Advanced.md

ResNet50 v1.5 FP32 training - Advanced Instructions

This document has advanced instructions for running ResNet50 v1.5 FP32 training, which provides more control over the individual parameters that are used. For more information on using /benchmarks/launch_benchmark.py, see the launch benchmark documentation.

Prior to using these instructions, please follow the setup instructions from the model's README and/or the AI Kit documentation to get your environment setup (if running on bare metal) and download the dataset, pretrained model, etc. If you are using AI Kit, please exclude the --docker-image flag from the commands below, since you will be running the the TensorFlow conda environment instead of docker.

Any of the launch_benchmark.py commands below can be run on bare metal by removing the --docker-image arg. Ensure that you have all of the required prerequisites installed in your environment before running without the docker container.

If you are new to docker and are running into issues with the container, see this document for troubleshooting tips.

Once your environment is setup, navigate to the benchmarks directory of the model zoo and set environment variables pointing to the directory for the dataset, number of MPI processes, and an output directory where log files will be written.

# cd to the benchmarks directory in the model zoo
cd benchmarks

export DATASET_DIR=<path to the dataset>
export OUTPUT_DIR=<directory where log files will be written>
export MPI_NUM_PROCESSES=<desired number of MPI processes (optional)>

ResNet50 v1.5 FP32 training can be run to test full training, single-epoch training, or a training demo. Use one of the following examples below, depending on your use case.

• For full training to convergence run the following command that uses the DATASET_DIR:

python launch_benchmark.py \
  --model-name=resnet50v1_5 \
  --precision=fp32 \
  --mode=training \
  --framework tensorflow \
  --checkpoint ${OUTPUT_DIR} \
  --data-location=${DATASET_DIR} \
  --output-dir ${OUTPUT_DIR} \
  --noinstall \
  --docker-image intel/intel-optimized-tensorflow:latest

• For a single epoch of training, use the command below that uses the DATASET_DIR, a batch size of 256, and a -- steps=100 argument:

python launch_benchmark.py \
  --model-name=resnet50v1_5 \
  --precision=fp32 \
  --mode=training \
  --framework tensorflow \
  --batch-size 256 \
  --checkpoint ${OUTPUT_DIR} \
  --data-location=${DATASET_DIR} \
  --output-dir ${OUTPUT_DIR} \
  --noinstall \
  --docker-image intel/intel-optimized-tensorflow:latest \
  -- steps=100 train_epochs=1 epochs_between_evals=1

• For an even shorter training demo (only 50 steps), use the command below that uses the DATASET_DIR, a batch size of 16, and a -- steps=50 argument:

python launch_benchmark.py \
  --model-name=resnet50v1_5 \
  --precision=fp32 \
  --mode=training \
  --framework tensorflow \
  --batch-size 16 \
  --checkpoint ${OUTPUT_DIR} \
  --data-location=${DATASET_DIR} \
  --output-dir ${OUTPUT_DIR} \
  --noinstall \
  --docker-image intel/intel-optimized-tensorflow:latest \
  -- steps=50 train_epochs=1 epochs_between_evals=1

If you run the script for more than 100 steps, you should see training loss decreasing like below:

I0816 basic_session_run_hooks.py:262] loss = 8.442491, step = 0
I0816 basic_session_run_hooks.py:260] loss = 8.373407, step = 100 (... sec)
...

Distributed Training

Training can be done in a distributed fashion. On a dual (or eight) socket system, one can create two (or eight) MPI processes (one socket each) to do the training. As an example, run the following command to start ResNet50v1.5 FP32 training run using 2 MPI processes.

python launch_benchmark.py \
  --model-name resnet50v1_5 \
  --precision fp32 \
  --mode training \
  --framework tensorflow \
  --data-location ${DATASET_DIR} \
  --output-dir ${OUTPUT_DIR} \
  --mpi_num_processes 2 \
  --docker-image intel/intel-optimized-tensorflow:latest

The above distributed training runs one MPI process per socket, to maximize performance. Users can run more than one (commonly two) MPI processes per socket. The following command achieves launching 4 MPI processes over 2 sockets. Note that in this case we need to reduce the OMP_NUM_THREADS and intra_op_parallelism_threads by half (minus one or two for performance sometimes, e.g. half of 28 becomes 14, and we can use 12 for good performance). This is controlled by "-a ". Batch size can remain the same for weak scaling or reduced by half as well for strong scaling.

python launch_benchmark.py \
  --model-name resnet50v1_5 \
  --precision fp32 \
  --mode training \
  --framework tensorflow \
  --data-location ${DATASET_DIR} \
  --output-dir ${OUTPUT_DIR} \
  --mpi_num_processes 4 \
  --mpi_num_processes_per_socket 2 \
  --docker-image intel/intel-optimized-tensorflow:latest \
  -a <half the amount of cores per socket or less>

Similarly, the following command achieves launching 2 MPI processes over 1 socket.

python launch_benchmark.py \
  --model-name resnet50v1_5 \
  --precision fp32 \
  --mode training \
  --framework tensorflow \
  --data-location ${DATASET_DIR} \
  --output-dir ${OUTPUT_DIR} \
  --mpi_num_processes 2 \
  --mpi_num_processes_per_socket 1 \
  --docker-image intel/intel-optimized-tensorflow:latest \
  -a <half the amount of cores per socket or less>

You can check output trained model accuracy by setting --eval=True in the command. After training is over, it automatically run inference and report accuracy results.

Finally, the following command runs MPI across multiple nodes on bare-metal, with 2 MPI processes per node. Each node must have passwordless ssh enabled for the user running the command below. All hosts should have these additional packages installed: (apt-get) openmpi-bin openmpi-common libopenmpi-dev, (pip) horovod==0.20.0

python launch_benchmark.py \
  --verbose \
  --model-name resnet50v1_5 \
  --precision fp32 \
  --mode training \
  --framework tensorflow \
  --noinstall \
  --checkpoint ${OUTPUT_DIR} \
  --data-location ${DATASET_DIR} \
  --mpi_hostnames 'host1,host2' \
  --mpi_num_processes 4 2>&1