def generate_dict():
# simulation dictionary has to have:
# N - SharedArray of one entity enumerating the step of the simulation
# stages - integer number of processing stages
# num_proc - integer number of processors that will be utilized for workers
# stage0 - a list of element parameters (dictionaries) for each execution element
# stage0_size - number of execution elements in the stage
# other parameters are model specific.
simulation_dict = PVM_Create.create_blank_dictionary()
simulation_dict['stages'] = 1
simulation_dict['num_proc'] = mp.cpu_count() / 2
simulation_dict['stage0'] = []
simulation_dict['execution_unit_module'] = 'PVM_models.demo00_unit'
unit = importlib.import_module(simulation_dict['execution_unit_module'])
simulation_dict['stage0_size'] = 100
arena = SharedArray.SharedNumpyArray(
(500, 500), np.uint8) # the arena will be 500x500 in this case
simulation_dict['arena'] = arena
for i in range(simulation_dict['stage0_size']):
unit_parameters = {}
unit_parameters['arena'] = arena
# block will define a sub-block of the arena that each unit will be processing
# it will contain (x0, y0, width, height)
unit_parameters['block'] = SharedArray.SharedNumpyArray((4, ), np.int)
# each unit gets a 100x100 sub-block of the arena
unit_parameters['block'][0] = (i / 10) * 50
unit_parameters['block'][1] = (i % 10) * 50
unit_parameters['block'][2] = 50
unit_parameters['block'][3] = 50
simulation_dict['stage0'].append(unit_parameters)
for i in range(simulation_dict['stage0_size']):
unit.ExecutionUnit.generate_missing_parameters(
simulation_dict['stage0'][i])
return simulation_dict
def upgrade_to_ver_1(parameters):
parameters['internal_buffers'] = []
parameters['output_min'] = SharedArray.SharedNumpyArray_like(
parameters['output_block'])
parameters['output_max'] = SharedArray.SharedNumpyArray_like(
parameters['output_block'])
for (block, delta, pred_block,
pred_block2) in parameters['signal_blocks']:
block01 = SharedArray.SharedNumpyArray_like(block)
block02 = SharedArray.SharedNumpyArray_like(block)
block03 = SharedArray.SharedNumpyArray_like(block)
parameters['internal_buffers'].append((block01, block02, block03))
def create_blank_dictionary(name="default",
description="default",
save_sources=False):
"""
Creates a minimal model dictionary.
:return: model dict
:rtype: dict
"""
import PVM_framework.SharedArray as SharedArray
import numpy as np
simulation_dict = {}
simulation_dict['N'] = SharedArray.SharedNumpyArray((1, ), np.int64)
simulation_dict['N'][0] = 0
simulation_dict['record_filename'] = SharedArray.SharedNumpyArray((256, ),
np.uint8)
simulation_dict['paused'] = SharedArray.SharedNumpyArray((1, ), np.int)
simulation_dict['paused'][0] = 0
simulation_dict['finished'] = SharedArray.SharedNumpyArray((1, ), np.int)
simulation_dict['finished'][0] = 0
simulation_dict['flags'] = SharedArray.SharedNumpyArray((16, ), np.uint8)
for i in range(16):
simulation_dict['flags'][i] = 0
simulation_dict['num_proc'] = mp.cpu_count() / 2
simulation_dict['debug_infrastructure'] = {}
simulation_dict['debug_infrastructure']['enabled_hookups'] = {}
simulation_dict['debug_infrastructure']['disabled_hookups'] = {}
logging.info("Created a blank dictionary")
simulation_dict['description'] = description
simulation_dict['name'] = name
simulation_dict['sources'] = {}
if save_sources:
for filename in glob.glob("*.py"):
f = open(filename, "r")
simulation_dict['sources'][filename] = f.read()
f.close()
logging.info("Saved source file %s into the dictionary" % filename)
for filename in glob.glob("*.pyx"):
f = open(filename, "r")
simulation_dict['sources'][filename] = f.read()
f.close()
logging.info("Saved source file %s into the dictionary" % filename)
simulation_dict['hash'] = "%08x" % random.getrandbits(32)
simulation_dict['timestamp'] = datetime.datetime.now().strftime(
'%Y_%m_%d_%H_%M_%S')
logging.info("Assigned hash %s to this simulation instance" %
simulation_dict['hash'])
logging.info("Assigned timestamp %s to this simulation instance" %
simulation_dict['timestamp'])
return simulation_dict
def test_sharedness_of_a_darray():
"""
Test if the double buffered array is actually shared between two processes.
"""
parity = SharedArray.SharedNumpyArray((1, ), np.uint32)
Array = SharedArray.DoubleBufferedSharedNumpyArray((10, 10), np.float,
parity)
p = multiprocessing.Process(target=worker_test_sharedness_of_a_darray,
args=(Array, parity))
p.start()
p.join()
assert (Array[5, 5] == 15)
parity[0] = 0
assert (Array[5, 5] == 10)
def upgrade(simulation_dict):
old_log = simulation_dict['error_log']
if old_log.shape[1] < 1000000:
new_log = SharedArray.SharedNumpyArray((old_log.shape[0], 1000000),
np.float)
new_log[:, 0:old_log.shape[1]] = old_log
simulation_dict['error_log'] = new_log
if "flags" not in simulation_dict.keys():
simulation_dict['flags'] = SharedArray.SharedNumpyArray((16, ),
np.uint8)
for i in range(16):
simulation_dict['flags'][i] = 0
if "record_filename" not in simulation_dict.keys():
simulation_dict['record_filename'] = SharedArray.SharedNumpyArray(
(256, ), np.uint8)
def get_layers(dims=None):
"""
Instantiate MLP layers with given dimensions
:param dims: list of ints
:return: layer dictinary
"""
if not dims:
dims = []
layers = []
for d in dims:
layers.append({
'activation': SharedArray.SharedNumpyArray(d, np.float),
'error': SharedArray.SharedNumpyArray(d, np.float),
'delta': SharedArray.SharedNumpyArray(d, np.float)
})
return layers
def test_pickling_darray(tmpdir):
"""
Pickles and unpickles a double buffered shared array
"""
parity = SharedArray.SharedNumpyArray((1, ), np.uint32)
A = SharedArray.DoubleBufferedSharedNumpyArray((10, 10), np.float, parity)
A[1, 1] = 15
parity[0] = 1
A[1, 1] = 12
save({"array": A, "parity": parity}, str(tmpdir) + "/test.p.gz")
ddict = load(str(tmpdir) + "/test.p.gz")
B = ddict["array"]
parity = ddict["parity"]
assert (B[1, 1] == 12)
parity[0] = 0
assert (B[1, 1] == 15)
def worker_test_IPC_attachability():
Array = SharedArray.SharedNumpyArray((10, 10),
np.float,
tag="abcdefgh312",
create=False)
Array[5, 5] = 10
return
def generate_dict():
# simulation dictionary has to have:
# N - SharedArray of one entity enumerating the step of the simulation
# stages - integer number of processing stages
# num_proc - integer number of processors that will be utilized for workers
# stage0 - a list of element parameters (dictionaries) for each execution element
# stage0_size - number of execution elements in the stage
# other parameters are model specific.
simulation_dict = PVM_Create.create_blank_dictionary()
simulation_dict['stages'] = 1
simulation_dict['num_proc'] = mp.cpu_count() / 2
simulation_dict['stage0'] = []
simulation_dict['execution_unit_module'] = 'PVM_models.demo02_unit'
unit = importlib.import_module(simulation_dict['execution_unit_module'])
blocks_per_dim = 40
block_size = 5
simulation_dict['stage0_size'] = blocks_per_dim * blocks_per_dim
yet_previous_array = SharedArray.SharedNumpyArray(
(block_size * blocks_per_dim, block_size * blocks_per_dim, 3),
np.uint8)
simulation_dict['yet_previous_array'] = yet_previous_array
previous_array = SharedArray.SharedNumpyArray(
(block_size * blocks_per_dim, block_size * blocks_per_dim, 3),
np.uint8)
simulation_dict['previous_array'] = previous_array
current_array = SharedArray.SharedNumpyArray(
(block_size * blocks_per_dim, block_size * blocks_per_dim, 3),
np.uint8)
simulation_dict['current_array'] = current_array
predicted_array = SharedArray.SharedNumpyArray(
(block_size * blocks_per_dim, block_size * blocks_per_dim, 3),
np.uint8)
simulation_dict['predicted_array'] = predicted_array
learning_rate = SharedArray.SharedNumpyArray((1, ), np.float)
learning_rate[0] = 0.01
simulation_dict['learning_rate'] = learning_rate
momentum = SharedArray.SharedNumpyArray((1, ), np.float)
momentum[0] = 0.5
simulation_dict['momentum'] = momentum
for i in range(simulation_dict['stage0_size']):
unit_parameters = {}
unit_parameters['learning_rate'] = learning_rate
unit_parameters['momentum'] = momentum
unit_parameters['yet_previous_array'] = yet_previous_array
unit_parameters['previous_array'] = previous_array
unit_parameters['current_array'] = current_array
unit_parameters['predicted_array'] = predicted_array
# block will define a sub-block of the arena that each unit will be processing
# it will contain (x0, y0, width, height)
unit_parameters['block'] = SharedArray.SharedNumpyArray((4, ), np.int)
# each unit gets a 100x100 sub-block of the arena
unit_parameters['block'][0] = (i / blocks_per_dim) * block_size
unit_parameters['block'][1] = (i % blocks_per_dim) * block_size
unit_parameters['block'][2] = block_size
unit_parameters['block'][3] = block_size
simulation_dict['stage0'].append(unit_parameters)
for i in range(simulation_dict['stage0_size']):
unit.ExecutionUnit.generate_missing_parameters(
simulation_dict['stage0'][i])
return simulation_dict
def test_pickling_array(tmpdir):
"""
Pickles and unpickles a shared array
"""
A = SharedArray.SharedNumpyArray((10, 10), np.float)
A[1, 1] = 15
save(A, str(tmpdir) + "/test.p.gz")
B = load(str(tmpdir) + "/test.p.gz")
assert (B[1, 1] == 15)
def test_IPC_attachability():
"""
Test if the array is actually shared between two processes.
"""
Array = SharedArray.SharedNumpyArray((10, 10),
np.float,
tag="abcdefgh312",
create=True)
p = multiprocessing.Process(target=worker_test_IPC_attachability, args=())
p.start()
p.join()
assert (Array[5, 5] == 10)
def upgrade_readout(simulation_dict):
"""
Upgrade the dictionary with a parameters for a three layer perceptron for readout
:param simulation_dict:
:return:
"""
logging.info("Upgrading readout network to a full perceptron")
import PVM_framework.SharedArray as SharedArray
import PVM_framework.MLP as MLP
simulation_dict['stage0_size'] = len(simulation_dict['stage0'])
needed = True
for i in range(simulation_dict['stage0_size']):
if len(simulation_dict['stage0'][i]["MLP_parameters_additional"]
['layers']) == 2 and len(
simulation_dict['stage0'][i]["MLP_parameters_additional"]
['weights']) == 1:
nhidden = simulation_dict['stage0'][i]["MLP_parameters_additional"][
'layers'][0]['activation'].shape[0] - 1
nadditional = simulation_dict['stage0'][i][
"MLP_parameters_additional"]['layers'][-1]['activation'].shape[
0] - 1
layer = {
'activation': SharedArray.SharedNumpyArray((nhidden + 1),
np.float),
'error': SharedArray.SharedNumpyArray((nhidden + 1), np.float),
'delta': SharedArray.SharedNumpyArray((nhidden + 1), np.float)
}
simulation_dict['stage0'][i]["MLP_parameters_additional"][
'layers'].insert(1, layer)
simulation_dict['stage0'][i]["MLP_parameters_additional"]['weights'] = \
[MLP.initialize_weights(SharedArray.SharedNumpyArray((nhidden+1, nhidden), np.float)),
MLP.initialize_weights(SharedArray.SharedNumpyArray((nhidden+1, nadditional), np.float)),
]
else:
needed = False
if needed:
logging.info("Upgrade complete")
else:
logging.info("Upgrade was not nescessary")
def test_sharedness_of_an_array():
"""
Test if the array is actually shared between two processes.
"""
for dtype in (np.float, np.float16, np.float32, np.float64, np.float128,
np.uint8, np.uint16, np.uint32, np.uint64, np.int8, np.int16,
np.int32, np.int64, np.complex, np.complex64):
Array = SharedArray.SharedNumpyArray((10, 10), dtype)
p = multiprocessing.Process(target=worker_test_sharednes_of_an_array,
args=(Array, None))
p.start()
p.join()
assert (Array[5, 5] == 10)
def get_weights(layers=None):
"""
Instantiate MLP weights
:param layers: layer dictionary
:return: list of weight matrices
"""
if not layers:
layers = []
mlp_weights = []
for l in range(len(layers) - 1):
l0 = layers[l]['activation'].shape[0]
l1 = layers[l + 1]['activation'].shape[0] - 1
mlp_weights.append(
initialize_weights(SharedArray.SharedNumpyArray((l0, l1),
np.float)))
return mlp_weights
def parrent_test_barrier(UseSpinLock=False):
"""
Parent (controller) process, instantiates the shared memory and
synchronization objects and spawns worker processes. Verifies that
the data consistency is assured as the code executes.
"""
test_barrier = SyncUtils.Barrier(4, timeout=0.001, UseSpinLock=UseSpinLock)
shared_data = SharedArray.SharedNumpyArray((4,), np.float)
procs = []
for i in xrange(4):
proc = multiprocessing.Process(target=worker_test_barrier, args=(test_barrier, shared_data, i))
procs.append(proc)
proc.start()
test_barrier.parent_barrier()
test_barrier.resume_workers()
for i in xrange(4):
test_barrier.parent_barrier()
assert (sum(shared_data) == (i + 1) * 4)
test_barrier.resume_workers()
test_barrier.parent_barrier(quit_workers=True)
for p in procs:
p.join()
def generate_missing_parameters(parameters):
"""
This method can be called to generate all the missing dictionary parameters when all
the other relevant variables are known. Leave empty if there is nothing more to generate
"""
ninputs1 = 3 * parameters['block'][2] * parameters['block'][3] * 3
nhidden = ninputs1 / 5
ninputs = ninputs1 + nhidden
noutputs = 2 * parameters['block'][2] * parameters['block'][3] * 3
parameters["MLP_parameters"] = {}
parameters["MLP_parameters"]['layers'] = [
{
'activation': SharedArray.SharedNumpyArray((ninputs + 1),
np.float),
'error': SharedArray.SharedNumpyArray((ninputs + 1), np.float),
'delta': SharedArray.SharedNumpyArray((ninputs + 1), np.float)
},
{
'activation': SharedArray.SharedNumpyArray((nhidden + 1),
np.float),
'error': SharedArray.SharedNumpyArray((nhidden + 1), np.float),
'delta': SharedArray.SharedNumpyArray((nhidden + 1), np.float)
},
{
'activation':
SharedArray.SharedNumpyArray((noutputs + 1), np.float),
'error':
SharedArray.SharedNumpyArray((noutputs + 1), np.float),
'delta':
SharedArray.SharedNumpyArray((noutputs + 1), np.float)
},
]
parameters["MLP_parameters"]['beta'] = SharedArray.SharedNumpyArray(
(1, ), np.float)
parameters["MLP_parameters"]['beta'][0] = 1.0
parameters["MLP_parameters"]['learning_rate'] = parameters[
'learning_rate']
parameters["MLP_parameters"]['momentum'] = parameters['momentum']
parameters["MLP_parameters"]['mse'] = SharedArray.SharedNumpyArray(
(1, ), np.float)
parameters["MLP_parameters"]['weights'] = [
MLP.initialize_weights(
SharedArray.SharedNumpyArray((ninputs + 1, nhidden),
np.float)),
MLP.initialize_weights(
SharedArray.SharedNumpyArray((nhidden + 1, noutputs),
np.float))
]
def upgrade_dictionary_to_ver1_0(simulation_dict):
upgrade(simulation_dict)
if "version_major" not in simulation_dict.keys(
) or simulation_dict["version_major"] < 1:
logging.info(
"Simulation dictionary of the old type, automatically upgrading to ver 1.0"
)
if 'learning_rates' not in simulation_dict.keys():
simulation_dict['learning_rates'] = []
if 'momenta' not in simulation_dict.keys():
simulation_dict['momenta'] = []
if 'taus' not in simulation_dict.keys():
simulation_dict['taus'] = []
if 'predicted_arrays' not in simulation_dict.keys():
simulation_dict['predicted_arrays'] = []
if 'predicted_arrays_t2' not in simulation_dict.keys():
simulation_dict['predicted_arrays_t2'] = []
if 'predicted_readout_arrays' not in simulation_dict.keys():
simulation_dict['predicted_readout_arrays'] = []
if 'readout_arrays' not in simulation_dict.keys():
simulation_dict['readout_arrays'] = []
if 'predicted_arrays' not in simulation_dict.keys():
simulation_dict['predicted_arrays'] = []
if 'state_arrays' not in simulation_dict.keys():
simulation_dict['state_arrays'] = []
if 'delta_arrays' not in simulation_dict.keys():
simulation_dict['delta_arrays'] = []
for i in range(PVM_MAX_LAYERS): # max number of layers
if "delta_array%02d" % i in simulation_dict.keys():
simulation_dict['delta_arrays'].append(
simulation_dict['delta_array%02d' % i])
del simulation_dict['delta_array%02d' % i]
if "learning_rate%02d" % i in simulation_dict.keys():
simulation_dict['learning_rates'].append(
simulation_dict['learning_rate%02d' % i])
del simulation_dict['learning_rate%02d' % i]
if "state_array%02d" % i in simulation_dict.keys():
simulation_dict['state_arrays'].append(
simulation_dict['state_array%02d' % i])
del simulation_dict['state_array%02d' % i]
if "predicted_readout_array_float%02d" % i in simulation_dict.keys(
):
simulation_dict['predicted_readout_arrays'].append(
simulation_dict['predicted_readout_array_float%02d' % i])
del simulation_dict['predicted_readout_array_float%02d' % i]
if "readout_array_float%02d" % i in simulation_dict.keys():
simulation_dict['readout_arrays'].append(
simulation_dict['readout_array_float%02d' % i])
del simulation_dict['readout_array_float%02d' % i]
if "predicted_array%02d" % i in simulation_dict.keys():
simulation_dict['predicted_arrays'].append(
simulation_dict['predicted_array%02d' % i])
simulation_dict['predicted_arrays_t2'].append(
SharedArray.SharedNumpyArray_like(
simulation_dict['predicted_array%02d' % i]))
del simulation_dict['predicted_array%02d' % i]
if "motor_delta_array_float" in simulation_dict.keys():
del simulation_dict["motor_delta_array_float"]
if "readout_array_float" in simulation_dict.keys():
del simulation_dict["readout_array_float"]
if "readout_array_float" in simulation_dict.keys():
del simulation_dict["readout_array_float"]
if "predicted_motor_derivative_array_float" in simulation_dict.keys():
del simulation_dict["predicted_motor_derivative_array_float"]
if "predicted_readout_array_float" in simulation_dict.keys():
del simulation_dict["predicted_readout_array_float"]
if "additional_learning_rate" in simulation_dict.keys():
simulation_dict["readout_learning_rate"] = simulation_dict[
"additional_learning_rate"]
del simulation_dict["additional_learning_rate"]
elif "readout_learning_rate" not in simulation_dict.keys():
simulation_dict[
"readout_learning_rate"] = SharedArray.SharedNumpyArray(
(1, ), np.float)
simulation_dict["readout_learning_rate"][:] = 0.00001
simulation_dict['execution_unit_module'] += "_v1"
if not simulation_dict['execution_unit_module'].startswith(
"PVM_models"):
simulation_dict[
'execution_unit_module'] = "PVM_models." + simulation_dict[
'execution_unit_module']
ex_unit = importlib.import_module(
simulation_dict['execution_unit_module'])
for s in range(simulation_dict['stages']):
stage = simulation_dict['stage%d' % s]
for unit in stage:
signal_blocks = unit['signal_blocks']
unit['signal_blocks'] = []
for block in signal_blocks:
# Each block: signal_block, delta_block, prediction_t+1, prediction_t+2
unit['signal_blocks'].append([
block[0], block[1], block[2],
SharedArray.SharedNumpyArray_like(block[2])
])
context_blocks = unit['context_blocks']
unit['context_blocks'] = []
for block in context_blocks:
# Each block: context_block, delta_block, switching_factor
unit['context_blocks'].append(
[block[0], block[1], block[2]])
readout_blocks = unit['predicted_blocks']
del unit['predicted_blocks']
unit['readout_blocks'] = []
for block in readout_blocks:
# Each block: teaching_signal, delta_block, readout_block
unit['readout_blocks'].append(
[block[0], block[1], block[2]])
# Delta blocks can remain unchanged
if "learning_rate" in unit.keys():
unit['primary_learning_rate'] = unit.pop("learning_rate")
if "momentum" in unit.keys():
unit['primary_momentum'] = unit.pop("momentum")
unit['readout_momentum'] = unit['primary_momentum']
if "additional_learning_rate" in unit.keys():
unit['readout_learning_rate'] = unit.pop(
"additional_learning_rate")
else:
unit['readout_learning_rate'] = simulation_dict[
"readout_learning_rate"]
# Output block may remain unchanged
if "MLP_parameters" in unit.keys():
unit['Primary_Predictor_params'] = unit.pop(
"MLP_parameters")
if "MLP_parameters_res" in unit.keys():
unit['Residual_Predictor_params'] = unit.pop(
"MLP_parameters_res")
if "MLP_parameters_additional" in unit.keys():
unit['Readout_Predictor_params'] = unit.pop(
"MLP_parameters_additional")
unit['flags'] = simulation_dict['flags']
ex_unit.ExecutionUnit.upgrade_to_ver_1(unit)
simulation_dict['version_major'] = 1
simulation_dict['version_minor'] = 0
# Remove all the old source files
simulation_dict['sources'] = {}
logging.info("Upgrade succesfull")
else:
for s in range(simulation_dict['stages']):
stage = simulation_dict['stage%d' % s]
for unit in stage:
unit['flags'] = simulation_dict['flags']
logging.info("Dictionary already ver 1.0 or above, no need to upgrade")
def generate_missing_parameters(parameters, options):
"""
This method can be called to generate all the missing dictionary parameters when all
the other relevant variables are known. Leave empty if there is nothing more to generate.
When complex_unit is False, a standard 3-layer MLP is used.
When complex_unit is True, an MLP with additional hidden layers is used.
There needs to be no return value, the method leaves a side effect by modifying the perameters dict.
:param parameters: parameter dictionary
:type parameters: dict
"""
complex_unit = options['unit_type'] == "complex"
polynomial = options['polynomial'] == '1'
autoencoder = options['autoencoder'] == '1'
nhidden = np.prod(parameters['output_block'].shape)
parameters['output_min'] = SharedArray.SharedNumpyArray_like(
parameters['output_block'])
parameters['output_max'] = SharedArray.SharedNumpyArray_like(
parameters['output_block'])
ninputs = 0
noutputs = 0
ncontext = 0
# Any additional memory buffers needed in the operation of the unit
parameters['internal_buffers'] = []
for (block, delta, pred_block,
pred_block2) in parameters['signal_blocks']:
block01 = SharedArray.SharedNumpyArray_like(block)
block02 = SharedArray.SharedNumpyArray_like(block)
block03 = SharedArray.SharedNumpyArray_like(block)
parameters['internal_buffers'].append((block01, block02, block03))
for (block, delta, pred_block,
pred_block2) in parameters['signal_blocks']:
ninputs += np.prod(block.shape) * len(
ExecutionUnit.UNSUPERVISED_SIGNAL_INPUTS)
for (block, delta, factor) in parameters['context_blocks']:
ncontext += np.prod(block.shape)
for (block, delta, pred_block,
pred_block2) in parameters['signal_blocks']:
noutputs += 2 * np.prod(
block.shape) # to predict two steps into the future
nadditional = 0
for (block, delta, pblock) in parameters['readout_blocks']:
nadditional += np.prod(block.shape)
parameters["Primary_Predictor_params"] = {}
parameters["Residual_Predictor_params"] = {}
parameters["Readout_Predictor_params"] = {}
if complex_unit: # 4 layer perceptron
parameters["Primary_Predictor_params"]['layers'] = MLP.get_layers(
[ncontext + 1, 3 * nhidden + 1, 2 * nhidden + 1, noutputs + 1])
parameters["Residual_Predictor_params"]['layers'] = MLP.get_layers(
[ninputs + 1, 2 * nhidden + 1, nhidden + 1, noutputs + 1])
parameters["complex"] = True
else: # 3 layer perceptron Simple MLP Unit (not complex unit)
parameters["Primary_Predictor_params"]['layers'] = MLP.get_layers(
[ncontext + 1, 2 * nhidden + 1, noutputs + 1])
parameters["Residual_Predictor_params"]['layers'] = MLP.get_layers(
[ninputs + 2 * nhidden + 1, nhidden + 1, noutputs + 1])
parameters["Primary_Predictor_params"][
'beta'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["Primary_Predictor_params"]['beta'][0] = 1.0
parameters["Primary_Predictor_params"]['learning_rate'] = parameters[
'primary_learning_rate']
parameters["Primary_Predictor_params"]['momentum'] = parameters[
'momentum']
parameters["Primary_Predictor_params"][
'mse'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["Primary_Predictor_params"]['weights'] = MLP.get_weights(
parameters["Primary_Predictor_params"]['layers'])
parameters["Residual_Predictor_params"][
'beta'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["Residual_Predictor_params"]['beta'][0] = 1.0
parameters["Residual_Predictor_params"]['learning_rate'] = parameters[
'primary_learning_rate']
parameters["Residual_Predictor_params"]['momentum'] = parameters[
'momentum']
parameters["Residual_Predictor_params"][
'mse'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["Residual_Predictor_params"]['weights'] = MLP.get_weights(
parameters["Residual_Predictor_params"]['layers'])
parameters["Readout_Predictor_params"]['layers'] = MLP.get_layers(
[nhidden + 1, 2 * nhidden + 1, nadditional + 1])
parameters["Readout_Predictor_params"][
'beta'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["Readout_Predictor_params"]['beta'][0] = 1.0
parameters["Readout_Predictor_params"]['learning_rate'] = parameters[
'readout_learning_rate']
parameters["Readout_Predictor_params"]['momentum'] = parameters[
'momentum']
parameters["Readout_Predictor_params"][
'mse'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["Readout_Predictor_params"]['weights'] = MLP.get_weights(
parameters["Readout_Predictor_params"]['layers'])
parameters["Primary_Predictor_params"]['polynomial'] = polynomial
parameters["Residual_Predictor_params"]['polynomial'] = polynomial
parameters["Readout_Predictor_params"]['polynomial'] = polynomial
parameters['autoencoder'] = autoencoder
def generate_missing_parameters(parameters, options):
"""
This method can be called to generate all the missing dictionary parameters when all
the other relevant variables are known. Leave empty if there is nothing more to generate.
When complex_unit is False, a standard 3-layer MLP is used.
When complex_unit is True, an MLP with additional hidden layers is used.
There needs to be no return value, the method leaves a side effect by modifying the perameters dict.
:param parameters: parameter dictionary
:type parameters: dict
"""
complex_unit = options['unit_type'] == "complex"
polynomial = options['polynomial'] == '1'
autoencoder = options['autoencoder'] == '1'
use_t_2_block = options['use_t_minus_2_block'] == '1'
use_derivative = options['use_derivative'] == '1'
use_integral = options['use_integral'] == '1'
use_error = options['use_error'] == '1'
predict_2_steps = options['predict_two_steps'] == '1'
use_global_backprop = options['use_global_backprop'] == '1'
complex_context_in_second_layer = options[
'feed_context_in_complex_layer'] == '1'
parameters['normalize_output'] = options["normalize_output"] == "1"
parameters['backpropagate_readout_error'] = options[
"backpropagate_readout_error"] == "1"
nhidden = np.prod(parameters['output_block'].shape)
parameters['output_min'] = SharedArray.SharedNumpyArray_like(
parameters['output_block'])
parameters['output_max'] = SharedArray.SharedNumpyArray_like(
parameters['output_block']) + 1
parameters['avg_delta'] = SharedArray.SharedNumpyArray_like(
parameters['output_block'])
ninputs = 0
noutputs = 0
ncontext = 0
# Any additional memory buffers needed in the operation of the unit
parameters['internal_buffers'] = []
parameters['integral_blocks'] = []
parameters['derivative_blocks'] = []
parameters['error_blocks'] = []
parameters['use_derivative'] = use_derivative
parameters['use_integral'] = use_integral
parameters['use_error'] = use_error
parameters['use_t_2_block'] = use_t_2_block
parameters['predict_2_steps'] = predict_2_steps
for (block, delta, pred_block,
pred_block2) in parameters['signal_blocks']:
block01 = SharedArray.SharedNumpyArray_like(block)
block02 = SharedArray.SharedNumpyArray_like(block)
block03 = SharedArray.SharedNumpyArray_like(block)
parameters['internal_buffers'].append((block01, block02, block03))
if use_derivative:
parameters['derivative_blocks'].append(
SharedArray.SharedNumpyArray_like(block))
if use_integral:
parameters['integral_blocks'].append(
SharedArray.SharedNumpyArray_like(block))
if use_error:
parameters['error_blocks'].append(
SharedArray.SharedNumpyArray_like(block))
input_block_features = 1
output_predictions = 1
if use_derivative:
input_block_features += 1
if use_integral:
input_block_features += 1
if use_error:
input_block_features += 1
if use_t_2_block:
input_block_features += 1
if predict_2_steps:
output_predictions += 1
for (block, delta, pred_block,
pred_block2) in parameters['signal_blocks']:
ninputs += np.prod(block.shape) * input_block_features
for (block, delta, factor) in parameters['context_blocks']:
ncontext += np.prod(block.shape)
for (block, delta, pred_block,
pred_block2) in parameters['signal_blocks']:
noutputs += np.prod(block.shape) * output_predictions
nreadout = 0
for (block, delta, pblock) in parameters['readout_blocks']:
nreadout += np.prod(block.shape)
parameters["Primary_Predictor_params"] = {}
parameters["Readout_Predictor_params"] = {}
if complex_unit and complex_context_in_second_layer: # 4 layer perceptron
parameters["Primary_Predictor_params"]['layers'] = MLP.get_layers([
ninputs + 1, 2 * nhidden + ncontext + 1, nhidden + 1,
noutputs + 1
])
elif complex_unit:
parameters["Primary_Predictor_params"]['layers'] = MLP.get_layers([
ninputs + ncontext + 1, 2 * nhidden + 1, nhidden + 1,
noutputs + 1
])
else: # 3 layer perceptron Simple MLP Unit (not complex unit)
parameters["Primary_Predictor_params"]['layers'] = MLP.get_layers(
[ninputs + ncontext + 1, nhidden + 1, noutputs + 1])
parameters["Primary_Predictor_params"][
'beta'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["Primary_Predictor_params"]['beta'][0] = 1.0
parameters["Primary_Predictor_params"]['learning_rate'] = parameters[
'primary_learning_rate']
parameters["Primary_Predictor_params"]['momentum'] = parameters[
'momentum']
parameters["Primary_Predictor_params"][
'mse'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["Primary_Predictor_params"]['weights'] = MLP.get_weights(
parameters["Primary_Predictor_params"]['layers'])
parameters["Readout_Predictor_params"]['layers'] = MLP.get_layers(
[nhidden + 1, 2 * nhidden + 1, nreadout + 1])
parameters["Readout_Predictor_params"][
'beta'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["Readout_Predictor_params"]['beta'][0] = 1.0
parameters["Readout_Predictor_params"]['learning_rate'] = parameters[
'readout_learning_rate']
parameters["Readout_Predictor_params"]['momentum'] = parameters[
'momentum']
parameters["Readout_Predictor_params"][
'mse'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["Readout_Predictor_params"]['weights'] = MLP.get_weights(
parameters["Readout_Predictor_params"]['layers'])
parameters["Primary_Predictor_params"]['polynomial'] = polynomial
parameters["Readout_Predictor_params"]['polynomial'] = polynomial
parameters['autoencoder'] = autoencoder
parameters['use_global_backprop'] = use_global_backprop
parameters[
"complex_context_in_second_layer"] = complex_context_in_second_layer
parameters["complex"] = complex_unit
def generate_v1(name, description, options):
input_block_size = int(options["input_block_size"])
hidden_size = int(options["hidden_block_size"])
layer_shape = map(lambda x: int(x), options["layer_shapes"])
readout_block_size = map(lambda x: int(x), options["readout_block_size"])
readout_layer = map(lambda x: x == "1", options["enable_readout"])
lateral_radius = float(options["lateral_radius"])
fan_in_square_size = int(options["fan_in_square_size"])
fan_in_radius = int(options["fan_in_radius"])
readout_depth = int(options["readout_depth"])
ex_module = options["ex_module"]
exclude_self = (options["context_exclude_self"] == '1')
last_layer_context_to_all = (options["last_layer_context_to_all"] == '1')
send_context_two_layers_back = (
options["send_context_two_layers_back"] == '1')
simulation_dict = create_blank_dictionary(
name=name,
description=description,
save_sources=(options["save_source_files"] == '1'))
simulation_dict['stages'] = 1
simulation_dict['num_proc'] = 2 * mp.cpu_count() / 3
simulation_dict['stage0'] = []
simulation_dict['execution_unit_module'] = ex_module
simulation_dict['version_major'] = 1
simulation_dict['version_minor'] = 0
unit = importlib.import_module(simulation_dict['execution_unit_module'])
blocks_per_dim = layer_shape
layers = len(blocks_per_dim)
error_log = SharedArray.SharedNumpyArray((layers + 1, 1000000), np.float)
simulation_dict['error_log'] = error_log
simulation_dict['input_block_size'] = input_block_size
simulation_dict['hidden_size'] = hidden_size
simulation_dict['learning_rates'] = []
simulation_dict['momenta'] = []
simulation_dict['taus'] = []
simulation_dict['predicted_arrays'] = []
simulation_dict['predicted_arrays_t2'] = []
simulation_dict['predicted_readout_arrays'] = []
simulation_dict['readout_arrays'] = []
simulation_dict['state_arrays'] = []
simulation_dict['delta_arrays'] = []
input_array = SharedArray.SharedNumpyArray(
(input_block_size * blocks_per_dim[0],
input_block_size * blocks_per_dim[0], 3), np.uint8)
simulation_dict['input_array'] = input_array
input_array_float = SharedArray.SharedNumpyArray(
(input_block_size * blocks_per_dim[0],
input_block_size * blocks_per_dim[0], 3), np.float)
simulation_dict['input_array_float'] = input_array_float
for (i, bpd) in enumerate(blocks_per_dim):
if readout_layer[i]:
readout_array_float00 = SharedArray.SharedNumpyArray(
(bpd * readout_block_size[i], bpd * readout_block_size[i],
readout_depth), np.float)
simulation_dict['readout_arrays'].append(readout_array_float00)
predicted_readout_array_float00 = SharedArray.SharedNumpyArray(
(bpd * readout_block_size[i], bpd * readout_block_size[i],
readout_depth), np.float)
simulation_dict['predicted_readout_arrays'].append(
predicted_readout_array_float00)
# input array 0 is a special case, 3 dimensions because of color input
predicted_array0 = SharedArray.SharedNumpyArray(
(input_block_size * blocks_per_dim[0],
input_block_size * blocks_per_dim[0], 3), np.float)
simulation_dict['predicted_arrays'].append(predicted_array0)
predicted_array2 = SharedArray.SharedNumpyArray(
(input_block_size * blocks_per_dim[0],
input_block_size * blocks_per_dim[0], 3), np.float)
simulation_dict['predicted_arrays_t2'].append(predicted_array2)
delta_array0 = SharedArray.SharedNumpyArray(
(input_block_size * blocks_per_dim[0],
input_block_size * blocks_per_dim[0], 3), np.float)
simulation_dict['delta_arrays'].append(delta_array0)
# All the rest is generic
for (i, bpd) in enumerate(blocks_per_dim[:-1]):
predicted_array1 = SharedArray.SharedNumpyArray(
(hidden_size * bpd, hidden_size * bpd), np.float)
simulation_dict['predicted_arrays'].append(predicted_array1)
predicted_array2 = SharedArray.SharedNumpyArray(
(hidden_size * bpd, hidden_size * bpd), np.float)
simulation_dict['predicted_arrays_t2'].append(predicted_array2)
delta_array1 = SharedArray.SharedNumpyArray(
(hidden_size * bpd, hidden_size * bpd), np.float)
simulation_dict['delta_arrays'].append(delta_array1)
for (i, bpd) in enumerate(blocks_per_dim):
state_array0 = SharedArray.SharedNumpyArray(
(hidden_size * bpd, hidden_size * bpd), np.float)
simulation_dict['state_arrays'].append(state_array0)
# Base learning rate
for (i, bpd) in enumerate(blocks_per_dim):
learning_rate = SharedArray.SharedNumpyArray((1, ), np.float)
learning_rate[0] = 0.0
simulation_dict['learning_rates'].append(learning_rate)
additional_learning_rate = SharedArray.SharedNumpyArray((1, ), np.float)
additional_learning_rate[0] = 0.0
simulation_dict['readout_learning_rate'] = additional_learning_rate
# Momentum is the same everywhere.
momentum = SharedArray.SharedNumpyArray((1, ), np.float)
momentum[0] = float(options["momentum"])
simulation_dict['momentum'] = momentum
# Tau is the integration constant for the signal integral
tau = SharedArray.SharedNumpyArray((1, ), np.float)
tau[0] = float(options['tau'])
simulation_dict['tau'] = tau
context_factor_lateral = SharedArray.SharedNumpyArray((1, ), np.float)
context_factor_lateral[0] = 0.0
simulation_dict['context_factor_lateral'] = context_factor_lateral
context_factor_feedback = SharedArray.SharedNumpyArray((1, ), np.float)
context_factor_feedback[0] = 0.0
simulation_dict['context_factor_feedback'] = context_factor_feedback
base_index = [0]
for bpd in blocks_per_dim:
base_index.append(base_index[-1] + bpd * bpd)
# Layer 0 is specific and has to be constructed separately
for i in xrange(blocks_per_dim[0] * blocks_per_dim[0]):
unit_parameters = create_basic_unit_v1(
simulation_dict['learning_rates'][0], momentum, tau,
additional_learning_rate)
x = (i / blocks_per_dim[0]) * input_block_size
y = (i % blocks_per_dim[0]) * input_block_size
dx = input_block_size
dy = input_block_size
input_block = SharedArray.DynamicView(input_array_float)[x:x + dx,
y:y + dy]
predicted_block = SharedArray.DynamicView(
simulation_dict['predicted_arrays'][0])[x:x + dx, y:y + dy]
predicted_block_2 = SharedArray.DynamicView(
simulation_dict['predicted_arrays_t2'][0])[x:x + dx, y:y + dy]
delta_block = SharedArray.DynamicView(
simulation_dict['delta_arrays'][0])[x:x + dx, y:y + dy]
if not (predicted_block.shape == (dx, dy, 3)):
print predicted_block.shape
raise Exception("Block sizes don't agree")
k = (i / blocks_per_dim[0]) * hidden_size
l = (i % blocks_per_dim[0]) * hidden_size
output_block = SharedArray.DynamicView(
simulation_dict['state_arrays'][0])[k:k + hidden_size,
l:l + hidden_size]
unit_parameters['signal_blocks'].append(
(input_block, delta_block, predicted_block, predicted_block_2))
unit_parameters['output_block'] = output_block
if readout_layer[0]:
# Motor heatmap prediction
layer = 0
bpd = blocks_per_dim[layer]
readout_teaching_block = SharedArray.DynamicView(
simulation_dict['readout_arrays']
[layer])[(i / bpd) * readout_block_size[0]:(i / bpd + 1) *
readout_block_size[0],
(i % bpd) * readout_block_size[0]:(i % bpd + 1) *
readout_block_size[0]]
readout_delta_block = SharedArray.SharedNumpyArray_like(
readout_teaching_block)
predicted_readout_block = SharedArray.DynamicView(
simulation_dict['predicted_readout_arrays']
[layer])[(i / bpd) * readout_block_size[0]:(i / bpd + 1) *
readout_block_size[0],
(i % bpd) * readout_block_size[0]:(i % bpd + 1) *
readout_block_size[0]]
unit_parameters['readout_blocks'] = [
(readout_teaching_block, readout_delta_block,
predicted_readout_block)
]
unit_parameters["layer"] = 0
# End motor heatmap prediction
simulation_dict['stage0'].append(unit_parameters)
# Layer 0 surround
gather_surround(simulation_dict, (base_index[0], blocks_per_dim[0]),
radius=lateral_radius,
context_factor=context_factor_lateral,
exclude_self=exclude_self)
# The following layers are more generic
for layer in range(1, layers):
for i in xrange(blocks_per_dim[layer] * blocks_per_dim[layer]):
unit_parameters = create_basic_unit_v1(
simulation_dict['learning_rates'][layer], momentum, tau,
additional_learning_rate)
k = (i / blocks_per_dim[layer]) * hidden_size
l = (i % blocks_per_dim[layer]) * hidden_size
output_block = SharedArray.DynamicView(
simulation_dict['state_arrays'][layer])[k:k + hidden_size,
l:l + hidden_size]
unit_parameters['output_block'] = output_block
if readout_layer[layer]:
# Motor heatmap prediction
bpd = blocks_per_dim[layer]
readout_teaching_block = SharedArray.DynamicView(
simulation_dict['readout_arrays'][layer])[
(i / bpd) * readout_block_size[layer]:(i / bpd + 1) *
readout_block_size[layer],
(i % bpd) * readout_block_size[layer]:(i % bpd + 1) *
readout_block_size[layer]]
readout_delta_block = SharedArray.SharedNumpyArray_like(
readout_teaching_block)
predicted_readout_block = SharedArray.DynamicView(
simulation_dict['predicted_readout_arrays'][layer])[
(i / bpd) * readout_block_size[layer]:(i / bpd + 1) *
readout_block_size[layer],
(i % bpd) * readout_block_size[layer]:(i % bpd + 1) *
readout_block_size[layer]]
unit_parameters['readout_blocks'] = [
(readout_teaching_block, readout_delta_block,
predicted_readout_block)
]
unit_parameters["layer"] = layer
# End motor heatmap prediction
simulation_dict['stage0'].append(unit_parameters)
# Connect to the previous layer
connect_forward_and_back_v1(
simulation_dict, (base_index[layer - 1], blocks_per_dim[layer - 1],
simulation_dict['predicted_arrays'][layer],
simulation_dict['predicted_arrays_t2'][layer]),
(base_index[layer], blocks_per_dim[layer]),
square_size=fan_in_square_size,
radius=fan_in_radius,
context_factor=context_factor_feedback)
# Layer surround
gather_surround(simulation_dict,
(base_index[layer], blocks_per_dim[layer]),
radius=lateral_radius,
context_factor=context_factor_lateral,
exclude_self=exclude_self)
if send_context_two_layers_back and layer > 1:
connect_back(simulation_dict,
(base_index[layer], blocks_per_dim[layer]),
(base_index[layer - 2], blocks_per_dim[layer - 2]),
square_size=2 * fan_in_square_size,
radius=2 * fan_in_radius,
context_factor=context_factor_feedback)
# Add the global feedback from the top layer
if last_layer_context_to_all:
logging.info("Connecting last layer back to everyone")
for to_idx in xrange(base_index[layers - 1]):
for from_idx in range(base_index[layers - 1],
len(simulation_dict["stage0"])):
context_block = simulation_dict['stage0'][from_idx][
'output_block']
delta_block2 = SharedArray.SharedNumpyArray_like(context_block)
simulation_dict['stage0'][from_idx]['delta_blocks'].append(
delta_block2)
# Connect the context block to the source
simulation_dict['stage0'][to_idx]['context_blocks'].append(
(context_block, delta_block2, context_factor_feedback))
simulation_dict['stage0_size'] = len(simulation_dict['stage0'])
for i in range(simulation_dict['stage0_size']):
simulation_dict['stage0'][i]['flags'] = simulation_dict['flags']
unit.ExecutionUnit.generate_missing_parameters(
simulation_dict['stage0'][i], options=options)
return simulation_dict
for y in range(blocks_per_dim1):
surround = get_fan_in((x, y),
dim_x_l=blocks_per_dim0,
dim_y_l=blocks_per_dim0,
dim_x_u=blocks_per_dim1,
dim_y_u=blocks_per_dim1,
block_x=square_size,
block_y=square_size,
radius=radius)
dest = index1 + x * (blocks_per_dim1) + y # destination unit
for xy in surround:
source = index0 + xy[0] * blocks_per_dim0 + xy[
1] # source unit
# Prepare the input and corresponding delta block at source
input_block = simulation_dict['stage0'][source]['output_block']
delta_block = SharedArray.SharedNumpyArray_like(input_block)
simulation_dict['stage0'][source]['delta_blocks'].append(
delta_block)
# Prepare the context and corresonding delta block at destination
context_block = simulation_dict['stage0'][dest]['output_block']
delta_block2 = SharedArray.SharedNumpyArray_like(context_block)
simulation_dict['stage0'][dest]['delta_blocks'].append(
delta_block2)
# Connect the context block to the source
simulation_dict['stage0'][source]['context_blocks'].append(
(context_block, delta_block2, context_factor))
# Prepare the predicted blocks
xx = xy[0] * hidden_size
yy = xy[1] * hidden_size
assert (predicted_array[xx:xx + dx, yy:yy +
dy].shape == context_block.shape)
def generate_missing_parameters(parameters):
"""
This method can be called to generate all the missing dictionary parameters when all
the other relevant variables are known. Leave empty if there is nothing more to generate
"""
ninputs = ((2 * parameters['array'].shape[0]) /
3) * parameters['block'][2] * parameters['block'][3] * 3
for block in parameters["context_blocks"]:
ninputs += np.prod(block.shape)
nhidden = parameters['internal_rep_size']
noutputs = (parameters['array'].shape[0] -
(2 * parameters['array'].shape[0]) /
3) * parameters['block'][2] * parameters['block'][3] * 3
parameters["MLP_parameters"] = {}
parameters["MLP_parameters"]['layers'] = [
{
'activation': SharedArray.SharedNumpyArray((ninputs + 1),
np.float),
'error': SharedArray.SharedNumpyArray((ninputs + 1), np.float),
'delta': SharedArray.SharedNumpyArray((ninputs + 1), np.float)
},
{
'activation': SharedArray.SharedNumpyArray((nhidden + 1),
np.float),
'error': SharedArray.SharedNumpyArray((nhidden + 1), np.float),
'delta': SharedArray.SharedNumpyArray((nhidden + 1), np.float)
},
{
'activation':
SharedArray.SharedNumpyArray((noutputs + 1), np.float),
'error':
SharedArray.SharedNumpyArray((noutputs + 1), np.float),
'delta':
SharedArray.SharedNumpyArray((noutputs + 1), np.float)
},
]
parameters["MLP_parameters"]['beta'] = SharedArray.SharedNumpyArray(
(1, ), np.float)
parameters["MLP_parameters"]['beta'][0] = 1.0
parameters["MLP_parameters"]['learning_rate'] = parameters[
'learning_rate']
parameters["MLP_parameters"]['momentum'] = parameters['momentum']
parameters["MLP_parameters"]['mse'] = SharedArray.SharedNumpyArray(
(1, ), np.float)
parameters["MLP_parameters"]['weights'] = [
MLP.initialize_weights(
SharedArray.SharedNumpyArray((ninputs + 1, nhidden),
np.float)),
MLP.initialize_weights(
SharedArray.SharedNumpyArray((nhidden + 1, noutputs),
np.float))
]
def generate_dict(blocks_per_dim=8, block_size=8):
# simulation dictionary has to have:
# N - SharedArray of one entity enumerating the step of the simulation
# stages - integer number of processing stages
# num_proc - integer number of processors that will be utilized for workers
# stage0 - a list of element parameters (dictionaries) for each execution element
# stage0_size - number of execution elements in the stage
# other parameters are model specific.
simulation_dict = PVM_Create.create_blank_dictionary()
simulation_dict['stages'] = 1
simulation_dict['num_proc'] = mp.cpu_count() / 2
simulation_dict['stage0'] = []
simulation_dict['execution_unit_module'] = 'PVM_models.demo03_unit'
unit = importlib.import_module(simulation_dict['execution_unit_module'])
internal_rep_size = 2 * block_size * block_size * 3 / 10
simulation_dict['stage0_size'] = 3 * blocks_per_dim * blocks_per_dim + 1
# Create the high resolution x 3 frame array
array = SharedArray.SharedNumpyArray(
(3, block_size * blocks_per_dim, block_size * blocks_per_dim, 3),
np.uint8)
simulation_dict['array'] = array
predicted_array = SharedArray.SharedNumpyArray(
(1, block_size * blocks_per_dim, block_size * blocks_per_dim, 3),
np.uint8)
simulation_dict['predicted_array'] = predicted_array
# Create the medium resolution x 12 frame array
array1 = SharedArray.SharedNumpyArray(
(4 * 3, block_size * blocks_per_dim / 2,
block_size * blocks_per_dim / 2, 3), np.uint8)
simulation_dict['array1'] = array1
predicted_array1 = SharedArray.SharedNumpyArray(
(4, block_size * blocks_per_dim / 2, block_size * blocks_per_dim / 2,
3), np.uint8)
simulation_dict['predicted_array1'] = predicted_array1
# Create the low resolution x 48 frame array
array2 = SharedArray.SharedNumpyArray(
(16 * 3, block_size * blocks_per_dim / 4,
block_size * blocks_per_dim / 4, 3), np.uint8)
simulation_dict['array2'] = array2
predicted_array2 = SharedArray.SharedNumpyArray(
(16, block_size * blocks_per_dim / 4, block_size * blocks_per_dim / 4,
3), np.uint8)
simulation_dict['predicted_array2'] = predicted_array2
# Create the low resolution x 3 frame array
array3 = SharedArray.SharedNumpyArray((3, block_size, block_size, 3),
np.uint8)
simulation_dict['array3'] = array3
predicted_array3 = SharedArray.SharedNumpyArray(
(1, block_size, block_size, 3), np.uint8)
simulation_dict['predicted_array3'] = predicted_array3
context_block = SharedArray.SharedNumpyArray((internal_rep_size, ),
np.float)
# Create some basic MLP parameters
learning_rate = SharedArray.SharedNumpyArray((1, ), np.float)
learning_rate[0] = 0.01
simulation_dict['learning_rate'] = learning_rate
momentum = SharedArray.SharedNumpyArray((1, ), np.float)
momentum[0] = 0.5
simulation_dict['momentum'] = momentum
# Run loops to create and connect all the units
# High spatial resolution, low temporal cover layer
for x in range(blocks_per_dim):
for y in range(blocks_per_dim):
output_block = SharedArray.SharedNumpyArray((internal_rep_size, ),
np.float)
block = SharedArray.SharedNumpyArray((4, ), np.int)
block[0] = x * block_size
block[1] = y * block_size
block[2] = block_size
block[3] = block_size
unit_parameters = create_unit_parameters(
learning_rate, momentum, block, array, predicted_array,
internal_rep_size, context_block, output_block)
simulation_dict['stage0'].append(unit_parameters)
# Medium spatial resolution, medium temporal cover layer
for x in range(blocks_per_dim):
for y in range(blocks_per_dim):
output_block = SharedArray.SharedNumpyArray((internal_rep_size, ),
np.float)
block = SharedArray.SharedNumpyArray((4, ), np.int)
block[0] = x * (block_size / 2)
block[1] = y * (block_size / 2)
block[2] = block_size / 2
block[3] = block_size / 2
unit_parameters = create_unit_parameters(
learning_rate, momentum, block, array1, predicted_array1,
internal_rep_size, context_block, output_block)
simulation_dict['stage0'].append(unit_parameters)
# Low spatial resolution, high temporal cover layer
for x in range(blocks_per_dim):
for y in range(blocks_per_dim):
output_block = SharedArray.SharedNumpyArray((internal_rep_size, ),
np.float)
block = SharedArray.SharedNumpyArray((4, ), np.int)
block[0] = x * (block_size / 4)
block[1] = y * (block_size / 4)
block[2] = block_size / 4
block[3] = block_size / 4
unit_parameters = create_unit_parameters(
learning_rate, momentum, block, array2, predicted_array2,
internal_rep_size, context_block, output_block)
simulation_dict['stage0'].append(unit_parameters)
# Each upper layer will provide its internal activations as context to the lower layer.
for x in range(blocks_per_dim):
for y in range(blocks_per_dim):
i = x * blocks_per_dim + y
simulation_dict['stage0'][
i + blocks_per_dim * blocks_per_dim]['context_blocks'].append(
simulation_dict['stage0'][i + 2 * blocks_per_dim *
blocks_per_dim]['output_block'])
simulation_dict['stage0'][i]['context_blocks'].append(
simulation_dict['stage0'][i + blocks_per_dim *
blocks_per_dim]['output_block'])
surround = get_neighbours(x, y, blocks_per_dim, blocks_per_dim)
surround_idx = map(lambda a: a[0] * blocks_per_dim + a[1],
surround)
for idx in surround_idx:
simulation_dict['stage0'][i]['context_blocks'].append(
simulation_dict['stage0'][idx]['output_block'])
simulation_dict['stage0'][
i +
blocks_per_dim * blocks_per_dim]['context_blocks'].append(
simulation_dict['stage0'][
idx +
blocks_per_dim * blocks_per_dim]['output_block'])
simulation_dict['stage0'][
i + 2 * blocks_per_dim *
blocks_per_dim]['context_blocks'].append(
simulation_dict['stage0']
[idx +
2 * blocks_per_dim * blocks_per_dim]['output_block'])
# This single unit will cover the entire scene and supply its internal activations as
# context to everyone.
for i in range(1):
output_block = context_block
block = SharedArray.SharedNumpyArray((4, ), np.int)
# each unit gets a 100x100 sub-block of the arena
block[0] = 0
block[1] = 0
block[2] = block_size
block[3] = block_size
unit_parameters = create_unit_parameters(learning_rate, momentum,
block, array3,
predicted_array3,
internal_rep_size,
context_block, output_block)
simulation_dict['stage0'].append(unit_parameters)
for i in range(simulation_dict['stage0_size']):
unit.ExecutionUnit.generate_missing_parameters(
simulation_dict['stage0'][i])
return simulation_dict
def generate_missing_parameters(parameters,
complex_unit=False,
complex_unit_extra_layer=False,
polynomial=False,
autoencoder=False):
"""
This method can be called to generate all the missing dictionary parameters when all
the other relevant variables are known. Leave empty if there is nothing more to generate.
When complex_unit is False, a standard 3-layer MLP is used.
When complex_unit is True, an MLP with additional hidden layers is used.
There needs to be no return value, the method leaves a side effect by modifying the perameters dict.
:param parameters: parameter dictionary
:type parameters: dict
"""
nhidden = np.prod(parameters['output_block'].shape)
ntohidden = np.prod(parameters['output_block'].shape)
ninputs = 0
noutputs = 0
ncontext = 0
for (block, delta, pred_block, past_block, dblock, iblock,
pred_block_local) in parameters['signal_blocks']:
ninputs += np.prod(block.shape) * len(
ExecutionUnit.UNSUPERVISED_SIGNAL_INPUTS)
for (block, delta, factor) in parameters['context_blocks']:
# ninputs += np.prod(block.shape)
ncontext += np.prod(block.shape)
for (block, delta, pred_block, past_block, dblock, iblock,
pred_block_local) in parameters['signal_blocks']:
noutputs += np.prod(block.shape)
nadditional = 0
for (block, delta, pblock, pdblock) in parameters['predicted_blocks']:
nadditional += np.prod(block.shape)
nmiddle = nhidden * 2
parameters["MLP_parameters"] = {}
parameters["MLP_parameters_additional"] = {}
if complex_unit:
mlp_layers = [{
'activation':
SharedArray.SharedNumpyArray((ninputs + 1), np.float),
'error':
SharedArray.SharedNumpyArray((ninputs + 1), np.float),
'delta':
SharedArray.SharedNumpyArray((ninputs + 1), np.float)
}, {
'activation':
SharedArray.SharedNumpyArray((nmiddle + ncontext + 1),
np.float),
'error':
SharedArray.SharedNumpyArray((nmiddle + ncontext + 1),
np.float),
'delta':
SharedArray.SharedNumpyArray((nmiddle + ncontext + 1),
np.float)
}, {
'activation':
SharedArray.SharedNumpyArray((nhidden + 1), np.float),
'error':
SharedArray.SharedNumpyArray((nhidden + 1), np.float),
'delta':
SharedArray.SharedNumpyArray((nhidden + 1), np.float)
}]
if complex_unit_extra_layer:
mlp_layers.append({
'activation':
SharedArray.SharedNumpyArray((nhidden + 1), np.float),
'error':
SharedArray.SharedNumpyArray((nhidden + 1), np.float),
'delta':
SharedArray.SharedNumpyArray((nhidden + 1), np.float)
})
mlp_layers.append({
'activation':
SharedArray.SharedNumpyArray((noutputs + 1), np.float),
'error':
SharedArray.SharedNumpyArray((noutputs + 1), np.float),
'delta':
SharedArray.SharedNumpyArray((noutputs + 1), np.float)
})
parameters["MLP_parameters"]['layers'] = mlp_layers
parameters["MLP_parameters"][
'beta'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["MLP_parameters"]['beta'][0] = 1.0
parameters["MLP_parameters"]['learning_rate'] = parameters[
'learning_rate']
parameters["MLP_parameters"]['momentum'] = parameters['momentum']
parameters["MLP_parameters"]['mse'] = SharedArray.SharedNumpyArray(
(1, ), np.float)
mlp_weights = [
MLP.MLP.initialize_weights(
SharedArray.SharedNumpyArray((ninputs + 1, nmiddle),
np.float)),
MLP.MLP.initialize_weights(
SharedArray.SharedNumpyArray(
(nmiddle + ncontext + 1, nhidden), np.float))
]
if complex_unit_extra_layer:
mlp_weights.append(
MLP.MLP.initialize_weights(
SharedArray.SharedNumpyArray((nhidden + 1, nhidden),
np.float)))
mlp_weights.append(
MLP.MLP.initialize_weights(
SharedArray.SharedNumpyArray((nhidden + 1, noutputs),
np.float)))
parameters["MLP_parameters"]['weights'] = mlp_weights
parameters["complex"] = True
else: # Simple MLP Unit (not complex unit)
parameters["MLP_parameters"]['layers'] = [
{
'activation':
SharedArray.SharedNumpyArray((ninputs + ncontext + 1),
np.float),
'error':
SharedArray.SharedNumpyArray((ninputs + ncontext + 1),
np.float),
'delta':
SharedArray.SharedNumpyArray((ninputs + ncontext + 1),
np.float)
},
{
'activation':
SharedArray.SharedNumpyArray((nhidden + 1), np.float),
'error':
SharedArray.SharedNumpyArray((nhidden + 1), np.float),
'delta':
SharedArray.SharedNumpyArray((nhidden + 1), np.float)
},
{
'activation':
SharedArray.SharedNumpyArray((noutputs + 1), np.float),
'error':
SharedArray.SharedNumpyArray((noutputs + 1), np.float),
'delta':
SharedArray.SharedNumpyArray((noutputs + 1), np.float)
},
]
parameters["MLP_parameters"][
'beta'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["MLP_parameters"]['beta'][0] = 1.0
parameters["MLP_parameters"]['learning_rate'] = parameters[
'learning_rate']
parameters["MLP_parameters"]['momentum'] = parameters['momentum']
parameters["MLP_parameters"]['mse'] = SharedArray.SharedNumpyArray(
(1, ), np.float)
parameters["MLP_parameters"]['weights'] = [
MLP.MLP.initialize_weights(
SharedArray.SharedNumpyArray(
(ninputs + ncontext + 1, nhidden), np.float)),
MLP.MLP.initialize_weights(
SharedArray.SharedNumpyArray((nhidden + 1, noutputs),
np.float)),
]
parameters["MLP_parameters_additional"]['layers'] = [
{
'activation': SharedArray.SharedNumpyArray((nhidden + 1),
np.float),
'error': SharedArray.SharedNumpyArray((nhidden + 1), np.float),
'delta': SharedArray.SharedNumpyArray((nhidden + 1), np.float)
},
{
'activation':
SharedArray.SharedNumpyArray((nadditional + 1), np.float),
'error':
SharedArray.SharedNumpyArray((nadditional + 1), np.float),
'delta':
SharedArray.SharedNumpyArray((nadditional + 1), np.float)
},
]
parameters["MLP_parameters_additional"][
'beta'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["MLP_parameters_additional"]['beta'][0] = 1.0
parameters["MLP_parameters_additional"]['learning_rate'] = parameters[
'additional_learning_rate']
parameters["MLP_parameters_additional"]['momentum'] = parameters[
'momentum']
parameters["MLP_parameters_additional"][
'mse'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["MLP_parameters_additional"]['weights'] = [
MLP.MLP.initialize_weights(
SharedArray.SharedNumpyArray((nhidden + 1, nadditional),
np.float)),
]
if polynomial:
parameters["MLP_parameters"]['polynomial'] = True
parameters["MLP_parameters_additional"]['polynomial'] = True
if autoencoder:
parameters['autoencoder'] = True
def upgrade_to_ver_1(parameters):
parameters['internal_buffers'] = []
parameters['output_min'] = SharedArray.SharedNumpyArray_like(
parameters['output_block'])
parameters['output_max'] = SharedArray.SharedNumpyArray_like(
parameters['output_block'])
parameters['avg_delta'] = SharedArray.SharedNumpyArray_like(
parameters['output_block'])
parameters['integral_blocks'] = []
parameters['derivative_blocks'] = []
parameters['error_blocks'] = []
parameters['use_derivative'] = True
parameters['use_integral'] = True
parameters['use_error'] = True
parameters['use_t_2_block'] = False
parameters['predict_2_steps'] = False
parameters['use_global_backprop'] = False
parameters['normalize_output'] = False
parameters["complex_context_in_second_layer"] = False
for (block, delta, pred_block,
pred_block2) in parameters['signal_blocks']:
block01 = SharedArray.SharedNumpyArray_like(block)
block02 = SharedArray.SharedNumpyArray_like(block)
block03 = SharedArray.SharedNumpyArray_like(block)
parameters['internal_buffers'].append((block01, block02, block03))
parameters['derivative_blocks'].append(
SharedArray.SharedNumpyArray_like(block))
parameters['integral_blocks'].append(
SharedArray.SharedNumpyArray_like(block))
parameters['error_blocks'].append(
SharedArray.SharedNumpyArray_like(block))
if "complex" not in parameters.keys():
parameters["complex"] = False
if len(parameters["Primary_Predictor_params"]['layers']) == 4:
parameters["complex"] = True
if "autoencoder" not in parameters.keys():
parameters["autoencoder"] = False
if "readout_learning_rate" not in parameters.keys():
parameters['readout_learning_rate'] = parameters[
"Primary_Predictor_params"]["learning_rate"]
if "momentum" not in parameters.keys():
parameters['momentum'] = parameters["Primary_Predictor_params"][
"momentum"]
nhidden = parameters["Primary_Predictor_params"]['layers'][-2][
'activation'].shape[0] - 1
nreadout = 0
nouputs = 0
for (block, delta, pred_block,
pred_block2) in parameters['signal_blocks']:
nouputs += np.prod(block.shape)
for (block, delta, pblock) in parameters['readout_blocks']:
nreadout += np.prod(block.shape)
if "Readout_Predictor_params" not in parameters.keys():
parameters["Readout_Predictor_params"] = {}
parameters["Readout_Predictor_params"]['layers'] = MLP.get_layers(
[nhidden + 1, nreadout + 1])
parameters["Readout_Predictor_params"][
'beta'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["Readout_Predictor_params"]['beta'][0] = 1.0
parameters["Readout_Predictor_params"][
'learning_rate'] = parameters['readout_learning_rate']
parameters["Readout_Predictor_params"]['momentum'] = parameters[
'momentum']
parameters["Readout_Predictor_params"][
'mse'] = SharedArray.SharedNumpyArray((1, ), np.float)
parameters["Readout_Predictor_params"][
'weights'] = MLP.get_weights(
parameters["Readout_Predictor_params"]['layers'])
parameters["Readout_Predictor_params"]['weights'][
0][:] = parameters["Primary_Predictor_params"]['weights'][
-1][:, nouputs:]
old_weight_matrix = parameters["Primary_Predictor_params"][
'weights'][-1]
parameters["Primary_Predictor_params"]['weights'][
-1] = SharedArray.SharedNumpyArray((nhidden + 1, nouputs),
np.float)
parameters["Primary_Predictor_params"]['weights'][
-1][:] = old_weight_matrix[:, :nouputs]
parameters["Primary_Predictor_params"]['layers'][-1] = {
'activation': SharedArray.SharedNumpyArray(nouputs, np.float),
'error': SharedArray.SharedNumpyArray(nouputs, np.float),
'delta': SharedArray.SharedNumpyArray(nouputs, np.float)
}
parameters['backpropagate_readout_error'] = True
