def main(config, coordinates, gid, field_width, peak_rate, input_features_path,
input_features_namespaces, output_features_namespace,
output_weights_path, output_features_path, initial_weights_path,
reference_weights_path, h5types_path, synapse_name,
initial_weights_namespace, reference_weights_namespace,
output_weights_namespace, reference_weights_are_delta,
connections_path, optimize_method, destination, sources, arena_id,
max_delta_weight, field_width_scale, max_iter, verbose, dry_run,
plot):
"""
:param config: str (path to .yaml file)
:param coordinates: tuple of float
:param gid: int
:param field_width: float
:param peak_rate: float
:param input_features_path: str (path to .h5 file)
:param input_features_namespaces: str
:param output_features_namespace: str
:param output_weights_path: str (path to .h5 file)
:param output_features_path: str (path to .h5 file)
:param initial_weights_path: str (path to .h5 file)
:param reference_weights_path: str (path to .h5 file)
:param h5types_path: str (path to .h5 file)
:param synapse_name: str
:param initial_weights_namespace: str
:param output_weights_namespace: str
:param reference_weights_are_delta: bool
:param connections_path: str (path to .h5 file)
:param destination: str (population name)
:param sources: list of str (population name)
:param arena_id: str
:param max_delta_weight: float
:param field_width_scale: float
:param max_iter: int
:param verbose: bool
:param dry_run: bool
:param interactive: bool
:param plot: bool
"""
utils.config_logging(verbose)
logger = utils.get_script_logger(__file__)
env = Env(config_file=config)
if not dry_run:
if output_weights_path is None:
raise RuntimeError(
'Missing required argument: output_weights_path.')
if not os.path.isfile(output_weights_path):
if initial_weights_path is not None and os.path.isfile(
initial_weights_path):
input_file_path = initial_weights_path
elif h5types_path is not None and os.path.isfile(h5types_path):
input_file_path = h5types_path
else:
raise RuntimeError(
'Missing required source for h5types: either an initial_weights_path or an '
'h5types_path must be provided.')
with h5py.File(output_weights_path, 'a') as output_file:
with h5py.File(input_file_path, 'r') as input_file:
input_file.copy('/H5Types', output_file)
this_input_features_namespaces = [
'%s %s' % (input_features_namespace, arena_id)
for input_features_namespace in input_features_namespaces
]
features_attr_names = ['Arena Rate Map']
spatial_resolution = env.stimulus_config['Spatial Resolution'] # cm
arena = env.stimulus_config['Arena'][arena_id]
default_run_vel = arena.properties['default run velocity'] # cm/s
arena_x, arena_y = stimulus.get_2D_arena_spatial_mesh(
arena, spatial_resolution)
dim_x = len(arena_x)
dim_y = len(arena_y)
if gid is None:
target_gids = []
else:
target_gids = [gid]
dst_input_features = defaultdict(dict)
num_fields = len(coordinates)
this_field_width = np.array([field_width] * num_fields, dtype=np.float32)
this_scaled_field_width = np.array([field_width * field_width_scale] *
num_fields,
dtype=np.float32)
this_peak_rate = np.array([peak_rate] * num_fields, dtype=np.float32)
this_x0 = np.array([x for x, y in coordinates], dtype=np.float32)
this_y0 = np.array([y for x, y in coordinates], dtype=np.float32)
this_rate_map = np.asarray(get_rate_map(this_x0, this_y0, this_field_width,
this_peak_rate, arena_x, arena_y),
dtype=np.float32)
target_map = np.asarray(get_rate_map(this_x0, this_y0,
this_scaled_field_width,
this_peak_rate, arena_x, arena_y),
dtype=np.float32)
selectivity_type = env.selectivity_types['place']
dst_input_features[destination][target_gid] = {
'Selectivity Type': np.array([selectivity_type], dtype=np.uint8),
'Num Fields': np.array([num_fields], dtype=np.uint8),
'Field Width': this_field_width,
'Peak Rate': this_peak_rate,
'X Offset': this_x0,
'Y Offset': this_y0,
'Arena Rate Map': this_rate_map.ravel()
}
initial_weights_by_syn_id_dict = dict()
selection = [target_gid]
if initial_weights_path is not None:
initial_weights_iter = \
read_cell_attribute_selection(initial_weights_path, destination, namespace=initial_weights_namespace,
selection=selection)
syn_weight_attr_dict = dict(initial_weights_iter)
syn_ids = syn_weight_attr_dict[target_gid]['syn_id']
weights = syn_weight_attr_dict[target_gid][synapse_name]
for (syn_id, weight) in zip(syn_ids, weights):
initial_weights_by_syn_id_dict[int(syn_id)] = float(weight)
logger.info(
'destination: %s; gid %i; read initial synaptic weights for %i synapses'
% (destination, target_gid, len(initial_weights_by_syn_id_dict)))
reference_weights_by_syn_id_dict = None
if reference_weights_path is not None:
reference_weights_by_syn_id_dict = dict()
reference_weights_iter = \
read_cell_attribute_selection(reference_weights_path, destination, namespace=reference_weights_namespace,
selection=selection)
syn_weight_attr_dict = dict(reference_weights_iter)
syn_ids = syn_weight_attr_dict[target_gid]['syn_id']
weights = syn_weight_attr_dict[target_gid][synapse_name]
for (syn_id, weight) in zip(syn_ids, weights):
reference_weights_by_syn_id_dict[int(syn_id)] = float(weight)
logger.info(
'destination: %s; gid %i; read reference synaptic weights for %i synapses'
% (destination, target_gid, len(reference_weights_by_syn_id_dict)))
source_gid_set_dict = defaultdict(set)
syn_ids_by_source_gid_dict = defaultdict(list)
initial_weights_by_source_gid_dict = dict()
if reference_weights_by_syn_id_dict is None:
reference_weights_by_source_gid_dict = None
else:
reference_weights_by_source_gid_dict = dict()
(graph, edge_attr_info) = read_graph_selection(file_name=connections_path,
selection=[target_gid],
namespaces=['Synapses'])
syn_id_attr_index = None
for source, edge_iter in viewitems(graph[destination]):
if source not in sources:
continue
this_edge_attr_info = edge_attr_info[destination][source]
if 'Synapses' in this_edge_attr_info and \
'syn_id' in this_edge_attr_info['Synapses']:
syn_id_attr_index = this_edge_attr_info['Synapses']['syn_id']
for (destination_gid, edges) in edge_iter:
assert destination_gid == target_gid
source_gids, edge_attrs = edges
syn_ids = edge_attrs['Synapses'][syn_id_attr_index]
count = 0
for i in range(len(source_gids)):
this_source_gid = int(source_gids[i])
source_gid_set_dict[source].add(this_source_gid)
this_syn_id = int(syn_ids[i])
if this_syn_id not in initial_weights_by_syn_id_dict:
this_weight = \
env.connection_config[destination][source].mechanisms['default'][synapse_name]['weight']
initial_weights_by_syn_id_dict[this_syn_id] = this_weight
syn_ids_by_source_gid_dict[this_source_gid].append(this_syn_id)
if this_source_gid not in initial_weights_by_source_gid_dict:
initial_weights_by_source_gid_dict[this_source_gid] = \
initial_weights_by_syn_id_dict[this_syn_id]
if reference_weights_by_source_gid_dict is not None:
reference_weights_by_source_gid_dict[this_source_gid] = \
reference_weights_by_syn_id_dict[this_syn_id]
count += 1
logger.info(
'destination: %s; gid %i; set initial synaptic weights for %d inputs from source population '
'%s' % (destination, destination_gid, count, source))
syn_count_by_source_gid_dict = dict()
for source_gid in syn_ids_by_source_gid_dict:
syn_count_by_source_gid_dict[source_gid] = len(
syn_ids_by_source_gid_dict[source_gid])
input_rate_maps_by_source_gid_dict = dict()
for source in sources:
source_gids = list(source_gid_set_dict[source])
for input_features_namespace in this_input_features_namespaces:
input_features_iter = read_cell_attribute_selection(
input_features_path,
source,
namespace=input_features_namespace,
mask=set(features_attr_names),
selection=source_gids)
count = 0
for gid, attr_dict in input_features_iter:
input_rate_maps_by_source_gid_dict[gid] = attr_dict[
'Arena Rate Map'].reshape((dim_x, dim_y))
count += 1
logger.info('Read %s feature data for %i cells in population %s' %
(input_features_namespace, count, source))
if is_interactive:
context.update(locals())
normalized_delta_weights_dict, arena_LS_map = \
synapses.generate_structured_weights(target_map=target_map,
initial_weight_dict=initial_weights_by_source_gid_dict,
input_rate_map_dict=input_rate_maps_by_source_gid_dict,
syn_count_dict=syn_count_by_source_gid_dict,
max_delta_weight=max_delta_weight, arena_x=arena_x, arena_y=arena_y,
reference_weight_dict=reference_weights_by_source_gid_dict,
reference_weights_are_delta=reference_weights_are_delta,
reference_weights_namespace=reference_weights_namespace,
optimize_method=optimize_method, verbose=verbose, plot=plot)
output_syn_ids = np.empty(len(initial_weights_by_syn_id_dict),
dtype='uint32')
output_weights = np.empty(len(initial_weights_by_syn_id_dict),
dtype='float32')
i = 0
for source_gid, this_weight in viewitems(normalized_delta_weights_dict):
for syn_id in syn_ids_by_source_gid_dict[source_gid]:
output_syn_ids[i] = syn_id
output_weights[i] = this_weight
i += 1
output_weights_dict = {
target_gid: {
'syn_id': output_syn_ids,
synapse_name: output_weights
}
}
logger.info('destination: %s; gid %i; generated %s for %i synapses' %
(destination, target_gid, output_weights_namespace,
len(output_weights)))
if not dry_run:
this_output_weights_namespace = '%s %s' % (output_weights_namespace,
arena_id)
logger.info('Destination: %s; appending %s ...' %
(destination, this_output_weights_namespace))
append_cell_attributes(output_weights_path,
destination,
output_weights_dict,
namespace=this_output_weights_namespace)
logger.info('Destination: %s; appended %s' %
(destination, this_output_weights_namespace))
output_weights_dict.clear()
if output_features_path is not None:
this_output_features_namespace = '%s %s' % (
output_features_namespace, arena_id)
cell_attr_dict = dst_input_features[destination]
cell_attr_dict[target_gid]['Arena State Map'] = np.asarray(
arena_LS_map.ravel(), dtype=np.float32)
logger.info('Destination: %s; appending %s ...' %
(destination, this_output_features_namespace))
append_cell_attributes(output_features_path,
destination,
cell_attr_dict,
namespace=this_output_features_namespace)
if is_interactive:
context.update(locals())
def main(config, coordinates, field_width, gid, input_features_path,
input_features_namespaces, initial_weights_path,
output_features_namespace, output_features_path, output_weights_path,
reference_weights_path, h5types_path, synapse_name,
initial_weights_namespace, output_weights_namespace,
reference_weights_namespace, connections_path, destination, sources,
non_structured_sources, non_structured_weights_namespace,
non_structured_weights_path, arena_id, field_width_scale,
max_delta_weight, max_opt_iter, max_weight_decay_fraction,
optimize_method, optimize_tol, optimize_grad, peak_rate,
reference_weights_are_delta, arena_margin, target_amplitude, io_size,
chunk_size, value_chunk_size, cache_size, write_size, verbose,
dry_run, plot, show_fig, save_fig):
"""
:param config: str (path to .yaml file)
:param input_features_path: str (path to .h5 file)
:param initial_weights_path: str (path to .h5 file)
:param initial_weights_namespace: str
:param output_weights_namespace: str
:param connections_path: str (path to .h5 file)
:param destination: str
:param sources: list of str
:param io_size:
:param chunk_size:
:param value_chunk_size:
:param cache_size:
:param write_size:
:param verbose:
:param dry_run:
:return:
"""
utils.config_logging(verbose)
logger = utils.get_script_logger(__file__)
comm = MPI.COMM_WORLD
rank = comm.rank
nranks = comm.size
if io_size == -1:
io_size = comm.size
if rank == 0:
logger.info('%s: %i ranks have been allocated' % (__file__, comm.size))
env = Env(comm=comm, config_file=config, io_size=io_size)
if plot and (not save_fig) and (not show_fig):
show_fig = True
if (not dry_run) and (rank == 0):
if not os.path.isfile(output_weights_path):
if initial_weights_path is not None:
input_file = h5py.File(initial_weights_path, 'r')
elif h5types_path is not None:
input_file = h5py.File(h5types_path, 'r')
else:
raise RuntimeError(
'h5types input path must be specified when weights path is not specified.'
)
output_file = h5py.File(output_weights_path, 'w')
input_file.copy('/H5Types', output_file)
input_file.close()
output_file.close()
env.comm.barrier()
LTD_output_weights_namespace = 'LTD %s %s' % (output_weights_namespace,
arena_id)
LTP_output_weights_namespace = 'LTP %s %s' % (output_weights_namespace,
arena_id)
this_input_features_namespaces = [
'%s %s' % (input_features_namespace, arena_id)
for input_features_namespace in input_features_namespaces
]
selectivity_type_index = {
i: n
for n, i in viewitems(env.selectivity_types)
}
target_selectivity_type_name = 'place'
target_selectivity_type = env.selectivity_types[
target_selectivity_type_name]
features_attrs = defaultdict(dict)
source_features_attr_names = [
'Selectivity Type', 'Num Fields', 'Field Width', 'Peak Rate',
'Module ID', 'Grid Spacing', 'Grid Orientation',
'Field Width Concentration Factor', 'X Offset', 'Y Offset'
]
target_features_attr_names = [
'Selectivity Type', 'Num Fields', 'Field Width', 'Peak Rate',
'X Offset', 'Y Offset'
]
local_random = np.random.RandomState()
seed_offset = int(
env.model_config['Random Seeds']['GC Structured Weights'])
spatial_resolution = env.stimulus_config['Spatial Resolution'] # cm
arena = env.stimulus_config['Arena'][arena_id]
default_run_vel = arena.properties['default run velocity'] # cm/s
gid_count = 0
start_time = time.time()
target_gid_set = None
if len(gid) > 0:
target_gid_set = set(gid)
all_sources = sources + non_structured_sources
connection_gen_list = [ NeuroH5ProjectionGen(connections_path, source, destination, namespaces=['Synapses'], comm=comm) \
for source in all_sources ]
output_features_dict = {}
LTP_output_weights_dict = {}
LTD_output_weights_dict = {}
for iter_count, attr_gen_package in enumerate(
zip_longest(*connection_gen_list)):
local_time = time.time()
this_gid = attr_gen_package[0][0]
if not all([
attr_gen_items[0] == this_gid
for attr_gen_items in attr_gen_package
]):
raise Exception(
'Rank: %i; destination: %s; this_gid not matched across multiple attribute '
'generators: %s' %
(rank, destination,
[attr_gen_items[0] for attr_gen_items in attr_gen_package]))
if (target_gid_set is not None) and (this_gid not in target_gid_set):
continue
if this_gid is None:
selection = []
logger.info('Rank: %i received None' % rank)
else:
selection = [this_gid]
local_random.seed(int(this_gid + seed_offset))
has_structured_weights = False
dst_input_features_attr_dict = {}
for input_features_namespace in this_input_features_namespaces:
input_features_iter = read_cell_attribute_selection(
input_features_path,
destination,
namespace=input_features_namespace,
mask=set(target_features_attr_names),
comm=env.comm,
selection=selection)
count = 0
for gid, attr_dict in input_features_iter:
dst_input_features_attr_dict[gid] = attr_dict
count += 1
if rank == 0:
logger.info(
'Read %s feature data for %i cells in population %s' %
(input_features_namespace, count, destination))
arena_margin_size = 0.
arena_margin = max(arena_margin, 0.)
target_selectivity_features_dict = {}
target_selectivity_config_dict = {}
target_field_width_dict = {}
for gid in selection:
target_selectivity_features_dict[
gid] = dst_input_features_attr_dict.get(gid, {})
target_selectivity_features_dict[gid][
'Selectivity Type'] = np.asarray([target_selectivity_type],
dtype=np.uint8)
num_fields = target_selectivity_features_dict[gid]['Num Fields'][0]
if coordinates[0] is not None:
num_fields = 1
target_selectivity_features_dict[gid]['X Offset'] = np.asarray(
[coordinates[0]], dtype=np.float32)
target_selectivity_features_dict[gid]['Y Offset'] = np.asarray(
[coordinates[1]], dtype=np.float32)
target_selectivity_features_dict[gid][
'Num Fields'] = np.asarray([num_fields], dtype=np.uint8)
if field_width is not None:
target_selectivity_features_dict[gid][
'Field Width'] = np.asarray([field_width] * num_fields,
dtype=np.float32)
else:
this_field_width = target_selectivity_features_dict[gid][
'Field Width']
target_selectivity_features_dict[gid][
'Field Width'] = this_field_width[:num_fields]
if peak_rate is not None:
target_selectivity_features_dict[gid][
'Peak Rate'] = np.asarray([peak_rate] * num_fields,
dtype=np.float32)
input_cell_config = stimulus.get_input_cell_config(
target_selectivity_type,
selectivity_type_index,
selectivity_attr_dict=target_selectivity_features_dict[gid])
if input_cell_config.num_fields > 0:
arena_margin_size = max(
arena_margin_size,
np.max(input_cell_config.field_width) * arena_margin)
target_field_width_dict[gid] = input_cell_config.field_width
target_selectivity_config_dict[gid] = input_cell_config
has_structured_weights = True
arena_x, arena_y = stimulus.get_2D_arena_spatial_mesh(
arena, spatial_resolution, margin=arena_margin_size)
for gid, input_cell_config in viewitems(
target_selectivity_config_dict):
target_map = np.asarray(input_cell_config.get_rate_map(
arena_x, arena_y, scale=field_width_scale),
dtype=np.float32)
target_selectivity_features_dict[gid][
'Arena Rate Map'] = target_map
if not has_structured_weights:
selection = []
initial_weights_by_syn_id_dict = defaultdict(lambda: dict())
initial_weights_by_source_gid_dict = defaultdict(lambda: dict())
if initial_weights_path is not None:
initial_weights_iter = \
read_cell_attribute_selection(initial_weights_path, destination,
namespace=initial_weights_namespace,
selection=selection)
initial_weights_gid_count = 0
initial_weights_syn_count = 0
for this_gid, syn_weight_attr_dict in initial_weights_iter:
syn_ids = syn_weight_attr_dict['syn_id']
weights = syn_weight_attr_dict[synapse_name]
for (syn_id, weight) in zip(syn_ids, weights):
initial_weights_by_syn_id_dict[this_gid][int(
syn_id)] = float(weight)
initial_weights_gid_count += 1
initial_weights_syn_count += len(syn_ids)
logger.info(
'destination: %s; read initial synaptic weights for %i gids and %i syns'
% (destination, initial_weights_gid_count,
initial_weights_syn_count))
if len(non_structured_sources) > 0:
non_structured_weights_by_syn_id_dict = defaultdict(lambda: dict())
non_structured_weights_by_source_gid_dict = defaultdict(
lambda: dict())
else:
non_structured_weights_by_syn_id_dict = None
if non_structured_weights_path is not None:
non_structured_weights_iter = \
read_cell_attribute_selection(initial_weights_path, destination,
namespace=non_structured_weights_namespace,
selection=selection)
non_structured_weights_gid_count = 0
non_structured_weights_syn_count = 0
for this_gid, syn_weight_attr_dict in non_structured_weights_iter:
syn_ids = syn_weight_attr_dict['syn_id']
weights = syn_weight_attr_dict[synapse_name]
for (syn_id, weight) in zip(syn_ids, weights):
non_structured_weights_by_syn_id_dict[this_gid][int(
syn_id)] = float(weight)
non_structured_weights_gid_count += 1
non_structured_weights_syn_count += len(syn_ids)
logger.info(
'destination: %s; read non-structured synaptic weights for %i gids and %i syns'
% (
destination,
non_structured_weights_gid_count,
non_structured_weights_syn_count,
))
reference_weights_by_syn_id_dict = None
reference_weights_by_source_gid_dict = defaultdict(lambda: dict())
if reference_weights_path is not None:
reference_weights_by_syn_id_dict = defaultdict(lambda: dict())
reference_weights_iter = \
read_cell_attribute_selection(reference_weights_path, destination, namespace=reference_weights_namespace,
selection=selection)
reference_weights_gid_count = 0
for this_gid, syn_weight_attr_dict in reference_weights_iter:
syn_ids = syn_weight_attr_dict['syn_id']
weights = syn_weight_attr_dict[synapse_name]
for (syn_id, weight) in zip(syn_ids, weights):
reference_weights_by_syn_id_dict[this_gid][int(
syn_id)] = float(weight)
logger.info(
'destination: %s; read reference synaptic weights for %i gids'
% (destination, reference_weights_gid_count))
syn_count_by_source_gid_dict = defaultdict(int)
source_gid_set_dict = defaultdict(set)
syn_ids_by_source_gid_dict = defaultdict(list)
structured_syn_id_count = 0
if has_structured_weights:
for source, (destination_gid, (source_gid_array,
conn_attr_dict)) in zip_longest(
all_sources, attr_gen_package):
syn_ids = conn_attr_dict['Synapses']['syn_id']
count = 0
this_initial_weights_by_syn_id_dict = None
this_initial_weights_by_source_gid_dict = None
this_reference_weights_by_syn_id_dict = None
this_reference_weights_by_source_gid_dict = None
this_non_structured_weights_by_syn_id_dict = None
this_non_structured_weights_by_source_gid_dict = None
if destination_gid is not None:
this_initial_weights_by_syn_id_dict = initial_weights_by_syn_id_dict[
destination_gid]
this_initial_weights_by_source_gid_dict = initial_weights_by_source_gid_dict[
destination_gid]
if reference_weights_by_syn_id_dict is not None:
this_reference_weights_by_syn_id_dict = reference_weights_by_syn_id_dict[
destination_gid]
this_reference_weights_by_source_gid_dict = reference_weights_by_source_gid_dict[
destination_gid]
this_non_structured_weights_by_syn_id_dict = non_structured_weights_by_syn_id_dict[
destination_gid]
this_non_structured_weights_by_source_gid_dict = non_structured_weights_by_source_gid_dict[
destination_gid]
for i in range(len(source_gid_array)):
this_source_gid = source_gid_array[i]
this_syn_id = syn_ids[i]
if this_syn_id in this_initial_weights_by_syn_id_dict:
this_syn_wgt = this_initial_weights_by_syn_id_dict[
this_syn_id]
if this_source_gid not in this_initial_weights_by_source_gid_dict:
this_initial_weights_by_source_gid_dict[
this_source_gid] = this_syn_wgt
if this_reference_weights_by_syn_id_dict is not None:
this_reference_weights_by_source_gid_dict[this_source_gid] = \
this_reference_weights_by_syn_id_dict[this_syn_id]
elif this_syn_id in this_non_structured_weights_by_syn_id_dict:
this_syn_wgt = this_non_structured_weights_by_syn_id_dict[
this_syn_id]
if this_source_gid not in this_non_structured_weights_by_source_gid_dict:
this_non_structured_weights_by_source_gid_dict[
this_source_gid] = this_syn_wgt
source_gid_set_dict[source].add(this_source_gid)
syn_ids_by_source_gid_dict[this_source_gid].append(
this_syn_id)
syn_count_by_source_gid_dict[this_source_gid] += 1
count += 1
if source not in non_structured_sources:
structured_syn_id_count += len(syn_ids)
logger.info(
'Rank %i; destination: %s; gid %i; %d edges from source population %s'
% (rank, destination, this_gid, count, source))
input_rate_maps_by_source_gid_dict = {}
if len(non_structured_sources) > 0:
non_structured_input_rate_maps_by_source_gid_dict = {}
else:
non_structured_input_rate_maps_by_source_gid_dict = None
for source in all_sources:
if has_structured_weights:
source_gids = list(source_gid_set_dict[source])
else:
source_gids = []
if rank == 0:
logger.info(
'Reading %s feature data for %i cells in population %s...'
% (input_features_namespace, len(source_gids), source))
for input_features_namespace in this_input_features_namespaces:
input_features_iter = read_cell_attribute_selection(
input_features_path,
source,
namespace=input_features_namespace,
mask=set(source_features_attr_names),
comm=env.comm,
selection=source_gids)
count = 0
for gid, attr_dict in input_features_iter:
this_selectivity_type = attr_dict['Selectivity Type'][0]
this_selectivity_type_name = selectivity_type_index[
this_selectivity_type]
input_cell_config = stimulus.get_input_cell_config(
this_selectivity_type,
selectivity_type_index,
selectivity_attr_dict=attr_dict)
this_arena_rate_map = np.asarray(
input_cell_config.get_rate_map(arena_x, arena_y),
dtype=np.float32)
if source in non_structured_sources:
non_structured_input_rate_maps_by_source_gid_dict[
gid] = this_arena_rate_map
else:
input_rate_maps_by_source_gid_dict[
gid] = this_arena_rate_map
count += 1
if rank == 0:
logger.info(
'Read %s feature data for %i cells in population %s' %
(input_features_namespace, count, source))
if has_structured_weights:
if is_interactive:
context.update(locals())
save_fig_path = None
if save_fig is not None:
save_fig_path = '%s/Structured Weights %s %d.png' % (
save_fig, destination, this_gid)
normalized_LTP_delta_weights_dict, LTD_delta_weights_dict, arena_LS_map = \
synapses.generate_structured_weights(target_map=target_selectivity_features_dict[this_gid]['Arena Rate Map'],
initial_weight_dict=this_initial_weights_by_source_gid_dict,
reference_weight_dict=this_reference_weights_by_source_gid_dict,
reference_weights_are_delta=reference_weights_are_delta,
reference_weights_namespace=reference_weights_namespace,
input_rate_map_dict=input_rate_maps_by_source_gid_dict,
non_structured_input_rate_map_dict=non_structured_input_rate_maps_by_source_gid_dict,
non_structured_weights_dict=this_non_structured_weights_by_source_gid_dict,
syn_count_dict=syn_count_by_source_gid_dict,
max_delta_weight=max_delta_weight,
max_opt_iter=max_opt_iter,
max_weight_decay_fraction=max_weight_decay_fraction,
target_amplitude=target_amplitude,
arena_x=arena_x, arena_y=arena_y,
optimize_method=optimize_method,
optimize_tol=optimize_tol,
optimize_grad=optimize_grad,
verbose=verbose, plot=plot, show_fig=show_fig,
save_fig=save_fig_path,
fig_kwargs={'gid': this_gid,
'field_width': target_field_width_dict[this_gid]})
gc.collect()
this_selectivity_dict = target_selectivity_features_dict[this_gid]
output_features_dict[this_gid] = {
fld: this_selectivity_dict[fld]
for fld in [
'Selectivity Type', 'Num Fields', 'Field Width',
'Peak Rate', 'X Offset', 'Y Offset'
]
}
output_features_dict[this_gid]['Arena State Map'] = np.asarray(
arena_LS_map.ravel(), dtype=np.float32)
output_syn_ids = np.empty(structured_syn_id_count, dtype='uint32')
LTD_output_weights = np.empty(structured_syn_id_count,
dtype='float32')
LTP_output_weights = np.empty(structured_syn_id_count,
dtype='float32')
i = 0
for source_gid in normalized_LTP_delta_weights_dict:
for syn_id in syn_ids_by_source_gid_dict[source_gid]:
output_syn_ids[i] = syn_id
LTP_output_weights[i] = normalized_LTP_delta_weights_dict[
source_gid]
LTD_output_weights[i] = LTD_delta_weights_dict[source_gid]
i += 1
LTP_output_weights_dict[this_gid] = {
'syn_id': output_syn_ids,
synapse_name: LTP_output_weights
}
LTD_output_weights_dict[this_gid] = {
'syn_id': output_syn_ids,
synapse_name: LTD_output_weights
}
logger.info(
'Rank %i; destination: %s; gid %i; generated structured weights for %i inputs in %.2f '
's' % (rank, destination, this_gid, len(output_syn_ids),
time.time() - local_time))
gid_count += 1
if iter_count % write_size == 0:
if not dry_run:
append_cell_attributes(output_weights_path,
destination,
LTD_output_weights_dict,
namespace=LTD_output_weights_namespace,
comm=env.comm,
io_size=env.io_size,
chunk_size=chunk_size,
value_chunk_size=value_chunk_size)
append_cell_attributes(output_weights_path,
destination,
LTP_output_weights_dict,
namespace=LTP_output_weights_namespace,
comm=env.comm,
io_size=env.io_size,
chunk_size=chunk_size,
value_chunk_size=value_chunk_size)
count = comm.reduce(len(LTP_output_weights_dict),
op=MPI.SUM,
root=0)
if rank == 0:
logger.info(
'Destination: %s; appended weights for %i cells' %
(destination, count))
if output_features_path is not None:
if output_features_namespace is None:
output_features_namespace = '%s Selectivity' % target_selectivity_type_name.title(
)
this_output_features_namespace = '%s %s' % (
output_features_namespace, arena_id)
logger.info(str(output_features_dict))
append_cell_attributes(
output_features_path,
destination,
output_features_dict,
namespace=this_output_features_namespace)
count = comm.reduce(len(output_features_dict),
op=MPI.SUM,
root=0)
if rank == 0:
logger.info(
'Destination: %s; appended selectivity features for %i cells'
% (destination, count))
LTP_output_weights_dict.clear()
LTD_output_weights_dict.clear()
output_features_dict.clear()
gc.collect()
env.comm.barrier()
if not dry_run:
append_cell_attributes(output_weights_path,
destination,
LTD_output_weights_dict,
namespace=LTD_output_weights_namespace,
comm=env.comm,
io_size=env.io_size,
chunk_size=chunk_size,
value_chunk_size=value_chunk_size)
append_cell_attributes(output_weights_path,
destination,
LTP_output_weights_dict,
namespace=LTP_output_weights_namespace,
comm=env.comm,
io_size=env.io_size,
chunk_size=chunk_size,
value_chunk_size=value_chunk_size)
count = comm.reduce(len(LTP_output_weights_dict), op=MPI.SUM, root=0)
if rank == 0:
logger.info('Destination: %s; appended weights for %i cells' %
(destination, count))
if output_features_path is not None:
if output_features_namespace is None:
output_features_namespace = 'Selectivity Features'
this_output_features_namespace = '%s %s' % (
output_features_namespace, arena_id)
append_cell_attributes(output_features_path,
destination,
output_features_dict,
namespace=this_output_features_namespace)
count = comm.reduce(len(output_features_dict), op=MPI.SUM, root=0)
if rank == 0:
logger.info(
'Destination: %s; appended selectivity features for %i cells'
% (destination, count))
env.comm.barrier()
global_count = comm.gather(gid_count, root=0)
if rank == 0:
logger.info(
'destination: %s; %i ranks assigned structured weights to %i cells in %.2f s'
% (destination, comm.size, np.sum(global_count),
time.time() - start_time))
def main(config, coordinates, field_width, gid, input_features_path,
input_features_namespaces, initial_weights_path,
output_features_namespace, output_features_path, output_weights_path,
reference_weights_path, h5types_path, synapse_name,
initial_weights_namespace, output_weights_namespace,
reference_weights_namespace, connections_path, destination, sources,
non_structured_sources, non_structured_weights_namespace,
non_structured_weights_path, arena_id, field_width_scale,
max_opt_iter, max_weight_decay_fraction, optimize_tol, peak_rate,
reference_weights_are_delta, arena_margin, target_amplitude, io_size,
chunk_size, value_chunk_size, cache_size, write_size, verbose,
dry_run, plot, show_fig, save_fig, debug):
"""
:param config: str (path to .yaml file)
:param input_features_path: str (path to .h5 file)
:param initial_weights_path: str (path to .h5 file)
:param initial_weights_namespace: str
:param output_weights_namespace: str
:param connections_path: str (path to .h5 file)
:param destination: str
:param sources: list of str
:param io_size:
:param chunk_size:
:param value_chunk_size:
:param write_size:
:param verbose:
:param dry_run:
:return:
"""
utils.config_logging(verbose)
script_name = __file__
logger = utils.get_script_logger(script_name)
comm = MPI.COMM_WORLD
rank = comm.rank
nranks = comm.size
if io_size == -1:
io_size = comm.size
if rank == 0:
logger.info(f'{comm.size} ranks have been allocated')
env = Env(comm=comm, config_file=config, io_size=io_size)
env.comm.barrier()
if plot and (not save_fig) and (not show_fig):
show_fig = True
if (not dry_run) and (rank == 0):
if not os.path.isfile(output_weights_path):
if initial_weights_path is not None:
input_file = h5py.File(initial_weights_path, 'r')
elif h5types_path is not None:
input_file = h5py.File(h5types_path, 'r')
else:
raise RuntimeError(
'h5types input path must be specified when weights path is not specified.'
)
output_file = h5py.File(output_weights_path, 'w')
input_file.copy('/H5Types', output_file)
input_file.close()
output_file.close()
env.comm.barrier()
LTD_output_weights_namespace = f'LTD {output_weights_namespace} {arena_id}'
LTP_output_weights_namespace = f'LTP {output_weights_namespace} {arena_id}'
this_input_features_namespaces = [
f'{input_features_namespace} {arena_id}'
for input_features_namespace in input_features_namespaces
]
selectivity_type_index = {
i: n
for n, i in viewitems(env.selectivity_types)
}
target_selectivity_type_name = 'place'
target_selectivity_type = env.selectivity_types[
target_selectivity_type_name]
features_attrs = defaultdict(dict)
source_features_attr_names = [
'Selectivity Type', 'Num Fields', 'Field Width', 'Peak Rate',
'Module ID', 'Grid Spacing', 'Grid Orientation',
'Field Width Concentration Factor', 'X Offset', 'Y Offset'
]
target_features_attr_names = [
'Selectivity Type', 'Num Fields', 'Field Width', 'Peak Rate',
'X Offset', 'Y Offset'
]
seed_offset = int(
env.model_config['Random Seeds']['GC Structured Weights'])
spatial_resolution = env.stimulus_config['Spatial Resolution'] # cm
arena = env.stimulus_config['Arena'][arena_id]
default_run_vel = arena.properties['default run velocity'] # cm/s
gid_count = 0
start_time = time.time()
target_gid_set = None
if len(gid) > 0:
target_gid_set = set(gid)
projections = [(source, destination) for source in sources]
graph_info = read_graph_info(connections_path,
namespaces=['Connections', 'Synapses'],
read_node_index=True)
for projection in projections:
if projection not in graph_info:
raise RuntimeError(
f'Projection {projection[0]} -> {projection[1]} is not present in connections file.'
)
if target_gid_set is None:
target_gid_set = set(graph_info[projection][1])
all_sources = sources + non_structured_sources
src_input_features_attr_dict = {source: {} for source in all_sources}
for source in sorted(all_sources):
this_src_input_features_attr_dict = {}
for this_input_features_namespace in this_input_features_namespaces:
logger.info(
f'Rank {rank}: Reading {this_input_features_namespace} feature data for cells in population {source}'
)
input_features_dict = scatter_read_cell_attributes(
input_features_path,
source,
namespaces=[this_input_features_namespace],
mask=set(source_features_attr_names),
comm=env.comm,
io_size=env.io_size)
for gid, attr_dict in input_features_dict[
this_input_features_namespace]:
this_src_input_features_attr_dict[gid] = attr_dict
src_input_features_attr_dict[
source] = this_src_input_features_attr_dict
source_gid_count = env.comm.reduce(
len(this_src_input_features_attr_dict), op=MPI.SUM, root=0)
if rank == 0:
logger.info(
f'Rank {rank}: Read feature data for {source_gid_count} cells in population {source}'
)
dst_gids = []
if target_gid_set is not None:
for i, gid in enumerate(target_gid_set):
if i % nranks == rank:
dst_gids.append(gid)
dst_input_features_attr_dict = {}
for this_input_features_namespace in this_input_features_namespaces:
feature_count = 0
gid_count = 0
logger.info(
f'Rank {rank}: reading {this_input_features_namespace} feature data for {len(dst_gids)} cells in population {destination}'
)
input_features_iter = scatter_read_cell_attribute_selection(
input_features_path,
destination,
namespace=this_input_features_namespace,
mask=set(target_features_attr_names),
selection=dst_gids,
io_size=env.io_size,
comm=env.comm)
for gid, attr_dict in input_features_iter:
gid_count += 1
if (len(coordinates) > 0) or (attr_dict['Num Fields'][0] > 0):
dst_input_features_attr_dict[gid] = attr_dict
feature_count += 1
logger.info(
f'Rank {rank}: read {this_input_features_namespace} feature data for '
f'{gid_count} / {feature_count} cells in population {destination}')
feature_count = env.comm.reduce(feature_count, op=MPI.SUM, root=0)
env.comm.barrier()
if rank == 0:
logger.info(
f'Read {this_input_features_namespace} feature data for {feature_count} cells in population {destination}'
)
feature_dst_gids = list(dst_input_features_attr_dict.keys())
all_feature_gids_per_rank = comm.allgather(feature_dst_gids)
all_feature_gids = sorted(
[item for sublist in all_feature_gids_per_rank for item in sublist])
request_dst_gids = []
for i, gid in enumerate(all_feature_gids):
if i % nranks == rank:
request_dst_gids.append(gid)
dst_input_features_attr_dict = exchange_input_features(
env.comm, request_dst_gids, dst_input_features_attr_dict)
dst_gids = list(dst_input_features_attr_dict.keys())
if rank == 0:
logger.info(
f"Rank {rank} feature dict is {dst_input_features_attr_dict}")
dst_count = env.comm.reduce(len(dst_gids), op=MPI.SUM, root=0)
logger.info(f"Rank {rank} has {len(dst_gids)} feature gids")
if rank == 0:
logger.info(f'Total {dst_count} feature gids')
max_dst_count = env.comm.allreduce(len(dst_gids), op=MPI.MAX)
env.comm.barrier()
max_iter_count = max_dst_count
output_features_dict = {}
LTP_output_weights_dict = {}
LTD_output_weights_dict = {}
non_structured_output_weights_dict = {}
for iter_count in range(max_iter_count):
gc.collect()
local_time = time.time()
selection = []
if len(dst_gids) > 0:
dst_gid = dst_gids.pop()
selection.append(dst_gid)
logger.info(f'Rank {rank} received gid {dst_gid}')
env.comm.barrier()
arena_margin_size = 0.
arena_margin = max(arena_margin, 0.)
target_selectivity_features_dict = {}
target_selectivity_config_dict = {}
target_field_width_dict = {}
for destination_gid in selection:
arena_margin_size = init_selectivity_config(
destination_gid,
spatial_resolution,
arena,
arena_margin,
arena_margin_size,
coordinates,
field_width,
field_width_scale,
peak_rate,
target_selectivity_type,
selectivity_type_index,
dst_input_features_attr_dict,
target_selectivity_features_dict,
target_selectivity_config_dict,
target_field_width_dict,
logger=logger)
arena_x, arena_y = stimulus.get_2D_arena_spatial_mesh(
arena, spatial_resolution, margin=arena_margin_size)
selection = list(target_selectivity_features_dict.keys())
initial_weights_by_source_gid_dict = defaultdict(lambda: dict())
initial_weights_by_syn_id_dict = \
read_weights(initial_weights_path, initial_weights_namespace, synapse_name,
destination, selection, env.comm, env.io_size, defaultdict(lambda: dict()),
logger=logger if rank == 0 else None)
non_structured_weights_by_source_gid_dict = defaultdict(lambda: dict())
non_structured_weights_by_syn_id_dict = None
if len(non_structured_sources) > 0:
non_structured_weights_by_syn_id_dict = \
read_weights(non_structured_weights_path, non_structured_weights_namespace, synapse_name,
destination, selection, env.comm, env.io_size, defaultdict(lambda: dict()),
logger=logger if rank == 0 else None)
reference_weights_by_syn_id_dict = None
reference_weights_by_source_gid_dict = defaultdict(lambda: dict())
if reference_weights_path is not None:
reference_weights_by_syn_id_dict = \
read_weights(reference_weights_path, reference_weights_namespace, synapse_name,
destination, selection, env.comm, env.io_size, defaultdict(lambda: dict()),
logger=logger if rank == 0 else None)
source_gid_set_dict = defaultdict(set)
syn_count_by_source_gid_dict = defaultdict(lambda: defaultdict(int))
syn_ids_by_source_gid_dict = defaultdict(lambda: defaultdict(list))
structured_syn_id_count = defaultdict(int)
non_structured_syn_id_count = defaultdict(int)
projections = [(source, destination) for source in all_sources]
edge_iter_dict, edge_attr_info = scatter_read_graph_selection(
connections_path,
selection=selection,
namespaces=['Synapses'],
projections=projections,
comm=env.comm,
io_size=env.io_size)
syn_counts_by_source = init_syn_weight_dicts(
destination, non_structured_sources, edge_iter_dict,
edge_attr_info, initial_weights_by_syn_id_dict,
initial_weights_by_source_gid_dict,
non_structured_weights_by_syn_id_dict,
non_structured_weights_by_source_gid_dict,
reference_weights_by_syn_id_dict,
reference_weights_by_source_gid_dict, source_gid_set_dict,
syn_count_by_source_gid_dict, syn_ids_by_source_gid_dict,
structured_syn_id_count, non_structured_syn_id_count)
for source in syn_counts_by_source:
for this_gid in syn_counts_by_source[source]:
count = syn_counts_by_source[source][this_gid]
logger.info(
f'Rank {rank}: destination: {destination}; gid {this_gid}; '
f'{count} edges from source population {source}')
input_rate_maps_by_source_gid_dict = {}
if len(non_structured_sources) > 0:
non_structured_input_rate_maps_by_source_gid_dict = {}
else:
non_structured_input_rate_maps_by_source_gid_dict = None
for source in all_sources:
source_gids = list(source_gid_set_dict[source])
if rank == 0:
logger.info(
f'Rank {rank}: getting feature data for {len(source_gids)} cells in population {source}'
)
this_src_input_features = exchange_input_features(
env.comm, source_gids, src_input_features_attr_dict[source])
count = 0
for this_gid in source_gids:
attr_dict = this_src_input_features[this_gid]
this_selectivity_type = attr_dict['Selectivity Type'][0]
this_selectivity_type_name = selectivity_type_index[
this_selectivity_type]
input_cell_config = stimulus.get_input_cell_config(
this_selectivity_type,
selectivity_type_index,
selectivity_attr_dict=attr_dict)
this_arena_rate_map = np.asarray(
input_cell_config.get_rate_map(arena_x, arena_y),
dtype=np.float32)
if source in non_structured_sources:
non_structured_input_rate_maps_by_source_gid_dict[
this_gid] = this_arena_rate_map
else:
input_rate_maps_by_source_gid_dict[
this_gid] = this_arena_rate_map
count += 1
for destination_gid in selection:
if is_interactive:
context.update(locals())
save_fig_path = None
if save_fig is not None:
save_fig_path = f'{save_fig}/Structured Weights {destination} {destination_gid}.png'
reference_weight_dict = None
if reference_weights_path is not None:
reference_weight_dict = reference_weights_by_source_gid_dict[
destination_gid]
LTP_delta_weights_dict, LTD_delta_weights_dict, arena_structured_map = \
synapses.generate_structured_weights(destination_gid,
target_map=target_selectivity_features_dict[destination_gid]['Arena Rate Map'],
initial_weight_dict=initial_weights_by_source_gid_dict[destination_gid],
#reference_weight_dict=reference_weight_dict,
#reference_weights_are_delta=reference_weights_are_delta,
#reference_weights_namespace=reference_weights_namespace,
input_rate_map_dict=input_rate_maps_by_source_gid_dict,
non_structured_input_rate_map_dict=non_structured_input_rate_maps_by_source_gid_dict,
non_structured_weights_dict=non_structured_weights_by_source_gid_dict[destination_gid],
syn_count_dict=syn_count_by_source_gid_dict[destination_gid],
max_opt_iter=max_opt_iter,
max_weight_decay_fraction=max_weight_decay_fraction,
target_amplitude=target_amplitude,
arena_x=arena_x, arena_y=arena_y,
optimize_tol=optimize_tol,
verbose=verbose if rank == 0 else False,
plot=plot, show_fig=show_fig,
save_fig=save_fig_path,
fig_kwargs={'gid': destination_gid,
'field_width': target_field_width_dict[destination_gid]})
input_rate_maps_by_source_gid_dict.clear()
target_map_flat = target_selectivity_features_dict[
destination_gid]['Arena Rate Map'].flat
arena_map_residual_mae = np.mean(
np.abs(arena_structured_map - target_map_flat))
output_features_dict[destination_gid] = \
{ fld: target_selectivity_features_dict[destination_gid][fld]
for fld in ['Selectivity Type',
'Num Fields',
'Field Width',
'Peak Rate',
'X Offset',
'Y Offset',]}
output_features_dict[destination_gid][
'Rate Map Residual Mean Error'] = np.asarray(
[arena_map_residual_mae], dtype=np.float32)
this_structured_syn_id_count = structured_syn_id_count[
destination_gid]
output_syn_ids = np.empty(this_structured_syn_id_count,
dtype='uint32')
LTD_output_weights = np.empty(this_structured_syn_id_count,
dtype='float32')
LTP_output_weights = np.empty(this_structured_syn_id_count,
dtype='float32')
i = 0
for source_gid in LTP_delta_weights_dict:
for syn_id in syn_ids_by_source_gid_dict[destination_gid][
source_gid]:
output_syn_ids[i] = syn_id
LTP_output_weights[i] = LTP_delta_weights_dict[source_gid]
LTD_output_weights[i] = LTD_delta_weights_dict[source_gid]
i += 1
LTP_output_weights_dict[destination_gid] = {
'syn_id': output_syn_ids,
synapse_name: LTP_output_weights
}
LTD_output_weights_dict[destination_gid] = {
'syn_id': output_syn_ids,
synapse_name: LTD_output_weights
}
this_non_structured_syn_id_count = non_structured_syn_id_count[
destination_gid]
i = 0
logger.info(
f'Rank {rank}; destination: {destination}; gid {destination_gid}; '
f'generated structured weights for {len(output_syn_ids)} inputs in {time.time() - local_time:.2f} s; '
f'residual error is {arena_map_residual_mae:.2f}')
gid_count += 1
gc.collect()
env.comm.barrier()
if (write_size > 0) and (iter_count % write_size == 0):
if not dry_run:
append_cell_attributes(output_weights_path,
destination,
LTD_output_weights_dict,
namespace=LTD_output_weights_namespace,
comm=env.comm,
io_size=env.io_size,
chunk_size=chunk_size,
value_chunk_size=value_chunk_size)
append_cell_attributes(output_weights_path,
destination,
LTP_output_weights_dict,
namespace=LTP_output_weights_namespace,
comm=env.comm,
io_size=env.io_size,
chunk_size=chunk_size,
value_chunk_size=value_chunk_size)
count = env.comm.reduce(len(LTP_output_weights_dict),
op=MPI.SUM,
root=0)
env.comm.barrier()
if rank == 0:
logger.info(
f'Destination: {destination}; appended weights for {count} cells'
)
if output_features_path is not None:
if output_features_namespace is None:
output_features_namespace = f'{target_selectivity_type_name.title()} Selectivity'
this_output_features_namespace = f'{output_features_namespace} {arena_id}'
append_cell_attributes(
output_features_path,
destination,
output_features_dict,
namespace=this_output_features_namespace)
count = env.comm.reduce(len(output_features_dict),
op=MPI.SUM,
root=0)
env.comm.barrier()
if rank == 0:
logger.info(
f'Destination: {destination}; appended selectivity features for {count} cells'
)
LTP_output_weights_dict.clear()
LTD_output_weights_dict.clear()
output_features_dict.clear()
gc.collect()
env.comm.barrier()
if (iter_count >= 10) and debug:
break
env.comm.barrier()
if not dry_run:
append_cell_attributes(output_weights_path,
destination,
LTD_output_weights_dict,
namespace=LTD_output_weights_namespace,
comm=env.comm,
io_size=env.io_size,
chunk_size=chunk_size,
value_chunk_size=value_chunk_size)
append_cell_attributes(output_weights_path,
destination,
LTP_output_weights_dict,
namespace=LTP_output_weights_namespace,
comm=env.comm,
io_size=env.io_size,
chunk_size=chunk_size,
value_chunk_size=value_chunk_size)
count = comm.reduce(len(LTP_output_weights_dict), op=MPI.SUM, root=0)
env.comm.barrier()
if rank == 0:
logger.info(
f'Destination: {destination}; appended weights for {count} cells'
)
if output_features_path is not None:
if output_features_namespace is None:
output_features_namespace = 'Selectivity Features'
this_output_features_namespace = f'{output_features_namespace} {arena_id}'
append_cell_attributes(output_features_path,
destination,
output_features_dict,
namespace=this_output_features_namespace)
count = env.comm.reduce(len(output_features_dict),
op=MPI.SUM,
root=0)
env.comm.barrier()
if rank == 0:
logger.info(
f'Destination: {destination}; appended selectivity features for {count} cells'
)
env.comm.barrier()
global_count = env.comm.gather(gid_count, root=0)
env.comm.barrier()
if rank == 0:
total_count = np.sum(global_count)
total_time = time.time() - start_time
logger.info(
f'Destination: {destination}; '
f'{env.comm.size} ranks assigned structured weights to {total_count} cells in {total_time:.2f} s'
)
def main(config, coordinates, gid, field_width, peak_rate, input_features_path, input_features_namespaces,
output_weights_path, output_features_path, weights_path, h5types_path, synapse_name, initial_weights_namespace,
structured_weights_namespace, connections_path, destination, sources, arena_id, baseline_weight,
field_width_scale, max_iter, verbose, dry_run, interactive):
"""
:param config: str (path to .yaml file)
:param weights_path: str (path to .h5 file)
:param initial_weights_namespace: str
:param structured_weights_namespace: str
:param connections_path: str (path to .h5 file)
:param destination: str
:param sources: list of str
:param verbose:
:param dry_run:
:return:
"""
utils.config_logging(verbose)
logger = utils.get_script_logger(__file__)
env = Env(config_file=config)
if output_weights_path is None:
if weights_path is None:
raise RuntimeError('Output weights path must be specified when weights path is not specified.')
output_weights_path = weights_path
if (not dry_run):
if not os.path.isfile(output_weights_path):
if weights_path is not None:
input_file = h5py.File(weights_path,'r')
elif h5types_path is not None:
input_file = h5py.File(h5types_path,'r')
else:
raise RuntimeError('h5types input path must be specified when weights path is not specified.')
output_file = h5py.File(output_weights_path,'w')
input_file.copy('/H5Types',output_file)
input_file.close()
output_file.close()
this_input_features_namespaces = ['%s %s' % (input_features_namespace, arena_id) for input_features_namespace in input_features_namespaces]
initial_weights_dict = None
if weights_path is not None:
logger.info('Reading initial weights data from %s...' % weights_path)
cell_attributes_dict = read_cell_attribute_selection(weights_path, destination,
namespaces=[initial_weights_namespace],
selection=[gid])
if initial_weights_namespace in cell_attributes_dict:
initial_weights_iter = cell_attributes_dict[initial_weights_namespace]
initial_weights_dict = { gid: attr_dict for gid, attr_dict in initial_weights_iter }
else:
raise RuntimeError('Initial weights namespace %s was not found in file %s' % (initial_weights_namespace, weights_path))
logger.info('Rank %i; destination: %s; read synaptic weights for %i cells' %
(env.comm.rank, destination, len(initial_weights_dict)))
features_attr_names = ['Num Fields', 'Field Width', 'Peak Rate', 'X Offset', 'Y Offset', 'Arena Rate Map']
local_random = np.random.RandomState()
seed_offset = int(env.model_config['Random Seeds']['GC Structured Weights'])
local_random.seed(int(gid + seed_offset))
spatial_resolution = env.stimulus_config['Spatial Resolution'] # cm
arena = env.stimulus_config['Arena'][arena_id]
default_run_vel = arena.properties['default run velocity'] # cm/s
x, y = stimulus.get_2D_arena_spatial_mesh(arena, spatial_resolution)
plasticity_kernel = lambda x, y, x_loc, y_loc, sx, sy: gauss2d(x-x_loc, y-y_loc, sx=sx, sy=sy)
plasticity_kernel = np.vectorize(plasticity_kernel, excluded=[2,3,4,5])
dst_input_features = defaultdict(dict)
num_fields = len(coordinates)
this_field_width = np.array([field_width]*num_fields, dtype=np.float32)
this_peak_rate = np.array([peak_rate]*num_fields, dtype=np.float32)
this_x0 = np.array([x for x, y in coordinates], dtype=np.float32)
this_y0 = np.array([y for x, y in coordinates], dtype=np.float32)
this_rate_map = np.asarray(get_rate_map(this_x0, this_y0, this_field_width, this_peak_rate, x, y),
dtype=np.float32)
selectivity_type = env.selectivity_types['place']
dst_input_features[destination][gid] = {
'Selectivity Type': np.array([selectivity_type], dtype=np.uint8),
'Num Fields': np.array([num_fields], dtype=np.uint8),
'Field Width': this_field_width,
'Peak Rate': this_peak_rate,
'X Offset': this_x0,
'Y Offset': this_y0,
'Arena Rate Map': this_rate_map.ravel() }
selection=[gid]
structured_weights_dict = {}
source_syn_dict = defaultdict(lambda: defaultdict(list))
syn_weight_dict = {}
if weights_path is not None:
initial_weights_iter = read_cell_attribute_selection(weights_path, destination,
namespace=initial_weights_namespace,
selection=selection)
syn_weight_attr_dict = dict(initial_weights_iter)
syn_ids = syn_weight_attr_dict[gid]['syn_id']
weights = syn_weight_attr_dict[gid][synapse_name]
for (syn_id, weight) in zip(syn_ids, weights):
syn_weight_dict[int(syn_id)] = float(weight)
logger.info('destination: %s; gid %i; received synaptic weights for %i synapses' %
(destination, gid, len(syn_weight_dict)))
(graph, edge_attr_info) = read_graph_selection(file_name=connections_path,
selection=[gid],
namespaces=['Synapses'])
syn_id_attr_index = None
for source, edge_iter in viewitems(graph[destination]):
this_edge_attr_info = edge_attr_info[destination][source]
if 'Synapses' in this_edge_attr_info and \
'syn_id' in this_edge_attr_info['Synapses']:
syn_id_attr_index = this_edge_attr_info['Synapses']['syn_id']
for (destination_gid, edges) in edge_iter:
assert destination_gid == gid
source_gids, edge_attrs = edges
syn_ids = edge_attrs['Synapses'][syn_id_attr_index]
this_source_syn_dict = source_syn_dict[source]
count = 0
for i in range(len(source_gids)):
this_source_gid = source_gids[i]
this_syn_id = syn_ids[i]
this_syn_wgt = syn_weight_dict.get(this_syn_id, 0.0)
this_source_syn_dict[this_source_gid].append((this_syn_id, this_syn_wgt))
count += 1
logger.info('destination: %s; gid %i; %d synaptic weights from source population %s' %
(destination, gid, count, source))
src_input_features = defaultdict(dict)
for source in sources:
source_gids = list(source_syn_dict[source].keys())
for input_features_namespace in this_input_features_namespaces:
input_features_iter = read_cell_attribute_selection(input_features_path, source,
namespace=input_features_namespace,
mask=set(features_attr_names),
selection=source_gids)
this_src_input_features = src_input_features[source]
count = 0
for gid, attr_dict in input_features_iter:
this_src_input_features[gid] = attr_dict
count += 1
logger.info('Read %s feature data for %i cells in population %s' % (input_features_namespace, count, source))
this_syn_weights = \
synapses.generate_structured_weights(destination_gid,
destination,
synapse_name,
sources,
dst_input_features,
src_input_features,
source_syn_dict,
spatial_mesh=(x,y),
plasticity_kernel=plasticity_kernel,
field_width_scale=field_width_scale,
baseline_weight=baseline_weight,
local_random=local_random,
interactive=interactive)
assert this_syn_weights is not None
structured_weights_dict[destination_gid] = this_syn_weights
logger.info('destination: %s; gid %i; generated structured weights for %i inputs'
% (destination, destination_gid, len(this_syn_weights['syn_id'])))
gc.collect()
if not dry_run:
logger.info('Destination: %s; appending structured weights...' % (destination))
this_structured_weights_namespace = '%s %s' % (structured_weights_namespace, arena_id)
append_cell_attributes(output_weights_path, destination, structured_weights_dict,
namespace=this_structured_weights_namespace)
logger.info('Destination: %s; appended structured weights' % (destination))
structured_weights_dict.clear()
if output_features_path is not None:
output_features_namespace = 'Place Selectivity %s' % arena_id
cell_attr_dict = dst_input_features[destination]
logger.info('Destination: %s; appending features...' % (destination))
append_cell_attributes(output_features_path, destination,
cell_attr_dict, namespace=output_features_namespace)
gc.collect()
del(syn_weight_dict)
del(src_input_features)
del(dst_input_features)