import copy import itertools from pathlib import Path import random import pickle import pandas as pd import numpy as np import torch from functionalities_test import is_identity_function, test_status 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 class SpawnLinspaceExperiment(SpawnExperiment): def spawn_and_continue(self, number_clones: int = None): number_clones = number_clones or self.nr_clones df = pd.DataFrame( columns=['clone', 'parent', 'parent2', 'MAE_pre', 'MAE_post', 'MSE_pre', 'MSE_post', 'MIM_pre', 'MIM_post', 'noise', 'status_pst']) # 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 = 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 # 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. net1.start_time = self.ST_steps - 150 net1_input_data = net1.input_weight_matrix().detach() net1_target_data = net1.create_target_weights(net1_input_data).detach() net2.start_time = self.ST_steps - 150 net2_input_data = net2.input_weight_matrix().detach() net2_target_data = net2.create_target_weights(net2_input_data).detach() 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.logspace(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, name=f"{net1.name}_{net2.name}_clone_{str(j)}", start_time=self.ST_steps + 100) clone.apply_weights(torch.as_tensor(in_between_weight)) clone.s_train_weights_history = copy.deepcopy(net1.s_train_weights_history) clone.number_trained = copy.deepcopy(net1.number_trained) # Pre Training distances (after noise application of course) clone_pre_weights = clone.create_target_weights(clone.input_weight_matrix()).detach() MAE_pre = MAE(net1_target_data, clone_pre_weights) MSE_pre = MSE(net1_target_data, clone_pre_weights) MIM_pre = mean_invariate_manhattan_distance(net1_target_data, clone_pre_weights) try: # 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) if any([torch.isnan(x).any() for x in clone.parameters()]): raise ValueError except ValueError: print("Ran into nan in 'in beetween weights' array.") df.loc[len(df)] = [j, net1.name, net2.name, MAE_pre, 0, MSE_pre, 0, MIM_pre, 0, self.noise, clone.is_fixpoint] continue # Post Training distances for comparison clone_post_weights = clone.create_target_weights(clone.input_weight_matrix()).detach() MAE_post = MAE(net1_target_data, clone_post_weights) MSE_post = MSE(net1_target_data, clone_post_weights) MIM_post = mean_invariate_manhattan_distance(net1_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} (between {net1.name} and {net2.name}) is fixpoint." f"\nMSE({net1.name},{j}): {MSE_post}" f"\nMAE({net1.name},{j}): {MAE_post}" f"\nMIM({net1.name},{j}): {MIM_post}\n") self.nets.append(clone) df.loc[len(df)] = [j, net1.name, net2.name, MAE_pre, MAE_post, MSE_pre, MSE_post, MIM_pre, MIM_post, self.noise, clone.is_fixpoint] for net1, net2 in pairwise_net_list: try: value = 'MAE' c_selector = [f'{value}_pre', f'{value}_post'] values = df.loc[(df['parent'] == net1.name) & (df['parent2'] == net2.name)][c_selector] this_min, this_max = values.values.min(), values.values.max() df.loc[(df['parent'] == net1.name) & (df['parent2'] == net2.name), c_selector] = (values - this_min) / (this_max - this_min) except ValueError: pass for parent in self.parents: 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 if __name__ == '__main__': NET_INPUT_SIZE = 4 NET_OUT_SIZE = 1 # Define number of runs & name: ST_runs = 1 ST_runs_name = "test-27" ST_steps = 2000 ST_epochs = 2 ST_log_step_size = 10 # Define number of networks & their architecture nr_clones = 25 ST_population_size = 10 ST_net_hidden_size = 2 ST_net_learning_rate = 0.04 ST_name_hash = random.getrandbits(32) print(f"Running the Spawn experiment:") exp = SpawnLinspaceExperiment( 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, nr_clones=nr_clones, noise=1e-8, directory=Path('output') / 'spawn_basin' / f'{ST_name_hash}' / f'linage' ) df = exp.df directory = Path('output') / 'spawn_basin' / f'{ST_name_hash}' / 'linage' with (directory / f"experiment_pickle_{ST_name_hash}.p").open('wb') as f: pickle.dump(exp, f) 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") # 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") # Pointplot with pre and after parent Distances import seaborn as sns from matplotlib import pyplot as plt, ticker # ptplt = sns.pointplot(data=exp.df, x='MAE_pre', y='MAE_post', join=False) ptplt = sns.scatterplot(x=exp.df['MAE_pre'], y=exp.df['MAE_post']) # ptplt.set(xscale='log', yscale='log') x0, x1 = ptplt.axes.get_xlim() y0, y1 = ptplt.axes.get_ylim() lims = [max(x0, y0), min(x1, y1)] # This is the x=y line using transforms ptplt.plot(lims, lims, 'w', linestyle='dashdot', transform=ptplt.axes.transData) ptplt.plot([0, 1], [0, 1], ':k', transform=ptplt.axes.transAxes) ptplt.set(xlabel='Mean Absolute Distance before Self-Training', ylabel='Mean Absolute Distance after Self-Training') # ptplt.axes.xaxis.set_major_formatter(ticker.FuncFormatter(lambda x, pos: round(float(x), 2))) # ptplt.xticks(rotation=45) #for ind, label in enumerate(ptplt.get_xticklabels()): # if ind % 10 == 0: # every 10th label is kept # label.set_visible(True) # else: # label.set_visible(False) filepath = exp.directory / 'mim_dist_plot.pdf' plt.tight_layout() plt.savefig(filepath, dpi=600, format='pdf', bbox_inches='tight')