pipelines/contrib/components/openvino/tf-slim

Slim models generator

This component is automating implementation of slim models. It can create a graph from slim models zoo, load the variables pre-trained checkpoint and export the model in the form of Tensorflow frozen graph and saved model.

The results of the component can be saved in a local path or in GCS cloud storage. The can be used to other ML pipeline components like OpenVINO model optimizer, OpenVINO predict or OpenVINO Model Server.

Building

docker build --build-arg http_proxy=$http_proxy --build-arg https_proxy=$https_proxy .

Using the component

python slim_model.py --help
usage: slim_model.py [-h] [--model_name MODEL_NAME] [--export_dir EXPORT_DIR]
                     [--batch_size BATCH_SIZE]
                     [--checkpoint_url CHECKPOINT_URL]
                     [--num_classes NUM_CLASSES]

Slim model generator

optional arguments:
  -h, --help            show this help message and exit
  --model_name MODEL_NAME
  --export_dir EXPORT_DIR
                        GCS or local path to save the generated model
  --batch_size BATCH_SIZE
                        batch size to be used in the exported model
  --checkpoint_url CHECKPOINT_URL
                        URL to the pretrained compressed checkpoint
  --num_classes NUM_CLASSES
                        number of model classes

Model name can be any model defined in the slim repository. The naming convention needs to match the key name from net_factory.pyhttps://github.com/tensorflow/models/blob/master/research/slim/nets/nets_factory.py#L39)

export dir can be a local path in the container or it might be GCS path to store generated files:

• model graph file in pb format
• frozen graph including weights from the provided checkpoint
• event file which can be imported in tensorboard
• saved model which will be stored in subfolder called 1.

batch size represent the batch used in the exported models. It can be natural number to represent fixed batch size or -1 value can be set for dynamic batch size.

checkpoint_url is the URL to a pre-trained checkpoint https://github.com/tensorflow/models/tree/master/research/slim#pre-trained-models It must match the model specified in model_name parameter.

num classes should include model specific number of classes in the outputs. For slim models it should be a value of 1000 or 1001. It must match the number of classes used in the requested model name.

Examples

python slim_model.py --model_name mobilenet_v1_050 --export_dir /tmp/mobilnet 
--batch_size 1 --num_classes=1001 \
--checkpoint_url http://download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_0.5_160.tgz

python slim_model.py --model_name resnet_v1_50 --export_dir gs://<bucket_name>/resnet \
--batch_size -1 --num_classes=1000 \
--checkpoint_url http://download.tensorflow.org/models/resnet_v1_50_2016_08_28.tar.gz

python slim_model.py --model_name inception_v4 --export_dir gs://<bucket_name>/inception \
--batch_size -1 --num_classes=1001 \
--checkpoint_url http://download.tensorflow.org/models/inception_v4_2016_09_09.tar.gz

python slim_model.py --model_name vgg_19 --export_dir /tmp/vgg \
--batch_size 1 --num_classes=1000 \
--checkpoint_url http://download.tensorflow.org/models/vgg_19_2016_08_28.tar.gz