Updated 'temporal_pattern' code to work with tensorflow 2.

Author: Bodo Rueckauer

snntoolbox/simulation/backends/inisim/temporal_pattern.py

@@ -75,7 +75,7 @@ def init_neurons(self, input_shape):
         self.spikerates = k.zeros(output_shape)
 
     def update_spikevars(self, x):
-        return [tf.assign(self.spikerates, x)]
+        return [self.spikerates.assign(x)]
 
 
 def spike_call(call):
@@ -123,7 +123,7 @@ def to_binary(x, num_bits):
         distributed across the first dimension of ``binary_array``.
     """
 
-    shape = k.shape(x)
+    shape = k.int_shape(x)
 
     binary_array = k.zeros([num_bits] + list(shape[1:]), k.floatx())
 
@@ -177,7 +177,7 @@ def iterate_powers(act_value, idx_p, idx_l, idx_m, idx_n):
             p = powers[idx_p]
             c = k.greater_equal(act_value, p)
             b = tf.cond(c, lambda: 1., lambda: 0.)
-            a = tf.assign(binary_array[idx_p, idx_l, idx_m, idx_n], b)
+            a = binary_array[idx_p, idx_l, idx_m, idx_n].assign(b)
             new_act_value = tf.cond(c, lambda: act_value - p,
                                     lambda: act_value)
             with tf.control_dependencies([a]):
@@ -210,7 +210,7 @@ def iterate_powers(act_value, idx_p, idx_l):
             p = powers[idx_p]
             c = k.greater_equal(act_value, p)
             b = tf.cond(c, lambda: 1., lambda: 0.)
-            a = tf.assign(binary_array[idx_p, idx_l], b)
+            a = binary_array[idx_p, idx_l].assign(b)
             new_act_value = tf.cond(c, lambda: act_value - p,
                                     lambda: act_value)
             with tf.control_dependencies([a]):
@@ -249,20 +249,20 @@ def to_binary_numpy(x, num_bits):
     powers = [2**-i for i in range(num_bits)]
 
     if len(x.shape) > 2:
-        for l in range(x.shape[1]):
+        for j in range(x.shape[1]):
             for m in range(x.shape[2]):
                 for n in range(x.shape[3]):
-                    f = x[0, l, m, n]
+                    f = x[0, j, m, n]
                     for i in range(num_bits):
                         if f >= powers[i]:
-                            binary_array[i, l, m, n] = 1
+                            binary_array[i, j, m, n] = 1
                             f -= powers[i]
     else:
-        for l in range(x.shape[1]):
-            f = x[0, l]
+        for j in range(x.shape[1]):
+            f = x[0, j]
             for i in range(num_bits):
                 if f >= powers[i]:
-                    binary_array[i, l] = 1
+                    binary_array[i, j] = 1
                     f -= powers[i]
     return binary_array
 

snntoolbox/simulation/target_simulators/INI_temporal_mean_rate_target_sim.py

@@ -103,17 +103,23 @@ def build_pooling(self, layer):
 
     def compile(self):
 
+        self.snn = keras.models.Model(
+            self._input_images,
+            self._spiking_layers[self.parsed_model.layers[-1].name])
+        self.snn.compile('sgd', 'categorical_crossentropy', ['accuracy'])
+
+        # Tensorflow 2 lists all variables as weights, including our state
+        # variables (membrane potential etc). So a simple
+        # snn.set_weights(parsed_model.get_weights()) does not work any more.
+        # Need to extract the actual weights here.
+
         def remove_name_counter(name_in):
             splits = str(name_in).split('_')
             name_out = splits[0] + '_' + splits[1]
             if len(splits) == 3:
-                name_out += re.sub('\d+/', '/', splits[2])
+                name_out += re.sub(r'\d+/', '/', splits[2])
             return name_out
 
-        self.snn = keras.models.Model(
-            self._input_images,
-            self._spiking_layers[self.parsed_model.layers[-1].name])
-        self.snn.compile('sgd', 'categorical_crossentropy', ['accuracy'])
         parameter_map = {remove_name_counter(p.name): v for p, v in
                          zip(self.parsed_model.weights,
                              self.parsed_model.get_weights())}
@@ -125,6 +131,7 @@ def remove_name_counter(name_in):
                 count += 1
         assert count == len(parameter_map), "Not all weights have been " \
                                             "transferred from ANN to SNN."
+
         for layer in self.snn.layers:
             if hasattr(layer, 'bias'):
                 # Adjust biases to time resolution of simulator.

snntoolbox/simulation/target_simulators/INI_temporal_pattern_target_sim.py

@@ -7,6 +7,8 @@
 from __future__ import division, absolute_import
 from __future__ import print_function, unicode_literals
 
+import re
+
 import keras
 import numpy as np
 from future import standard_library
@@ -45,11 +47,36 @@ def __init__(self, config, queue=None):
         self.num_bits = self.config.getint('conversion', 'num_bits')
 
     def compile(self):
+
         self.snn = keras.models.Model(
             self._input_images,
             self._spiking_layers[self.parsed_model.layers[-1].name])
         self.snn.compile('sgd', 'categorical_crossentropy', ['accuracy'])
-        self.snn.set_weights(self.parsed_model.get_weights())
+
+        # Tensorflow 2 lists all variables as weights, including our state
+        # variables (membrane potential etc). So a simple
+        # snn.set_weights(parsed_model.get_weights()) does not work any more.
+        # Need to extract the actual weights here:
+
+        def remove_name_counter(name_in):
+            splits = str(name_in).split('_')
+            name_out = splits[0] + '_' + splits[1]
+            if len(splits) == 3:
+                name_out += re.sub(r'\d+/', '/', splits[2])
+            return name_out
+
+        parameter_map = {remove_name_counter(p.name): v for p, v in
+                         zip(self.parsed_model.weights,
+                             self.parsed_model.get_weights())}
+        count = 0
+        for p in self.snn.weights:
+            name = remove_name_counter(p.name)
+            if name in parameter_map:
+                keras.backend.set_value(p, parameter_map[name])
+                count += 1
+        assert count == len(parameter_map), "Not all weights have been " \
+                                            "transferred from ANN to SNN."
+
         for layer in self.snn.layers:
             if hasattr(layer, 'bias'):
                 # Adjust biases to time resolution of simulator.