Merge remote-tracking branch 'origin/journal' into journal
# Conflicts: # journal_basins.py
This commit is contained in:
steffen-illium
commit
800a2c8f6b
@ -78,3 +78,18 @@ 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):
|
||||
net.is_fixpoint = "divergent"
|
||||
elif is_identity_function(net): # is default value
|
||||
net.is_fixpoint = "identity_func"
|
||||
elif is_zero_fixpoint(net):
|
||||
net.is_fixpoint = "fix_zero"
|
||||
elif is_secondary_fixpoint(net):
|
||||
net.is_fixpoint = "fix_sec"
|
||||
else:
|
||||
net.is_fixpoint = "other_func"
|
||||
|
||||
return net
|
||||
@ -1,18 +1,21 @@
|
||||
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
|
||||
from functionalities_test import is_identity_function, test_status
|
||||
from network import Net
|
||||
from visualization import plot_3d_self_train, plot_loss
|
||||
import numpy as np
|
||||
from tabulate import tabulate
|
||||
from sklearn.metrics import mean_absolute_error as MAE
|
||||
from sklearn.metrics import mean_squared_error as MSE
|
||||
|
||||
import pandas as pd
|
||||
import seaborn as sns
|
||||
from matplotlib import pyplot as plt
|
||||
|
||||
def prng():
|
||||
return random.random()
|
||||
@ -120,8 +123,8 @@ class SpawnExperiment:
|
||||
self.spawn_and_continue()
|
||||
self.weights_evolution_3d_experiment()
|
||||
# self.visualize_loss()
|
||||
self.distance_matrix = distance_matrix(self.nets)
|
||||
self.parent_clone_distances = distance_from_parent(self.nets)
|
||||
self.distance_matrix = distance_matrix(self.nets, print_it=False)
|
||||
self.parent_clone_distances = distance_from_parent(self.nets, print_it=False)
|
||||
|
||||
self.save()
|
||||
|
||||
@ -136,13 +139,13 @@ class SpawnExperiment:
|
||||
for _ in range(self.ST_steps):
|
||||
net.self_train(1, self.log_step_size, self.net_learning_rate)
|
||||
|
||||
# print(f"\nLast weight matrix (epoch: {self.epochs}):\n
|
||||
# {net.input_weight_matrix()}\nLossHistory: {net.loss_history[-10:]}")
|
||||
self.nets.append(net)
|
||||
|
||||
def spawn_and_continue(self, number_clones: int = None):
|
||||
number_clones = number_clones or self.nr_clones
|
||||
|
||||
df = pd.DataFrame(columns=['parent', 'MAE_pre','MAE_post', 'MSE_pre', 'MSE_post', 'MIM_pre', 'MIM_post', 'noise', 'status_post'])
|
||||
|
||||
# For every initial net {i} after populating (that is fixpoint after first epoch);
|
||||
for i in range(self.population_size):
|
||||
net = self.nets[i]
|
||||
@ -168,27 +171,46 @@ class SpawnExperiment:
|
||||
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)
|
||||
|
||||
# Pre Training distances (after noise application of course)
|
||||
clone_pre_weights = clone.create_target_weights(clone.input_weight_matrix())
|
||||
MAE_pre = MAE(net_target_data, clone_pre_weights)
|
||||
MSE_pre = MSE(net_target_data, clone_pre_weights)
|
||||
MIM_pre = mean_invariate_manhattan_distance(net_target_data, clone_pre_weights)
|
||||
|
||||
# Then finish training each clone {j} (for remaining epoch-1 * ST_steps)
|
||||
# and add to nets for plotting if they are fixpoints themselves;
|
||||
# Then finish training each clone {j} (for remaining epoch-1 * ST_steps) ..
|
||||
for _ in range(self.epochs - 1):
|
||||
for _ in range(self.ST_steps):
|
||||
# soup Evolve
|
||||
clone.self_train(1, self.log_step_size, self.net_learning_rate)
|
||||
|
||||
# Post Training distances for comparison
|
||||
clone_post_weights = clone.create_target_weights(clone.input_weight_matrix())
|
||||
MAE_post = MAE(net_target_data, clone_post_weights)
|
||||
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.
|
||||
test_status(clone)
|
||||
if is_identity_function(clone):
|
||||
input_data = clone.input_weight_matrix()
|
||||
target_data = clone.create_target_weights(input_data)
|
||||
print(f"Clone {j} (of net_{i}) is fixpoint. \nMSE(j,i): "
|
||||
f"{MSE(net_target_data, target_data)}, \nMAE(j,i): {MAE(net_target_data, target_data)}\n")
|
||||
self.nets.append(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)
|
||||
|
||||
df.loc[clone.name] = [net.name, MAE_pre, MAE_post, MSE_pre, MSE_post, MIM_pre, MIM_post, self.noise, clone.is_fixpoint]
|
||||
|
||||
# Finally take parent net {i} and finish it's training for comparison to clone development.
|
||||
for _ in range(self.epochs - 1):
|
||||
for _ in range(self.ST_steps):
|
||||
net.self_train(1, self.log_step_size, self.net_learning_rate)
|
||||
net_weights_after = net.create_target_weights(net.input_weight_matrix())
|
||||
print(f"Parent net's distance to original position."
|
||||
f"\nMSE(OG,new): {MAE(net_target_data, net_weights_after)}"
|
||||
f"\nMAE(OG,new): {MSE(net_target_data, net_weights_after)}"
|
||||
f"\nMIM(OG,new): {mean_invariate_manhattan_distance(net_target_data, net_weights_after)}\n")
|
||||
|
||||
else:
|
||||
print("No fixpoints found.")
|
||||
self.df = df
|
||||
|
||||
def weights_evolution_3d_experiment(self):
|
||||
exp_name = f"ST_{str(len(self.nets))}_nets_3d_weights_PCA"
|
||||
@ -213,20 +235,21 @@ if __name__ == "__main__":
|
||||
# Define number of runs & name:
|
||||
ST_runs = 1
|
||||
ST_runs_name = "test-27"
|
||||
ST_steps = 2000
|
||||
ST_steps = 2500
|
||||
ST_epochs = 2
|
||||
ST_log_step_size = 10
|
||||
|
||||
# Define number of networks & their architecture
|
||||
nr_clones = 50
|
||||
nr_clones = 5
|
||||
ST_population_size = 1
|
||||
ST_net_hidden_size = 2
|
||||
ST_net_learning_rate = 0.04
|
||||
ST_name_hash = random.getrandbits(32)
|
||||
|
||||
print(f"Running the Spawn experiment:")
|
||||
for noise_factor in [9]:
|
||||
SpawnExperiment(
|
||||
exp_list = []
|
||||
for noise_factor in range(2,5):
|
||||
exp = SpawnExperiment(
|
||||
population_size=ST_population_size,
|
||||
log_step_size=ST_log_step_size,
|
||||
net_input_size=NET_INPUT_SIZE,
|
||||
@ -237,5 +260,16 @@ if __name__ == "__main__":
|
||||
st_steps=ST_steps,
|
||||
nr_clones=nr_clones,
|
||||
noise=pow(10, -noise_factor),
|
||||
directory=Path('output') / 'spawn_basin' / f'{ST_name_hash}_10e-{noise_factor}'
|
||||
directory=Path('output') / 'spawn_basin' / f'{ST_name_hash}' / f'10e-{noise_factor}'
|
||||
)
|
||||
exp_list.append(exp)
|
||||
|
||||
# Boxplot with counts of nr_fixpoints, nr_other, nr_etc. on y-axis
|
||||
df = pd.concat([exp.df for exp in exp_list])
|
||||
sns.countplot(data=df, x="noise", hue="status_post")
|
||||
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')
|
||||
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")
|
||||
304
journal_soup_basins.py
Normal file
304
journal_soup_basins.py
Normal file
@ -0,0 +1,304 @@
|
||||
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
|
||||
from network import Net
|
||||
from visualization import plot_3d_self_train, plot_loss, plot_3d_soup
|
||||
import numpy as np
|
||||
from tabulate import tabulate
|
||||
from sklearn.metrics import mean_absolute_error as MAE
|
||||
from sklearn.metrics import mean_squared_error as MSE
|
||||
import pandas as pd
|
||||
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)
|
||||
|
||||
|
||||
def mean_invariate_manhattan_distance(x, y):
|
||||
# One of these one-liners that might be smart or really dumb. Goal is to find pairwise
|
||||
# distances of ascending values, ie. sum (abs(min1_X-min1_Y), abs(min2_X-min2Y) ...) / mean.
|
||||
# Idea was to find weight sets that have same values but just in different positions, that would
|
||||
# make this distance 0.
|
||||
return np.mean(list(map(l1, zip(sorted(x.numpy()), sorted(y.numpy())))))
|
||||
|
||||
|
||||
def distance_matrix(nets, distance="MIM", print_it=True):
|
||||
matrix = [[0 for _ in range(len(nets))] for _ in range(len(nets))]
|
||||
for net in range(len(nets)):
|
||||
weights = nets[net].input_weight_matrix()[:, 0]
|
||||
for other_net in range(len(nets)):
|
||||
other_weights = nets[other_net].input_weight_matrix()[:, 0]
|
||||
if distance in ["MSE"]:
|
||||
matrix[net][other_net] = MSE(weights, other_weights)
|
||||
elif distance in ["MAE"]:
|
||||
matrix[net][other_net] = MAE(weights, other_weights)
|
||||
elif distance in ["MIM"]:
|
||||
matrix[net][other_net] = mean_invariate_manhattan_distance(weights, other_weights)
|
||||
|
||||
if print_it:
|
||||
print(f"\nDistance matrix (all to all) [{distance}]:")
|
||||
headers = [i.name for i in nets]
|
||||
print(tabulate(matrix, showindex=headers, headers=headers, tablefmt='orgtbl'))
|
||||
return matrix
|
||||
|
||||
|
||||
def distance_from_parent(nets, distance="MIM", print_it=True):
|
||||
list_of_matrices = []
|
||||
parents = list(filter(lambda x: "clone" not in x.name and is_identity_function(x), nets))
|
||||
distance_range = range(10)
|
||||
for parent in parents:
|
||||
parent_weights = parent.create_target_weights(parent.input_weight_matrix())
|
||||
clones = list(filter(lambda y: parent.name in y.name and parent.name != y.name, nets))
|
||||
matrix = [[0 for _ in distance_range] for _ in range(len(clones))]
|
||||
|
||||
for dist in distance_range:
|
||||
for idx, clone in enumerate(clones):
|
||||
clone_weights = clone.create_target_weights(clone.input_weight_matrix())
|
||||
if distance in ["MSE"]:
|
||||
matrix[idx][dist] = MSE(parent_weights, clone_weights) < pow(10, -dist)
|
||||
elif distance in ["MAE"]:
|
||||
matrix[idx][dist] = MAE(parent_weights, clone_weights) < pow(10, -dist)
|
||||
elif distance in ["MIM"]:
|
||||
matrix[idx][dist] = mean_invariate_manhattan_distance(parent_weights, clone_weights) < pow(10,
|
||||
-dist)
|
||||
|
||||
if print_it:
|
||||
print(f"\nDistances from parent {parent.name} [{distance}]:")
|
||||
col_headers = [str(f"10e-{d}") for d in distance_range]
|
||||
row_headers = [str(f"clone_{i}") for i in range(len(clones))]
|
||||
print(tabulate(matrix, showindex=row_headers, headers=col_headers, tablefmt='orgtbl'))
|
||||
|
||||
list_of_matrices.append(matrix)
|
||||
|
||||
return list_of_matrices
|
||||
|
||||
|
||||
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:
|
||||
self.population_size = population_size
|
||||
self.log_step_size = log_step_size
|
||||
self.net_input_size = net_input_size
|
||||
self.net_hidden_size = net_hidden_size
|
||||
self.net_out_size = net_out_size
|
||||
self.net_learning_rate = net_learning_rate
|
||||
self.epochs = epochs
|
||||
self.ST_steps = st_steps
|
||||
self.attack_chance = attack_chance
|
||||
self.loss_history = []
|
||||
self.nr_clones = nr_clones
|
||||
self.noise = noise or 10e-5
|
||||
print("\nNOISE:", self.noise)
|
||||
|
||||
self.directory = Path(directory)
|
||||
self.directory.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
# Populating environment & evolving entities
|
||||
self.nets = []
|
||||
self.populate_environment()
|
||||
self.evolve()
|
||||
|
||||
self.spawn_and_continue()
|
||||
self.weights_evolution_3d_experiment()
|
||||
# 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.save()
|
||||
|
||||
def populate_environment(self):
|
||||
loop_population_size = tqdm(range(self.population_size))
|
||||
for i in loop_population_size:
|
||||
loop_population_size.set_description("Populating experiment %s" % i)
|
||||
|
||||
net_name = f"soup_net_{str(i)}"
|
||||
net = Net(self.net_input_size, self.net_hidden_size, self.net_out_size, net_name)
|
||||
|
||||
self.nets.append(net)
|
||||
|
||||
def evolve(self):
|
||||
loop_epochs = tqdm(range(self.epochs))
|
||||
for i in loop_epochs:
|
||||
loop_epochs.set_description("Evolving soup %s" % i)
|
||||
|
||||
# A network attacking another network with a given percentage
|
||||
if random.randint(1, 100) <= self.attack_chance:
|
||||
random_net1, random_net2 = random.sample(range(self.population_size), 2)
|
||||
random_net1 = self.nets[random_net1]
|
||||
random_net2 = self.nets[random_net2]
|
||||
print(f"\n Attack: {random_net1.name} -> {random_net2.name}")
|
||||
random_net1.attack(random_net2)
|
||||
|
||||
# Self-training each network in the population
|
||||
for j in range(self.population_size):
|
||||
net = self.nets[j]
|
||||
|
||||
for _ in range(self.ST_steps):
|
||||
net.self_train(1, self.log_step_size, self.net_learning_rate)
|
||||
|
||||
def spawn_and_continue(self, number_clones: int = None):
|
||||
number_clones = number_clones or self.nr_clones
|
||||
|
||||
df = pd.DataFrame(
|
||||
columns=['parent', 'MAE_pre', 'MAE_post', 'MSE_pre', 'MSE_post', 'MIM_pre', 'MIM_post', 'noise',
|
||||
'status_post'])
|
||||
|
||||
# For every initial net {i} after populating (that is fixpoint after first epoch);
|
||||
for i in range(self.population_size):
|
||||
net = self.nets[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.
|
||||
net.start_time = self.ST_steps - 150
|
||||
net_input_data = net.input_weight_matrix()
|
||||
net_target_data = net.create_target_weights(net_input_data)
|
||||
|
||||
if is_identity_function(net):
|
||||
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!
|
||||
# To make sure PCA will plot the same trajectory up until this point, we clone the
|
||||
# parent-net's weight history as well.
|
||||
for j in range(number_clones):
|
||||
clone = Net(net.input_size, net.hidden_size, net.out_size,
|
||||
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.s_train_weights_history = copy.deepcopy(net.s_train_weights_history)
|
||||
clone.number_trained = copy.deepcopy(net.number_trained)
|
||||
|
||||
# Pre Training distances (after noise application of course)
|
||||
clone_pre_weights = clone.create_target_weights(clone.input_weight_matrix())
|
||||
MAE_pre = MAE(net_target_data, clone_pre_weights)
|
||||
MSE_pre = MSE(net_target_data, clone_pre_weights)
|
||||
MIM_pre = mean_invariate_manhattan_distance(net_target_data, clone_pre_weights)
|
||||
|
||||
# Then finish training each clone {j} (for remaining epoch-1 * ST_steps) ..
|
||||
for _ in range(self.epochs - 1):
|
||||
for _ in range(self.ST_steps):
|
||||
clone.self_train(1, self.log_step_size, self.net_learning_rate)
|
||||
|
||||
# Post Training distances for comparison
|
||||
clone_post_weights = clone.create_target_weights(clone.input_weight_matrix())
|
||||
MAE_post = MAE(net_target_data, clone_post_weights)
|
||||
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.
|
||||
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)
|
||||
|
||||
df.loc[clone.name] = [net.name, MAE_pre, MAE_post, MSE_pre, MSE_post, MIM_pre, MIM_post, self.noise,
|
||||
clone.is_fixpoint]
|
||||
|
||||
# Finally take parent net {i} and finish it's training for comparison to clone development.
|
||||
for _ in range(self.epochs - 1):
|
||||
for _ in range(self.ST_steps):
|
||||
net.self_train(1, self.log_step_size, self.net_learning_rate)
|
||||
net_weights_after = net.create_target_weights(net.input_weight_matrix())
|
||||
print(f"Parent net's distance to original position."
|
||||
f"\nMSE(OG,new): {MAE(net_target_data, net_weights_after)}"
|
||||
f"\nMAE(OG,new): {MSE(net_target_data, net_weights_after)}"
|
||||
f"\nMIM(OG,new): {mean_invariate_manhattan_distance(net_target_data, net_weights_after)}\n")
|
||||
|
||||
self.df = df
|
||||
|
||||
def weights_evolution_3d_experiment(self):
|
||||
exp_name = f"soup_basins_{str(len(self.nets))}_nets_3d_weights_PCA"
|
||||
return plot_3d_soup(self.nets, exp_name, self.directory)
|
||||
|
||||
def visualize_loss(self):
|
||||
for i in range(len(self.nets)):
|
||||
net_loss_history = self.nets[i].loss_history
|
||||
self.loss_history.append(net_loss_history)
|
||||
plot_loss(self.loss_history, self.directory)
|
||||
|
||||
def save(self):
|
||||
pickle.dump(self, open(f"{self.directory}/experiment_pickle.p", "wb"))
|
||||
print(f"\nSaved experiment to {self.directory}.")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
NET_INPUT_SIZE = 4
|
||||
NET_OUT_SIZE = 1
|
||||
|
||||
# Define number of runs & name:
|
||||
ST_runs = 1
|
||||
ST_runs_name = "test-27"
|
||||
soup_ST_steps = 2500
|
||||
soup_epochs = 2
|
||||
soup_log_step_size = 10
|
||||
|
||||
# Define number of networks & their architecture
|
||||
nr_clones = 15
|
||||
soup_population_size = 2
|
||||
soup_net_hidden_size = 2
|
||||
soup_net_learning_rate = 0.04
|
||||
soup_attack_chance = 10
|
||||
soup_name_hash = random.getrandbits(32)
|
||||
|
||||
print(f"Running the Soup-Spawn experiment:")
|
||||
exp_list = []
|
||||
for noise_factor in range(2, 5):
|
||||
exp = SoupSpawnExperiment(
|
||||
population_size=soup_population_size,
|
||||
log_step_size=soup_log_step_size,
|
||||
net_input_size=NET_INPUT_SIZE,
|
||||
net_hidden_size=soup_net_hidden_size,
|
||||
net_out_size=NET_OUT_SIZE,
|
||||
net_learning_rate=soup_net_learning_rate,
|
||||
epochs=soup_epochs,
|
||||
st_steps=soup_ST_steps,
|
||||
attack_chance=soup_attack_chance,
|
||||
nr_clones=nr_clones,
|
||||
noise=pow(10, -noise_factor),
|
||||
directory=Path('output') / 'soup_spawn_basin' / f'{soup_name_hash}' / f'10e-{noise_factor}'
|
||||
)
|
||||
exp_list.append(exp)
|
||||
|
||||
# Boxplot with counts of nr_fixpoints, nr_other, nr_etc. on y-axis
|
||||
df = pd.concat([exp.df for exp in exp_list])
|
||||
sns.countplot(data=df, x="noise", hue="status_post")
|
||||
plt.savefig(f"output/soup_spawn_basin/{soup_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')
|
||||
sns.catplot(data=mlt, x="time", y="Average Distance", col="noise", kind="point", col_wrap=5, sharey=False)
|
||||
plt.savefig(f"output/soup_spawn_basin/{soup_name_hash}/clone_distance_catplot.png")
|
||||
266
journal_soup_robustness.py
Normal file
266
journal_soup_robustness.py
Normal file
@ -0,0 +1,266 @@
|
||||
import copy
|
||||
import random
|
||||
import os.path
|
||||
import pickle
|
||||
from pathlib import Path
|
||||
from typing import Union
|
||||
|
||||
import numpy as np
|
||||
import pandas as pd
|
||||
import seaborn as sns
|
||||
from tqdm import tqdm
|
||||
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
|
||||
|
||||
|
||||
def prng():
|
||||
return random.random()
|
||||
|
||||
|
||||
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__()
|
||||
self.population_size = population_size
|
||||
|
||||
self.net_input_size = net_i_size
|
||||
self.net_hidden_size = net_h_size
|
||||
self.net_out_size = net_o_size
|
||||
self.net_learning_rate = learning_rate
|
||||
self.attack_chance = attack_chance
|
||||
self.train_nets = train_nets
|
||||
# self.SA_steps = SA_steps
|
||||
self.ST_steps = ST_steps
|
||||
self.epochs = epochs
|
||||
self.log_step_size = log_step_size
|
||||
|
||||
self.loss_history = []
|
||||
|
||||
self.fixpoint_counters = {
|
||||
"identity_func": 0,
|
||||
"divergent": 0,
|
||||
"fix_zero": 0,
|
||||
"fix_weak": 0,
|
||||
"fix_sec": 0,
|
||||
"other_func": 0
|
||||
}
|
||||
# <self.fixpoint_counters_history> is used for keeping track of the amount of fixpoints in %
|
||||
self.fixpoint_counters_history = []
|
||||
self.id_functions = []
|
||||
|
||||
self.directory = Path(directory)
|
||||
self.directory.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
self.population = []
|
||||
self.populate_environment()
|
||||
|
||||
self.evolve()
|
||||
self.fixpoint_percentage()
|
||||
self.weights_evolution_3d_experiment()
|
||||
self.count_fixpoints()
|
||||
self.visualize_loss()
|
||||
|
||||
self.time_to_vergence, self.time_as_fixpoint = self.test_robustness()
|
||||
|
||||
def populate_environment(self):
|
||||
loop_population_size = tqdm(range(self.population_size))
|
||||
for i in tqdm(range(self.population_size)):
|
||||
loop_population_size.set_description("Populating soup experiment %s" % i)
|
||||
|
||||
net_name = f"soup_network_{i}"
|
||||
net = Net(self.net_input_size, self.net_hidden_size, self.net_out_size, net_name)
|
||||
self.population.append(net)
|
||||
|
||||
def evolve(self):
|
||||
""" Evolving consists of attacking & self-training. """
|
||||
|
||||
loop_epochs = tqdm(range(self.epochs))
|
||||
for i in loop_epochs:
|
||||
loop_epochs.set_description("Evolving soup %s" % i)
|
||||
|
||||
# A network attacking another network with a given percentage
|
||||
if random.randint(1, 100) <= self.attack_chance:
|
||||
random_net1, random_net2 = random.sample(range(self.population_size), 2)
|
||||
random_net1 = self.population[random_net1]
|
||||
random_net2 = self.population[random_net2]
|
||||
print(f"\n Attack: {random_net1.name} -> {random_net2.name}")
|
||||
random_net1.attack(random_net2)
|
||||
|
||||
# Self-training each network in the population
|
||||
for j in range(self.population_size):
|
||||
net = self.population[j]
|
||||
|
||||
for _ in range(self.ST_steps):
|
||||
net.self_train(1, self.log_step_size, self.net_learning_rate)
|
||||
|
||||
# Testing for fixpoints after each batch of ST steps to see relevant data
|
||||
if i % self.ST_steps == 0:
|
||||
test_for_fixpoints(self.fixpoint_counters, self.population)
|
||||
fixpoints_percentage = round(self.fixpoint_counters["identity_func"] / self.population_size, 1)
|
||||
self.fixpoint_counters_history.append(fixpoints_percentage)
|
||||
|
||||
# Resetting the fixpoint counter. Last iteration not to be reset -
|
||||
# it is important for the bar_chart_fixpoints().
|
||||
if i < self.epochs:
|
||||
self.reset_fixpoint_counters()
|
||||
|
||||
def test_robustness(self, print_it=True, noise_levels=10, seeds=10):
|
||||
# assert (len(self.id_functions) == 1 and seeds > 1) or (len(self.id_functions) > 1 and seeds == 1)
|
||||
is_synthetic = True if len(self.id_functions) > 1 and seeds == 1 else False
|
||||
avg_time_to_vergence = [[0 for _ in range(noise_levels)] for _ in
|
||||
range(seeds if is_synthetic else len(self.id_functions))]
|
||||
avg_time_as_fixpoint = [[0 for _ in range(noise_levels)] for _ in
|
||||
range(seeds if is_synthetic else len(self.id_functions))]
|
||||
row_headers = []
|
||||
data_pos = 0
|
||||
# This checks wether to use synthetic setting with multiple seeds
|
||||
# or multi network settings with a singlee seed
|
||||
|
||||
df = pd.DataFrame(columns=['seed', 'noise_level', 'application_step', 'absolute_loss'])
|
||||
for i, fixpoint in enumerate(self.id_functions): # 1 / n
|
||||
row_headers.append(fixpoint.name)
|
||||
for seed in range(seeds): # n / 1
|
||||
for noise_level in range(noise_levels):
|
||||
self_application_steps = 1
|
||||
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 = self.apply_noise(clone, rand_noise)
|
||||
|
||||
while not is_zero_fixpoint(clone) and not is_divergent(clone):
|
||||
if is_identity_function(clone):
|
||||
avg_time_as_fixpoint[i][noise_level] += 1
|
||||
|
||||
# -> before
|
||||
clone_weight_pre_application = clone.input_weight_matrix()
|
||||
target_data_pre_application = clone.create_target_weights(clone_weight_pre_application)
|
||||
|
||||
clone.self_application(1, self.log_step_size)
|
||||
avg_time_to_vergence[i][noise_level] += 1
|
||||
# -> after
|
||||
clone_weight_post_application = clone.input_weight_matrix()
|
||||
target_data_post_application = clone.create_target_weights(clone_weight_post_application)
|
||||
|
||||
absolute_loss = F.l1_loss(target_data_pre_application, target_data_post_application).item()
|
||||
|
||||
setting = i if is_synthetic else seed
|
||||
|
||||
df.loc[data_pos] = [setting, noise_level, self_application_steps, absolute_loss]
|
||||
data_pos += 1
|
||||
self_application_steps += 1
|
||||
|
||||
# calculate the average:
|
||||
df = df.replace([np.inf, -np.inf], np.nan)
|
||||
df = df.dropna()
|
||||
# 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'
|
||||
filename = f"absolute_loss_perapplication_boxplot_grid.png"
|
||||
filepath = directory / filename
|
||||
|
||||
plt.savefig(str(filepath))
|
||||
|
||||
if print_it:
|
||||
col_headers = [str(f"10e-{d}") for d in range(noise_levels)]
|
||||
|
||||
print(f"\nAppplications steps until divergence / zero: ")
|
||||
print(tabulate(avg_time_to_vergence, showindex=row_headers, headers=col_headers, tablefmt='orgtbl'))
|
||||
|
||||
print(f"\nTime as fixpoint: ")
|
||||
print(tabulate(avg_time_as_fixpoint, showindex=row_headers, headers=col_headers, tablefmt='orgtbl'))
|
||||
|
||||
return avg_time_as_fixpoint, avg_time_to_vergence
|
||||
|
||||
def weights_evolution_3d_experiment(self):
|
||||
exp_name = f"soup_{self.population_size}_nets_{self.ST_steps}_training_{self.epochs}_epochs"
|
||||
return plot_3d_soup(self.population, exp_name, self.directory)
|
||||
|
||||
def count_fixpoints(self):
|
||||
self.id_functions = test_for_fixpoints(self.fixpoint_counters, self.population)
|
||||
exp_details = f"Evolution steps: {self.epochs} epochs"
|
||||
bar_chart_fixpoints(self.fixpoint_counters, self.population_size, self.directory, self.net_learning_rate,
|
||||
exp_details)
|
||||
|
||||
def fixpoint_percentage(self):
|
||||
runs = self.epochs / self.ST_steps
|
||||
SA_steps = None
|
||||
line_chart_fixpoints(self.fixpoint_counters_history, runs, self.ST_steps, SA_steps, self.directory,
|
||||
self.population_size)
|
||||
|
||||
def visualize_loss(self):
|
||||
for i in range(len(self.population)):
|
||||
net_loss_history = self.population[i].loss_history
|
||||
self.loss_history.append(net_loss_history)
|
||||
|
||||
plot_loss(self.loss_history, self.directory)
|
||||
|
||||
def reset_fixpoint_counters(self):
|
||||
self.fixpoint_counters = {
|
||||
"identity_func": 0,
|
||||
"divergent": 0,
|
||||
"fix_zero": 0,
|
||||
"fix_weak": 0,
|
||||
"fix_sec": 0,
|
||||
"other_func": 0
|
||||
}
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
NET_INPUT_SIZE = 4
|
||||
NET_OUT_SIZE = 1
|
||||
|
||||
soup_epochs = 100
|
||||
soup_log_step_size = 5
|
||||
soup_ST_steps = 20
|
||||
# soup_SA_steps = 10
|
||||
|
||||
# Define number of networks & their architecture
|
||||
soup_population_size = 20
|
||||
soup_net_hidden_size = 2
|
||||
soup_net_learning_rate = 0.04
|
||||
|
||||
# soup_attack_chance in %
|
||||
soup_attack_chance = 10
|
||||
|
||||
# not used yet: soup_train_nets has 3 possible values "no", "before_SA", "after_SA".
|
||||
soup_train_nets = "no"
|
||||
soup_name_hash = random.getrandbits(32)
|
||||
soup_synthetic = True
|
||||
|
||||
print(f"Running the robustness comparison experiment:")
|
||||
SoupRobustnessExperiment(
|
||||
population_size=soup_population_size,
|
||||
net_i_size=NET_INPUT_SIZE,
|
||||
net_h_size=soup_net_hidden_size,
|
||||
net_o_size=NET_OUT_SIZE,
|
||||
learning_rate=soup_net_learning_rate,
|
||||
attack_chance=soup_attack_chance,
|
||||
train_nets=soup_train_nets,
|
||||
ST_steps=soup_ST_steps,
|
||||
epochs=soup_epochs,
|
||||
log_step_size=soup_log_step_size,
|
||||
directory=Path('output') / 'robustness' / f'{soup_name_hash}'
|
||||
)
|
||||
Loading…
x
Reference in New Issue
Block a user