def _CompressHistogram(self, histo_ev):
"""Callback for _ProcessHistogram."""
return CompressedHistogramEvent(
histo_ev.wall_time,
histo_ev.step,
compressor.compress_histogram_proto(
histo_ev.histogram_value, self._compression_bps))
def _CompressHistogram(self, histo_ev):
"""Callback for _ProcessHistogram."""
return CompressedHistogramEvent(
histo_ev.wall_time,
histo_ev.step,
compressor.compress_histogram_proto(
histo_ev.histogram_value, self._compression_bps))
def aggregate(path, run_wanted=0):
keys = [
#'Eval/Reward',
# 'Eval/Length',
# 'Eval/MCTS_Confidence',
#'Train/AvgLoss'
# 'Actor/Sample_length',
# 'Train/Samples'
'Train/Shark_Speed_H'
]
agg_keys = {k: [] for k in keys}
path = glob.glob(os.path.join(path, "events.out.tfevents.*"))
if not path:
return None, None
path = path[0]
import tensorflow as tf
i = 0
j = 0
str_ = ''
for event in my_summary_iterator(path):
val = event.summary.value
if not val:
continue
tag = event.summary.value[0].tag
for key in keys:
if tag.startswith(key):
# run = tag[tag.rfind('/')+1:]
# if int(run) != run_wanted:
# continue
x = compress_histogram_proto(event.summary.value[0].histo)
# import pdb; pdb.set_trace()
agg_keys[key].append([y.value for y in x])
if i in [1, 744, 988, 2288, 2971]:
arr = tf.make_ndarray(event.summary.value[0].tensor)
X = (arr[:, 0] + arr[:, 1]) / 2.
Y = np.vstack([X, arr[:, 2]]).T.tolist()
str_ += '\n\\addplot3 [area plot] coordinates {'
for x_ in Y:
str_ += str((x_[0], j, x_[1]))
str_ += '};\n'
j += 1
i += 1
print(str_)
import pdb
pdb.set_trace()
return agg_keys
def _read_tb_data(self):
"""Read data from a tensorboard file"""
files = self._get_tb_files()
x, y = [], []
scalar = None
# Read TB file
for file in files:
# Traverse events/summaries
for e in tf.train.summary_iterator(file):
# Traverse every value in the summary
for v in e.summary.value:
if self.tb_tag == v.tag:
if scalar is None:
if v.HasField("simple_value"):
scalar = True
elif v.HasField("histo"):
scalar = False
else:
raise ValueError(
"Only simple_value and histo data can be processed from TB file"
)
x.append(e.step)
if scalar:
y.append(v.simple_value)
else:
# Convert to compressed histogram
y.append([
chv.value
for chv in compress_histogram_proto(v.histo)
])
# Check for correct parsing
assert len(x) > 0 and len(y) > 0, "Parsing TB incorrect:"
"len(x)=%d, len(y)=%d\nx: %s\ny: %s" % (len(x), len(y), x, y)
# Sort the lists by step
x, y = (list(t) for t in zip(*sorted(zip(x, y), key=lambda t: t[0])))
return dict(x=x, y=y, i=None)
def aggregate(path, run_wanted=0):
keys = [
#'Eval/Reward',
# 'Eval/Length',
# 'Eval/MCTS_Confidence',
#'Train/AvgLoss'
'Actor/Sample_length',
'Train/Samples'
]
agg_keys = {k: [] for k in keys}
path = os.path.join(path, "event.tfevents")
for event in my_summary_iterator(path):
tag = event.summary.value[0].tag
for key in keys:
if tag.startswith(key):
run = tag[tag.rfind('/') + 1:]
if int(run) != run_wanted:
continue
x = compress_histogram_proto(event.summary.value[0].histo)
agg_keys[key].append([y.value for y in x])
return agg_keys