diff --git a/README.md b/README.md index 7be6661..5b2f39d 100644 --- a/README.md +++ b/README.md @@ -1,6 +1,11 @@ # self-rep NN paper - ALIFE journal edition -- [x] Plateau / Pillar sizeWhat does happen to the fixpoints after noise introduction and retraining?Options beeing: Same Fixpoint, Similar Fixpoint (Basin), Different Fixpoint? Do they do the clustering thingy? +- [x] Plateau / Pillar sizeWhat does happen to the fixpoints after noise introduction and retraining?Options beeing: Same Fixpoint, Similar Fixpoint (Basin), + - Different Fixpoint? + Yes, we did not found same (10-5) + - Do they do the clustering thingy? + Kind of: Small movement towards (MIM-Distance getting smaller) parent fixpoint. + Small movement for everyone? -> Distribution - see `journal_basins.py` for the "train -> spawn with noise -> train again and see where they end up" functionality. Apply noise follows the `vary` function that was used in the paper robustness test with `+- prng() * eps`. Change if desired. @@ -9,6 +14,9 @@ - [ ] Same Thing with Soup interactionWe would expect the same behaviour...Influence of interaction with near and far away particles. +- [ ] How are basins / "attractor areas" shaped? + - Weired.... tbc... + - [x] Robustness test with a trained NetworkTraining for high quality fixpoints, compare with the "perfect" fixpoint. Average Loss per application step - see `journal_robustness.py` for robustness test modeled after cristians robustness-exp (with the exeption that we put noise on the weights). Has `synthetic` bool to switch to hand-modeled perfect fixpoint instead of naturally trained ones. @@ -19,7 +27,7 @@ - [ ] Adjust Self Training so that it favors second order fixpoints-> Second order test implementation (?) -- [ ] Barplot over clones -> how many become a fixpoint cs how many diverge per noise level +- [x] Barplot over clones -> how many become a fixpoint cs how many diverge per noise level - [ ] Box-Plot of Avg. Distance of clones from parent diff --git a/experiments/self_train_exp.py b/experiments/self_train_exp.py index f422eda..1a467b9 100644 --- a/experiments/self_train_exp.py +++ b/experiments/self_train_exp.py @@ -55,8 +55,6 @@ class SelfTrainExperiment: net = Net(self.net_input_size, self.net_hidden_size, self.net_out_size, net_name) for _ in range(self.epochs): - input_data = net.input_weight_matrix() - target_data = net.create_target_weights(input_data) 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:]}") @@ -113,5 +111,6 @@ def run_ST_experiment(population_size, batch_size, net_input_size, net_hidden_si summary_fixpoint_experiment(runs, population_size, epochs, experiments, net_learning_rate, summary_directory_name, summary_pre_title) + if __name__ == '__main__': raise NotImplementedError('Test this here!!!') diff --git a/journal_robustness.py b/journal_robustness.py index b9f44e2..4330614 100644 --- a/journal_robustness.py +++ b/journal_robustness.py @@ -4,7 +4,6 @@ import pandas as pd import torch import random import copy -import numpy as np from pathlib import Path from tqdm import tqdm @@ -21,6 +20,7 @@ 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], @@ -28,15 +28,32 @@ def generate_perfekt_synthetic_fixpoint_weights(): ], 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 """ + # 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 + # 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: @@ -55,7 +72,7 @@ class RobustnessComparisonExperiment: self.epochs = epochs self.ST_steps = st_steps self.loss_history = [] - self.synthetic = synthetic + self.is_synthetic = synthetic self.fixpoint_counters = { "identity_func": 0, "divergent": 0, @@ -71,14 +88,14 @@ class RobustnessComparisonExperiment: self.id_functions = [] self.nets = self.populate_environment() self.count_fixpoints() - self.time_to_vergence, self.time_as_fixpoint = self.test_robustness() + self.time_to_vergence, self.time_as_fixpoint = self.test_robustness( + seeds=population_size if self.is_synthetic else 1) self.save() def populate_environment(self): - loop_population_size = tqdm(range(self.population_size)) nets = [] - if self.synthetic: + 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) @@ -86,6 +103,7 @@ class RobustnessComparisonExperiment: 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) @@ -95,62 +113,75 @@ class RobustnessComparisonExperiment: 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) - 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))] + 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 = [] - 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) + df = pd.DataFrame(columns=['setting', 'noise_level', 'steps', 'absolute_loss', '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) + for seed in range(seeds): # n / 1 + setting = seed if self.is_synthetic else i - while not is_zero_fixpoint(clone) and not is_divergent(clone): - if is_identity_function(clone): - avg_time_as_fixpoint[i][noise_level] += 1 + for noise_level in range(noise_levels): + steps = 0 + 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) - # -> before - clone_weight_pre_application = clone.input_weight_matrix() - target_data_pre_application = clone.create_target_weights(clone_weight_pre_application) + 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) - 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) + 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() + 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 + 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, 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('steps', ascending=False).drop_duplicates(['setting', 'noise_level']) + df_melted = df_sorted.reset_index().melt(id_vars=['setting', 'noise_level', 'steps'], + value_vars=['time_to_vergence', 'time_as_fixpoint'], + var_name="Measurement", + value_name="Steps") + # Plotting + sns.set(style='whitegrid') + bf = sns.boxplot(data=df_melted, y='Steps', x='noise_level', hue='Measurement', palette=PALETTE) + bf.set_title('Robustness as self application steps per noise level') + plt.tight_layout() - # 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) + # 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 @@ -160,13 +191,11 @@ class RobustnessComparisonExperiment: 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(tabulate(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 - + # 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}" @@ -174,14 +203,12 @@ class RobustnessComparisonExperiment: 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) - def save(self): pickle.dump(self, open(f"{self.directory}/experiment_pickle.p", "wb")) print(f"\nSaved experiment to {self.directory}.") @@ -194,7 +221,7 @@ if __name__ == "__main__": ST_steps = 1000 ST_epochs = 5 ST_log_step_size = 10 - ST_population_size = 5 + ST_population_size = 100 ST_net_hidden_size = 2 ST_net_learning_rate = 0.04 ST_name_hash = random.getrandbits(32) @@ -211,5 +238,5 @@ if __name__ == "__main__": epochs=ST_epochs, st_steps=ST_steps, synthetic=ST_synthetic, - directory=Path('output') / 'robustness' / f'{ST_name_hash}' + directory=Path('output') / 'journal_robustness' / f'{ST_name_hash}' )