run_debug.sh

XLA_FLAGS="--xla_gpu_cuda_data_dir=/public/software/compiler/rocm/dtk-21.10.1/amdgcn/bitcode/ --xla_dump_hlo_pass_re=.* --xla_dump_hlo_as_html --xla_dump_to=./tmp" TF_DUMP_GRAPH_PREFIX="./tf_graph" hipprof --hip-trace python3 ./resnet_ctl_imagenet_main.py \
--base_learning_rate=10.0 \
--batch_size=32 \
--nocache_decoded_image \
--data_dir=/public/software/apps/DeepLearning/Data/ImageNet-tensorflow \
--device_warmup_steps=1 \
--dtype=fp32 \
--noenable_checkpoint_and_export \
--noenable_device_warmup \
--enable_eager \
--epochs_between_evals=4 \
--noeval_dataset_cache \
--eval_offset_epochs=2 \
--label_smoothing=0.1 \
--lars_epsilon=0 \
--log_steps=125 \
--lr_schedule=polynomial \
--optimizer=LARS \
--noreport_accuracy_metrics \
--single_l2_loss_op \
--steps_per_loop=25 \
--train_epochs=1 \
--notraining_dataset_cache \
--notrace_warmup \
--nouse_synthetic_data \
--use_tf_function \
--verbosity=0 \
--warmup_epochs=5 \
--weight_decay=0.0002 \
--target_accuracy=0.759 \
--momentum=0.9 \
--num_replicas=64 \
--num_accumulation_steps=2 \
--num_classes=1000 \
--noskip_eval