统计 OpenHands 手语模型 计算量 FLOPs && 参数量 Params

Introduction

本文 对 OpenHands – Docs 统计FLOPs && Params 的相关事项 进行 补充说明

官方源码传送门: OpenHands – Github

Prepare

Dataset

以 GSLDataset 为例，下载 && 解压 GSL.zip，目录如下：

.GSL
├── GSL_pose
│   ├── health1_signer1_rep1_glosses
│   ├── health1_signer1_rep2_glosses
│   ├── health1_signer1_rep3_glosses
│   ├── health1_signer1_rep4_glosses
│   ├── health1_signer1_rep5_glosses
... ...
│   ├── police5_signer7_rep1_glosses
│   ├── police5_signer7_rep2_glosses
│   ├── police5_signer7_rep3_glosses
│   ├── police5_signer7_rep4_glosses
│   └── police5_signer7_rep5_glosses
├── GSL_pose.zip
├── GSL_split
│   ├── GSL_continuous
│   │   ├── GSL-SD-test.txt
│   │   ├── GSL-SD-train.txt
│   │   ├── GSL-SD-val.txt
│   │   ├── bbox_for_gsl_continuous.txt
│   │   ├── greek_classes.txt
│   │   ├── gsl_split_SI_dev.csv
│   │   ├── gsl_split_SI_test.csv
│   │   └── gsl_split_SI_train.csv
│   └── GSL_isolated
│       ├── bbox_for_gsl_isolated.txt
│       ├── iso_classes.csv
│       ├── test_greek_iso.csv
│       ├── train_greek_iso.csv
│       └── val_greek_iso.csv
└── GSL_split.zip

Train checkpoints

以 gsl_lstm.zip 为例，下载 && 解压，目录如下：

gsl
└── lstm
    ├── config.yaml
    └── epoch=107-step=118043.ckpt

PS：should 自行 于 lstm/config.yaml 补充 test_pipeline 参数，如下：

data:
	...	...
	
    test_pipeline:   # self-add.
        dataset:
            _target_: openhands.datasets.isolated.GSLDataset
            split_file: "GSL/GSL_split/GSL_isolated/test_greek_iso.csv"  # 解压`GSL.zip`可得
            root_dir: "GSL/GSL_pose"  	# 解压`GSL.zip`可得
            class_mappings_file_path: "GSL/GSL_split/GSL_isolated/iso_classes.csv"   # 解压`GSL.zip`可得
            splits: "test"
            modality: "pose"
            # inference_mode: true  # self-add
            inference_mode: false  # self-add

        transforms:
            - PoseSelect:
                preset: mediapipe_holistic_minimal_27
            # - PoseTemporalSubsample:
            #       num_frames: 32
            - CenterAndScaleNormalize:
                reference_points_preset: shoulder_mediapipe_holistic_minimal_27
                scale_factor: 1

FLOPs && Params

import omegaconf
from openhands.apis.inference import InferenceModel

def profile_func(model, input_dim=(1, 3, 224, 224)):  # batchsize - 1
    import torch
    from thop import profile
    input = torch.randn(input_dim)
    flops, params = profile(model, inputs=(input, ))
    print("flops: {} params: {} ".format(flops, params))
    print("%.2fG" % (flops/1e9), "%.2fM" % (params/1e6))

if __name__ == "__main__":
        
    cfg = omegaconf.OmegaConf.load("gsl/lstm/config.yaml")
    model = InferenceModel(cfg=cfg)   
	"""
    model: <class 'openhands.apis.inference.InferenceModel'>
	model.model: <class 'openhands.models.network.Network'>
	"""
	
    profile_func(model.model, input_dim=(1, 2, 100, 27))   # 输入维度可自定义

Outputs

[WARN] Cannot find rule for <class 'openhands.models.encoder.graph.pose_flattener.PoseFlattener'>. Treat it as zero Macs and zero Params.
[INFO] Register count_lstm() for <class 'torch.nn.modules.rnn.LSTM'>.
[INFO] Register count_linear() for <class 'torch.nn.modules.linear.Linear'>.
[WARN] Cannot find rule for <class 'openhands.models.decoder.utils.AttentionBlock'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'openhands.models.decoder.rnn.RNNClassifier'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'openhands.models.network.Network'>. Treat it as zero Macs and zero Params.
flops: 145004032.0 params: 1650486.0 
0.15G 1.65M

Extension

INCLUDE

源码：https://github.com/AI4Bharat/INCLUDE 修改 runner.py 文件

# ... ...
args = parser.parse_args()

def profile_func(model, input_dim=(1, 3, 224, 224)):  # batchsize - 1
    import torch
    from thop import profile
    input = torch.randn(input_dim)
    flops, params = profile(model, inputs=(input, ))
    print("flops: {} params: {} ".format(flops, params))
    print("%.2fG" % (flops/1e9), "%.2fM" % (params/1e6))

if __name__ == "__main__":

    from configs import TransformerConfig
    from models import Transformer
    config = TransformerConfig(size=args.transformer_size)
    model = Transformer(config=config, n_classes=226)
    # print(model)
    
    profile_func(model, input_dim=(3, 256, 134))  

$ python runner.py --dataset include --use_augs --model transformer --data_dir data_dir 
[INFO] Register count_linear() for <class 'torch.nn.modules.linear.Linear'>.
[WARN] Cannot find rule for <class 'torch.nn.modules.sparse.Embedding'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'torch.nn.modules.normalization.LayerNorm'>. Treat it as zero Macs and zero Params.
[INFO] Register zero_ops() for <class 'torch.nn.modules.dropout.Dropout'>.
[WARN] Cannot find rule for <class 'models.transformer.PositionEmbedding'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'transformers.models.bert.modeling_bert.BertSelfAttention'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'transformers.models.bert.modeling_bert.BertSelfOutput'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'transformers.models.bert.modeling_bert.BertAttention'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'transformers.models.bert.modeling_bert.BertIntermediate'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'transformers.models.bert.modeling_bert.BertOutput'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'transformers.models.bert.modeling_bert.BertLayer'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'torch.nn.modules.container.ModuleList'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'models.transformer.Transformer'>. Treat it as zero Macs and zero Params.
flops: 2845091328.0 params: 3771362.0 
2.85G 3.77M

写在最后：若本文章对您有帮助，请点个赞啦 ٩(๑•̀ω•́๑)۶