PROCESSOR-SDK-J721E: edgeai-tidl-tools HOW TO generate new model artifact based on new deep learning model using run_python_examples.sh ??

Carlos Yeverino

Hello all,

I'm trying to generate NEW model artifact (FOR LISTED SUPPORT MODEL IN TIDL MODEL ZOO ITEM #8 )using the run_python_examples.sh script from edgeai-tidl-tools repo (https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/scripts/run_python_examples.sh) in order to use the generated tflite delegate. Even though I am able to generate the artifact, I observed wrong results when running the inference. Below the details.

1. I added the following in the script https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/examples/osrt_python/common_utils.py to add the model source and other parameters for the generation.

'od-tfl-ssd_mobilenet_v1_1_default_1' : {
        'model_path' : os.path.join(models_base_path,'ssd_mobilenet_v1_1_default_1.tflite'),
        'source' : {'model_url': 'tfhub.dev/.../1, 'opt': True},
        'mean': [127.5, 127.5, 127.5],
        'std' : [1/127.5, 1/127.5, 1/127.5],
        'num_images' : numImages,
        'num_classes': 91,
        'model_type': 'od',
        'session_name' : 'tflitert',
        'od_type' : 'HasDetectionPostProcLayer'
    },

Note: The model url is obtained from LISTED SUPPORT MODEL IN TIDL MODEL ZOO ITEM #8 link (tfhub) and has same parameter as the model od-tfl-ssd_mobilenet_v2_300_float already included in python examples.

Reference: TI Deep Learning Product User Guide: TIDL-RT Model Zoo

2. I also added the highlighted change in https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/examples/osrt_python/tfl/tflrt_delegate.py#L221
models = ['cl-tfl-mobilenet_v1_1.0_224', 'ss-tfl-deeplabv3_mnv2_ade20k_float', 'od-tfl-ssd_mobilenet_v2_300_float', 'od-tfl-ssd_mobilenet_v1_1_default_1']

EXPECTED RESULT (similar to py_out_od-tfl-ssd_mobilenet_v2_300_float_ADE_val_00001801.jpg):

OBTAINED RESULT (py_out_od-tfl-ssd_mobilenet_v1_1_default_1_ADE_val_00001801.jpg):

Here console output log:

run_python_examples NEW ARTIFACT console output.txt

***HOW CAN I PROPERLY GENERATE NEW MODEL ARTIFACT USING THE SCRIPTS FROM edgeai-tidl-tools REPO?

Thanks in advance!

over 3 years ago

0 Anand Pathak over 3 years ago

TI__Genius 9065 points

Hi,

Can you set "tensor_bits" : 32 and check if it generates the correct output?

One tip for running tflite specific script and required model only : You can run only the tflrt_delegate.py related scripts from run_python_examples.sh and comment out the other framework scripts execution. Also, you can set models = ['od-tfl-ssd_mobilenet_v1_1_default_1'] so script runs only the required model.

Regards,

Anand

0 Carlos Yeverino over 3 years ago in reply to Anand Pathak

Prodigy 55 points

Hello Anand,

thanks for your comment.

I tried your suggestion and set tensor_bits to 32, however, I still got wrong inference result (see image below).

Here my changes:
In https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/examples/osrt_python/common_utils.py#L33

tensor_bits = 32

I also added parameter in my model artifact configuration just to be sure
'od-tfl-ssd_mobilenet_v1_1_default_1' : {
        'model_path' : os.path.join(models_base_path,'ssd_mobilenet_v1_1_default_1.tflite'),
        'source' : {'model_url': 'tfhub.dev/.../1, 'opt': True},
        'mean': [127.5, 127.5, 127.5],
        'std' : [1/127.5, 1/127.5, 1/127.5],
        'num_images' : numImages,
        'num_classes': 91,
        'model_type': 'od',
        'session_name' : 'tflitert',
        "tensor_bits" : 32,
        'od_type' : 'HasDetectionPostProcLayer'
    },

Do you have an idea of what could be happening? Any suggestion to get expected inference result?

Thanks!

0 Carlos Yeverino over 3 years ago in reply to Anand Pathak

Prodigy 55 points

Hello Anand,

I UPDATED DESCRIPTION TO INDICATE THE EXACT TOPOLOGY I USED.

Regards.

0 Anand Pathak over 3 years ago in reply to Carlos Yeverino

TI__Genius 9065 points

Hi Carlos, we are looking into this issue. Need some more time, will keep you updated.

Regards,

Anand

0 John H. Gardner over 2 years ago in reply to Anand Pathak

TI__Guru 53201 points

Unlocking the thread for comment.

0 Fadil Muhammed over 2 years ago

TI__Prodigy 565 points

Hi Carlos,

We have identified a bug in case of pre quantized models in OSRT. We are working on this bug and plan to fix it soon after the next TIDL release (SDK 8.5 ), which will be available by Dec mid.

0 Carlos Yeverino over 2 years ago in reply to Fadil Muhammed

Prodigy 55 points

Hi Fadil,

thanks for the update!

Processors

Processors forum

PROCESSOR-SDK-J721E: edgeai-tidl-tools HOW TO generate new model artifact based on new deep learning model using run_python_examples.sh ??