multi_instance_training.sh

#!/usr/bin/env bash
#
# Copyright (c) 2023 Intel Corporation
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#

MODEL_DIR=${MODEL_DIR-$PWD}

if [ -z "${OUTPUT_DIR}" ]; then
  echo "The required environment variable OUTPUT_DIR has not been set"
  exit 1
fi

# Create the output directory in case it doesn't already exist
mkdir -p ${OUTPUT_DIR}

if [ -z "${DATASET_DIR}" ]; then
  echo "The required environment variable DATASET_DIR has not been set"
  exit 1
fi

if [ ! -d "${DATASET_DIR}" ]; then
  echo "The DATASET_DIR '${DATASET_DIR}' does not exist"
  exit 1
fi

if [ -z "${PRECISION}" ]; then
  echo "The required environment variable PRECISION has not been set"
  echo "Please set PRECISION to fp32 or bfloat16 or bfloat32 or fp16."
  exit 1
elif [ ${PRECISION} != "fp32" ] && [ ${PRECISION} != "bfloat16" ] && [ ${PRECISION} != "bfloat32" ] && [ ${PRECISION} != "fp16" ]; then
  echo "The specified precision '${PRECISION}' is unsupported."
  echo "Supported precisions are: fp32, bfloat16, bfloat32 and fp16"
  exit 1
fi

# Get number of cores per socket line from lscpu
cores_per_socket=$(lscpu |grep 'Core(s) per socket:' |sed 's/[^0-9]//g')
cores_per_socket="${cores_per_socket//[[:blank:]]/}"

CORES=`lscpu | grep Core | awk '{print $4}'`
SOCKETS=`lscpu | grep Socket | awk '{print $2}'`
NUMAS=`lscpu | grep 'NUMA node(s)' | awk '{print $3}'`
CORES_PER_NUMA=`expr $CORES \* $SOCKETS / $NUMAS`

NUM_INSTANCES=`expr $cores_per_socket / $CORES_PER_NUMA`

#Set up env variable for bfloat32
if [[ $PRECISION == "bfloat32" ]]; then
  export ONEDNN_DEFAULT_FPMATH_MODE=BF16
  PRECISION="fp32"
fi

# If batch size env is not mentioned, then the workload will run with the default batch size.
if [ -z "${BATCH_SIZE}" ]; then
  BATCH_SIZE="1024"
  echo "Running with default batch size of ${BATCH_SIZE}"
fi

source "${MODEL_DIR}/quickstart/common/utils.sh"
_ht_status_spr
_command python ${MODEL_DIR}/benchmarks/launch_benchmark.py \
  --model-name=resnet50v1_5 \
  --precision=${PRECISION} \
  --mode=training \
  --framework tensorflow \
  --checkpoint ${OUTPUT_DIR} \
  --data-location=${DATASET_DIR} \
  --output-dir ${OUTPUT_DIR} \
  --mpi_num_processes=${NUM_INSTANCES} \
  --mpi_num_processes_per_socket=${NUM_INSTANCES} \
  --batch-size ${BATCH_SIZE} \
  --num-intra-threads $CORES_PER_NUMA \
  --num-inter-threads 2 \
  $@ \
  -- \
  train_epochs=1 epochs_between_evals=1