Releasing 0.3.2 (Brian2 extensions).

RELEASE_NOTES.rst → changelog.rst

@@ -1,6 +1,16 @@
 SNN Toolbox: Release Notes
 ==========================
 
+Version 0.3.2
+-------------
+
+Simulation with Brian2 backend now supports:
+    - Constant input currents (less noisy than Poisson input)
+    - Reset-by-subtraction (more accurate than reset-to-zero).
+    - Bias currents
+
+Thanks to wilkieolin for this contribution.
+
 Version 0.3.1
 -------------
 

examples/mnist_keras_brian2.py

@@ -13,7 +13,8 @@
 
 import keras
 from keras import Input, Model
-from keras.layers import Conv2D, AveragePooling2D, Flatten, Dense, Dropout
+from keras.layers import Conv2D, AveragePooling2D, Flatten, Dense, Dropout, \
+    BatchNormalization, Activation
 from keras.datasets import mnist
 from keras.utils import np_utils
 
@@ -67,24 +68,21 @@
 
 layer = Conv2D(filters=16,
                kernel_size=(5, 5),
-               strides=(2, 2),
-               activation='relu',
-               use_bias=False)(input_layer)
+               strides=(2, 2))(input_layer)
+layer = BatchNormalization(axis=axis)(layer)
+layer = Activation('relu')(layer)
 layer = Conv2D(filters=32,
                kernel_size=(3, 3),
-               activation='relu',
-               use_bias=False)(layer)
+               activation='relu')(layer)
 layer = AveragePooling2D()(layer)
 layer = Conv2D(filters=8,
                kernel_size=(3, 3),
                padding='same',
-               activation='relu',
-               use_bias=False)(layer)
+               activation='relu')(layer)
 layer = Flatten()(layer)
 layer = Dropout(0.01)(layer)
 layer = Dense(units=10,
-              activation='softmax',
-              use_bias=False)(layer)
+              activation='softmax')(layer)
 
 model = Model(input_layer, layer)
 
@@ -123,11 +121,11 @@
     'duration': 50,                 # Number of time steps to run each sample.
     'num_to_test': 5,               # How many test samples to run.
     'batch_size': 1,                # Batch size for simulation.
-    'dt': 0.1                       # Time interval for the differential equations to be solved over.
+    'dt': 0.1                       # Time resolution for ODE solving.
 }
 
 config['input'] = {
-    'poisson_input': True           # Images are encodes as spike trains.
+    'poisson_input': False          # Images are encodes as spike trains.
 }
 
 config['output'] = {

examples/mnist_keras_nest.py

@@ -68,23 +68,19 @@
 layer = Conv2D(filters=16,
                kernel_size=(5, 5),
                strides=(2, 2),
-               activation='relu',
-               use_bias=False)(input_layer)
+               activation='relu')(input_layer)
 layer = Conv2D(filters=32,
                kernel_size=(3, 3),
-               activation='relu',
-               use_bias=False)(layer)
+               activation='relu')(layer)
 layer = AveragePooling2D()(layer)
 layer = Conv2D(filters=8,
                kernel_size=(3, 3),
                padding='same',
-               activation='relu',
-               use_bias=False)(layer)
+               activation='relu')(layer)
 layer = Flatten()(layer)
 layer = Dropout(0.01)(layer)
 layer = Dense(units=10,
-              activation='softmax',
-              use_bias=False)(layer)
+              activation='softmax')(layer)
 
 model = Model(input_layer, layer)
 
@@ -125,10 +121,6 @@
     'batch_size': 1,                # Batch size for simulation.
 }
 
-config['input'] = {
-    'poisson_input': True           # Images are encodes as spike trains.
-}
-
 config['output'] = {
     'plot_vars': {                  # Various plots (slows down simulation).
         'spiketrains',              # Leave section empty to turn off plots.

setup.py

@@ -14,7 +14,7 @@
 
 setup(
     name='snntoolbox',
-    version='0.3.1',  # see https://www.python.org/dev/peps/pep-0440/
+    version='0.3.2',  # see https://www.python.org/dev/peps/pep-0440/
     description='Spiking neural network conversion toolbox',
     long_description=long_description,
     author='Bodo Rueckauer',

snntoolbox/simulation/target_simulators/brian2_target_sim.py

@@ -65,8 +65,8 @@ def __init__(self, config, queue=None):
             self.v_reset = 'v = v - v_thresh'
         else:
             self.v_reset = 'v = v_reset'
-        self.eqs = ''' dv/dt = bias : 1
-                       bias : hertz'''
+        self.eqs = '''dv/dt = bias : 1
+                      bias : hertz'''
         self.spikemonitors = []
         self.statemonitors = []
         self.snn = None
@@ -219,14 +219,14 @@ def compile(self):
 
     def simulate(self, **kwargs):
 
+        inputs = kwargs[str('x_b_l')].flatten() / self.sim.ms
         if self._poisson_input:
-            self._input_layer.rates = kwargs[str('x_b_l')].flatten() * 1000 / \
-                                      self.rescale_fac * self.sim.Hz
+            self._input_layer.rates = inputs / self.rescale_fac
         elif self._dataset_format == 'aedat':
             # TODO: Implement by using brian2.SpikeGeneratorGroup.
             raise NotImplementedError
         else:
-            self._input_layer.bias = kwargs[str('x_b_l')].flatten() / self.sim.ms
+            self._input_layer.bias = inputs
 
         self.snn.run(self._duration * self.sim.ms, namespace=self._cell_params,
                      report='stdout', report_period=10 * self.sim.ms)
@@ -385,8 +385,8 @@ def set_spiketrain_stats_input(self):
         AbstractSNN.set_spiketrain_stats_input(self)
 
     def set_biases(self, biases):
-        """Set biases.
-        """
+        """Set biases."""
         if any(biases):
-            assert self.layers[-1].bias.shape == biases.shape, "Shape of biases and network do not match."
-            self.layers[-1].bias = biases * 1000 * self.sim.Hz
+            assert self.layers[-1].bias.shape == biases.shape, \
+                "Shape of biases and network do not match."
+            self.layers[-1].bias = biases / self.sim.ms