Updated on 2023-10-27 GMT+08:00

Training a TensorFlow Model

After Kubeflow is deployed, it is easy to use the ps-worker mode to train TensorFlow models. This section describes an official TensorFlow training example provided by Kubeflow. For details, see TensorFlow Training (TFJob).

Running the Mnist Example

  1. Deploy the TFJob resource to start training.

    Create the tf-mnist.yaml file. The following is an example:
    apiVersion: "kubeflow.org/v1"
    kind: TFJob
    metadata:
      name: tfjob-simple
      namespace: kubeflow
    spec:
      tfReplicaSpecs:
        Worker:
          replicas: 2
          restartPolicy: OnFailure
          template:
            spec:
              containers:
                - name: tensorflow
                  image: kubeflow/tf-mnist-with-summaries:latest
                  command:
                    - "python"
                    - "/var/tf_mnist/mnist_with_summaries.py"

  2. Create the TFJob.

    kubectl apply -f tf-mnist.yaml

  3. View the logs after the worker running is complete.

    kubectl -n kubeflow logs tfjob-simple-worker-0

    Information similar to the following is displayed:

    ...
    Accuracy at step 900: 0.964
    Accuracy at step 910: 0.9653
    Accuracy at step 920: 0.9665
    Accuracy at step 930: 0.9681
    Accuracy at step 940: 0.9664
    Accuracy at step 950: 0.9667
    Accuracy at step 960: 0.9694
    Accuracy at step 970: 0.9683
    Accuracy at step 980: 0.9687
    Accuracy at step 990: 0.966
    Adding run metadata for 999

  4. Delete the TFJob.

    kubectl delete -f tf-mnist.yaml

Using a GPU

The training can be performed in the GPU scenario. In this scenario, the cluster must contain GPU nodes and proper drivers must be installed.

  1. Specify the GPU resources in the TFJob.

    Create the tf-gpu.yaml file. The following is an example:

    This example runs in the TensorFlow distributed architecture. The ResNet50 model in the convolutional neural network (CNN) is used to train randomly generated images. A total of 32 (batch_size) images are trained each time, and the images are trained 100 times in total. Additionally, the performance (image/sec) of each training is recorded.

    apiVersion: "kubeflow.org/v1"
    kind: "TFJob"
    metadata:
      name: "tf-smoke-gpu"
    spec:
      tfReplicaSpecs:
        PS:
          replicas: 1
          template:
            metadata:
              creationTimestamp: null
            spec:
              containers:
                - args:
                    - python
                    - tf_cnn_benchmarks.py
                    - --batch_size=32
                    - --model=resnet50
                    - --variable_update=parameter_server
                    - --flush_stdout=true
                    - --num_gpus=1
                    - --local_parameter_device=cpu
                    - --device=cpu
                    - --data_format=NHWC
                  image: docker.io/kubeflow/tf-benchmarks-cpu:v20171202-bdab599-dirty-284af3
                  name: tensorflow
                  ports:
                    - containerPort: 2222
                      name: tfjob-port
                  resources:
                    limits:
                      cpu: "1"
                  workingDir: /opt/tf-benchmarks/scripts/tf_cnn_benchmarks
              restartPolicy: OnFailure
        Worker:
          replicas: 1
          template:
            metadata:
              creationTimestamp: null
            spec:
              containers:
                - args:
                    - python
                    - tf_cnn_benchmarks.py
                    - --batch_size=32
                    - --model=resnet50
                    - --variable_update=parameter_server
                    - --flush_stdout=true
                    - --num_gpus=1
                    - --local_parameter_device=cpu
                    - --device=gpu
                    - --data_format=NHWC
                  image: docker.io/kubeflow/tf-benchmarks-gpu:v20171202-bdab599-dirty-284af3
                  name: tensorflow
                  ports:
                    - containerPort: 2222
                      name: tfjob-port
                  resources:
                    limits:
                      nvidia.com/gpu: 1       # Number of GPUs
                  workingDir: /opt/tf-benchmarks/scripts/tf_cnn_benchmarks
              restartPolicy: OnFailure

  2. Create the TFJob.

    kubectl apply -f tf-gpu.yaml

  3. After the worker runs the job (about 5 minutes if a GPU is used), run the following command to view the result.

    kubectl logs tf-smoke-gpu-worker-0

    Information similar to the following is displayed:

    ...
    INFO|2023-09-02T12:04:25|/opt/launcher.py|27| Running warm up
    INFO|2023-09-02T12:08:55|/opt/launcher.py|27| Done warm up
    INFO|2023-09-02T12:08:55|/opt/launcher.py|27| Step      Img/sec loss
    INFO|2023-09-02T12:08:56|/opt/launcher.py|27| 1 images/sec: 68.8 +/- 0.0 (jitter = 0.0) 8.777
    INFO|2023-09-02T12:09:00|/opt/launcher.py|27| 10        images/sec: 70.4 +/- 0.4 (jitter = 1.8) 8.557
    INFO|2023-09-02T12:09:04|/opt/launcher.py|27| 20        images/sec: 70.5 +/- 0.3 (jitter = 1.5) 8.090
    INFO|2023-09-02T12:09:09|/opt/launcher.py|27| 30        images/sec: 70.3 +/- 0.3 (jitter = 1.6) 8.041
    INFO|2023-09-02T12:09:13|/opt/launcher.py|27| 40        images/sec: 70.1 +/- 0.2 (jitter = 1.7) 9.464
    INFO|2023-09-02T12:09:18|/opt/launcher.py|27| 50        images/sec: 70.1 +/- 0.2 (jitter = 1.6) 7.797
    INFO|2023-09-02T12:09:23|/opt/launcher.py|27| 60        images/sec: 70.1 +/- 0.2 (jitter = 1.6) 8.595
    INFO|2023-09-02T12:09:27|/opt/launcher.py|27| 70        images/sec: 70.0 +/- 0.2 (jitter = 1.7) 7.853
    INFO|2023-09-02T12:09:32|/opt/launcher.py|27| 80        images/sec: 69.9 +/- 0.2 (jitter = 1.7) 7.849
    INFO|2023-09-02T12:09:36|/opt/launcher.py|27| 90        images/sec: 69.8 +/- 0.2 (jitter = 1.7) 7.911
    INFO|2023-09-02T12:09:41|/opt/launcher.py|27| 100       images/sec: 69.7 +/- 0.1 (jitter = 1.7) 7.853
    INFO|2023-09-02T12:09:41|/opt/launcher.py|27| ----------------------------------------------------------------
    INFO|2023-09-02T12:09:41|/opt/launcher.py|27| total images/sec: 69.68
    INFO|2023-09-02T12:09:41|/opt/launcher.py|27| ----------------------------------------------------------------
    INFO|2023-09-02T12:09:42|/opt/launcher.py|80| Finished: python tf_cnn_benchmarks.py --batch_size=32 --model=resnet50 --variable_update=parameter_server --flush_stdout=true --num_gpus=1 --local_parameter_device=cpu --device=gpu --data_format=NHWC --job_name=worker --ps_hosts=tf-smoke-gpu-ps-0.default.svc:2222 --worker_hosts=tf-smoke-gpu-worker-0.default.svc:2222 --task_index=0
    INFO|2023-09-02T12:09:42|/opt/launcher.py|84| Command ran successfully sleep for ever.

    The training performance of a single GPU is 69.68 images per second.