def compress_activations(activations, out, compression_method,
compression_latch, thresholding_method):
#separate the activations into spatial subdivisions
activation_divisions = []
if (compression_method["type"] == "max"):
activations = activations.reshape((activations.shape[0], -1))
activation_divisions = [np.max(activations, axis=1)]
elif (compression_method["type"] == "even"):
step = activations.shape[1] / compression_method["value"]
for n in range(0, activations.shape[1], step):
for m in range(0, activations.shape[2], step):
restruct = activations[:, n:n + step, m:m + step].reshape(
(activations.shape[0], -1))
activation_divisions.append(np.max(restruct, axis=1))
elif (compression_method["type"] == "peaks"):
#print("start_gen")
activation_divisions = peak_finding.peak_finding_fast(
activations, n=compression_method["value"])
#print("stop_gen")
#print("activation_divisions: ", len(activation_divisions))
#perform thresholding on each of the spatial subdivisions
thresholded_activations = []
for i in range(len(activation_divisions)):
thresholded_activations.append([])
activation_latch = CountDownLatch(len(activation_divisions))
list_of_threads = []
for i in range(len(activation_divisions)):
#print("start")
if (len(activation_divisions[i]) > 0):
t = threading.Thread(target=threshold_activations,
args=(
activation_divisions[i],
thresholded_activations[i],
activation_latch,
thresholding_method,
))
# threshold_activations(activation_divisions[i], thresholded_activations[i], None, thresholding_method)
list_of_threads.append(t)
t.start()
else:
#print("err: ", i, activation_divisions[i])
activation_latch.count_down()
# wait for all threads to finish
activation_latch. await ()
#print("finished")
#for i in activation_divisions:
# print(i)
out.append(np.array(thresholded_activations).squeeze())
#print("out_shape:", np.array(thresholded_activations).squeeze().shape)
compression_latch.count_down()
def compress_activations(activations,
out,
compression_method,
compression_latch,
thresholding_method,
index=None,
max_val=None,
min_val=None):
#separate the activations into spatial subdivisions
activation_divisions = []
if (compression_method["type"] == "max"):
activations = activations.reshape((activations.shape[0], -1))
activation_divisions = [np.max(activations, axis=1)]
elif (compression_method["type"] == "even"):
step = activations.shape[1] / compression_method["value"]
for n in range(0, activations.shape[1], step):
for m in range(0, activations.shape[2], step):
restruct = activations[:, n:n + step, m:m + step].reshape(
(activations.shape[0], -1))
activation_divisions.append(np.max(restruct, axis=1))
elif (compression_method["type"] == "peaks"):
activation_divisions = peak_finding.peak_finding_fast(
activations, n=compression_method["value"])
#perform thresholding on each of the spatial subdivisions
thresholded_activations = []
if (thresholding_method != "none"):
for i in range(len(activation_divisions)):
thresholded_activations.append([])
activation_latch = CountDownLatch(len(activation_divisions))
list_of_threads = []
for i in range(len(activation_divisions)):
if (len(activation_divisions[i]) > 0):
t = None
if (thresholding_method == "min_max_norm"):
assert max_val != None, "max_val variable must be set"
assert min_val != None, "max_val variable must be set"
t = threading.Thread(target=threshold_activations,
args=(
activation_divisions[i],
thresholded_activations[i],
activation_latch,
thresholding_method,
index + i,
max_val,
min_val,
))
else:
t = threading.Thread(target=threshold_activations,
args=(
activation_divisions[i],
thresholded_activations[i],
activation_latch,
thresholding_method,
))
list_of_threads.append(t)
t.start()
else:
activation_latch.count_down()
# wait for all threads to finish
activation_latch. await ()
else:
thresholded_activations = activation_divisions
out.append(np.array(thresholded_activations).squeeze())
compression_latch.count_down()