* added UNet for medical image segmentation * added TF-AMP support for RN50 * small updates for other models (READMEs, benchmark & testing scripts)
19 lines
848 B
Bash
19 lines
848 B
Bash
CKPT_DIR=${1:-"/results/SSD320_FP32_1GPU"}
|
|
PIPELINE_CONFIG_PATH=${2:-"/workdir/models/research/configs"}"/ssd320_bench.config"
|
|
GPUS=1
|
|
|
|
TENSOR_OPS=0
|
|
export TF_ENABLE_CUBLAS_TENSOR_OP_MATH_FP32=${TENSOR_OPS}
|
|
export TF_ENABLE_CUDNN_TENSOR_OP_MATH_FP32=${TENSOR_OPS}
|
|
export TF_ENABLE_CUDNN_RNN_TENSOR_OP_MATH_FP32=${TENSOR_OPS}
|
|
|
|
TRAIN_LOG=$(python -u ./object_detection/model_main.py \
|
|
--pipeline_config_path=${PIPELINE_CONFIG_PATH} \
|
|
--model_dir=${CKPT_DIR} \
|
|
--alsologtostder \
|
|
"${@:3}" 2>&1)
|
|
PERF=$(echo "$TRAIN_LOG" | awk -v GPUS=$GPUS '/global_step\/sec/{ array[num++]=$2 } END { for (x = 3*num/4; x < num; ++x) { sum += array[x] }; print GPUS*32*4*sum/num " img/s"}')
|
|
|
|
mkdir -p $CKPT_DIR
|
|
echo "Single GPU single precision training performance: $PERF" | tee $CKPT_DIR/train_log
|
|
echo "$TRAIN_LOG" >> $CKPT_DIR/train_log
|