Skip to content

Latest commit

 

History

History
90 lines (66 loc) · 2.99 KB

02-tensorflow-serving.md

File metadata and controls

90 lines (66 loc) · 2.99 KB
Raw

10.2 TensorFlow Serving

To build the app we need to convert the keras model HDF5 into special format called tensorflow SavedModel. To do that, we download a prebuilt model and save it in the working directory:

wget https://github.com/DataTalksClub/machine-learning-zoomcamp/releases/download/chapter7-model/xception_v4_large_08_0.894.h5 -O clothing-model.h5

Then convert the model to SavedModel format:

import tensorflow as tf
from tensorflow import keras

model = keras.models.load_model('./clothing-model.h5')

tf.saved_model.save(model, 'clothing-model')

We can inspect what's inside the saved model using the utility (saved_model_cli) from TensorFlow and the following command:

saved_model_cli show --dir clothing-model --all

Running the command outputs a few things but we are interested in the signature, specifically the following one. For instance:

signature_def['serving_default']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['input_8'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 299, 299, 3)
        name: serving_default_input_8:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['dense_7'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 10)
        name: StatefulPartitionedCall:0
  Method name is: tensorflow/serving/predict

Alternatively one can also use the following command to output just the desired part:

saved_model_cli show --dir clothing-model --tag_set serve --signature_def serving_default

We can run the model (clothing-model) with the prebuilt docker image tensorflow/serving:2.7.0:

docker run -it --rm \
  -p 8500:8500 \
  -v $(pwd)/clothing-model:/models/clothing-model/1 \
  -e MODEL_NAME="clothing-model" \
  tensorflow/serving:2.7.0
  • docker run -it --rm (to run the docker)
  • -p 8500:8500 (port mapping)
  • -v $(pwd)/clothing-model:/models/clothing-model/1 (volumn mapping of absolute model directory to model directory inside the docker image)
  • -e MODEL_NAME="clothing-model" (set environment variable for docker image)
  • tensorflow/serving:2.7.0 (name of the image to run)

Tensorflow uses specical serving called gRPC protocol which is optimized to use binary data format. We need to convert our prediction into protobuf.

Notes

Add notes from the video (PRs are welcome)

⚠️ The notes are written by the community.
If you see an error here, please create a PR with a fix.

Navigation