pipelines/contrib/samples/openvino/tf-slim

OpenVINO workflow demo with TensorFlow slim models optimization and evaluation

This pipeline integrates three components:

• TensorFlow slim models generator - it creates slim models which are considered as exemplary to demonstrate OpenVINO model optimization and execution
• OpenVINO model optimizer - it converts the model from TensorFlow or other frameworks into Intermediate Representation format. It is one time operation. The result can be used in the inference execution using Inference Engine.
• Inference Engine prediction - it is running the inference evaluation based on input and label data in numpy files

The pipeline can be used to assess the results of model optimization in terms of inference execution time and classification accuracy.

During the pipeline execution there are generated several artifacts.

In the path set by tf-export-dir parameter:

• Graph definition in protobuffer format
• Frozen graph created based on imported checkpoint

In the path set by generated-model-dir parameter:

• TensorFlow model in a frozen saved_model format. It can be consumed by TensorFlow based inference scripts or TensorFlow serving. This format can't be used for additional training because all variables are changed to constants.
• Intermediate Representation of the optimized model created based on the exported TF saved model. It can be used to run inference operation using OpenVINO Inference Engine or to be imported by OpenVINO Model Server
• TensorFlow even file which can be used to visualize the graph in tensorboard

generated-model-dir value should be pointed to a shared location like google storage, persistent storage or NFS so it could be passed between components.

Parameters

• model_name = slim model name which match the allowed topologies from slim repo
• num_classes = how many classes is defined in the model. For slim models it could be 1000 or 1001 depending on the model
• checkpoint_url = URL to the compressed checkpoint for the pre-trained models -batch_size - batch size for the generated IR model
• export_dir - location for storing auxiliary model formats
• generated_model_dir - location for storing pipeline results - google storage path. If set to a numerical subfolder it could be used by TensorFlow Serving or OpenVINO Model Server
• mo_options - a list of options to be passed to model optimizer, it assumes the source model in saved_model folder will be downloaed to the 'current directory' so use --saved_model_dir . to point to the model generated in first step. Refer to mo.py --help for details.
• input_numpy_file - gs path to the numpy file including evaluation dataset in the NCHW format.
• label_numpy_file - gs path to the numpy file with labels for evaluation images. They will be used to access model accuracy.

Examples of the parameters

model-name - resnet_v1_50<br />
model-name - 1000
checkpoint-url - http://download.tensorflow.org/models/resnet_v1_50_2016_08_28.tar.gz
batch-size - 1
tf-export-dir = /tmp/export
generated-model-dir = gs://your-bucket/folder
mo_options - saved_model_dir .
input_numpy_file - gs://intelai_public_models/images/imgs.npy
label-numpy-file - gs://intelai_public_models/images/lbs.npy