improve demo time

image_classification/run_demo.sh

@@ -38,48 +38,43 @@ fi
 
 export COMPLIANCE_FILE=$(pwd)/mlperf_compliance.log
 
+sed -i "s/NUM_TRAIN_FILES = 1024/NUM_TRAIN_FILES = 1/g" tensorflow2/imagenet_preprocessing.py
+sed -i "s/        os.path.join(data_dir, 'train-%05d-of-01024' % i)/        os.path.join(data_dir, 'train-%05d-of-00001' % i)/g" tensorflow2/imagenet_preprocessing.py
+
 # run benchmark
 echo "running benchmark"
 
 # run training
+TF_XLA_FLAGS='--tf_xla_auto_jit=2' \
 python3 tensorflow2/resnet_ctl_imagenet_main.py \
   --distribution_strategy="one_device" \
   --base_learning_rate=8.5 \
-  --batch_size=150 \
+  --batch_size=200 \
   --data_dir=${DATASET_DIR} \
-  --datasets_num_private_threads=1 \
-  --dtype=fp32 \
-  --device_warmup_steps=1 \
+  --datasets_num_private_threads=4 \
+  --dtype=fp16 \
   --noenable_device_warmup \
   --enable_eager \
-  --noenable_xla \
-  --epochs_between_evals=1 \
+  --enable_xla \
   --noeval_dataset_cache \
-  --eval_offset_epochs=1 \
-  --eval_prefetch_batchs=8 \
   --label_smoothing=0.1 \
-  --lars_epsilon=0 \
   --log_steps=125 \
-  --lr_schedule=polynomial \
   --model_dir=${OUTPUT_MODEL_DIR} \
-  --momentum=0.9 \
-  --num_accumulation_steps=1 \
+  --num_accumulation_steps=32 \
   --num_classes=1000 \
   --num_gpus=1 \
-  --optimizer=LARS \
+  --optimizer=SGD \
   --noreport_accuracy_metrics \
   --single_l2_loss_op \
   --skip_eval \
-  --steps_per_loop=1 \
-  --target_accuracy=0.759 \
+  --steps_per_loop=20 \
   --notf_data_experimental_slack \
   --tf_gpu_thread_mode=gpu_private \
   --notrace_warmup \
   --train_epochs=1 \
   --notraining_dataset_cache \
-  --training_prefetch_batchs=8 \
+  --training_prefetch_batchs=128 \
   --nouse_synthetic_data \
-  --warmup_epochs=1 \
   --weight_decay=0.0002 |& tee "$LOG_DIR/train_console.log"
 
 # Copy log file to MLCube log folder
@@ -102,4 +97,4 @@ echo "ENDING TIMING RUN AT $end_fmt"
 result=$(($end - $start))
 result_name="IMAGE_CLASSIFICATION"
 
-echo "RESULT,$result_name,,$result,nvidia,$start_fmt"
+echo "RESULT,$result_name,,$result,nvidia,$start_fmt"