代码拉取完成，页面将自动刷新
CNN Direction Model is a model designed to perform binary classification of text images on whether the text in the image is going from left-to-right or right-to-left.
CNN Direction Model's composition consists of 1 convolutional layer and 4 residual blocks for feature extraction. The feature extraction stage is then followed by 3 dense layers to perform the classification.
For training and evaluation, we use the French Street Name Signs (FSNS) released by Google as the training data, which contains approximately 1 million training images and their corresponding ground truth words. Note that these datasets are very large.
you can find how to generate dataset in crnn_seq2seq_ocr
After installing MindSpore via the official website, you can start training and evaluation as follows:
# enter script dir, train CNNDirectionModel
bash scripts/run_standalone_train.sh [DEVICE_ID] [DATASET_PATH] [PRETRAINED_CKPT_PATH](optional)
# example: bash run_standalone_train_ascend.sh 0 /home/DataSet/FSNS/train/
├── cv
├── cnn_direction_model
├── README.md // descriptions about cnn_direction_model
├── ascend310_infer // application for 310 inference
├── requirements.txt // packages needed
├── scripts
│ ├──run_distribute_train_ascend.sh // distributed training in ascend
│ ├──run_standalone_eval_ascend.sh // evaluate in ascend
│ ├──run_distribute_train_gpu.sh // distributed training in gpu
│ ├──run_standalone_eval_gpush // evaluate in gpu
│ ├──run_eval.sh // shell script for evaluation on Ascend
│ ├──run_standalone_train_ascend.sh // train standalone in ascend
│ ├──run_standalone_train_gpu.sh // train standalone in gpu
├── src
│ ├──dataset.py // creating dataset
│ ├──cnn_direction_model.py // cnn_direction_model architecture
│ ├──create_mindrecord.py // convert raw data to mindrecords
├── model_utils
├──config.py // Parameter config
├──moxing_adapter.py // modelarts device configuration
├──device_adapter.py // Device Config
├──local_adapter.py // local device config
├── train.py // training script
├── eval.py // evaluation script
├── default_config.yaml // config file
├── postprogress.py // post process for 310 inference
├── export.py // export checkpoint files into air/mindir
Major parameters in default_config.yaml as follows:
--data_root: The path to the raw training data images for conversion to mindrecord script.
--val_data_root: The path to the raw test data images for conversion to mindrecord script.
--annotation_file: The path to the raw training annotation file.
--val_annotation_file: The path to the raw test annotation file.
--mindrecord_dir: The path to which create_mindrecord.py uses to save the resulting mindrecords for training and testing.
--epoch_size: Total training epochs.
--batch_size: Training batch size.
--im_size_h: Image height used as input to the model.
--im_size_w: Image width used as input the model.
running on Ascend
bash scripts/run_standalone_train_ascend.sh device_id path-to-train-mindrecords pre-trained-chkpt(optional)
# example: bash run_standalone_train_ascend.sh 0 /home/DataSet/FSNS/train/
The model checkpoint will be saved script/train.
Before running the command below, please check the checkpoint path used for evaluation.
running on Ascend
bash scripts/run_standalone_eval_ascend.sh device_id path-to-test-mindrecords trained-chkpt-path
# example: bash run_standalone_eval_ascend.sh /home/DataSet/FSNS/test/ /home/model/cnn_direction_model/ckpt/cnn_direction_model-10_32644.ckpt
Results of evaluation will be printed after evaluation process is completed.
For distributed training, a hccl configuration file with JSON format needs to be created in advance.
Please follow the instructions in the link below:
https://gitee.com/mindspore/models/tree/master/utils/hccl_tools.
Run scripts/run_distribute_train_ascend.sh to train the model distributed. The usage of the script is:
bash scripts/run_distribute_train_ascend.sh [rank_table] [train_dataset_path] [PRETRAINED_CKPT_PATH(optional)]
For example, you can run the shell command below to launch the training procedure.
bash run_distribute_train_ascend.sh ~/hccl_8p.json /home/DataSet/FSNS/train/
# Example of using distributed training dpn on modelarts :
# Data set storage method
# ├── FSNS # dir
# ├── train # train dir
# ├── train.zip # mindrecord train dataset zip
# ├── pre_trained # predtrained dir if exists
# ├── eval # eval dir
# ├── test.zip # mindrecord eval dataset zip
# ├── checkpoint # ckpt files dir
# (1) Choose either a (modify yaml file parameters) or b (modelArts create training job to modify parameters) 。
# a. set "enable_modelarts=True" 。
# set "run_distribute=True"
# set "save_checkpoint_path=/cache/train/outputs/"
# set "train_dataset_path=/cache/data/train/"
# set "pre_trained=/cache/data/pre_trained/pred file name" Without pre-training weights pre_trained=""
#
# b. add "enable_modelarts=True" Parameters are on the interface of modearts。
# Set the parameters required by method a on the modelarts interface
# Note: The path parameter does not need to be quoted
# (2) Set the path of the network configuration file "_config_path=/The path of config in default_config.yaml/"
# (3) Set the code path on the modelarts interface "/path/cnn_direction_model"。
# (4) Set the model's startup file on the modelarts interface "train.py" 。
# (5) Set the data path of the model on the modelarts interface ".../FSNS/train"(choices FSNS/train Folder path) ,
# The output path of the model "Output file path" and the log path of the model "Job log path" 。
# (6) start trainning the model。
# Example of using model inference on modelarts
# (1) Place the trained model to the corresponding position of the bucket。
# (2) chocie a or b。
# a. set "enable_modelarts=True" 。
# set "eval_dataset_path=/cache/data/test/"
# set "checkpoint_path=/cache/data/checkpoint/checkpoint file name"
# b. Add "enable_modelarts=True" parameter on the interface of modearts。
# Set the parameters required by method a on the modelarts interface
# Note: The path parameter does not need to be quoted
# (3) Set the path of the network configuration file "_config_path=/The path of config in default_config.yaml/"
# (4) Set the code path on the modelarts interface "/path/cnn_direction_model"。
# (5) Set the model's startup file on the modelarts interface "eval.py" 。
# (6) Set the data path of the model on the modelarts interface ".../FSNS/eval"(choices FSNS/eval Folder path) ,
# The output path of the model "Output file path" and the log path of the model "Job log path" 。
# (7) Start model inference。
python export.py --ckpt_file [CKPT_PATH] --device_target [DEVICE_TARGET] --file_format[EXPORT_FORMAT]
EXPORT_FORMAT should be in ["AIR", "MINDIR"]
Export MindIR example on ModelArts
Data storage method is the same as training
# (1) Choose either a (modify yaml file parameters) or b (modelArts create training job to modify parameters)。
# a. set "enable_modelarts=True"
# set "file_name=cnn_direction_model"
# set "file_format=MINDIR"
# set "ckpt_file=/cache/data/checkpoint file name"
# b. Add "enable_modelarts=True" parameter on the interface of modearts。
# Set the parameters required by method a on the modelarts interface
# Note: The path parameter does not need to be quoted
# (2)Set the path of the network configuration file "_config_path=/The path of config in default_config.yaml/"
# (3) Set the code path on the modelarts interface "/path/crnn"。
# (4) Set the model's startup file on the modelarts interface "export.py" 。
# (5) Set the data path of the model on the modelarts interface ".../crnn_dataset/eval/checkpoint"(choices crnn_dataset/eval/checkpoint Folder path) ,
# The output path of the model "Output file path" and the log path of the model "Job log path" 。
Before inference, please refer to MindSpore Inference with C++ Deployment Guide to set environment variables.
Before performing inference, we need to export model first. Air model can only be exported in Ascend 910 environment, mindir model can be exported in any environment.
# Ascend310 inference
bash run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [DEVICE_ID]
Inference result is saved in current path, you can find result like this in acc.log file.
top1_correct=10096, total=10202, acc=98.96%
top1_correct=8888, total=10202, acc=87.12%
| Parameters | Ascend |
|---|---|
| Resource | Ascend 910; CPU 2.60GHz, 192cores; Memory 755G; OS Euler2.8 |
| uploaded Date | 01/15/2021 (month/day/year) |
| MindSpore Version | 1.1 |
| Dataset | FSNS |
| Training Parameters | epoch=5, steps=104,477, batch_size = 40, lr=5e-4 |
| Optimizer | Adam |
| Loss Function | Softmax Cross Entropy |
| outputs | top 1 accuracy |
| Overall accuracy | 91.72% |
| Speed | 583 ms/step |
| Total time | 17 hours |
In train.py, we set some seeds before training.
Note: This model will be move to the /models/research/ directory in r1.8.
Please check the official homepage.
此处可能存在不合适展示的内容，页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容，可点击提交进行申诉，我们将尽快为您处理。
马建仓 AI 助手