def modify_parameter_precision(weights, biases, config, attributes):
if config.getboolean('cell', 'binarize_weights'):
from snntoolbox.utils.utils import binarize
print("Binarizing weights.")
weights = binarize(weights)
elif config.getboolean('cell', 'quantize_weights'):
assert 'Qm.f' in attributes, \
"In the [cell] section of the configuration file, " \
"'quantize_weights' was set to True. For this to " \
"work, the layer needs to specify the fixed point " \
"number format 'Qm.f'."
from snntoolbox.utils.utils import reduce_precision
m, f = attributes.get('Qm.f')
print("Quantizing weights to Q{}.{}.".format(m, f))
weights = reduce_precision(weights, m, f)
if attributes.get('quantize_bias', False):
biases = reduce_precision(biases, m, f)
# These attributes are not needed any longer and would not be
# understood by Keras when building the parsed model.
attributes.pop('quantize_bias', None)
attributes.pop('Qm.f', None)
return weights, biases
def parse(self):
"""Extract the essential information about a neural network.
This method serves to abstract the conversion process of a network from
the language the input model was built in (e.g. Keras or Lasagne).
The methods iterates over all layers of the input model and writes the
layer specifications and parameters into `_layer_list`. The keys are
chosen in accordance with Keras layer attributes to facilitate
instantiation of a new, parsed Keras model (done in a later step by
`build_parsed_model`).
This function applies several simplifications and adaptations to prepare
the model for conversion to spiking. These modifications include:
- Removing layers only used during training (Dropout,
BatchNormalization, ...)
- Absorbing the parameters of BatchNormalization layers into the
parameters of the preceeding layer. This does not affect performance
because batch-norm-parameters are constant at inference time.
- Removing ReLU activation layers, because their function is inherent to
the spike generation mechanism. The information which nonlinearity was
used in the original model is preserved in the ``activation`` key in
`_layer_list`. If the output layer employs the softmax function, a
spiking version is used when testing the SNN in INIsim or MegaSim
simulators.
- Inserting a Flatten layer between Conv and FC layers, if the input
model did not explicitly include one.
"""
layers = self.get_layer_iterable()
snn_layers = eval(self.config.get('restrictions', 'snn_layers'))
name_map = {}
idx = 0
inserted_flatten = False
for layer in layers:
layer_type = self.get_type(layer)
# Absorb BatchNormalization layer into parameters of previous layer
if layer_type == 'BatchNormalization':
parameters_bn = list(self.get_batchnorm_parameters(layer))
inbound = self.get_inbound_layers_with_parameters(layer)
assert len(inbound) == 1, \
"Could not find unique layer with parameters " \
"preceeding BatchNorm layer."
prev_layer = inbound[0]
prev_layer_idx = name_map[str(id(prev_layer))]
parameters = list(
self._layer_list[prev_layer_idx]['parameters'])
print("Absorbing batch-normalization parameters into " +
"parameters of previous {}.".format(
self.get_type(prev_layer)))
self._layer_list[prev_layer_idx]['parameters'] = \
absorb_bn_parameters(*(parameters + parameters_bn))
if layer_type == 'GlobalAveragePooling2D':
print("Replacing GlobalAveragePooling by AveragePooling "
"plus Flatten.")
pool_size = [layer.input_shape[-2], layer.input_shape[-1]]
self._layer_list.append({
'layer_type':
'AveragePooling2D',
'name':
self.get_name(layer, idx, 'AveragePooling2D'),
'input_shape':
layer.input_shape,
'pool_size':
pool_size,
'inbound':
self.get_inbound_names(layer, name_map)
})
name_map['AveragePooling2D' + str(idx)] = idx
idx += 1
num_str = str(idx) if idx > 9 else '0' + str(idx)
shape_string = str(np.prod(layer.output_shape[1:]))
self._layer_list.append({
'name':
num_str + 'Flatten_' + shape_string,
'layer_type':
'Flatten',
'inbound': [self._layer_list[-1]['name']]
})
name_map['Flatten' + str(idx)] = idx
idx += 1
inserted_flatten = True
if layer_type not in snn_layers:
print("Skipping layer {}.".format(layer_type))
continue
if not inserted_flatten:
inserted_flatten = self.try_insert_flatten(
layer, idx, name_map)
idx += inserted_flatten
print("Parsing layer {}.".format(layer_type))
if layer_type == 'MaxPooling2D' and \
self.config.getboolean('conversion', 'max2avg_pool'):
print("Replacing max by average pooling.")
layer_type = 'AveragePooling2D'
if inserted_flatten:
inbound = [self._layer_list[-1]['name']]
inserted_flatten = False
else:
inbound = self.get_inbound_names(layer, name_map)
attributes = self.initialize_attributes(layer)
attributes.update({
'layer_type': layer_type,
'name': self.get_name(layer, idx),
'inbound': inbound
})
if layer_type == 'Dense':
self.parse_dense(layer, attributes)
if layer_type == 'Conv2D':
self.parse_convolution(layer, attributes)
if layer_type in {'Dense', 'Conv2D'}:
weights, bias = attributes['parameters']
if self.config.getboolean('cell', 'binarize_weights'):
from snntoolbox.utils.utils import binarize
print("Binarizing weights.")
weights = binarize(weights)
elif self.config.getboolean('cell', 'quantize_weights'):
assert 'Qm.f' in attributes, \
"In the [cell] section of the configuration file, "\
"'quantize_weights' was set to True. For this to " \
"work, the layer needs to specify the fixed point " \
"number format 'Qm.f'."
from snntoolbox.utils.utils import reduce_precision
m, f = attributes.get('Qm.f')
print("Quantizing weights to Q{}.{}.".format(m, f))
weights = reduce_precision(weights, m, f)
if attributes.get('quantize_bias', False):
bias = reduce_precision(bias, m, f)
attributes['parameters'] = (weights, bias)
# These attributes are not needed any longer and would not be
# understood by Keras when building the parsed model.
attributes.pop('quantize_bias', None)
attributes.pop('Qm.f', None)
self.absorb_activation(layer, attributes)
if 'Pooling' in layer_type:
self.parse_pooling(layer, attributes)
if layer_type == 'Concatenate':
self.parse_concatenate(layer, attributes)
self._layer_list.append(attributes)
# Map layer index to layer id. Needed for inception modules.
name_map[str(id(layer))] = idx
idx += 1
print('')
