address review comments

dual_net.py

@@ -685,7 +685,7 @@ def freeze_graph(model_path, use_trt=False, trt_max_batch_size=8,
             max_workspace_size_bytes=1 << 29,
             precision_mode=trt_precision)
 
-    metadata = get_model_metadata({
+    metadata = make_model_metadata({
         'engine': 'tf',
         'use_trt': bool(use_trt),
     })
@@ -733,14 +733,14 @@ def freeze_graph_tpu(model_path):
     out_graph = tf.graph_util.convert_variables_to_constants(
         sess, sess.graph.as_graph_def(), output_names)
 
-    metadata = get_model_metadata({
+    metadata = make_model_metadata({
         'engine': 'tpu',
     })
 
     atomic_write_model(out_graph, metadata, model_path)
 
 
-def get_model_metadata(metadata):
+def make_model_metadata(metadata):
     for f in ['conv_width', 'fc_width', 'trunk_layers', 'use_SE', 'use_SE_bias',
               'use_swish', 'bool_features', 'input_features']:
         metadata[f] = getattr(FLAGS, f)

ml_perf/scripts/bootstrap.sh

@@ -36,7 +36,11 @@ for var_name in flag_dir golden_chunk_dir holdout_dir log_dir model_dir \
                 selfplay_dir sgf_dir work_dir; do
   dir="${!var_name}"
   if [[ "${dir}" == gs://* ]]; then
+    # `gsutil rm -f` "helpfully" returns a non-zero error code if the requested file
+    # target files don't exist.
+    set +e
     gsutil -m rm -rf "${dir}"/*
+    set -e
   else
     mkdir -p "${dir}"
     rm -rf "${dir}"/*

ml_perf/scripts/start_selfplay.sh

@@ -33,8 +33,8 @@ for device in {0..7}; do
   CUDA_VISIBLE_DEVICES="${device}" \
   ./bazel-bin/cc/concurrent_selfplay \
     --flagfile="${flag_dir}/selfplay.flags" \
-    --output_dir="${data_dir}/selfplay/\$MODEL/${device}" \
-    --holdout_dir="${data_dir}/holdout/\$MODEL/${device}" \
+    --output_dir="${selfplay_dir}/\$MODEL/${device}" \
+    --holdout_dir="${holdout_dir}/\$MODEL/${device}" \
     --model="${model_dir}/%d.pb" \
     --run_forever=1 \
     --abort_file=${abort_file} \

ml_perf/train_loop.py

@@ -61,7 +61,7 @@
                      'the training chunks as more threads are used to '
                      'compress the data. Using too many threads however could '
                      'slow down training time if each shard gets much smaller '
-                     'than around 100MB'.
+                     'than around 100MB.')
 
 flags.DEFINE_string('golden_chunk_dir', None, 'Training example directory.')
 flags.DEFINE_string('holdout_dir', None, 'Holdout example directory.')

train.py

@@ -42,8 +42,9 @@
                      'once over training data.')
 
 flags.DEFINE_integer('num_examples', None,
-                     'Total number of examples passed. Used to calculate '
-                     'steps_to_train if it isn\'t set.')
+                     'Total number of input examples. This is only used if '
+                     'steps_to_train is not set. Requires that filter_amount '
+                     'is 1.0.')
 
 flags.DEFINE_integer('window_size', 500000,
                      'Number of games to include in the window')
@@ -68,6 +69,13 @@
     lambda flags: flags['use_tpu'] if flags['use_bt'] else True,
     '`use_bt` flag only valid with `use_tpu` as well')
 
+@flags.multi_flags_validator(
+    ['num_examples', 'steps_to_train', 'filter_amount'],
+    '`num_examples` requires `steps_to_train==0` and `filter_amount==1.0`')
+def _example_flags_validator(flags_dict):
+    if not flags_dict['num_examples']:
+        return True
+    return not flags_dict['steps_to_train'] and flags_dict['filter_amount'] == 1.0
 
 @flags.multi_flags_validator(
     ['use_bt', 'cbt_project', 'cbt_instance', 'cbt_table'],