Hi, Lutz,

Thanks for supporting Tengine!
Actaully, this tmfile model on google drive is out of date compare to the latest version of Tengine v1.12.
The data here you found is aligned by

>     int32_t dim_0;
>     int32_t dim_1;
>     int32_t dim_2;
>     int32_t dim_3;
>     int32_t dim_size;
>     int32_t axis;`

04b6791 has changed the TM2_ReshapeParam here:

The model on google drive used the previous format. So it makes sense with 0x0001 * 0x0006 * 0x005B, which equals 1 * 6 * 91.

With the latest model:

You could use this model: https://github.com/pierricklee/tmfile-sample/blob/master/detect_tflite.tmfile

@LEMON-LM Please update this model to google drive.

@pierricklee Users will have copies of the old file and the same happens for other files. Users might use older versions of Tengine and might still open older files in Netron which is now going to crash when the old format is loaded. You can't call it the old format because it looks identical to the new format and tools can't tell the difference.

This is not acceptable. If you change the format and want things to continue working you need to update version numbers (either the version for the operator or the version number for the entire model) so tools can support both versions. Tools can't be crashing because there is no way to handle this.

@lutzroeder That is a good suggestion! We are planning add an opset version and magic number into tmfile model. This will be on Tengine roadmap as soon as possible. I will keep you updated.

@pierricklee op_ver exists on TM2_Operator already (there is also model version). It needs to be updated for breaking changes like this and all new files need to be regenerated to include the new version number. The longer you wait, the more these incompatible identical looking files will spread causing problems.

@lutzroeder That's the point. We are rewriting the op_ver part and will regenerate model files.