Freezing Models

Preparation

Download the checkpoint files to ckpt folder here. (Use the MobilenetV1 224 pretrained checkpoint from here)

Export the inference graph:

python export_inference_graph.py \
    --alsologtostderr --model_name=mobilenet_v1 \
    --image_size=224 --output_file=mobilenet.pb

Freeze the graph:

python freeze_graph.py --input_graph=mobilenet.pb \
                       --input_checkpoint=ckpt/mobilenet_v1_1.0_224.ckpt \
                       --output_graph=mobilenet_f.pb --input_binary=true \
                       --output_node_names=MobilenetV1/Predictions/Reshape_1

Edit pb_to_pbtxt.py to use the name of your PB file as graph. Make pbtxt:

python pb_to_pbtxt.py

Serving

See api.py for example. Change input and output tensor names accordingly.

Credits

Got the export and freezing scripts from tensorflow. Basically moving the utilities out of the repo to allow them to be used without bazel.

Models from tensorflow/models repo.

Datasets, pretrained checkpoints, etc. from tensorflow/models/research/slim.

Working as of 3/4/2018.