def make_session_end_summary(status: str, end_time_secs: Optional[int] = None):
"""
Args:
status: outcome of this run, one of of 'UNKNOWN', 'SUCCESS', 'FAILURE', 'RUNNING'
end_time_secs: optional ending time in seconds
Returns:
"""
status = Status.DESCRIPTOR.values_by_name[
f"STATUS_{status.upper()}"].number
if end_time_secs is None:
import time
end_time_secs = int(time.time())
session_end_summary = SessionEndInfo(status=status,
end_time_secs=end_time_secs)
session_end_content = HParamsPluginData(
session_end_info=session_end_summary, version=PLUGIN_DATA_VERSION)
session_end_summary_metadata = SummaryMetadata(
plugin_data=SummaryMetadata.PluginData(
plugin_name=PLUGIN_NAME,
content=session_end_content.SerializeToString()))
session_end_summary = Summary(value=[
Summary.Value(tag=SESSION_END_INFO_TAG,
metadata=session_end_summary_metadata)
])
return session_end_summary
def make_session_start_summary(
hparam_values,
group_name: Optional[str] = None,
start_time_secs: Optional[int] = None,
):
"""Assign values to the hyperparameters in the context of this session.
Args:
hparam_values: a dict of ``hp_name`` -> ``hp_value`` mappings
group_name: optional group name for this session
start_time_secs: optional starting time in seconds
Returns:
"""
if start_time_secs is None:
import time
start_time_secs = int(time.time())
session_start_info = SessionStartInfo(group_name=group_name,
start_time_secs=start_time_secs)
for hp_name, hp_value in hparam_values.items():
# Logging a None would raise an exception when setting session_start_info.hparams[hp_name].number_value = None.
# Logging a float.nan instead would work, but that run would not show at all in the tensorboard hparam plugin.
# The best thing to do here is to skip that value, it will show as a blank cell in the table view of the
# tensorboard plugin. However, that run would not be shown in the parallel coord or in the scatter plot view.
if hp_value is None:
loguru.warning(
f"Hyper parameter {hp_name} is `None`: the tensorboard hp plugin "
f"will show this run in table view, but not in parallel coordinates "
f"view or in scatter plot matrix view")
continue
if isinstance(hp_value, string_types):
session_start_info.hparams[hp_name].string_value = hp_value
continue
if isinstance(hp_value, bool):
session_start_info.hparams[hp_name].bool_value = hp_value
continue
if not isinstance(hp_value, (int, float)):
hp_value = make_np(hp_value)[0]
session_start_info.hparams[hp_name].number_value = hp_value
session_start_content = HParamsPluginData(
session_start_info=session_start_info, version=PLUGIN_DATA_VERSION)
session_start_summary_metadata = SummaryMetadata(
plugin_data=SummaryMetadata.PluginData(
plugin_name=PLUGIN_NAME,
content=session_start_content.SerializeToString()))
session_start_summary = Summary(value=[
Summary.Value(tag=SESSION_START_INFO_TAG,
metadata=session_start_summary_metadata)
])
return session_start_summary
def add_video(name, imgs, epoch, tb_writer):
# Make a video from the given imgs and output to tensorboard.
strt = time.time()
vid_imgs = np.array(imgs) # TxHxWxC, C=RGBA
video = tensorboardX.summary.make_video(vid_imgs, fps=12) # slower: fps<24
vs = Summary(value=[Summary.Value(tag=name+'_video', image=video)])
tb_writer.file_writer.add_summary(vs, global_step=epoch)
# add_video takes NxTxHxWxC and fails on RGBA
#tb_writer.add_video(name+'_video', vid_tensor=vid_imgs, fps=4)
print('{:s} video took {:2f}'.format(name, time.time()-strt))
def add_video(name, imgs, epoch, tb_writer):
strt = time.time()
from tensorboardX import summary
from tensorboardX.proto.summary_pb2 import Summary
vid_imgs = np.array(imgs) # TxHxWxC, C=RGBA
video = summary.make_video(vid_imgs, fps=24)
vs = Summary(value=[Summary.Value(tag=name + '_video', image=video)])
tb_writer.file_writer.add_summary(vs, global_step=epoch)
# add_video takes NxTxHxWxC and fails on RGBA
#tb_writer.add_video(name+'_video', vid_tensor=vid_imgs, fps=4)
print('{:s} video took {:2f}'.format(name, time.time() - strt))
def add_video(writer,
tag,
tensor_thwc,
global_step=None,
fps=30,
walltime=None):
"""found that add_video from tbX is buggy"""
tag = _clean_tag(tag)
video = make_video(tensor_thwc, fps)
summary = Summary(value=[Summary.Value(tag=tag, image=video)])
writer.file_writer.add_summary(summary, global_step, walltime)
def add_image(self, tag, img_np, global_step=None):
"""Add image data to summary.
Note that this requires the ``pillow`` package.
Args:
tag (string): Data identifier
img_np (np.ndarray): Image data, uint8
global_step (int): Global step value to record
Shape:
img_tensor: :math:`(H, W, 3)`.
"""
tag = tensorboardX.summary._clean_tag(tag)
image = tensorboardX.summary.make_image(img_np)
self.file_writer.add_summary(Summary(value=[Summary.Value(tag=tag, image=image)]), global_step)
def clips_to_video(clips, h, w, c):
# encode sequence of images into gif string
clip = concatenate_videoclips(clips)
filename = tempfile.NamedTemporaryFile(suffix=".gif", delete=False).name
# moviepy >= 1.0.0 use logger=None to suppress output.
try:
clip.write_gif(filename, verbose=False, logger=None)
except TypeError:
get_logger().warning(
"Upgrade to moviepy >= 1.0.0 to suppress the progress bar.")
clip.write_gif(filename, verbose=False)
with open(filename, "rb") as f:
tensor_string = f.read()
try:
os.remove(filename)
except OSError:
get_logger().warning(
"The temporary file used by moviepy cannot be deleted.")
return TBXSummary.Image(height=h,
width=w,
colorspace=c,
encoded_image_string=tensor_string)
def image(tag, tensor, rescale=1, dataformats="CHW"):
"""Outputs a `Summary` protocol buffer with images. The summary has up to
`max_images` summary values containing images. The images are built from
`tensor` which must be 3-D with shape `[height, width, channels]` and where
`channels` can be:
* 1: `tensor` is interpreted as Grayscale.
* 3: `tensor` is interpreted as RGB.
* 4: `tensor` is interpreted as RGBA.
Args:
tag: A name for the generated node. Will also serve as a series name in
TensorBoard.
tensor: A 3-D `uint8` or `float32` `Tensor` of shape `[height, width,
channels]` where `channels` is 1, 3, or 4.
'tensor' can either have values in [0, 1] (float32) or [0, 255] (uint8).
The image() function will scale the image values to [0, 255] by applying
a scale factor of either 1 (uint8) or 255 (float32).
Returns:
A scalar `Tensor` of type `string`. The serialized `Summary` protocol
buffer.
"""
tag = tbxsummary._clean_tag(tag)
tensor = tbxmake_np(tensor)
tensor = convert_to_HWC(tensor, dataformats)
# Do not assume that user passes in values in [0, 255], use data type to detect
if tensor.dtype != np.uint8:
tensor = (tensor * 255.0).astype(np.uint8)
image = tbxsummary.make_image(tensor, rescale=rescale)
return TBXSummary(value=[TBXSummary.Value(tag=tag, image=image)])
def text(tag, text):
plugin_data = SummaryMetadata.PluginData(
plugin_name="text",
content=TextPluginData(version=0).SerializeToString())
smd = SummaryMetadata(plugin_data=plugin_data)
string_val = []
for item in text:
string_val.append(item.encode(encoding="utf_8"))
tensor = TensorProto(
dtype="DT_STRING",
string_val=string_val,
tensor_shape=TensorShapeProto(
dim=[TensorShapeProto.Dim(size=len(text))]),
)
return Summary(value=[Summary.Value(tag=tag, metadata=smd, tensor=tensor)])
def test_event_file_writer_roundtrip(self):
_TAGNAME = 'dummy'
_DUMMY_VALUE = 42
logdir = self.get_temp_dir()
w = EventFileWriter(logdir)
summary = Summary(
value=[Summary.Value(tag=_TAGNAME, simple_value=_DUMMY_VALUE)])
fakeevent = event_pb2.Event(summary=summary)
w.add_event(fakeevent)
w.close()
event_files = sorted(glob.glob(os.path.join(logdir, '*')))
self.assertEqual(len(event_files), 1)
r = PyRecordReader_New(event_files[0])
r.GetNext() # meta data, so skip
r.GetNext()
self.assertEqual(fakeevent.SerializeToString(), r.record())
def tf_summary_to_dict(tf_summary_str_or_pb, namespace=""):
"""Convert a Tensorboard Summary to a dictionary
Accepts either a tensorflow.summary.Summary
or one encoded as a string.
"""
values = {}
if isinstance(tf_summary_str_or_pb, Summary):
summary_pb = tf_summary_str_or_pb
elif isinstance(tf_summary_str_or_pb, Event):
summary_pb = tf_summary_str_or_pb.summary
values["global_step"] = tf_summary_str_or_pb.step
values["_timestamp"] = tf_summary_str_or_pb.wall_time
else:
summary_pb = Summary()
summary_pb.ParseFromString(tf_summary_str_or_pb)
for value in summary_pb.value:
kind = value.WhichOneof("value")
if kind == "simple_value":
values[namespaced_tag(value.tag, namespace)] = value.simple_value
elif kind == "image":
from PIL import Image
image = wandb.Image(
Image.open(six.BytesIO(value.image.encoded_image_string)))
tag_idx = value.tag.rsplit('/', 1)
if len(tag_idx) > 1 and tag_idx[1].isdigit():
tag, idx = tag_idx
values.setdefault(history_image_key(tag), []).append(image)
else:
values[history_image_key(value.tag)] = image
# Coming soon...
# elif kind == "audio":
# audio = wandb.Audio(six.BytesIO(value.audio.encoded_audio_string),
# sample_rate=value.audio.sample_rate, content_type=value.audio.content_type)
elif kind == "histo":
first = value.histo.bucket_limit[0] + \
value.histo.bucket_limit[0] - value.histo.bucket_limit[1]
last = value.histo.bucket_limit[-2] + \
value.histo.bucket_limit[-2] - value.histo.bucket_limit[-3]
np_histogram = (list(value.histo.bucket),
[first] + value.histo.bucket_limit[:-1] + [last])
values[namespaced_tag(
value.tag)] = wandb.Histogram(np_histogram=np_histogram)
return values
def summary_image(img):
output = io.BytesIO()
img.save(output, format='PNG')
encoded = output.getvalue()
output.close()
return Summary.Image(height=img.height,
width=img.width,
colorspace=len(img.getbands()),
encoded_image_string=encoded)
def make_video(tensor, fps):
t, h, w, c = tensor.shape
with tempfile.NamedTemporaryFile() as f:
filename = f.name + '.gif'
images = [tensor[t] for t in range(tensor.shape[0])]
imageio.mimwrite(filename, images, duration=0.04)
with open(filename, 'rb') as f:
tensor_string = f.read()
try:
os.remove(filename)
except OSError:
pass
return Summary.Image(height=h,
width=w,
colorspace=c,
encoded_image_string=tensor_string)
def make_experiment_summary(hparam_infos, metric_infos, experiment):
"""Define hyperparameters and metrics.
Args:
hparam_infos: information about all hyperparameters (name, description, type etc.),
list of dicts containing 'name' (required), 'type', 'description', 'display_name',
'domain_discrete', 'domain_interval'
metric_infos: information about all metrics (tag, description etc.),
list of dicts containing 'tag' (required), 'dataset_type', 'description', 'display_name'
experiment: dict containing 'name' (required), 'description', 'time_created_secs', 'user'
Returns:
"""
def make_hparam_info(hparam):
data_type = hparam.get("type")
if hparam.get("type") is None:
data_type = DataType.DATA_TYPE_UNSET
elif hparam.get("type") in string_types:
data_type = DataType.DATA_TYPE_STRING
elif hparam.get("type") is bool:
data_type = DataType.DATA_TYPE_BOOL
elif hparam.get("type") in (float, int):
data_type = DataType.DATA_TYPE_FLOAT64
return HParamInfo(
name=hparam["name"],
type=data_type,
description=hparam.get("description"),
display_name=hparam.get("display_name"),
domain_discrete=hparam.get("domain_discrete"),
domain_interval=hparam.get("domain_interval"),
)
def make_metric_info(metric):
return MetricInfo(
name=MetricName(tag=metric["tag"]),
dataset_type=DatasetType.Value(
f'DATASET_{metric.get("dataset_type", "UNKNOWN").upper()}'),
description=metric.get("description"),
display_name=metric.get("display_name"),
)
def make_experiment_info(experiment, metric_infos, hparam_infos):
return Experiment(
name=experiment["name"],
description=experiment.get("description"),
time_created_secs=experiment.get("time_created_secs"),
user=experiment.get("user"),
metric_infos=metric_infos,
hparam_infos=hparam_infos,
)
metric_infos = [make_metric_info(m) for m in metric_infos]
hparam_infos = [make_hparam_info(hp) for hp in hparam_infos]
experiment = make_experiment_info(experiment, metric_infos, hparam_infos)
experiment_content = HParamsPluginData(experiment=experiment,
version=PLUGIN_DATA_VERSION)
experiment_summary_metadata = SummaryMetadata(
plugin_data=SummaryMetadata.PluginData(
plugin_name=PLUGIN_NAME,
content=experiment_content.SerializeToString()))
experiment_summary = Summary(value=[
Summary.Value(tag=EXPERIMENT_TAG, metadata=experiment_summary_metadata)
])
return experiment_summary
def _video(tag, vid):
# noinspection PyProtectedMember
tag = tbxsummary._clean_tag(tag)
return TBXSummary(value=[TBXSummary.Value(tag=tag, image=vid)])
def tf_summary_to_dict(tf_summary_str_or_pb, namespace=""):
"""Convert a Tensorboard Summary to a dictionary
Accepts either a tensorflow.summary.Summary
or one encoded as a string.
"""
values = {}
if hasattr(tf_summary_str_or_pb, "summary"):
summary_pb = tf_summary_str_or_pb.summary
values[namespaced_tag("global_step",
namespace)] = tf_summary_str_or_pb.step
values["_timestamp"] = tf_summary_str_or_pb.wall_time
elif isinstance(tf_summary_str_or_pb, (str, bytes, bytearray)):
summary_pb = Summary()
summary_pb.ParseFromString(tf_summary_str_or_pb)
else:
if not hasattr(tf_summary_str_or_pb, "value"):
raise ValueError(
"Can't log %s, only Event, Summary, or Summary proto buffer strings are accepted"
)
else:
summary_pb = tf_summary_str_or_pb
for value in summary_pb.value:
kind = value.WhichOneof("value")
if kind in IGNORE_KINDS:
continue
if kind == "simple_value":
values[namespaced_tag(value.tag, namespace)] = value.simple_value
elif kind == "tensor":
values[namespaced_tag(value.tag,
namespace)] = make_ndarray(value.tensor)
elif kind == "image":
from PIL import Image
img_str = value.image.encoded_image_string
# Supports gifs from TboardX
if img_str.startswith(b"GIF"):
image = wandb.Video(six.BytesIO(img_str), format="gif")
else:
image = wandb.Image(Image.open(six.BytesIO(img_str)))
tag_idx = value.tag.rsplit('/', 1)
if len(tag_idx) > 1 and tag_idx[1].isdigit():
tag, idx = tag_idx
values.setdefault(history_image_key(tag, namespace),
[]).append(image)
else:
values[history_image_key(value.tag, namespace)] = [image]
# Coming soon...
# elif kind == "audio":
# audio = wandb.Audio(six.BytesIO(value.audio.encoded_audio_string),
# sample_rate=value.audio.sample_rate, content_type=value.audio.content_type)
elif kind == "histo":
tag = namespaced_tag(value.tag, namespace)
if len(value.histo.bucket_limit) >= 3:
first = value.histo.bucket_limit[0] + \
value.histo.bucket_limit[0] - value.histo.bucket_limit[1]
last = value.histo.bucket_limit[-2] + \
value.histo.bucket_limit[-2] - value.histo.bucket_limit[-3]
np_histogram = (list(value.histo.bucket), [first] +
value.histo.bucket_limit[:-1] + [last])
try:
#TODO: we should just re-bin if there are too many buckets
values[tag] = wandb.Histogram(np_histogram=np_histogram)
except ValueError:
wandb.termwarn(
"Not logging key \"{}\". Histograms must have fewer than {} bins"
.format(tag, wandb.Histogram.MAX_LENGTH),
repeat=False)
else:
#TODO: is there a case where we can render this?
wandb.termwarn(
"Not logging key \"{}\". Found a histogram with only 2 bins."
.format(tag),
repeat=False)
elif value.tag == "_hparams_/session_start_info":
if wandb.util.get_module("tensorboard.plugins.hparams"):
from tensorboard.plugins.hparams import plugin_data_pb2
plugin_data = plugin_data_pb2.HParamsPluginData()
plugin_data.ParseFromString(value.metadata.plugin_data.content)
for key, param in six.iteritems(
plugin_data.session_start_info.hparams):
if not wandb.run.config.get(key):
wandb.run.config[
key] = param.number_value or param.string_value or param.bool_value
else:
wandb.termerror(
"Received hparams tf.summary, but could not import the hparams plugin from tensorboard"
)
return values
def tf_summary_to_dict(tf_summary_str_or_pb, namespace=""):
"""Convert a Tensorboard Summary to a dictionary
Accepts either a tensorflow.summary.Summary
or one encoded as a string.
"""
values = {}
if hasattr(tf_summary_str_or_pb, "summary"):
summary_pb = tf_summary_str_or_pb.summary
values[namespaced_tag("global_step",
namespace)] = tf_summary_str_or_pb.step
values["_timestamp"] = tf_summary_str_or_pb.wall_time
elif isinstance(tf_summary_str_or_pb, (str, bytes, bytearray)):
summary_pb = Summary()
summary_pb.ParseFromString(tf_summary_str_or_pb)
else:
if not hasattr(tf_summary_str_or_pb, "value"):
raise ValueError(
"Can't log %s, only Event, Summary, or Summary proto buffer strings are accepted"
)
else:
summary_pb = tf_summary_str_or_pb
for value in summary_pb.value:
kind = value.WhichOneof("value")
if kind == "simple_value":
values[namespaced_tag(value.tag, namespace)] = value.simple_value
elif kind == "image":
from PIL import Image
image = wandb.Image(
Image.open(six.BytesIO(value.image.encoded_image_string)))
tag_idx = value.tag.rsplit('/', 1)
if len(tag_idx) > 1 and tag_idx[1].isdigit():
tag, idx = tag_idx
values.setdefault(history_image_key(tag, namespace),
[]).append(image)
else:
values[history_image_key(value.tag, namespace)] = image
# Coming soon...
# elif kind == "audio":
# audio = wandb.Audio(six.BytesIO(value.audio.encoded_audio_string),
# sample_rate=value.audio.sample_rate, content_type=value.audio.content_type)
elif kind == "histo":
first = value.histo.bucket_limit[0] + \
value.histo.bucket_limit[0] - value.histo.bucket_limit[1]
last = value.histo.bucket_limit[-2] + \
value.histo.bucket_limit[-2] - value.histo.bucket_limit[-3]
np_histogram = (list(value.histo.bucket),
[first] + value.histo.bucket_limit[:-1] + [last])
values[namespaced_tag(
value.tag,
namespace)] = wandb.Histogram(np_histogram=np_histogram)
elif value.tag == "_hparams_/session_start_info":
if wandb.util.get_module("tensorboard.plugins.hparams"):
from tensorboard.plugins.hparams import plugin_data_pb2
plugin_data = plugin_data_pb2.HParamsPluginData()
plugin_data.ParseFromString(value.metadata.plugin_data.content)
for key, param in six.iteritems(
plugin_data.session_start_info.hparams):
if not wandb.run.config.get(key):
wandb.run.config[
key] = param.number_value or param.string_value or param.bool_value
else:
wandb.termerror(
"Received hparams tf.summary, but could not import the hparams plugin from tensorboard"
)
return values
def add_image(writer, tag, img):
summary = Summary(value=[Summary.Value(tag=tag, image=summary_image(img))])
writer.file_writer.add_summary(summary)
def _video(self, tag, vid):
tag = tbxsummary._clean_tag(tag)
return TBXSummary(value=[TBXSummary.Value(tag=tag, image=vid)])