class LWPRFA(FA): parametric = False def __init__(self, indim, outdim): FA.__init__(self, indim, outdim) self.filename = None def reset(self): FA.reset(self) # initialize the LWPR function self.lwpr = LWPR(self.indim, self.outdim) self.lwpr.init_D = 10.*np.eye(self.indim) self.lwpr.init_alpha = 0.1*np.ones([self.indim, self.indim]) self.lwpr.meta = True def predict(self, inp): """ predict the output for the given input. """ # the next 3 lines fix a bug when lwpr models are pickled and unpickled again # without it, a TypeError is thrown "Expected a double precision numpy array." # even though the numpy array is double precision. inp = self._asFlatArray(inp) inp_tmp = np.zeros(inp.shape) inp_tmp[:] = inp return self.lwpr.predict(inp_tmp) def train(self): for i, t in self.dataset: i = self._asFlatArray(i) t = self._asFlatArray(t) self.lwpr.update(i, t) def _cleanup(self): if self.filename and os.path.exists(self.filename): os.remove(self.filename) def __getstate__(self): """ required for pickle. removes the lwpr model from the dictionary and saves it to file explicitly. """ # create unique hash key for filename and write lwpr to file hashkey = hashlib.sha1(str(self.lwpr) + time.ctime() + str(np.random.random())).hexdigest()[:8] if not os.path.exists('.lwprmodels'): os.makedirs('.lwprmodels') # remove any old files if existing if self.filename: os.remove(self.filename) self.filename = '.lwprmodels/lwpr_%s.binary'%hashkey self.lwpr.write_binary(self.filename) # remove lwpr from dictionary and return state state = self.__dict__.copy() del state['lwpr'] return state def __setstate__(self, state): """ required for pickle. loads the stored lwpr model explicitly. """ self.__dict__.update(state) self.lwpr = LWPR(self.filename)
def train_lwpr(datafile, resultfolder, max_num_train, patience_list, improvement_threshold, init_lwpr_setting, hist_window, start_epoch=0, cmd_scaler=1.0, modelfile='lwpr_model'): curr_path = os.getcwd() if resultfolder in os.listdir(curr_path): print "subfolder exists" else: print "Not Exist, so make subfolder" os.mkdir(resultfolder) # Load Data dataset = loadmat(datafile) train_data_x, train_data_y = dataset['train_data_x'], dataset['train_data_y'] valid_data_x, valid_data_y = dataset['valid_data_x'], dataset['valid_data_y'] num_data, num_valid = train_data_x.shape[0], valid_data_x.shape[0] speed_hw, cmd_hw = hist_window[0], hist_window[1] input_dim = 2*(speed_hw+cmd_hw) # normalize command part train_data_x[:, 2*speed_hw:] = train_data_x[:, 2*speed_hw:] * cmd_scaler valid_data_x[:, 2*speed_hw:] = valid_data_x[:, 2*speed_hw:] * cmd_scaler # Set-up Parameters/Model for Training Procedure max_num_trials = max_num_train improvement_threshold = improvement_threshold error_hist, best_model_error, prev_train_time = [], np.inf, 0 initD, initA, penalty = init_lwpr_setting[0], init_lwpr_setting[1], init_lwpr_setting[2] w_gen, w_prune = init_lwpr_setting[3], init_lwpr_setting[4] best_model_epoch = 0 if start_epoch < 1: # Initialize Two 1-Dimensional Models LWPR_model_left = LWPR(input_dim, 1) #LWPR_model_left.init_D = initD * np.eye(input_dim) tmp_arr = np.ones(input_dim) tmp_arr[input_dim-2*cmd_hw:input_dim] = init_lwpr_setting[5] LWPR_model_left.init_D = initD * np.diag(tmp_arr) LWPR_model_left.update_D = False # True #LWPR_model_left.init_alpha = initA * np.eye(input_dim) tmp_arr = np.ones(input_dim) tmp_arr[input_dim-2*cmd_hw:input_dim] = init_lwpr_setting[5] LWPR_model_left.init_alpha = initA * np.diag(tmp_arr) LWPR_model_left.penalty = penalty LWPR_model_left.meta = True LWPR_model_left.meta_rate = 20 LWPR_model_left.w_gen = w_gen LWPR_model_left.w_prune = w_prune LWPR_model_right = LWPR(input_dim, 1) #LWPR_model_right.init_D = initD * np.eye(input_dim) tmp_arr = np.ones(input_dim) tmp_arr[input_dim-2*cmd_hw:input_dim] = init_lwpr_setting[5] LWPR_model_right.init_D = initD * np.diag(tmp_arr) LWPR_model_right.update_D = False # True #LWPR_model_right.init_alpha = initA * np.eye(input_dim) tmp_arr = np.ones(input_dim) tmp_arr[input_dim-2*cmd_hw:input_dim] = init_lwpr_setting[5] LWPR_model_right.init_alpha = initA * np.diag(tmp_arr) LWPR_model_right.penalty = penalty LWPR_model_right.meta = True LWPR_model_right.meta_rate = 20 LWPR_model_right.w_gen = w_gen LWPR_model_right.w_prune = w_prune patience = patience_list[0] else: modelfile_name = './' + resultfolder + '/' + modelfile + '_left_epoch' + str(start_epoch-1) + '.bin' LWPR_model_left = LWPR(modelfile_name) print '\tRead LWPR model for left wheel(%d)' % (LWPR_model_left.num_rfs[0]) modelfile_name = './' + resultfolder + '/' + modelfile + '_right_epoch' + str(start_epoch-1) + '.bin' LWPR_model_right = LWPR(modelfile_name) print '\tRead LWPR model for right wheel(%d)' % (LWPR_model_right.num_rfs[0]) result_file_name = './' + resultfolder + '/Result_of_training_epoch' + str(start_epoch-1) + '.mat' result_file = loadmat(result_file_name) prev_train_time = result_file['train_time'] patience = result_file['patience'] best_model_error = result_file['best_model_error'] for cnt in range(start_epoch): error_hist.append([result_file['history_validation_error'][cnt][0], result_file['history_validation_error'][cnt][1], result_file['history_validation_error'][cnt][2]]) # Training Part model_prediction = np.zeros(valid_data_y.shape) tmp_x, tmp_y = np.zeros((input_dim, 1)), np.zeros((1,1)) print 'start training' start_train_time = timeit.default_timer() for train_cnt in range(start_epoch, max_num_trials): if patience < train_cnt: break rand_ind = np.random.permutation(num_data) for data_cnt in range(num_data): tmp_x[:,0] = train_data_x[rand_ind[data_cnt], 0:input_dim] tmp_y[0,0] = train_data_y[rand_ind[data_cnt], 0] _ = LWPR_model_left.update(tmp_x, tmp_y) tmp_y[0,0] = train_data_y[rand_ind[data_cnt], 1] _ = LWPR_model_right.update(tmp_x, tmp_y) if data_cnt % 5000 == 0: print '\ttrain epoch %d, data index %d, #rfs=%d/%d' % (train_cnt, data_cnt, LWPR_model_left.num_rfs, LWPR_model_right.num_rfs) for data_cnt in range(num_valid): tmp_x[:,0] = valid_data_x[data_cnt, 0:input_dim] model_prediction[data_cnt, 0], _ = LWPR_model_left.predict_conf(tmp_x) model_prediction[data_cnt, 1], _ = LWPR_model_right.predict_conf(tmp_x) diff = abs(valid_data_y - model_prediction) new_error = np.asarray([np.sum(diff)/float(num_valid), np.sqrt(np.sum(diff**2)/float(num_valid)), np.max(diff)]) error_hist.append([new_error[0], new_error[1], new_error[2]]) # save result of one training epoch modelfile_name = './' + resultfolder + '/' + modelfile + '_left_epoch' + str(train_cnt) + '.bin' LWPR_model_left.write_binary(modelfile_name) modelfile_name = './' + resultfolder + '/' + modelfile + '_right_epoch' + str(train_cnt) + '.bin' LWPR_model_right.write_binary(modelfile_name) if new_error[1] < best_model_error * improvement_threshold: best_model_epoch = train_cnt best_model_error = new_error[1] patience = max(patience, min(train_cnt+10, int(train_cnt * patience_list[1])) ) modelfile_name = './' + resultfolder + '/' + modelfile + '_best_left_epoch' + str(train_cnt) + '.bin' LWPR_model_left.write_binary(modelfile_name) modelfile_name = './' + resultfolder + '/' + modelfile + '_best_right_epoch' + str(train_cnt) + '.bin' LWPR_model_right.write_binary(modelfile_name) result_file_name = './' + resultfolder + '/Result_of_training_epoch' + str(train_cnt) + '.mat' result = {} result['train_time'] = timeit.default_timer() - start_train_time + prev_train_time result['best_model_error'] = best_model_error result['history_validation_error'] = error_hist result['patience'] = patience result['improvement_threshold'] = improvement_threshold result['init_D'] = initD result['init_alpha'] = initA result['penalty'] = penalty result['w_generate_criterion'] = w_gen result['w_prune_criterion'] = w_prune result['number_speed_in_input'] = 2*speed_hw result['number_cmd_in_input'] = 2*cmd_hw savemat(result_file_name, result) print '\n\tSave Intermediate Result Successfully' print '\t%d-th learning : #Data=%d/%d, #rfs=%d/%d, error=%f\n' %(train_cnt, LWPR_model_left.n_data, LWPR_model_right.n_data, LWPR_model_left.num_rfs, LWPR_model_right.num_rfs, error_hist[train_cnt][1]) print 'end training' return best_model_epoch