def histogram(self, tag, values, bins, step=None):
"""Saves histogram of values.
Args:
tag: str: label for this data
values: ndarray: will be flattened by this routine
bins: number of bins in histogram, or array of bins for onp.histogram
step: int: training step
"""
if step is None:
step = self._step
else:
self._step = step
values = onp.array(values)
bins = onp.array(bins)
values = onp.reshape(values, -1)
counts, limits = onp.histogram(values, bins=bins)
# boundary logic
cum_counts = onp.cumsum(onp.greater(counts, 0, dtype=onp.int32))
start, end = onp.searchsorted(cum_counts, [0, cum_counts[-1] - 1],
side='right')
start, end = int(start), int(end) + 1
counts = (counts[start - 1:end]
if start > 0 else onp.concatenate([[0], counts[:end]]))
limits = limits[start:end + 1]
sum_sq = values.dot(values)
histo = HistogramProto(min=values.min(),
max=values.max(),
num=len(values),
sum=values.sum(),
sum_squares=sum_sq,
bucket_limit=limits.tolist(),
bucket=counts.tolist())
summary = Summary(value=[Summary.Value(tag=tag, histo=histo)])
self.add_summary(summary, step)
def text(self, tag, textdata, step=None):
"""Saves a text summary.
Args:
tag: str: label for this data
textdata: string, or 1D/2D list/numpy array of strings
step: int: training step
Note: markdown formatting is rendered by tensorboard.
"""
if step is None:
step = self._step
else:
self._step = step
smd = SummaryMetadata(plugin_data=SummaryMetadata.PluginData(
plugin_name='text'))
if isinstance(textdata, (str, bytes)):
tensor = tf.make_tensor_proto(
values=[textdata.encode(encoding='utf_8')], shape=(1, ))
else:
textdata = onp.array(textdata) # convert lists, jax arrays, etc.
datashape = onp.shape(textdata)
if len(datashape) == 1:
tensor = tf.make_tensor_proto(
values=[td.encode(encoding='utf_8') for td in textdata],
shape=(datashape[0], ))
elif len(datashape) == 2:
tensor = tf.make_tensor_proto(values=[
td.encode(encoding='utf_8')
for td in onp.reshape(textdata, -1)
],
shape=(datashape[0],
datashape[1]))
summary = Summary(
value=[Summary.Value(tag=tag, metadata=smd, tensor=tensor)])
self.add_summary(summary, step)
def image(self, tag, image, step=None):
"""Saves RGB image summary from onp.ndarray [H,W], [H,W,1], or [H,W,3].
Args:
tag: str: label for this data
image: ndarray: [H,W], [H,W,1], [H,W,3] save image in greyscale or colors/
step: int: training step
"""
image = onp.array(image)
if step is None:
step = self._step
else:
self._step = step
if len(onp.shape(image)) == 2:
image = image[:, :, onp.newaxis]
if onp.shape(image)[-1] == 1:
image = onp.repeat(image, 3, axis=-1)
image_strio = io.BytesIO()
plt.imsave(image_strio, image, format='png')
image_summary = Summary.Image(
encoded_image_string=image_strio.getvalue(),
colorspace=3,
height=image.shape[0],
width=image.shape[1])
summary = Summary(value=[Summary.Value(tag=tag, image=image_summary)])
self.add_summary(summary, step)
def plot(self, tag, mpl_plt, step=None, close_plot=True):
"""Saves matplotlib plot output to summary image.
Args:
tag: str: label for this data
mpl_plt: matplotlib stateful pyplot object with prepared plotting state
step: int: training step
close_plot: bool: automatically closes plot
"""
if step is None:
step = self._step
else:
self._step = step
fig = mpl_plt.get_current_fig_manager()
img_w, img_h = fig.canvas.get_width_height()
image_buf = io.BytesIO()
mpl_plt.savefig(image_buf, format='png')
image_summary = Summary.Image(
encoded_image_string=image_buf.getvalue(),
colorspace=4, # RGBA
height=img_h,
width=img_w)
summary = Summary(value=[Summary.Value(tag=tag, image=image_summary)])
self.add_summary(summary, step)
if close_plot:
mpl_plt.close()
def _log_reward(self, reward):
if self.show_profiling:
self._log_performance()
self.logger.info("Intermediate reward {:.4f}".format(reward))
stat_reward = Summary(
value=[Summary.Value(tag="reward", simple_value=reward)])
self._tf_board_writer.add_summary(stat_reward)
def log_histogram(tag, values, step, bins=1000):
"""
log_histogram
Logs the histogram of a list/vector of values.
"""
# Convert to a numpy array
values = np.array(values)
# Create histogram using numpy
counts, bin_edges = np.histogram(values, bins=bins)
# Fill fields of histogram proto
hist = HistogramProto()
hist.min = float(np.min(values))
hist.max = float(np.max(values))
hist.num = int(np.prod(values.shape))
hist.sum = float(np.sum(values))
hist.sum_squares = float(np.sum(values**2))
# Requires equal number as bins, where the first goes from -DBL_MAX to bin_edges[1]
# See https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/framework/summary.proto#L30
# Thus, we drop the start of the first bin
bin_edges = bin_edges[1:]
# Add bin edges and counts
for edge in bin_edges:
hist.bucket_limit.append(edge)
for c in counts:
hist.bucket.append(c)
return Summary(value=[Summary.Value(tag=tag, histo=hist)])
def log_colorimages(tag, images, tagsuffix=''):
img = images
s = StringIO()
plt.imsave(s, img, format='png')
img_sum = Summary.Image(encoded_image_string=s.getvalue(),
height=img.shape[0],
width=img.shape[1])
return Summary(
value=[Summary.Value(tag='%s%s' % (tag, tagsuffix), image=img_sum)])
def log_images(tag, images, tagsuffix=''):
"""
log_images
Logs a list of images.
"""
def convert_to_uint8(img):
return np.uint8(img * 255)
if not type(images) == list:
img = images
s = StringIO()
Image.fromarray(convert_to_uint8(img), mode='L').save(s, 'png')
# Create an Image object
img_res = Summary.Image(encoded_image_string=s.getvalue(),
height=img.shape[0],
width=img.shape[1],
colorspace=1)
return Summary(value=[
Summary.Value(tag='%s%s' % (tag, tagsuffix), image=img_res)
])
else:
im_summaries = []
for nr, img in enumerate(images):
# Write the image to a string
s = StringIO()
Image.fromarray(convert_to_uint8(img), mode='L').save(s, 'png')
img_sum = Summary.Image(
encoded_image_string=s.getvalue(),
height=img.shape[0],
width=img.shape[1],
colorspace=1
) #https://github.com/tensorflow/tensorflow/blob/r1.3/tensorflow/core/framework/summary.proto
# Create a Summary value
im_summaries.append(
Summary.Value(tag='%s/%d%s' % (tag, nr, tagsuffix),
image=img_sum))
return Summary(value=im_summaries)
def to_summary(fig, tag):
"""
Convert a matplotlib figure ``fig`` into a TensorFlow Summary object
that can be directly fed into ``Summary.FileWriter``.
Example:
>>> fig, ax = ... # (as above)
>>> summary = to_summary(fig, tag='MyFigure/image')
>>> type(summary)
tensorflow.core.framework.summary_pb2.Summary
>>> summary_writer.add_summary(summary, global_step=global_step)
Args:
fig: A ``matplotlib.figure.Figure`` object.
tag (string): The tag name of the created summary.
Returns:
A TensorFlow ``Summary`` protobuf object containing the plot image
as a image summary.
"""
if not isinstance(tag, six.string_types):
raise TypeError("tag must be a string type")
# attach a new agg canvas
_old_canvas = fig.canvas
try:
canvas = FigureCanvasAgg(fig)
canvas.draw()
w, h = canvas.get_width_height()
# get PNG data from the figure
png_buffer = BytesIO()
canvas.print_png(png_buffer)
png_encoded = png_buffer.getvalue()
png_buffer.close()
summary_image = Summary.Image(
height=h,
width=w,
colorspace=4, # RGB-A
encoded_image_string=png_encoded)
summary = Summary(value=[Summary.Value(tag=tag, image=summary_image)])
return summary
finally:
fig.canvas = _old_canvas
def scalar(self, tag, value, step=None):
"""Saves scalar value.
Args:
tag: str: label for this data
value: int/float: number to log
step: int: training step
"""
value = float(onp.array(value))
if step is None:
step = self._step
else:
self._step = step
summary = Summary(value=[Summary.Value(tag=tag, simple_value=value)])
self.add_summary(summary, step)
def log_image(file_writer, tensor, epoch_no, tag):
height, width, channel = tensor.shape
tensor = ((tensor + 1) * 255)
tensor = tensor.astype('uint8')
image = Image.fromarray(tensor)
import io
output = io.BytesIO()
image.save(output, format='PNG')
image_string = output.getvalue()
output.close()
tf_img = Summary.Image(height=height,
width=width,
colorspace=channel,
encoded_image_string=image_string)
summary = Summary(value=[Summary.Value(tag=tag, image=tf_img)])
file_writer.add_summary(summary, epoch_no)
file_writer.flush()
def log_vector(tag, values):
"""
log_histogram
Logs a vector of values.
"""
values = np.array(values).flatten()
# Fill fields of histogram proto
hist = HistogramProto()
hist.min = 0
hist.max = len(values) - 1
hist.num = len(values)
hist.sum = float(np.sum(np.arange(hist.num)))
hist.sum_squares = float(np.sum(np.arange(hist.num)**2))
for idx, c in enumerate(values):
hist.bucket_limit.append(idx)
hist.bucket.append(c)
return Summary(value=[Summary.Value(tag=tag, histo=hist)])
def audio(self, tag, audiodata, step=None, sample_rate=44100):
"""Saves audio.
NB: single channel only right now.
Args:
tag: str: label for this data
audiodata: ndarray [Nsamples,]: data between (-1.0,1.0) to save as wave
step: int: training step
sample_rate: sample rate of passed in audio buffer
"""
audiodata = onp.array(audiodata)
if step is None:
step = self.step
else:
self.step = step
audiodata = onp.clip(onp.squeeze(audiodata), -1, 1)
if audiodata.ndim != 1:
raise ValueError('Audio data must be 1D.')
sample_list = (32767.0 * audiodata).astype(int).tolist()
wio = io.BytesIO()
wav_buf = wave.open(wio, 'wb')
wav_buf.setnchannels(1)
wav_buf.setsampwidth(2)
wav_buf.setframerate(sample_rate)
enc = b''.join([struct.pack('<h', v) for v in sample_list])
wav_buf.writeframes(enc)
wav_buf.close()
encoded_audio_bytes = wio.getvalue()
wio.close()
audio = Summary.Audio(
sample_rate=sample_rate,
num_channels=1,
length_frames=len(sample_list),
encoded_audio_string=encoded_audio_bytes,
content_type='audio/wav')
summary = Summary(value=[Summary.Value(tag=tag, audio=audio)])
self.writer.add_summary(summary, step)
imgs, labels = batch
if len(args.gpus) > 0:
imgs, labels = imgs.cuda(), labels.cuda()
imgs, labels = Variable(imgs), Variable(labels)
preds = net(imgs)
loss = loss_fn(preds, labels)
loss.backward()
# Write training summary
if args.write_summary and step % args.report_interval == 0:
param_norm = global_norm(net.parameters())
grad_norm = global_norm(
[param.grad for param in net.parameters()])
summary_proto = Summary(value=[
Summary.Value(tag='train/loss', simple_value=loss.data[0]),
Summary.Value(tag='train/param_norm',
simple_value=param_norm.data[0]),
Summary.Value(tag='train/grad_norm',
simple_value=grad_norm.data[0]),
])
writer.add_summary(summary_proto,
global_step=step * args.batch_size)
# Update parameters
optimizer.step()
step += 1
epoch += 1
# Evaluate on test
def log_scalar(tag, value):
"""
log_scalar
Logs a scalar.
"""
return Summary(value=[Summary.Value(tag=tag, simple_value=value)])
