journal linspace basins

.gitignore

@@ -0,0 +1 @@
+/output/

journal_basin_linspace_clones.py

@@ -8,15 +8,12 @@ import numpy as np
 import torch
 
 from functionalities_test import is_identity_function, test_status
-from journal_basins import SpawnExperiment, prng, mean_invariate_manhattan_distance
+from journal_basins import SpawnExperiment, mean_invariate_manhattan_distance
 from network import Net
 
 from sklearn.metrics import mean_absolute_error as MAE
 from sklearn.metrics import mean_squared_error as MSE
 
-import seaborn as sns
-from matplotlib import pyplot as plt
-
 
 class SpawnLinspaceExperiment(SpawnExperiment):
 
@@ -28,6 +25,12 @@ class SpawnLinspaceExperiment(SpawnExperiment):
                      'status_post'])
 
         # For every initial net {i} after populating (that is fixpoint after first epoch);
+        # parent = self.parents[0]
+        # parent_clone = clone = Net(parent.input_size, parent.hidden_size, parent.out_size,
+        #                         name=f"{parent.name}_clone_{0}", start_time=self.ST_steps)
+        # parent_clone.apply_weights(torch.as_tensor(parent.create_target_weights(parent.input_weight_matrix())))
+        # parent_clone = parent_clone.apply_noise(self.noise)
+        # self.parents.append(parent_clone)
         pairwise_net_list = itertools.combinations(self.parents, 2)
         for net1, net2 in pairwise_net_list:
             # We set parent start_time to just before this epoch ended, so plotting is zoomed in. Comment out to
@@ -42,11 +45,12 @@ class SpawnLinspaceExperiment(SpawnExperiment):
             net2_target_data = net2.create_target_weights(net2_input_data)
 
             if is_identity_function(net1) and is_identity_function(net2):
+            # if True:
                 # Clone the fixpoint x times and add (+-)self.noise to weight-sets randomly;
                 # To plot clones starting after first epoch (z=ST_steps), set that as start_time!
                 # To make sure PCA will plot the same trajectory up until this point, we clone the
                 # parent-net's weight history as well.
-                in_between_weights = np.linspace(net1_target_data, net2_target_data, number_clones,endpoint=False)
+                in_between_weights = np.linspace(net1_target_data, net2_target_data, number_clones, endpoint=False)
 
                 for j, in_between_weight in enumerate(in_between_weights):
                     clone = Net(net1.input_size, net1.hidden_size, net1.out_size,
@@ -89,7 +93,6 @@ class SpawnLinspaceExperiment(SpawnExperiment):
             for _ in range(self.epochs - 1):
                 for _ in range(self.ST_steps):
                     parent.self_train(1, self.log_step_size, self.net_learning_rate)
-            
 
         self.df = df
 
@@ -106,7 +109,7 @@ if __name__ == '__main__':
     ST_log_step_size = 10
 
     # Define number of networks & their architecture
-    nr_clones = 3
+    nr_clones = 20
     ST_population_size = 3
     ST_net_hidden_size = 2
     ST_net_learning_rate = 0.04
@@ -123,7 +126,7 @@ if __name__ == '__main__':
         epochs=ST_epochs,
         st_steps=ST_steps,
         nr_clones=nr_clones,
-        noise=None,
+        noise=1e-8,
         directory=Path('output') / 'spawn_basin' / f'{ST_name_hash}' / f'linage'
     )
     df = exp.df
@@ -133,10 +136,10 @@ if __name__ == '__main__':
     print(f"\nSaved experiment to {directory}.")
 
     # Boxplot with counts of nr_fixpoints, nr_other, nr_etc. on y-axis
-    sns.countplot(data=df, x="noise", hue="status_post")
+    # sns.countplot(data=df, x="noise", hue="status_post")
-    plt.savefig(f"output/spawn_basin/{ST_name_hash}/fixpoint_status_countplot.png")
+    # plt.savefig(f"output/spawn_basin/{ST_name_hash}/fixpoint_status_countplot.png")
 
     # Catplot (either kind="point" or "box") that shows before-after training distances to parent
-    mlt = df[["MIM_pre", "MIM_post", "noise"]].melt("noise", var_name="time", value_name='Average Distance')
+    # mlt = df[["MIM_pre", "MIM_post", "noise"]].melt("noise", var_name="time", value_name='Average Distance')
-    sns.catplot(data=mlt, x="time", y="Average Distance", col="noise", kind="point", col_wrap=5, sharey=False)
+    # sns.catplot(data=mlt, x="time", y="Average Distance", col="noise", kind="point", col_wrap=5, sharey=False)
-    plt.savefig(f"output/spawn_basin/{ST_name_hash}/clone_distance_catplot.png")
+    # plt.savefig(f"output/spawn_basin/{ST_name_hash}/clone_distance_catplot.png")

journal_basins.py

@@ -84,21 +84,6 @@ def distance_from_parent(nets, distance="MIM", print_it=True):
 
 class SpawnExperiment:
 
-    @staticmethod
-    def apply_noise(network, noise: int):
-        """ Changing the weights of a network to values + noise """
-
-        for layer_id, layer_name in enumerate(network.state_dict()):
-            for line_id, line_values in enumerate(network.state_dict()[layer_name]):
-                for weight_id, weight_value in enumerate(network.state_dict()[layer_name][line_id]):
-                    # network.state_dict()[layer_name][line_id][weight_id] = weight_value + noise
-                    if prng() < 0.5:
-                        network.state_dict()[layer_name][line_id][weight_id] = weight_value + noise
-                    else:
-                        network.state_dict()[layer_name][line_id][weight_id] = weight_value - noise
-
-        return network
-
     def __init__(self, population_size, log_step_size, net_input_size, net_hidden_size, net_out_size, net_learning_rate,
                  epochs, st_steps, nr_clones, noise, directory) -> None:
         self.population_size = population_size
@@ -171,7 +156,7 @@ class SpawnExperiment:
                                 f"ST_net_{str(i)}_clone_{str(j)}", start_time=self.ST_steps)
                     clone.load_state_dict(copy.deepcopy(net.state_dict())) 
                     rand_noise = prng() * self.noise
-                    clone = self.apply_noise(clone, rand_noise)
+                    clone = clone.apply_noise(rand_noise)
                     clone.s_train_weights_history = copy.deepcopy(net.s_train_weights_history)
                     clone.number_trained = copy.deepcopy(net.number_trained)
 

journal_robustness.py

@@ -91,7 +91,6 @@ class RobustnessComparisonExperiment:
         self.time_to_vergence, self.time_as_fixpoint = self.test_robustness(
             seeds=population_size if self.is_synthetic else 1)
 
-
     def populate_environment(self):
         nets = []
         if self.is_synthetic:
@@ -125,8 +124,8 @@ class RobustnessComparisonExperiment:
         # This checks wether to use synthetic setting with multiple seeds
         #   or multi network settings with a singlee seed
 
-        df = pd.DataFrame(columns=['setting', 'Noise Level', 'steps', 'absolute_loss',
+        df = pd.DataFrame(columns=['setting', 'Noise Level', 'Self Train Steps', 'absolute_loss',
-                                   'time_to_vergence', 'time_as_fixpoint'])
+                                   'Time to vergence', 'Time as fixpoint'])
         with tqdm(total=max(len(self.id_functions), seeds)) as pbar:
             for i, fixpoint in enumerate(self.id_functions):  # 1 / n
                 row_headers.append(fixpoint.name)
@@ -138,8 +137,7 @@ class RobustnessComparisonExperiment:
                         clone = Net(fixpoint.input_size, fixpoint.hidden_size, fixpoint.out_size,
                                     f"{fixpoint.name}_clone_noise10e-{noise_level}")
                         clone.load_state_dict(copy.deepcopy(fixpoint.state_dict()))
-                        rand_noise = prng() * pow(10, -noise_level)  # n / 1
+                        clone = clone.apply_noise(pow(10, -noise_level))
-                        clone = self.apply_noise(clone, rand_noise)
 
                         while not is_zero_fixpoint(clone) and not is_divergent(clone):
                             # -> before
@@ -154,7 +152,6 @@ class RobustnessComparisonExperiment:
 
                             absolute_loss = F.l1_loss(target_data_pre_application, target_data_post_application).item()
 
-
                             if is_identity_function(clone):
                                 time_as_fixpoint[setting][noise_level] += 1
                                 # When this raises a Type Error, we found a second order fixpoint!
@@ -166,26 +163,24 @@ class RobustnessComparisonExperiment:
                     pbar.update(1)
 
         # Get the measuremts at the highest time_time_to_vergence
-        df_sorted = df.sort_values('Steps', ascending=False).drop_duplicates(['setting', 'Noise Level'])
+        df_sorted = df.sort_values('Self Train Steps', ascending=False).drop_duplicates(['setting', 'Noise Level'])
-        df_melted = df_sorted.reset_index().melt(id_vars=['setting', 'Noise Level', 'Steps'],
+        df_melted = df_sorted.reset_index().melt(id_vars=['setting', 'Noise Level', 'Self Train Steps'],
                                                  value_vars=['Time to vergence', 'Time as fixpoint'],
                                                  var_name="Measurement",
-                                                 value_name="Steps")
+                                                 value_name="Steps").sort_values('Noise Level')
         # Plotting
         sns.set(style='whitegrid', font_scale=2)
         bf = sns.boxplot(data=df_melted, y='Steps', x='Noise Level', hue='Measurement', palette=PALETTE)
         synthetic = 'synthetic' if self.is_synthetic else 'natural'
-        bf.set_title(f'Robustness as self application steps per noise level for {synthetic} fixpoints.')
+        # bf.set_title(f'Robustness as self application steps per noise level for {synthetic} fixpoints.')
         plt.tight_layout()
 
         # sns.set(rc={'figure.figsize': (10, 50)})
         # bx = sns.catplot(data=df[df['absolute_loss'] < 1], y='absolute_loss', x='application_step', kind='box',
         #                  col='noise_level', col_wrap=3, showfliers=False)
-        directory = Path('output') / 'robustness'
-        directory.mkdir(parents=True, exist_ok=True)
-        filename = f"absolute_loss_perapplication_boxplot_grid.png"
-        filepath = directory / filename
 
+        filename = f"absolute_loss_perapplication_boxplot_grid_{'synthetic' if self.is_synthetic else 'wild'}.png"
+        filepath = self.directory / filename
         plt.savefig(str(filepath))
 
         if print_it:
@@ -219,11 +214,11 @@ if __name__ == "__main__":
     ST_steps = 1000
     ST_epochs = 5
     ST_log_step_size = 10
-    ST_population_size = 2
+    ST_population_size = 500
     ST_net_hidden_size = 2
     ST_net_learning_rate = 0.004
     ST_name_hash = random.getrandbits(32)
-    ST_synthetic = True
+    ST_synthetic = False
 
     print(f"Running the robustness comparison experiment:")
     exp = RobustnessComparisonExperiment(

journal_soup_basins.py

@@ -1,14 +1,12 @@
-import os
 from pathlib import Path
 import pickle
-from torch import mean
 
 from tqdm import tqdm
 import random
 import copy
-from functionalities_test import is_identity_function, test_status, test_for_fixpoints, is_zero_fixpoint, is_divergent, is_secondary_fixpoint
+from functionalities_test import is_identity_function, test_status, is_zero_fixpoint, is_divergent, is_secondary_fixpoint
 from network import Net
-from visualization import plot_3d_self_train, plot_loss, plot_3d_soup
+from visualization import plot_loss, plot_3d_soup
 import numpy as np
 from tabulate import tabulate
 from sklearn.metrics import mean_absolute_error as MAE
@@ -18,10 +16,6 @@ import seaborn as sns
 from matplotlib import pyplot as plt
 
 
-def prng():
-    return random.random()
-
-
 def l1(tup):
     a, b = tup
     return abs(a - b)
@@ -88,20 +82,6 @@ def distance_from_parent(nets, distance="MIM", print_it=True):
 
 class SoupSpawnExperiment:
 
-    @staticmethod
-    def apply_noise(network, noise: int):
-        """ Changing the weights of a network to values + noise """
-
-        for layer_id, layer_name in enumerate(network.state_dict()):
-            for line_id, line_values in enumerate(network.state_dict()[layer_name]):
-                for weight_id, weight_value in enumerate(network.state_dict()[layer_name][line_id]):
-                    # network.state_dict()[layer_name][line_id][weight_id] = weight_value + noise
-                    if prng() < 0.5:
-                        network.state_dict()[layer_name][line_id][weight_id] = weight_value + noise
-                    else:
-                        network.state_dict()[layer_name][line_id][weight_id] = weight_value - noise
-
-        return network
 
     def __init__(self, population_size, log_step_size, net_input_size, net_hidden_size, net_out_size, net_learning_rate,
                  epochs, st_steps, attack_chance, nr_clones, noise, directory) -> None:
@@ -220,8 +200,7 @@ class SoupSpawnExperiment:
                 clone = Net(net.input_size, net.hidden_size, net.out_size,
                             f"net_{str(i)}_clone_{str(j)}", start_time=self.ST_steps)
                 clone.load_state_dict(copy.deepcopy(net.state_dict()))
-                rand_noise = prng() * self.noise
+                clone = clone.apply_noise(self.noise)
-                clone = self.apply_noise(clone, rand_noise)
                 clone.s_train_weights_history = copy.deepcopy(net.s_train_weights_history)
                 clone.number_trained = copy.deepcopy(net.number_trained)
 
@@ -262,9 +241,9 @@ class SoupSpawnExperiment:
                           f"\nMSE({i},{j}): {MSE_post}"
                           f"\nMAE({i},{j}): {MAE_post}"
                           f"\nMIM({i},{j}): {MIM_post}\n")
-                    self.parents_clones_id_functions.append(clone):
+                    self.parents_clones_id_functions.append(clone)
 
-                df.loc[df.name==clone.name, ["MAE_post", "MSE_post", "MIM_post", "status_post"]] = [MAE_post, MSE_post, MIM_post, clone.is_fixpoint]
+                df.loc[df.name == clone.name, ["MAE_post", "MSE_post", "MIM_post", "status_post"]] = [MAE_post, MSE_post, MIM_post, clone.is_fixpoint]
 
             # Finally take parent net {i} and finish it's training for comparison to clone development.
             for _ in range(self.epochs - 1):

journal_soup_robustness.py

@@ -1,7 +1,6 @@
 import copy
 import random
-import os.path
+
-import pickle
 from pathlib import Path
 from typing import Union
 
@@ -13,7 +12,6 @@ from matplotlib import pyplot as plt
 from torch.nn import functional as F
 from tabulate import tabulate
 
-from experiments.helpers import check_folder, summary_fixpoint_percentage, summary_fixpoint_experiment
 from functionalities_test import test_for_fixpoints, is_zero_fixpoint, is_divergent, is_identity_function
 from network import Net
 from visualization import plot_loss, bar_chart_fixpoints, plot_3d_soup, line_chart_fixpoints
@@ -25,20 +23,6 @@ def prng():
 
 class SoupRobustnessExperiment:
 
-    @staticmethod
-    def apply_noise(network, noise: int):
-        """ Changing the weights of a network to values + noise """
-        for layer_id, layer_name in enumerate(network.state_dict()):
-            for line_id, line_values in enumerate(network.state_dict()[layer_name]):
-                for weight_id, weight_value in enumerate(network.state_dict()[layer_name][line_id]):
-                    # network.state_dict()[layer_name][line_id][weight_id] = weight_value + noise
-                    if prng() < 0.5:
-                        network.state_dict()[layer_name][line_id][weight_id] = weight_value + noise
-                    else:
-                        network.state_dict()[layer_name][line_id][weight_id] = weight_value - noise
-
-        return network
-
     def __init__(self, population_size, net_i_size, net_h_size, net_o_size, learning_rate, attack_chance,
                  train_nets, ST_steps, epochs, log_step_size, directory: Union[str, Path]):
         super().__init__()
@@ -146,8 +130,7 @@ class SoupRobustnessExperiment:
                     clone = Net(fixpoint.input_size, fixpoint.hidden_size, fixpoint.out_size,
                                 f"{fixpoint.name}_clone_noise10e-{noise_level}")
                     clone.load_state_dict(copy.deepcopy(fixpoint.state_dict()))
-                    rand_noise = prng() * pow(10, -noise_level)  # n / 1
+                    clone = clone.apply_noise(pow(10, -noise_level))
-                    clone = self.apply_noise(clone, rand_noise)
 
                     while not is_zero_fixpoint(clone) and not is_divergent(clone):
                         if is_identity_function(clone):

network.py

@@ -1,5 +1,6 @@
 # from __future__ import annotations
 import copy
+import random
 from typing import Union
 
 import torch
@@ -9,7 +10,12 @@ import numpy as np
 from torch import optim, Tensor
 
 
+def prng():
+    return random.random()
+
+
 class Net(nn.Module):
+
     @staticmethod
     def create_target_weights(input_weight_matrix: Tensor) -> Tensor:
         """ Outputting a tensor with the target weights. """
@@ -171,3 +177,16 @@ class Net(nn.Module):
         SA_steps = 1
 
         return other_net.apply_weights(my_evaluation)
+
+    def apply_noise(self, noise_size: float):
+        """ Changing the weights of a network to values + noise """
+        for layer_id, layer_name in enumerate(self.state_dict()):
+            for line_id, line_values in enumerate(self.state_dict()[layer_name]):
+                for weight_id, weight_value in enumerate(self.state_dict()[layer_name][line_id]):
+                    # network.state_dict()[layer_name][line_id][weight_id] = weight_value + noise
+                    if prng() < 0.5:
+                        self.state_dict()[layer_name][line_id][weight_id] = weight_value + noise_size * prng()
+                    else:
+                        self.state_dict()[layer_name][line_id][weight_id] = weight_value - noise_size * prng()
+
+        return self

visualization.py

@@ -9,6 +9,9 @@ from sklearn.decomposition import PCA
 import random
 import string
 
+from matplotlib import rcParams
+rcParams['axes.labelpad'] = 20
+
 
 def plot_output(output):
     """ Plotting the values of the final output """
@@ -65,6 +68,7 @@ def bar_chart_fixpoints(fixpoint_counter: Dict, population_size: int, directory:
     plt.xticks(range(len(fixpoint_counter)), list(fixpoint_counter.keys()))
 
     directory = Path(directory)
+    directory.mkdir(parents=True, exist_ok=True)
     filename = f"{str(population_size)}_nets_fixpoints_barchart.png"
     filepath = directory / filename
     plt.savefig(str(filepath))
@@ -139,19 +143,19 @@ def plot_3d(matrices_weights_history, directory: Union[str, Path], population_si
 
     #steps = mpatches.Patch(color="white", label=f"{z_axis_legend}: {len(matrices_weights_history)} steps")
     population_size = mpatches.Patch(color="white", label=f"Population: {population_size} networks")
-
+    if False:
-    if z_axis_legend == "Self-application":
+        if z_axis_legend == "Self-application":
-        if is_trained == '_trained':
+            if is_trained == '_trained':
-            trained = mpatches.Patch(color="white", label=f"Trained: true")
+                trained = mpatches.Patch(color="white", label=f"Trained: true")
+            else:
+                trained = mpatches.Patch(color="white", label=f"Trained: false")
+            ax.legend(handles=[population_size, trained])
         else:
-            trained = mpatches.Patch(color="white", label=f"Trained: false")
+            ax.legend(handles=[population_size])
-        ax.legend(handles=[population_size, trained])
-    else:
-        ax.legend(handles=[population_size])
 
     ax.set_title(f"PCA Transformed Weight Trajectories")
-    ax.set_xlabel("PCA Transformed X-Axis")
+    # ax.set_xlabel("PCA Transformed X-Axis")
-    ax.set_ylabel("PCA Transformed Y-Axis")
+    # ax.set_ylabel("PCA Transformed Y-Axis")
     ax.set_zlabel(f"Self Training Steps")
 
     # FIXME: Replace this kind of operation with pathlib.Path() object interactions