def stream_tfevents(path, file_api, run, step=0, namespace=""):
"""Parses and streams a tfevents file to the server"""
last_step = 0
row = {}
buffer = []
last_row = {}
global_step_key = namespaced_tag("global_step", namespace)
try:
for summary in tf.train.summary_iterator(path):
parsed = tf_summary_to_dict(summary, namespace=namespace)
if last_step != parsed[global_step_key]:
last_step = parsed[global_step_key]
if len(row) > 3: # Must have more than _timestamp, _step, and global_step
step += 1
row["_step"] = step
last_row = history_dict_to_json(run, deepcopy(row))
file_api.push("wandb-history.jsonl", util.json_dumps_safer_history(last_row))
row = {}
row.update(parsed)
except tf.errors.DataLossError:
wandb.termwarn("Found a truncated record in tfevents file, stopping parse")
step += 1
row["_step"] = step
last_row = history_dict_to_json(run, deepcopy(row))
file_api.push("wandb-history.jsonl", util.json_dumps_safer_history(last_row))
return last_row
def publish_history(self, data, step=None, run=None):
run = run or self._run
data = data_types.history_dict_to_json(run, data, step=step)
history = wandb_internal_pb2.HistoryRecord()
for k, v in six.iteritems(data):
item = history.item.add()
item.key = k
item.value_json = json_dumps_safer_history(v)
self._publish_history(history)
def send_history(self, data):
rec = wandb_internal_pb2.Record()
data = data_types.history_dict_to_json(self._run, data)
history = rec.history
for k, v in six.iteritems(data):
item = history.item.add()
item.key = k
item.value_json = json_dumps_safer_history(v)
self._queue_process(rec)
def publish_history(self, data, step=None, run=None, publish_step=True):
run = run or self._run
data = data_types.history_dict_to_json(run, data, step=step)
history = pb.HistoryRecord()
if publish_step:
assert step is not None
history.step.num = step
data.pop("_step", None)
for k, v in six.iteritems(data):
item = history.item.add()
item.key = k
item.value_json = json_dumps_safer_history(v) # type: ignore
self._publish_history(history)
def stream_tfevents(path, file_api, run, step=0, namespace=""):
"""Parses and streams a tfevents file to the server"""
last_step = 0
row = {}
buffer = []
last_row = {}
global_step_key = namespaced_tag("global_step", namespace)
for summary in tf.train.summary_iterator(path):
parsed = tf_summary_to_dict(summary, namespace=namespace)
if last_step != parsed[global_step_key]:
step += 1
row["_step"] = step
last_step = parsed[global_step_key]
# TODO: handle time
if len(row) > 0:
last_row = history_dict_to_json(run, row)
file_api.push("wandb-history.jsonl",
util.json_dumps_safer_history(last_row))
row.update(parsed)
# TODO: It's not clear to me why we still have wandb.data_types in last_row here,
# but we do so we convert again
return history_dict_to_json(run, last_row)
def _transform(self):
"""Transforms special classes into the proper format before writing"""
self.row = data_types.history_dict_to_json(self._run, self.row)
