Text Recognition (CRNN)

Training Parameters and Value Ranges

Table 1 describes the names, types, value ranges, default values, and descriptions of the CRNN training parameters.

**Table 1** Training parameters (**crnn_mindspore**/**model_train.py**)
Parameter	Type	Value Range	Default Value	Description
--save_best_ckpt	Bool	True or False	True	Save the best checkpoint during training and verification. (This parameter will be deleted in later versions.) Currently, the optimal CKPT is saved by default when run_eval is set to True.
--eval_start_epoch	Integer	[1,10000]	5	Sequence number of epochs from which evaluation is performed during training.
--eval_interval	Integer	[1,10000]	5	Evaluation interval. Retain the default value.
--split	Float	(0,1)	0.8	Ratio of the training dataset to the validation dataset.
--device_id	Integer	[0,7]	0	ID of the NPU used for training.
--device_num	Integer	Currently, only one device is supported, that is, the value can only be 1.	1	Number of NPUs used for training.
--epoch_size	Integer	[1,10000]	10	Number of training epochs.
--batch_size	Integer	Less than or equal to the number of images in the dataset. [1, 8192]	4	Batch size for training.
--pretrained_ckpt_path	String	-	"pre_trained_ckpt"	Path for storing the pre-trained model.
--input_width	Integer	[1, 10000)	100	Model input width.
--init_lr	Float	(0,1)	0.02	Training learning rate.
--train_dataset_path	String	-	None	Path of the training dataset.
--train_output_path	String	-	"train_output_path"	Training output path.
--run_eval	Bool	True or False	True	Whether to perform evaluation during training.

Training Command Reference

Run the following command to start the CRNN model training:

python3 model_train.py --train_dataset_path={train_dataset_path} -- train_output_path =./train_output_path  --input_width=100 --batch_size=4 --epoch_size=10 --init_lr=0.02  --run_eval=True --device_id=1 --device_num=1

The model training process is random. Use the evaluated accuracy. Figure 1 shows the training accuracy results.

Figure 1 Training accuracy results of the CRNN model

Figure 2 shows the log information after the training is complete.

Figure 2 Log information of the CRNN model training

After the training is complete, the .ckpt, .a310.om, .a310p.om, and .air model files are generated in the output directory specified by the --train_output_path parameter.

Evaluation Parameters and Value Ranges

Table 2 describes the names, types, value ranges, default values, and descriptions of the CRNN evaluation parameters.

**Table 2** Evaluation parameters (**crnn_mindspore**/**model_eval.py**)
Parameter	Type	Value Range	Default Value	Description
--eval_dataset_path	String	-	None	Path of the evaluation dataset.
--eval_ckpt_path	String	-	None	CKPT path used for evaluation.
--device_id	Integer	[0,7]	0	ID of the NPU for evaluation.
--eval_output_path	String	-	"eval_output_path"	Evaluation output path.

Evaluation Command Reference

Run the following command to evaluate the CRNN model:

python3 model_eval.py --eval_dataset_path={eval_dataset_path} --eval_ckpt_path=./train_output_path --eval_output_path=./eval_output_path --device_id=0

Use the checkpoint output during the training to evaluate the model accuracy. See Figure 3.

Figure 3 Evaluation accuracy results of the CRNN model

The files shown in Figure 4 are generated in the evaluation directory. statistics.csv collects statistics on the evaluation accuracy, statistics_ng_samples.csv collects statistics on the images with incorrect evaluation and prediction values, and statistics_ok_samples.csv collects statistics on the images with correct evaluation and prediction values.

Figure 4 Files generated during the CRNN model evaluation

Parent topic: Training Component