246 lines
10 KiB
Python
246 lines
10 KiB
Python
import pickle
|
|
|
|
import pandas as pd
|
|
import torch
|
|
import random
|
|
import copy
|
|
|
|
from pathlib import Path
|
|
|
|
from matplotlib.ticker import ScalarFormatter
|
|
from tqdm import tqdm
|
|
from tabulate import tabulate
|
|
|
|
from functionalities_test import is_identity_function, is_zero_fixpoint, test_for_fixpoints, is_divergent
|
|
from network import Net
|
|
from torch.nn import functional as F
|
|
from visualization import plot_loss, bar_chart_fixpoints
|
|
import seaborn as sns
|
|
from matplotlib import pyplot as plt
|
|
|
|
|
|
def prng():
|
|
return random.random()
|
|
|
|
|
|
def generate_perfekt_synthetic_fixpoint_weights():
|
|
return torch.tensor([[1.0], [0.0], [0.0], [0.0], [0.0], [0.0], [0.0], [0.0],
|
|
[1.0], [0.0], [0.0], [0.0],
|
|
[1.0], [0.0]
|
|
], dtype=torch.float32)
|
|
|
|
|
|
PALETTE = 10 * (
|
|
"#377eb8",
|
|
"#4daf4a",
|
|
"#984ea3",
|
|
"#e41a1c",
|
|
"#ff7f00",
|
|
"#a65628",
|
|
"#f781bf",
|
|
"#888888",
|
|
"#a6cee3",
|
|
"#b2df8a",
|
|
"#cab2d6",
|
|
"#fb9a99",
|
|
"#fdbf6f",
|
|
)
|
|
|
|
|
|
class RobustnessComparisonExperiment:
|
|
|
|
@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, synthetic, 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.loss_history = []
|
|
self.is_synthetic = synthetic
|
|
self.fixpoint_counters = {
|
|
"identity_func": 0,
|
|
"divergent": 0,
|
|
"fix_zero": 0,
|
|
"fix_weak": 0,
|
|
"fix_sec": 0,
|
|
"other_func": 0
|
|
}
|
|
|
|
self.directory = Path(directory)
|
|
self.directory.mkdir(parents=True, exist_ok=True)
|
|
|
|
self.id_functions = []
|
|
self.nets = self.populate_environment()
|
|
self.count_fixpoints()
|
|
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:
|
|
''' Either use perfect / hand-constructed fixpoint ... '''
|
|
net_name = f"net_{str(0)}_synthetic"
|
|
net = Net(self.net_input_size, self.net_hidden_size, self.net_out_size, net_name)
|
|
net.apply_weights(generate_perfekt_synthetic_fixpoint_weights())
|
|
nets.append(net)
|
|
|
|
else:
|
|
loop_population_size = tqdm(range(self.population_size))
|
|
for i in loop_population_size:
|
|
loop_population_size.set_description("Populating experiment %s" % i)
|
|
|
|
''' .. or use natural approach to train fixpoints from random initialisation. '''
|
|
net_name = f"net_{str(i)}"
|
|
net = Net(self.net_input_size, self.net_hidden_size, self.net_out_size, net_name)
|
|
for _ in range(self.epochs):
|
|
net.self_train(self.ST_steps, self.log_step_size, self.net_learning_rate)
|
|
nets.append(net)
|
|
return nets
|
|
|
|
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)
|
|
time_to_vergence = [[0 for _ in range(noise_levels)] for _ in
|
|
range(seeds if self.is_synthetic else len(self.id_functions))]
|
|
time_as_fixpoint = [[0 for _ in range(noise_levels)] for _ in
|
|
range(seeds if self.is_synthetic else len(self.id_functions))]
|
|
row_headers = []
|
|
|
|
# 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', 'Self Train Steps', 'absolute_loss',
|
|
'Time to convergence', '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)
|
|
for seed in range(seeds): # n / 1
|
|
setting = seed if self.is_synthetic else i
|
|
|
|
for noise_level in range(noise_levels):
|
|
steps = 0
|
|
clone = Net(fixpoint.input_size, fixpoint.hidden_size, fixpoint.out_size,
|
|
f"{fixpoint.name}_clone_noise_1e-{noise_level}")
|
|
clone.load_state_dict(copy.deepcopy(fixpoint.state_dict()))
|
|
clone = clone.apply_noise(pow(10, -noise_level))
|
|
|
|
while not is_zero_fixpoint(clone) and not is_divergent(clone):
|
|
# -> 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)
|
|
time_to_vergence[setting][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()
|
|
|
|
if is_identity_function(clone):
|
|
time_as_fixpoint[setting][noise_level] += 1
|
|
# When this raises a Type Error, we found a second order fixpoint!
|
|
steps += 1
|
|
|
|
df.loc[df.shape[0]] = [setting, f'$\mathregular{{10^{{-{noise_level}}}}}$',
|
|
steps, absolute_loss,
|
|
time_to_vergence[setting][noise_level],
|
|
time_as_fixpoint[setting][noise_level]]
|
|
pbar.update(1)
|
|
|
|
# Get the measuremts at the highest time_time_to_vergence
|
|
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', 'Self Train Steps'],
|
|
value_vars=['Time to convergence', 'Time as fixpoint'],
|
|
var_name="Measurement",
|
|
value_name="Steps").sort_values('Noise Level')
|
|
# Plotting
|
|
# plt.rcParams.update({
|
|
# "text.usetex": True,
|
|
# "font.family": "sans-serif",
|
|
# "font.size": 12,
|
|
# "font.weight": 'bold',
|
|
# "font.sans-serif": ["Helvetica"]})
|
|
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'
|
|
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)
|
|
|
|
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:
|
|
col_headers = [str(f"1e-{d}") for d in range(noise_levels)]
|
|
|
|
print(f"\nAppplications steps until divergence / zero: ")
|
|
# print(tabulate(time_to_vergence, showindex=row_headers, headers=col_headers, tablefmt='orgtbl'))
|
|
|
|
print(f"\nTime as fixpoint: ")
|
|
# print(tabulate(time_as_fixpoint, showindex=row_headers, headers=col_headers, tablefmt='orgtbl'))
|
|
return time_as_fixpoint, time_to_vergence
|
|
|
|
def count_fixpoints(self):
|
|
exp_details = f"ST steps: {self.ST_steps}"
|
|
self.id_functions = test_for_fixpoints(self.fixpoint_counters, self.nets)
|
|
bar_chart_fixpoints(self.fixpoint_counters, self.population_size, self.directory, self.net_learning_rate,
|
|
exp_details)
|
|
|
|
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)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
NET_INPUT_SIZE = 4
|
|
NET_OUT_SIZE = 1
|
|
|
|
ST_steps = 1000
|
|
ST_epochs = 5
|
|
ST_log_step_size = 10
|
|
ST_population_size = 10
|
|
ST_net_hidden_size = 2
|
|
ST_net_learning_rate = 0.004
|
|
ST_name_hash = random.getrandbits(32)
|
|
ST_synthetic = False
|
|
|
|
print(f"Running the robustness comparison experiment:")
|
|
exp = RobustnessComparisonExperiment(
|
|
population_size=ST_population_size,
|
|
log_step_size=ST_log_step_size,
|
|
net_input_size=NET_INPUT_SIZE,
|
|
net_hidden_size=ST_net_hidden_size,
|
|
net_out_size=NET_OUT_SIZE,
|
|
net_learning_rate=ST_net_learning_rate,
|
|
epochs=ST_epochs,
|
|
st_steps=ST_steps,
|
|
synthetic=ST_synthetic,
|
|
directory=Path('output') / 'journal_robustness' / f'{ST_name_hash}'
|
|
)
|
|
|
|
directory = Path('output') / 'journal_robustness' / f'{ST_name_hash}'
|
|
pickle.dump(exp, open(f"{directory}/experiment_pickle_{ST_name_hash}.p", "wb"))
|
|
print(f"\nSaved experiment to {directory}.") |