Merge branch 'master' of gitlab.lrz.de:mobile-ifi/bannana-networks

code/network.py

@@ -674,8 +674,6 @@ if __name__ == '__main__':
                         run_exp(net)
 
     if False:
-        # this does not work as the aggregation function screws over the fixpoint computation....
-        # TODO: check for fixpoint in aggregated space...
         with FixpointExperiment() as exp:
             run_count = 1000
             net = TrainingNeuralNetworkDecorator(AggregatingNeuralNetwork(4, width=2, depth=2)).with_params(epsilon=0.1e-6)

code/setups/fixpoint-density.py

@@ -0,0 +1,55 @@
+import sys
+
+sys.path += ['../', './']
+
+from util import *
+from experiment import *
+from network import *
+
+def generate_counters():
+    return {'divergent': 0, 'fix_zero': 0, 'fix_other': 0, 'fix_sec': 0, 'other': 0}
+
+def count(counters, net, notable_nets=[]):
+    if net.is_diverged():
+        counters['divergent'] += 1
+    elif net.is_fixpoint():
+        if net.is_zero():
+            counters['fix_zero'] += 1
+        else:
+            counters['fix_other'] += 1
+            notable_nets += [net]
+    elif net.is_fixpoint(2):
+        counters['fix_sec'] += 1
+        notable_nets += [net]
+    else:
+        counters['other'] += 1
+    return counters, notable_nets
+
+with Experiment('fixpoint-density') as exp:
+    exp.trials = 1000
+    exp.epsilon = 1e-4
+    net_generators = []
+    for activation in ['linear', 'sigmoid', 'relu']:
+        net_generators += [lambda activation=activation: WeightwiseNeuralNetwork(width=2, depth=2).with_keras_params(activation=activation, use_bias=False)]
+        net_generators += [lambda activation=activation: AggregatingNeuralNetwork(aggregates=4, width=2, depth=2).with_keras_params(activation=activation, use_bias=False)]
+        net_generators += [lambda activation=activation: RecurrentNeuralNetwork(width=2, depth=2).with_keras_params(activation=activation, use_bias=False)]
+    all_counters = []
+    all_notable_nets = []
+    all_names = []
+    for net_generator_id, net_generator in enumerate(net_generators):
+        counters = generate_counters()
+        notable_nets = []
+        for _ in tqdm(range(exp.trials)):
+            net = net_generator().with_params(epsilon=exp.epsilon)
+            name = str(net.__class__.__name__) + " activiation='" + str(net.get_keras_params().get('activation')) + "' use_bias='" + str(net.get_keras_params().get('use_bias')) + "'"
+            count(counters, net, notable_nets)
+        all_counters += [counters]
+        all_notable_nets += [notable_nets]
+        all_names += [name]
+    exp.save(all_counters=all_counters)
+    exp.save(all_notable_nets=all_notable_nets)
+    exp.save(all_names=all_names)
+    for exp_id, counter in enumerate(all_counters):
+        exp.log(all_names[exp_id])
+        exp.log(all_counters[exp_id])
+        exp.log('\n')

code/setups/known-fixpoint-variation.py

@@ -0,0 +1,76 @@
+import sys
+
+sys.path += ['../', './']
+
+from util import *
+from experiment import *
+from network import *
+from soup import prng
+
+from statistics import mean
+avg = mean
+    
+def generate_fixpoint_weights():
+    return [
+        np.array([[1.0, 0.0], [0.0, 0.0], [0.0, 0.0], [0.0, 0.0]], dtype=np.float32),
+        np.array([[1.0, 0.0], [0.0, 0.0]], dtype=np.float32),
+        np.array([[1.0], [0.0]], dtype=np.float32)
+    ]
+    
+def generate_fixpoint_net():
+    net = WeightwiseNeuralNetwork(width=2, depth=2).with_keras_params(activation='sigmoid')
+    net.set_weights(generate_fixpoint_weights())
+    return net
+
+def vary(old_weights, e=1.0):
+    new_weights = copy.deepcopy(old_weights)
+    for layer_id, layer in enumerate(new_weights):
+        for cell_id, cell in enumerate(layer):
+            for weight_id, weight in enumerate(cell):
+                if prng() < 0.5:
+                    new_weights[layer_id][cell_id][weight_id] = weight + prng() * e
+                else:
+                    new_weights[layer_id][cell_id][weight_id] = weight - prng() * e
+    return new_weights
+
+with Experiment('known-fixpoint-variation') as exp:
+    exp.depth = 10
+    exp.trials = 100
+    exp.max_steps = 100
+    exp.epsilon = 1e-4
+    exp.xs = []
+    exp.ys = []
+    exp.zs = []
+    exp.notable_nets = []
+    current_scale = 1.0
+    for _ in range(exp.depth):
+        print('variation scale ' + str(current_scale))
+        for _ in tqdm(range(exp.trials)):
+            net = generate_fixpoint_net().with_params(epsilon=exp.epsilon)
+            net.set_weights(vary(net.get_weights(), current_scale))
+            time_to_something = 0
+            time_as_fixpoint = 0
+            still_fixpoint = True
+            for _ in range(exp.max_steps):
+                net.self_attack()
+                if net.is_zero() or net.is_diverged():
+                    break
+                if net.is_fixpoint():
+                    if still_fixpoint:
+                        time_as_fixpoint += 1
+                    else:
+                        print('remarkable')
+                        exp.notable_nets += [net.get_weights()]
+                        still_fixpoint = True
+                else:
+                    still_fixpoint = False
+                time_to_something += 1
+            exp.xs += [current_scale]
+            exp.ys += [time_to_something]
+            exp.zs += [time_as_fixpoint]
+        current_scale /= 10.0
+    for d in range(exp.depth):
+        exp.log('variation 10e-' + str(d))
+        exp.log('avg time to vergence ' + str(avg(exp.ys[d*exp.trials:(d+1)*exp.trials])))
+        exp.log('avg time as fixpoint ' + str(avg(exp.zs[d*exp.trials:(d+1)*exp.trials])))
+        

code/setups/mixed-self-fixpoints.py

@@ -0,0 +1,72 @@
+import sys
+
+sys.path += ['../', './']
+
+from util import *
+from experiment import *
+from network import *
+
+import keras.backend
+
+
+def generate_counters():
+    return {'divergent': 0, 'fix_zero': 0, 'fix_other': 0, 'fix_sec': 0, 'other': 0}
+
+def count(counters, net, notable_nets=[]):
+    if net.is_diverged():
+        counters['divergent'] += 1
+    elif net.is_fixpoint():
+        if net.is_zero():
+            counters['fix_zero'] += 1
+        else:
+            counters['fix_other'] += 1
+            notable_nets += [net]
+    elif net.is_fixpoint(2):
+        counters['fix_sec'] += 1
+        notable_nets += [net]
+    else:
+        counters['other'] += 1
+    return counters, notable_nets
+
+with Experiment('training_fixpoint') as exp:
+    exp.trials = 20
+    exp.selfattacks = 4
+    exp.trains_per_selfattack_values = [100 * i for i in range(11)]
+    exp.epsilon = 1e-4
+    net_generators = []
+    for activation in ['linear', 'sigmoid', 'relu']:
+        for use_bias in [False]:
+            net_generators += [lambda activation=activation, use_bias=use_bias: WeightwiseNeuralNetwork(width=2, depth=2).with_keras_params(activation=activation, use_bias=use_bias)]
+            # net_generators += [lambda activation=activation, use_bias=use_bias: AggregatingNeuralNetwork(aggregates=4, width=2, depth=2).with_keras_params(activation=activation, use_bias=use_bias)]
+            # net_generators += [lambda activation=activation, use_bias=use_bias: RecurrentNeuralNetwork(width=2, depth=2).with_keras_params(activation=activation, use_bias=use_bias)]
+
+    all_names = []
+    all_data = []
+    for net_generator_id, net_generator in enumerate(net_generators):
+        xs = []
+        ys = []
+        for trains_per_selfattack in exp.trains_per_selfattack_values:
+            counters = generate_counters()
+            notable_nets = []
+            for _ in tqdm(range(exp.trials)):
+                net = TrainingNeuralNetworkDecorator(net_generator()).with_params(epsilon=exp.epsilon)
+                name = str(net.net.__class__.__name__) + " activiation='" + str(net.get_keras_params().get('activation')) + "' use_bias=" + str(net.get_keras_params().get('use_bias'))
+                for selfattack_id in range(exp.selfattacks):
+                    net.self_attack()
+                    for train_id in range(trains_per_selfattack):
+                        loss = net.compiled().train(epoch=selfattack_id*trains_per_selfattack+train_id)
+                    if net.is_diverged() or net.is_fixpoint():
+                        break
+                count(counters, net, notable_nets)
+                keras.backend.clear_session()
+            xs += [trains_per_selfattack]
+            ys += [float(counters['fix_zero'] + counters['fix_other']) / float(exp.trials)]
+        all_names += [name]
+        all_data += [{'xs':xs, 'ys':ys}]
+
+    exp.save(all_names=all_names)
+    exp.save(all_data=all_data)
+    for exp_id, name in enumerate(all_names):
+        exp.log(all_names[exp_id])
+        exp.log(all_data[exp_id])
+        exp.log('\n')

code/setups/training-fixpoints.py

@@ -0,0 +1,59 @@
+import sys
+
+sys.path += ['../', './']
+
+from util import *
+from experiment import *
+from network import *
+
+def generate_counters():
+    return {'divergent': 0, 'fix_zero': 0, 'fix_other': 0, 'fix_sec': 0, 'other': 0}
+
+def count(counters, net, notable_nets=[]):
+    if net.is_diverged():
+        counters['divergent'] += 1
+    elif net.is_fixpoint():
+        if net.is_zero():
+            counters['fix_zero'] += 1
+        else:
+            counters['fix_other'] += 1
+            notable_nets += [net]
+    elif net.is_fixpoint(2):
+        counters['fix_sec'] += 1
+        notable_nets += [net]
+    else:
+        counters['other'] += 1
+    return counters, notable_nets
+
+with Experiment('training_fixpoint') as exp:
+    exp.trials = 5
+    exp.run_count = 500
+    exp.epsilon = 1e-4
+    net_generators = []
+    for activation in ['linear', 'sigmoid', 'relu']:
+        for use_bias in [False]:
+            net_generators += [lambda activation=activation, use_bias=use_bias: WeightwiseNeuralNetwork(width=2, depth=2).with_keras_params(activation=activation, use_bias=use_bias)]
+            net_generators += [lambda activation=activation, use_bias=use_bias: AggregatingNeuralNetwork(aggregates=4, width=2, depth=2).with_keras_params(activation=activation, use_bias=use_bias)]
+            net_generators += [lambda activation=activation, use_bias=use_bias: RecurrentNeuralNetwork(width=2, depth=2).with_keras_params(activation=activation, use_bias=use_bias)]
+    all_counters = []
+    all_notable_nets = []
+    all_names = []
+    for net_generator_id, net_generator in enumerate(net_generators):
+        counters = generate_counters()
+        notable_nets = []
+        for _ in tqdm(range(exp.trials)):
+            net = TrainingNeuralNetworkDecorator(net_generator()).with_params(epsilon=exp.epsilon)
+            name = str(net.net.__class__.__name__) + " activiation='" + str(net.get_keras_params().get('activation')) + "' use_bias=" + str(net.get_keras_params().get('use_bias'))
+            for run_id in range(exp.run_count):
+                loss = net.compiled().train(epoch=run_id+1)
+            count(counters, net, notable_nets)
+        all_counters += [counters]
+        all_notable_nets += [notable_nets]
+        all_names += [name]
+    exp.save(all_counters=all_counters)
+    exp.save(all_notable_nets=all_notable_nets)
+    exp.save(all_names=all_names)
+    for exp_id, counter in enumerate(all_counters):
+        exp.log(all_names[exp_id])
+        exp.log(all_counters[exp_id])
+        exp.log('\n')