def get_debug_data(self, x, module, output):
image = None
activations = None
if isinstance(output, tuple) and len(output) > 0:
output = output[0]
if hasattr(output, 'shape'):
activations = self.get_histogram(x, output)
if len(output.shape) > 1:
# outputs come in batch usually, so pick first
sample = output[0].cpu().detach().numpy()
if len(sample.shape) == 3:
if sample.shape[0] == 3:
image = PIL.Image.fromarray(
get_layer_vis_square(sample))
else:
image = PIL.Image.fromarray(get_image_tales(sample))
elif len(sample.shape) > 1:
image = PIL.Image.fromarray(get_layer_vis_square(sample))
elif len(sample.shape) == 1:
if sample.shape[0] == 1:
# we got a single number
output = sample[0]
else:
image = make_image_from_dense(sample)
# elif isinstance(output[0], (float, str, int)):
# image = output
whistogram = None
bhistogram = None
if hasattr(module, 'weight') and module.weight is not None:
whistogram = self.get_histogram(x, module.weight)
if hasattr(module, 'bias') and module.bias is not None:
bhistogram = self.get_histogram(x, module.bias)
output_rep = None
if isinstance(image, PIL.Image.Image):
output_rep = image
elif isinstance(output, (float, np.floating)):
output_rep = float(output)
elif isinstance(output, (int, np.integer)):
output_rep = int(output)
return output_rep, activations, whistogram, bhistogram
def _image_and_histogram(self, x, output):
image = None
histogram = None
if hasattr(output, 'shape'):
# tf is not batch per default
sample = np.copy(output)
shape = output.shape
if self.is_batch:
# display only first item in batch
sample = np.copy(output[0])
shape = output.shape[1:] # first is batch shizzle
if len(shape) == 3:
if keras.backend.image_data_format() == 'channels_last':
sample = np.transpose(sample, (2, 0, 1))
if sample.shape[0] == 3:
image = PIL.Image.fromarray(get_layer_vis_square(sample))
else:
image = PIL.Image.fromarray(get_image_tales(sample))
elif len(shape) > 1:
image = PIL.Image.fromarray(get_layer_vis_square(sample))
elif len(shape) == 1:
if shape[0] == 1:
# we got a single number
output = sample[0]
else:
image = make_image_from_dense(sample)
h = np.histogram(sample, bins=20)
histogram = pack('<BIH', 1, int(x), h[0].size) + h[1].astype(
'<f').tobytes() + h[0].astype('<I').tobytes()
output_rep = None
if isinstance(image, PIL.Image.Image):
output_rep = image
elif isinstance(output, (float, np.floating)):
output_rep = float(output)
elif isinstance(output, (int, np.integer)):
output_rep = int(output)
return output_rep, histogram
def get_image_from_layer(self, layer_name: str):
layer = self.model.get_layer(layer_name)
model = Model(self.model.input, layer.output)
y = model.predict(self.debug_x, steps=1)
data = y[0] # we pick only the first in the batch
if isinstance(layer, (
keras.layers.Conv2D,
keras.layers.UpSampling2D,
keras.layers.MaxPooling2D,
keras.layers.AveragePooling2D,
keras.layers.GlobalMaxPool2D,
)):
data = np.transpose(data, (2, 0, 1))
image = PIL.Image.fromarray(get_image_tales(data))
else:
if len(data.shape) > 1:
if len(data.shape) == 3:
data = np.transpose(data, (2, 0, 1))
image = PIL.Image.fromarray(get_layer_vis_square(data))
else:
image = self.make_image_from_dense(data)
return self.pil_image_to_jpeg(image)
def log_insight(self,
*data,
name: str,
x=None,
image_convertion=True,
meta=None):
if x is None:
if self.job_steps > 0:
x = self.job_iteration + (self.job_step / self.job_steps)
elif self.job_iteration > 0:
x = self.job_iteration
else:
if name not in self.auto_x_of_insight:
self.auto_x_of_insight[name] = 0
self.auto_x_of_insight[name] += 1
x = self.auto_x_of_insight[name]
if not isinstance(x, (int, float)):
raise Exception('x needs to be integer or float')
if x not in self.created_insights:
self.created_insights[x] = True
self.client.job_action_threadsafe('addInsight', [
x,
time.time(),
self.job_iteration,
self.job_step,
])
for i, d in enumerate(data):
file_type = ''
if isinstance(d, PIL.Image.Image):
file_type = 'png'
d = pil_image_to_jpeg(d)
elif isinstance(d, np.ndarray):
# tf is not batch per default
if image_convertion:
sample = np.copy(d)
shape = d.shape
image = False
if len(shape) == 3:
try:
if 'keras' in sys.modules:
import keras
if keras.backend.image_data_format(
) == 'channels_last':
sample = np.transpose(sample, (2, 0, 1))
elif 'tensorflow.keras' in sys.modules:
import tensorflow.keras as keras
if keras.backend.image_data_format(
) == 'channels_last':
sample = np.transpose(sample, (2, 0, 1))
except:
pass
if sample.shape[0] == 3:
d = PIL.Image.fromarray(
get_layer_vis_square(sample))
image = True
else:
d = PIL.Image.fromarray(get_image_tales(sample))
image = True
elif len(shape) > 1:
d = PIL.Image.fromarray(get_layer_vis_square(sample))
image = True
elif len(shape) == 1:
if shape[0] != 1:
# we got a single number
d = sample[0]
else:
d = make_image_from_dense(sample)
image = True
if image:
file_type = 'png'
d = pil_image_to_jpeg(d)
else:
file_type = 'npy'
d = numpy_to_binary(d)
else:
file_type = 'npy'
d = numpy_to_binary(d)
else:
file_type = 'json'
d = bytes(json.dumps(d), encoding='utf-8')
if len(data) > 1:
file_name = name + '_' + str(i) + '.' + file_type
else:
file_name = name + '.' + file_type
self.client.job_action_threadsafe('addInsightEntry', [
x,
file_name,
datetime.utcnow().isoformat(),
{
'type': file_type,
'meta': meta
},
base64.b64encode(d).decode(),
])