Inspection

mindrecord

The following table describes the names, types, value ranges, default values, and descriptions of evaluation parameters of the mindrecord model.

**Table 1** Parameter description (**wafer_mindspore/inspection/generate_mindrecords.py**)
Parameter	Type	Value Range	Default Value	Description
--config	String	-	None	(Optional) Parameter configuration file path, whose permission must be set to 400.
--ratio	Float	(0 ,1)	0.8	Proportion of the training dataset.
--input_chip_id	String	-	None	Processor ID.
--input_setup_id	String	-	""	Processor layer.
--minddata_train	String	-	""	Path that stores the MindRecord training dataset.
--minddata_val	String	-	""	Path that stores the MindRecord validation dataset.
--rgb_dir	String	-	rgb_ref/	Path that stores the RGB reference images.
--gray_dir	String	-	gray_ref/	Path for saving the grayscale images.
--mask_dir	String	-	mask_ref/	Path that stores mask images.
--train_flag	Bool	True or False	True	Whether to balance training dataset.
--data_txt	String	-	None	Registered .txt file.
--num_threads	Integer	[1,8]	2	Number of used processes.

Images in rgb_dir, gray_dir, and mask_dir must be named in the specified format. The following is an example:

When input_chip_id is set to 001 and input_setup_id is set to M1, the reference image in rgb_dir must be named 001_M1_rgb_golden_die.png, the grayscale image in gray_dir must be named 001_M1_gray_golden_die.png, and the mask image must be named 001_M1_viaconductor_mask.png.

data_txt must be a TXT file, which describes information about images photographed at a machine and must comply with the following rules:

Each line corresponds to an image.
Each line contains 12 fields.
Fields are separated by a space.

For details about the fields, see Table 2. The following is a compilation example.

Create a data.txt file, which contains the following three lines of character strings:

img_1.jpg 001 M1 11224.435 12925.229 0 wafer_id1 01-08-22-13:09:22; device_id1 L21510106 model1 20
img_2.jpg 001 M1 13817.899 17822.569 0 wafer_id2 12-29-21-08:17:05; device_id1 L21510104 model2 20
img_3.jpg 001 M1 970.857 24809.092 0 wafer_id3 12-21-21-06:02:08; device_id3 L21510206 model3 20

**Table 2** Field description
Index	Type	Value Range	Description
0	String	-	Image path
1	String	-	Processor ID.
2	String	-	Processor layer.
3	Float	[0, 100000]	X coordinate of an image
4	Float	[0, 100000]	Y coordinate of an image
5	Integer	0 or 2	Image label
6	String	-	Wafer ID
7	String	-	Timestamp
8	String	-	Device ID
9	String	-	Batch number
10	String	-	Device model
11	Integer	[1, 1000]	Magnification ratio

The command for generating the mindrecord files is as follows:

python3 generate_mindrecords.py --config <config_path>

<config_path>: parameter list required for generating the mindrecord files.

Figure 1 Generating the mindrecord files

Before training the inspection model, you need to convert the dataset into mindrecord files. Images cannot be directly used as the training dataset. Instead, you need to use the generated mindrecord files for training.

Figure 2 Generated mindrecord files

Training Parameters and Value Ranges

Table 3 describes the names, types, value ranges, default values, and descriptions of training parameters of the inspection model.

**Table 3** Training parameters (**wafer_mindspore/inspection/inspection_train.py**)
Parameter	Type	Value Range	Default Value	Description
--config	String	-	None	(Optional) Parameter configuration file path, whose permission must be set to 400.
--training_data	String	-	./xxx/yyy/zzz"	Path of the MindRecord training dataset. If there are multiple datasets, separate them separated with commas (,).
--validation_data	String	-	./xxx/yyy/zzz"	Path of the MindRecord validation datasets. If there are multiple datasets, separate them separated with commas (,).
--save_checkpoint_dir	String	-	'./ckpt'	Path for saving the checkpoint.
--save_fnfp	String	"yes", "no"	no	Whether to save images that are incorrectly predicted.
--fnfp_path	String	-	./saved_fnfp_images	Path for saving fn and fp images.
--recall	Float	(0 ,1)	0.99	Recall rate.
--device_target	String	"Ascend" or "GPU"	"Ascend"	Training platform.
--steps_per_eval	Integer	[1,10000]	1000	Number of evaluation intervals.
--epochs	Integer	[1,10000]	300	Number of training epochs.
--batch_size	Integer	[1, 128]	16	Batch size for training.
--ref_type	String	"rgb", "gray"	'rgb'	Type of reference images.
--ref_dir	String	-	'./xxx/yyy/zzz'	Path that stores the reference images.
--bytes_dataset	String	"yes", "no"	'yes'	Whether to stitch reference images.
--chip_setups	String	-	""	Processor ID and layer. If there are multiple IDs and layers, separate them with commas (,).
--auto_balance	String	"filter", "sampler", "yes", "no"	'filter'	Data balancing mode.
--use_val_filter	Integer	[0, 10000]	2000	Maximum number of passes used by the validation dataset.
--input_res	Integer	[256, 1024]	256	Image resolution.

Create two subdirectories rgb_ref and gray_ref in ref_dir. The files in the subdirectories must be named in the specified format. The following is an example:

When chip_setups is set to 001M1, the reference image in rgb_ref must be named 001_M1_rgb_golden_die.png, and the grayscale image in gray_ref must be named 001_M1_gray_golden_die.png.

Training Command Reference

The command for training the inspection model is as follows:

python3 inspection_train.py  --config <config_path>

<config_path> indicates the list of parameters required for inspecting model training.

The model training process is random. Use the evaluated accuracy. The training accuracy is displayed as follows:

Figure 3 Training accuracy

After the training is complete, a .ckpt model file is generated in the output directory specified by the --save_checkpoint_dir parameter. Generate an AIR model by referring to the following code, and then use the model conversion tool to convert the AIR model into an OM model for inference.

Figure 4 AIR model conversion

Parent topic: Quality Inspection on Semiconductor Wafer