TDA4VMXEVM: [TDA4M] [TIDL] how to generating frozen graph from checkpoint

Nigel Peng

Part Number: TDA4VMXEVM

Hello TI,

I am using TIDL_00_09_00_00 to import the following pre-trained TensorFlow models: inceptionNetv1 ( inception_v1.ckpt ) and installed the env of tensorflow ok

but refer the comments "Refer the export_inference_graph for generating frozen graph from checkpoint. Also Refer this as example. "

I can't how to generate the frozen model,can you give me some information of this？

just like I can use optimize tool and dont't need modify any code in optimize_for_inference.py as below:

python "tensorflow\python\tools\optimize_for_inference.py" --input=mobilenet_v1_1.0_224_frozen.pb --output=mobilenet_v1_1.0_224_final.pb --input_names=input --output_names="MobilenetV1/Predictions/Softmax"

BR.

Nigel

over 5 years ago

0 kumar.desappan over 5 years ago

TI__Mastermind 22145 points

Optimization of mobilenet_v1 is different from inception_v1. The frozen graph is not supported for inception_v1, only the checkpoint file is available. You need to follow the steps mentioned in below to freeze the graph before optimizing.

"Refer the export_inference_graph for generating frozen graph from checkpoint. Also Refer this as example. "

We recommend taking help from TensorFlow community on this.

0 Nigel Peng over 5 years ago in reply to kumar.desappan

Intellectual 525 points

Hi Kumar

1、 Please provide an available command for freeze the graph for inception_v1 ,TKS.

2、 by the way,my sdk version num is psdk_rtos_auto_j7_06_01_01_12,which version of tensorflow do you recommend？my tensorflow vensor is 1.11.0

when I run the command to optimize mobilenet_v1_1.0_224.pb,there're some warning:

WARNING:tensorflow:Didn't find expected Conv2D input to 'MobilenetV2/expanded_conv/depthwise/BatchNorm/FusedBatchNorm'
WARNING:tensorflow:Didn't find expected Conv2D input to 'MobilenetV2/expanded_conv_1/depthwise/BatchNorm/FusedBatchNorm'
WARNING:tensorflow:Didn't find expected Conv2D input to 'MobilenetV2/expanded_conv_2/depthwise/BatchNorm/FusedBatchNorm'
WARNING:tensorflow:Didn't find expected Conv2D input to 'MobilenetV2/expanded_conv_3/depthwise/BatchNorm/FusedBatchNorm'
WARNING:tensorflow:Didn't find expected Conv2D input to 'MobilenetV2/expanded_conv_4/depthwise/BatchNorm/FusedBatchNorm'
WARNING:tensorflow:Didn't find expected Conv2D input to 'MobilenetV2/expanded_conv_5/depthwise/BatchNorm/FusedBatchNorm'

BR.

0 kumar.desappan over 5 years ago in reply to Nigel Peng

TI__Mastermind 22145 points

Find below for supported version

http://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/latest/exports/docs/tidl_j7_01_01_00_10/ti_dl/docs/user_guide_html/md_tidl_model_import.html#tidl_model_import_1

These warnings are OK.

Looks like this tread id duplicate of below. Avoid duplicate.

https://e2e.ti.com/support/processors/f/791/t/901553

Processors

Processors forum

TDA4VMXEVM: [TDA4M] [TIDL] how to generating frozen graph from checkpoint