steffen/self-replicating-neural-networks
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

debugged journal_soup_basin experiment

Browse Source
This commit is contained in:
ru43zex
2021-06-07 11:12:59 +03:00
parent 7e231b5b50
commit 987d7b95f3
3 changed files with 44 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
functionalities_test.py
View File

@@ -79,6 +79,7 @@ def test_for_fixpoints(fixpoint_counter: Dict, nets: List, id_functions=None):
def changing_rate(x_new, x_old):
return x_new - x_old
def test_status(net: Net) -> Net:
if is_divergent(net):

65
journal_soup_basins.py
View File

@@ -6,7 +6,7 @@ from torch import mean
from tqdm import tqdm
import random
import copy
from functionalities_test import is_identity_function, test_status, test_for_fixpoints
from functionalities_test import is_identity_function, test_status, test_for_fixpoints, is_zero_fixpoint, is_divergent, is_secondary_fixpoint
from network import Net
from visualization import plot_3d_self_train, plot_loss, plot_3d_soup
import numpy as np
@@ -123,17 +123,22 @@ class SoupSpawnExperiment:
self.directory.mkdir(parents=True, exist_ok=True)
# Populating environment & evolving entities
self.nets = []
self.id_functions = []
self.clone_soup = []
self.parents = []
self.clones = []
self.parents_with_clones = []
self.parents_clones_id_functions = []
self.populate_environment()
self.spawn_and_continue()
self.weights_evolution_3d_experiment(self.nets, "parents")
self.weights_evolution_3d_experiment(self.clone_soup, "clones")
self.weights_evolution_3d_experiment(self.parents, "only_parents")
self.weights_evolution_3d_experiment(self.clones, "only_clones")
self.weights_evolution_3d_experiment(self.parents_with_clones, "parents_with_clones")
self.weights_evolution_3d_experiment(self.parents_clones_id_functions, "id_f_with_parents")
# self.visualize_loss()
self.distance_matrix = distance_matrix(self.nets, print_it=False)
self.parent_clone_distances = distance_from_parent(self.nets, print_it=False)
self.distance_matrix = distance_matrix(self.parents_clones_id_functions, print_it=False)
self.parent_clone_distances = distance_from_parent(self.parents_clones_id_functions, print_it=False)
self.save()
@@ -148,10 +153,21 @@ class SoupSpawnExperiment:
for _ in range(self.ST_steps):
net.self_train(1, self.log_step_size, self.net_learning_rate)
self.nets.append(net)
self.parents.append(net)
self.parents_with_clones.append(net)
if is_identity_function(net):
self.id_functions.append(net)
self.parents_clones_id_functions.append(net)
print(f"\nNet {net.name} is identity function")
if is_divergent(net):
print(f"\nNet {net.name} is divergent")
if is_zero_fixpoint(net):
print(f"\nNet {net.name} is zero fixpoint")
if is_secondary_fixpoint(net):
print(f"\nNet {net.name} is secondary fixpoint")
def evolve(self, population):
print(f"Clone soup has a population of {len(population)} networks")
@@ -185,8 +201,8 @@ class SoupSpawnExperiment:
# MAE_pre, MSE_pre, MIM_pre = 0, 0, 0
# For every initial net {i} after populating (that is fixpoint after first epoch);
for i in range(len(self.id_functions)):
net = self.nets[i]
for i in range(len(self.parents)):
net = self.parents[i]
# We set parent start_time to just before this epoch ended, so plotting is zoomed in. Comment out to
# to see full trajectory (but the clones will be very hard to see).
# Make one target to compare distances to clones later when they have trained.
@@ -194,7 +210,7 @@ class SoupSpawnExperiment:
net_input_data = net.input_weight_matrix()
net_target_data = net.create_target_weights(net_input_data)
print(f"\nNet {i} is fixpoint")
# print(f"\nNet {i} is fixpoint")
# 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!
@@ -216,12 +232,15 @@ class SoupSpawnExperiment:
MIM_pre = mean_invariate_manhattan_distance(net_target_data, clone_pre_weights)
net.children.append(clone)
self.clone_soup.append(clone)
self.clones.append(clone)
self.parents_with_clones.append(clone)
self.evolve(self.clone_soup)
self.evolve(self.clones)
# evolve also with the parents together
# self.evolve(self.parents_with_clones)
for i in range(len(self.id_functions)):
net = self.nets[i]
for i in range(len(self.parents)):
net = self.parents[i]
net_input_data = net.input_weight_matrix()
net_target_data = net.create_target_weights(net_input_data)
@@ -234,14 +253,14 @@ class SoupSpawnExperiment:
MSE_post = MSE(net_target_data, clone_post_weights)
MIM_post = mean_invariate_manhattan_distance(net_target_data, clone_post_weights)
# .. log to data-frame and add to nets for 3d plotting if they are fixpoints themselves.
# .. log to data-frame and add to nets for 3d plotting if they are fixpoints themselves.
test_status(clone)
if is_identity_function(clone):
print(f"Clone {j} (of net_{i}) is fixpoint."
f"\nMSE({i},{j}): {MSE_post}"
f"\nMAE({i},{j}): {MAE_post}"
f"\nMIM({i},{j}): {MIM_post}\n")
self.nets.append(clone)
self.parents_clones_id_functions.append(clone)
df.loc[clone.name] = [net.name, MAE_pre, MAE_post, MSE_pre, MSE_post, MIM_pre, MIM_post, self.noise,
clone.is_fixpoint]
@@ -259,12 +278,12 @@ class SoupSpawnExperiment:
self.df = df
def weights_evolution_3d_experiment(self, nets_population, suffix):
exp_name = f"soup_basins_{str(len(self.nets))}_nets_3d_weights_PCA_{suffix}"
exp_name = f"soup_basins_{str(len(nets_population))}_nets_3d_weights_PCA_{suffix}"
return plot_3d_soup(nets_population, exp_name, self.directory)
def visualize_loss(self):
for i in range(len(self.nets)):
net_loss_history = self.nets[i].loss_history
for i in range(len(self.parents)):
net_loss_history = self.parents[i].loss_history
self.loss_history.append(net_loss_history)
plot_loss(self.loss_history, self.directory)
@@ -286,7 +305,7 @@ if __name__ == "__main__":
soup_log_step_size = 10
# Define number of networks & their architecture
nr_clones = 2
nr_clones = 3
soup_population_size = 2
soup_net_hidden_size = 2
soup_net_learning_rate = 0.04

2
visualization.py
View File

@@ -133,7 +133,7 @@ def plot_3d(matrices_weights_history, directory: Union[str, Path], population_si
ydata.append(weight_matrix_pca[j][1])
zdata = np.arange(start_time, len(ydata)*batch_size+start_time, batch_size)
ax.plot3D(xdata, ydata, zdata, label=f"net {i}")
ax.plot3D(xdata, ydata, zdata, label=f"net {i}", c="b")
if "parent" in nets_array[i].name:
ax.scatter(np.asarray(xdata), np.asarray(ydata), zdata, s=3, c="b")
else: