add image input in checkpoint file before converting to IR model

[hung__steven]

Hi 

i would like to convert tensorflow model based on Inception V3 to IR model.

but i need to add image input placeholder because the input of model is tf records

i use the code below and it works(inception_v3.py is in the file):

import tensorflow as tf
import tensorflow.contrib.slim as slim
import tensorflow.contrib.slim.nets as nets
import inception_v3 as network


with tf.Graph().as_default():
    images_input=tf.placeholder(tf.float32,[None,299,299,3],'input')
    init_op_global=tf.global_variables_initializer()
    
    with slim.arg_scope(network.inception_v3_arg_scope()):
        net,endpoint=network.inception_v3(images_input,num_classes=26,is_training=False)
    
    saver=tf.train.Saver(tf.global_variables())
    with tf.Session() as sess:
        sess.run(init_op_global)
        saver=tf.train.import_meta_graph('model.ckpt-1399202.meta')
        saver.restore(sess,'model.ckpt-1399202')

        saver.save(sess,'/my_path/my_model.ckpt')
        tf.train.write_graph(sess.graph,'.' ,'graph.pbtxt')

however, the output of the model with image input is significantly below my expectation

how do i correctly add placeholder in my model ?

Thanks 

[Monique_J_Intel]

Hi Steven,

When you say that the output of the model is below your expectations what specifically are you speaking about(output of the model when running inference?)?

Kind Regards,

Monique Jones

[hung__steven]

Hi Jone:

yes, there are different results when running inference

i think there is something wrong when adding placeholder in my model

if i do not add image placeholder and directly convert my model to IR model, the model optimizer cannot find the input

there are the information of two models from summarize_graph below  

Original frozen model:

No inputs spotted.
No variables spotted.
Found 1 possible outputs: (name=InceptionV3/Predictions/Reshape_1, op=Reshape) 
Found 21838868 (21.84M) const parameters, 0 (0) variable parameters, and 0 control_edges
Op types used: 493 Const, 190 Identity, 95 Conv2D, 94 Relu, 94 FusedBatchNorm, 15 ConcatV2, 10 AvgPool, 4 MaxPool, 2 Add, 2 Mul, 2 Reshape, 1 Floor, 1 FIFOQueueV2, 1 QueueDequeueV2, 1 RandomUniform, 1 RealDiv, 1 BiasAdd, 1 Softmax, 1 Squeeze, 1 Sub
To use with tensorflow/tools/benchmark:benchmark_model try these arguments:
bazel run tensorflow/tools/benchmark:benchmark_model -- --graph=/volume/inception_v3_frozen.pb --show_flops --input_layer= --input_layer_type= --input_layer_shape= --output_layer=InceptionV3/Predictions/Reshape_1

with image placeholder frozen model:

Found 1 possible inputs: (name=input, type=float(1), shape=[?,299,299,3]) 
No variables spotted.
Found 1 possible outputs: (name=InceptionV3/Predictions/Reshape_1, op=Reshape) 
Found 21873291 (21.87M) const parameters, 0 (0) variable parameters, and 0 control_edges
Op types used: 488 Const, 379 Identity, 95 Conv2D, 94 FusedBatchNorm, 94 Relu, 15 ConcatV2, 10 AvgPool, 4 MaxPool, 2 Reshape, 1 BiasAdd, 1 Placeholder, 1 Shape, 1 Softmax, 1 Squeeze
To use with tensorflow/tools/benchmark:benchmark_model try these arguments:
bazel run tensorflow/tools/benchmark:benchmark_model -- --graph=/volume/theme_detection/theme_detection_frozen_2.pb --show_flops --input_layer=input --input_layer_type=float --input_layer_shape=-1,299,299,3 --output_layer=InceptionV3/Predictions/Reshape_1

how do i correctly add image placeholder in my model or directly convert my model to IR model? 

THX

Besr Regards,

Steven Hung

[Monique_J_Intel · Accepted solution]

Hi Steven,

Since Model Optimizer doesn't support -1 as a dimension for your placeholder's shapes so you have to explicitly supply the shape for example python3 mo_tf.py --input_model <frozenmodel.pb> --input_shape [1,299,299,3] --input <input/placeholder layername>. For more information on the parameters that can be used to convert the model you can reference this page.

Kind Regards,

Monique Jones