 'backbone.layer4.1.conv1_(0, 512)_512': 1.0,\n",
" 'backbone.layer4.2.conv1_(0, 512)_512': 1.0\n",
" }),\n",
" mutator_cfg=dict(\n",
" type='ChannelMutator',\n",
" channel_unit_cfg=dict(\n",
" type='L1MutableChannelUnit',\n",
" default_args=dict(choice_mode='ratio')),\n",
" parse_cfg=dict(\n",
" type='ChannelAnalyzer',\n",
" tracer_type='FxTracer',\n",
" demo_input=dict(type='DefaultDemoInput', scope='mmcls'))))\n",
"dataset_type = 'ImageNet'\n",
"data_preprocessor = None\n",
"train_pipeline = [\n",
" dict(type='LoadImageFromFile', _scope_='mmcls'),\n",
" dict(type='RandomResizedCrop', scale=224, _scope_='mmcls'),\n",
" dict(type='RandomFlip', prob=0.5, direction='horizontal', _scope_='mmcls'),\n",
" dict(type='PackClsInputs', _scope_='mmcls')\n",
"]\n",
"test_pipeline = [\n",
" dict(type='LoadImageFromFile', _scope_='mmcls'),\n",
" dict(type='ResizeEdge', scale=256, edge='short', _scope_='mmcls'),\n",
" dict(type='CenterCrop', crop_size=224, _scope_='mmcls'),\n",
" dict(type='PackClsInputs', _scope_='mmcls')\n",
"]\n",
"train_dataloader = dict(\n",
" batch_size=32,\n",
" num_workers=5,\n",
" dataset=dict(\n",
" type='ImageNet',\n",
" data_root='data/imagenet',\n",
" ann_file='meta/train.txt',\n",
" data_prefix='train',\n",
" pipeline=[\n",
" dict(type='LoadImageFromFile'),\n",
" dict(type='RandomResizedCrop', scale=224),\n",
" dict(type='RandomFlip', prob=0.5, direction='horizontal'),\n",
" dict(type='PackClsInputs')\n",
" ],\n",
" _scope_='mmcls'),\n",
" sampler=dict(type='DefaultSampler', shuffle=True, _scope_='mmcls'),\n",
" persistent_workers=True)\n",
"val_dataloader = dict(\n",
" batch_size=32,\n",
" num_workers=5,\n",
" dataset=dict(\n",
" type='ImageNet',\n",
" data_root='data/imagenet',\n",
" ann_file='meta/val.txt',\n",
" data_prefix='val',\n",
" pipeline=[\n",
" dict(type='LoadImageFromFile'),\n",
" dict(type='ResizeEdge', scale=256, edge='short'),\n",
" dict(type='CenterCrop', crop_size=224),\n",
" dict(type='PackClsInputs')\n",
" ],\n",
" _scope_='mmcls'),\n",
" sampler=dict(type='DefaultSampler', shuffle=False, _scope_='mmcls'),\n",
" persistent_workers=True)\n",
"val_evaluator = dict(type='Accuracy', topk=(1, 5), _scope_='mmcls')\n",
"test_dataloader = dict(\n",
" batch_size=32,\n",
" num_workers=5,\n",
" dataset=dict(\n",
" type='ImageNet',\n",
" data_root='data/imagenet',\n",
" ann_file='meta/val.txt',\n",
" data_prefix='val',\n",
" pipeline=[\n",
" dict(type='LoadImageFromFile'),\n",
" dict(type='ResizeEdge', scale=256, edge='short'),\n",
" dict(type='CenterCrop', crop_size=224),\n",
" dict(type='PackClsInputs')\n",
" ],\n",
" _scope_='mmcls'),\n",
" sampler=dict(type='DefaultSampler', shuffle=False, _scope_='mmcls'),\n",
" persistent_workers=True)\n",
"test_evaluator = dict(type='Accuracy', topk=(1, 5), _scope_='mmcls')\n",
"optim_wrapper = dict(\n",
" optimizer=dict(\n",
" type='SGD', lr=0.1, momentum=0.9, weight_decay=0.0001,\n",
" _scope_='mmcls'))\n",
"param_scheduler = dict(\n",
" type='MultiStepLR',\n",
" by_epoch=True,\n",
" milestones=[30, 60, 90],\n",
" gamma=0.1,\n",
" _scope_='mmcls')\n",
"train_cfg = dict(by_epoch=True, max_epochs=100, val_interval=1)\n",
"val_cfg = dict()\n",
"test_cfg = dict()\n",
"auto_scale_lr = dict(base_batch_size=256)\n",
"default_scope = 'mmcls'\n",
"default_hooks = dict(\n",
" timer=dict(type='IterTimerHook', _scope_='mmcls'),\n",
" logger=dict(type='LoggerHook', interval=100, _scope_='mmcls'),\n",
" param_scheduler=dict(type='ParamSchedulerHook', _scope_='mmcls'),\n",
" checkpoint=dict(type='CheckpointHook', interval=1, _scope_='mmcls'),\n",
" sampler_seed=dict(type='DistSamplerSeedHook', _scope_='mmcls'),\n",
" visualization=dict(\n",
" type='VisualizationHook', enable=False, _scope_='mmcls'))\n",
"env_cfg = dict(\n",
" cudnn_benchmark=False,\n",
" mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),\n",
" dist_cfg=dict(backend='nccl'))\n",
"vis_backends = [dict(type='LocalVisBackend', _scope_='mmcls')]\n",
"visualizer = dict(\n",
" type='ClsVisualizer',\n",
" vis_backends=[dict(type='LocalVisBackend')],\n",
" _scope_='mmcls')\n",
"log_level = 'INFO'\n",
"load_from = None\n",
"resume = False\n"
]
}
],
"source": [
"! python ./tools/pruning/get_l1_prune_config.py $work_dir/pretrain.py --checkpoint $checkpoint_path -o $prune_config_path &> /dev/null\n",
"! cat $prune_config_path"
]
},
{
"cell_type": "code",
"execution_count": 15,
"metadata": {},
"outputs": [],
"source": [
"# 清理临时文件\n",
"! rm -r $work_dir"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3.9.13 ('lab2max')",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.9.13"
},
"orig_nbformat": 4,
"vscode": {
"interpreter": {
"hash": "e31a827d0913016ad78e01c7b97f787f4b9e53102dd62d238e8548bcd97ff875"
}
}
},
"nbformat": 4,
"nbformat_minor": 2
}

2
docs/en/advanced_guides/tutorials/how_to_prune_your_model.md
View File

@ -81,7 +81,7 @@ model = dict(
**Specific arguments**:
A algorithm may have its specific arguments. You need to read their documents to know how to config. Here, we only introduce the specific arguments of ItePruneAlgorithm.
- target_pruning_ratio: target_pruning_ratio is a dict that uses the name of units as keys and the choice values as values.. It indicates how many channels remain after pruning. You can use python ./tools/get_channel_units.py --choice {config_file} to get the choice template. Please refer to [How to Use our Config Tool for Pruning](./how_to_use_config_tool_of_pruning.md).
- target_pruning_ratio: target_pruning_ratio is a dict that uses the name of units as keys and the choice values as values.. It indicates how many channels remain after pruning. You can use python ./tools/pruning/get_channel_units.py --choice {config_file} to get the choice template. Please refer to [How to Use our Config Tool for Pruning](./how_to_use_config_tool_of_pruning.md).
- step_epoch: the step between two pruning operations.
- prune_times: the times to prune to reach the pruning target. Here, we prune resnet34 once, so we set it to 1.

10
docs/en/advanced_guides/tutorials/how_to_use_config_tool_of_pruning.md
View File

@ -121,9 +121,9 @@ mutator2.prepare_from_supernet(resnet34())
To make your development more fluent, we provide a command tool to parse a model and return the config template.
```shell
$ python ./tools/get_channel_units.py -h
$ python ./tools/pruning/get_channel_units.py -h
usage: get_channel_units.py [-h] [-c] [-i] [--choice] [-o OUTPUT_PATH] config
usage: pruning/get_channel_units.py [-h] [-c] [-i] [--choice] [-o OUTPUT_PATH] config
Get channel unit of a model.
@ -142,7 +142,7 @@ optional arguments:
Take the algorithm Slimmable Network as an example.
```shell
python ./tools/get_channel_units.py ./configs/pruning/mmcls/autoslim/autoslim_mbv2_1.5x_slimmable_subnet_8xb256_in1k.py
python ./tools/pruning/get_channel_units.py ./configs/pruning/mmcls/autoslim/autoslim_mbv2_1.5x_slimmable_subnet_8xb256_in1k.py
# {
# "type":"SlimmableChannelMutator",
@ -171,7 +171,7 @@ python ./tools/get_channel_units.py ./configs/pruning/mmcls/autoslim/autoslim_mb
The '-i' flag will return the config with the initialization arguments.
```shell
python ./tools/get_channel_units.py -i ./configs/pruning/mmcls/autoslim/autoslim_mbv2_1.5x_slimmable_subnet_8xb256_in1k.py
python ./tools/pruning/get_channel_units.py -i ./configs/pruning/mmcls/autoslim/autoslim_mbv2_1.5x_slimmable_subnet_8xb256_in1k.py
# {
# "type":"SlimmableChannelMutator",
@ -207,7 +207,7 @@ python ./tools/get_channel_units.py -i ./configs/pruning/mmcls/autoslim/autoslim
With "--choice" flag, it will return the choice template, a dict which uses unit_name as key, and use the choice value as value.
```shell
python ./tools/get_channel_units.py -i ./configs/pruning/mmcls/autoslim/autoslim_mbv2_1.5x_slimmable_subnet_8xb256_in1k.py --choice
python ./tools/pruning/get_channel_units.py -i ./configs/pruning/mmcls/autoslim/autoslim_mbv2_1.5x_slimmable_subnet_8xb256_in1k.py --choice
# {
# "backbone.conv1.conv_(0, 48)_48":32,

2
docs/en/user_guides/pruning_user_guide.md
View File

@ -146,3 +146,5 @@ Please refer to the following documents for more details.
- [MutableChannel](../../../mmrazor/models/mutables/mutable_channel/MutableChannel.md)
- [ChannelMutator](../../../mmrazor/models/mutables/mutable_channel/units/mutable_channel_unit.ipynb)
- [MutableChannelUnit](../../../mmrazor/models/mutators/channel_mutator/channel_mutator.ipynb)
- Demos
- [Config pruning](../../../demo/config_pruning.ipynb)

2
requirements/tests.txt
View File

@ -2,6 +2,8 @@ codecov
flake8
interrogate
isort==4.3.21
nbconvert
nbformat
pytest
xdoctest >= 0.10.0
yapf

32
tests/test_doc.py 100644
View File

@ -0,0 +1,32 @@
# Copyright (c) OpenMMLab. All rights reserved.
import os
from unittest import TestCase
import nbformat
from nbconvert.preprocessors import ExecutePreprocessor
TEST_DOC = os.getenv('TEST_DOC') == 'true'
notebook_paths = [
'./mmrazor/models/mutators/channel_mutator/channel_mutator.ipynb',
'./mmrazor/models/mutables/mutable_channel/units/mutable_channel_unit.ipynb', # noqa
'./demo/config_pruning.ipynb'
]
class TestDocs(TestCase):
def setUp(self) -> None:
if not TEST_DOC:
self.skipTest('disabled')
def test_notebooks(self):
for path in notebook_paths:
with self.subTest(path=path):
with open(path) as file:
nb_in = nbformat.read(file, nbformat.NO_CONVERT)
ep = ExecutePreprocessor(
timeout=600, kernel_name='python3')
try:
_ = ep.preprocess(nb_in)
except Exception:
self.fail()

1
tests/test_tools/__init__.py 100644
View File

@ -0,0 +1 @@
# Copyright (c) OpenMMLab. All rights reserved.

101
tests/test_tools/test_tools.py 100644
View File

@ -0,0 +1,101 @@
# Copyright (c) OpenMMLab. All rights reserved.
import os
import shutil
import subprocess
from unittest import TestCase
import torch
from mmrazor import digit_version
TEST_TOOLS = os.getenv('TEST_TOOLS') == 'true'
class TestTools(TestCase):
_config_path = None
def setUp(self) -> None:
if not TEST_TOOLS:
self.skipTest('disabled')
@property
def config_path(self):
if self._config_path is None:
self._config_path = self._get_config_path()
return self._config_path
def _setUp(self) -> None:
self.workdir = os.path.dirname(__file__) + '/tmp/'
if not os.path.exists(self.workdir):
os.mkdir(self.workdir)
def save_to_config(self, name, content):
with open(self.workdir + f'/{name}', 'w') as f:
f.write(content)
def test_get_channel_unit(self):
if digit_version(torch.__version__) < digit_version('1.12.0'):
self.skipTest('version of torch < 1.12.0')
for path in self.config_path:
with self.subTest(path=path):
self._setUp()
self.save_to_config('pretrain.py', f"""_base_=['{path}']""")
try:
subprocess.run([
'python', './tools/pruning/get_channel_units.py',
f'{self.workdir}/pretrain.py', '-o',
f'{self.workdir}/unit.json'
])
except Exception as e:
self.fail(f'{e}')
self.assertTrue(os.path.exists(f'{self.workdir}/unit.json'))
self._tearDown()
def test_get_prune_config(self):
if digit_version(torch.__version__) < digit_version('1.12.0'):
self.skipTest('version of torch < 1.12.0')
for path in self.config_path:
with self.subTest(path=path):
self._setUp()
self.save_to_config('pretrain.py', f"""_base_=['{path}']""")
try:
subprocess.run([
'python',
'./tools/pruning/get_l1_prune_config.py',
f'{self.workdir}/pretrain.py',
'-o',
f'{self.workdir}/prune.py',
])
pass
except Exception as e:
self.fail(f'{e}')
self.assertTrue(os.path.exists(f'{self.workdir}/prune.py'))
self._tearDown()
def _tearDown(self) -> None:
print('delete')
shutil.rmtree(self.workdir)
pass
def _get_config_path(self):
config_paths = []
paths = [
('mmcls', 'mmcls::resnet/resnet34_8xb32_in1k.py'),
('mmdet', 'mmdet::retinanet/retinanet_r18_fpn_1x_coco.py'),
(
'mmseg',
'mmseg::deeplabv3plus/deeplabv3plus_r50-d8_4xb4-20k_voc12aug-512x512.py' # noqa
),
('mmyolo',
'mmyolo::yolov5/yolov5_m-p6-v62_syncbn_fast_8xb16-300e_coco.py')
]
for repo_name, path in paths:
try:
__import__(repo_name)
config_paths.append(path)
except Exception:
pass
return config_paths

19
tools/get_channel_units.py → tools/pruning/get_channel_units.py
View File

@ -1,6 +1,7 @@
# Copyright (c) OpenMMLab. All rights reserved.
import argparse
import json
import sys
import torch.nn as nn
from mmengine import MODELS
@ -9,6 +10,8 @@ from mmengine.config import Config
from mmrazor.models import BaseAlgorithm
from mmrazor.models.mutators import ChannelMutator
sys.setrecursionlimit(int(pow(2, 20)))
def parse_args():
parser = argparse.ArgumentParser(
@ -40,11 +43,25 @@ def parse_args():
def main():
args = parse_args()
config = Config.fromfile(args.config)
default_scope = config['default_scope']
model = MODELS.build(config['model'])
if isinstance(model, BaseAlgorithm):
mutator = model.mutator
elif isinstance(model, nn.Module):
mutator = ChannelMutator()
mutator: ChannelMutator = ChannelMutator(
channel_unit_cfg=dict(
type='L1MutableChannelUnit',
default_args=dict(choice_mode='ratio'),
),
parse_cfg={
'type': 'ChannelAnalyzer',
'demo_input': {
'type': 'DefaultDemoInput',
'scope': default_scope
},
'tracer_type': 'FxTracer'
})
mutator.prepare_from_supernet(model)
if args.choice:
config = mutator.choice_template

127
tools/pruning/get_l1_prune_config.py 100644
View File

@ -0,0 +1,127 @@
# Copyright (c) OpenMMLab. All rights reserved.
import argparse
import copy
from typing import Dict
from mmengine import Config, fileio
from mmrazor.models.mutators import ChannelMutator
from mmrazor.registry import MODELS
def parse_args():
parser = argparse.ArgumentParser(
description='Get the config to prune a model.')
parser.add_argument('config', help='config of the model')
parser.add_argument(
'--checkpoint',
default=None,
type=str,
help='checkpoint path of the model')
parser.add_argument(
'--subnet',
default=None,
type=str,
help='pruning structure for the model')
parser.add_argument(
'-o',
type=str,
default='./prune.py',
help='output path to store the pruning config.')
args = parser.parse_args()
return args
def wrap_prune_config(config: Config, prune_target: Dict,
checkpoint_path: str):
config = copy.deepcopy(config)
default_scope = config['default_scope']
arch_config: Dict = config['model']
# update checkpoint_path
if checkpoint_path is not None:
arch_config.update({
'init_cfg': {
'type': 'Pretrained',
'checkpoint': checkpoint_path # noqa
},
})
# deal with data_preprocessor
if 'data_preprocessor' in config:
data_preprocessor = config['data_preprocessor']
arch_config.update({'data_preprocessor': data_preprocessor})
config['data_preprocessor'] = None
else:
data_preprocessor = None
# prepare algorithm
algorithm_config = dict(
_scope_='mmrazor',
type='ItePruneAlgorithm',
architecture=arch_config,
target_pruning_ratio=prune_target,
mutator_cfg=dict(
type='ChannelMutator',
channel_unit_cfg=dict(
type='L1MutableChannelUnit',
default_args=dict(choice_mode='ratio')),
parse_cfg=dict(
type='ChannelAnalyzer',
tracer_type='FxTracer',
demo_input=dict(type='DefaultDemoInput',
scope=default_scope))))
config['model'] = algorithm_config
return config
def change_config(config):
scope = config['default_scope']
config['model']['_scope_'] = scope
return config
if __name__ == '__main__':
args = parse_args()
config_path = args.config
checkpoint_path = args.checkpoint
target_path = args.o
origin_config = Config.fromfile(config_path)
origin_config = change_config(origin_config)
default_scope = origin_config['default_scope']
# get subnet config
model = MODELS.build(copy.deepcopy(origin_config['model']))
mutator: ChannelMutator = ChannelMutator(
channel_unit_cfg=dict(
type='L1MutableChannelUnit',
default_args=dict(choice_mode='ratio'),
),
parse_cfg={
'type': 'ChannelAnalyzer',
'demo_input': {
'type': 'DefaultDemoInput',
'scope': default_scope
},
'tracer_type': 'FxTracer'
})
mutator.prepare_from_supernet(model)
if args.subnet is None:
choice_template = mutator.choice_template
else:
input_choices = fileio.load(args.subnet)
try:
mutator.set_choices(input_choices)
choice_template = input_choices
except Exception as e:
print(f'error when apply input subnet: {e}')
choice_template = mutator.choice_template
# prune and finetune
prune_config: Config = wrap_prune_config(origin_config, choice_template,
checkpoint_path)
prune_config.dump(target_path)