#!/usr/bin/env python3 import pandas as pd import sys from os import listdir from os.path import isfile, join import numpy as np from process_mission_helpers import findBatchStartAndChecks, computeCosts, getParam def main(): if len(sys.argv) != 2: print("Usage: python process_mission2.py path_to_results_files") sys.exit(1) data_dir = sys.argv[1] print('Processing files in ' + data_dir) files_to_process = [f for f in listdir(data_dir) if isfile(join(data_dir, f))] results_dict = {} for filename in files_to_process: print(filename) try: df = pd.read_csv(data_dir + '/' +filename) # process this df # step 1 sort by date in the df.sort_values(by='eval_' + filename[:-4] + '_job', inplace=True) # reset the indexes df.reset_index(drop=True, inplace=True) # get the back indexs and check that all vehicles had the same params batch_start_indx, all_vehicles_same_params = findBatchStartAndChecks(df) if (all_vehicles_same_params): # compute costs/rewards for each row in df costs = computeCosts(df) new_dict = {} new_dict['df'] = df new_dict['batch_start_indx'] = batch_start_indx new_dict['all_good'] = all_vehicles_same_params new_dict['costs'] = costs results_dict[filename] = new_dict else: print('Error: This file did not pass the checks: ' + 'eval_' + filename[:-4] + '_job') except FileNotFoundError: print("Error: File not found.") np.savez('results_dict.npz',results_dict=results_dict) # Now loop through each variable in the set of runs, determine slope and adjust # print the average costs for fun print(' -------------------------------------------') batches = len(results_dict['cent.csv']['batch_start_indx']) for i in range(batches): batch_start = results_dict['cent.csv']['batch_start_indx'][i] if (i == batches - 1): batch_end = -1 else: batch_end = results_dict['cent.csv']['batch_start_indx'][i+1] -1 costs_this_batch = results_dict['cent.csv']['costs'][batch_start:batch_end] average_costs = np.mean(costs_this_batch) variance_costs = np.var(costs_this_batch) print(' batch = ' + str(i)) print(' ave cost = ' + str(average_costs) + ' variance = ' + str(variance_costs)) #now update the parameters params_to_update = ['ALL_INPUT_GAIN','SEARCH_VALUE_IMPRV_GAIN', 'EXPLORE_BIAS'] new_vals_dict = {} # load the adam states from last time saved_adam_states = np.load('adam_states.npz', allow_pickle=True) adam_param_state = saved_adam_states['adam_param_state'].tolist() beta1 = saved_adam_states['beta1'] beta2 = saved_adam_states['beta2'] epsilon = saved_adam_states['epsilon'] for param in params_to_update: print(' *************************** ') print(' Updating param: ' + param) # Step 1 # what is the param value for the center? # find cost at the center of this group cent_last_batch_start_index = results_dict['cent.csv']['batch_start_indx'][-1] print(' Starting at index = ' + str(cent_last_batch_start_index)) cent_costs_this_batch = results_dict['cent.csv']['costs'][cent_last_batch_start_index:] cent_ave_costs = sum(cent_costs_this_batch)/len(cent_costs_this_batch) cent_var_costs = np.var(cent_costs_this_batch) param_val_at_cent = getParam(param, cent_last_batch_start_index, results_dict['cent.csv']['df']) print(' Where the param val at the cent was = ' + str(param_val_at_cent)) # Step 2 # what are the costs in the other runs where this param was adjusted? # first find the other runs # then load the values into: slopes = [] slope_vars = [] for file_name, results in results_dict.items(): # then did the last batch start in this run? this_set_last_batch_start_index = results['batch_start_indx'][-1] # what was the param value in this run? param_val_this_run = getParam(param, this_set_last_batch_start_index, results['df']) # is it different? delta = param_val_this_run - param_val_at_cent if (delta == 0.0): continue print(' Found a perturbation in file ' + file_name + ' at index = ' + str(this_set_last_batch_start_index)) # it is different, so compute the cost this_set_costs_this_batch = results['costs'][this_set_last_batch_start_index:] this_set_ave_costs = sum(this_set_costs_this_batch)/len(this_set_costs_this_batch) this_set_var_costs = np.var(this_set_costs_this_batch) if (this_set_var_costs == 0.0): this_set_var_costs = 1.0 delta_cost = this_set_ave_costs - cent_ave_costs slope = delta_cost / delta print('Found estimated slope to be = ' + str(slope)) print('Found associated variance to be = ' + str(this_set_var_costs)) print('Using ' + str(len(this_set_costs_this_batch)) + ' samples') slopes.append(slope) slope_vars.append(this_set_var_costs) if ((len(slopes) < 2) or (len(slope_vars) < 2)): print('Error: Did not find at least two other sets of params with perturbed vals for ' + param) #quit() # Step 3. Now compute the new values num = 0.0 denom = 0.0 for i in range( len(slopes)): num += slopes[i]/slope_vars[i] denom += 1.0 / slope_vars[i] if (denom == 0.0): denom = 1.0 best_estimate_of_slope = num / denom print('Best estimate of slope ' + str(best_estimate_of_slope)) #update the adam states print('Running ADAM with these params') print(adam_param_state[param]) new_m = beta1 * float(adam_param_state[param]['m']) + ( 1.0 - beta1 ) * best_estimate_of_slope new_v = beta2 * float(adam_param_state[param]['v']) + ( 1.0 - beta2 ) * best_estimate_of_slope**2 m_hat = new_m / ( 1.0 - beta1**( adam_param_state[param]['t'] + 1.0)) v_hat = new_v / ( 1.0 - beta2**( adam_param_state[param]['t'] + 1.0)) print('new_m =' + str(new_m)) print('new_v =' + str(new_v)) print('m_hat =' + str(m_hat)) print('v_hat =' + str(v_hat)) # learning rate up_gain = 1.0 print('total update = ' + str(up_gain * m_hat / (np.sqrt(v_hat) + epsilon))) # update param via ADAM new_val = param_val_at_cent - up_gain * m_hat / (np.sqrt(v_hat) + epsilon) # clip it max_change = 0.9 if (new_val > (1.0 + max_change) * param_val_at_cent): new_val = (1.0 + max_change) * param_val_at_cent elif (new_val < (1.0 - max_change) * param_val_at_cent): new_val = (1.0 - max_change) * param_val_at_cent print('Old val = ' + str( param_val_at_cent) + ', new val = ' + str(new_val)) # and save it! new_vals_dict[param] = new_val # and also update the adam states adam_param_state[param]['m'] = new_m adam_param_state[param]['v'] = new_v adam_param_state[param]['t'] = adam_param_state[param]['t'] + 1.0 print('Updating ADAM with these params') print(adam_param_state[param]) # Now that we are done updating, we need to write out the adam states # for next time np.savez('adam_states.npz',adam_param_state=adam_param_state, beta1=beta1, beta2=beta2, epsilon=epsilon) # Now that we have the new params, we just need to write them out # Open the file in write mode ('w') with open("params/params_cent.txt", "w") as f: # Write each variable and its value to the file for key, val in new_vals_dict.items(): val_str = "{:.3f}".format(val) f.write(key + '=' + val_str + '\n') quit() if __name__ == "__main__": main()