def main(): if len(sys.argv) != 5: print "Error: exactly 4 arguments are required" MC_dir = sys.argv[1] setting_dir = sys.argv[2] training_path = sys.argv[3] data_outpath = sys.argv[4] # files to which this discriminant should be augmented #data_files = ["ggH125", "VBFH125", "ZH125", "WplusH125", "WminusH125"] data_files = [ "ggH125", "VBFH125", "ZH125", "WplusH125", "WminusH125", "ttH125" ] #data_files = ["ggH125", "VBFH125", "ZH125", "WplusH125", "WminusH125", "ttH125", "ZZTo4l", "ggTo2e2mu_Contin_MCFM701", "ggTo2mu2tau_Contin_MCFM701", "ggTo4mu_Contin_MCFM701", "ggTo2e2tau_Contin_MCFM701", "ggTo4e_Contin_MCFM701", "ggTo4tau_Contin_MCFM701"] confhandler = ModelCollectionConfigFileHandler() confhandler.load_configuration(setting_dir + "settings.conf") mcolls = confhandler.GetModelCollection(weightpath=training_path) for data_file in data_files: augment_file(MC_dir, data_outpath, data_file, mcolls)
def main(): if len(sys.argv) != 5: print "Error: exactly 4 arguments are required" in_folder = sys.argv[1] out_folder = sys.argv[2] tree_name = sys.argv[3] run_dir = sys.argv[4] confhandler = ModelCollectionConfigFileHandler() confhandler.load_configuration(os.path.join(run_dir, "settings.conf")) mcolls = confhandler.GetModelCollection(weightpath = os.path.join(run_dir, "training/")) augment_file(in_folder, out_folder, tree_name, mcolls)
def main(): if len(sys.argv) != 3: print "Error: exactly 2 arguments are required" #/data_CMS/cms/wind/CJLST_NTuples/ #MC_dir = sys.argv[1] setting_dir = sys.argv[1] training_dir = sys.argv[2] confhandler = ModelCollectionConfigFileHandler() confhandler.load_configuration(setting_dir + "settings.conf") mcolls = confhandler.GetModelCollection() train = Trainer(training_dir) opt = optimizers.SGD(lr=0.01, momentum=0.9, decay=1e-6) #opt = optimizers.Adam(lr = 0.001, beta_1 = 0.9, beta_2 = 0.999, epsilon = K.epsilon(), decay = 0.0) for mcoll in mcolls: train.train(mcoll, optimizer=opt, MC_weighting=False)
def distribute_training_settings(run_path): # load the configuration that is sitting there confhandler = ModelCollectionConfigFileHandler() confhandler.load_configuration(run_path + "settings.conf") # these are all the model collections that need to be trained mcolls = confhandler.GetModelCollection() # create the folder holding the settings for the individual models and their training settings_dir = run_path + "settings_training/" if not os.path.exists(settings_dir): os.makedirs(settings_dir) # iterate over these models and make a separate config file for each of them for mcoll in mcolls: training_settings_dir = settings_dir + mcoll.name + "/" if not os.path.exists(training_settings_dir): os.makedirs(training_settings_dir) outconf = ModelCollectionConfigFileHandler() outconf.ToConfiguration([mcoll]) outconf.save_configuration(training_settings_dir + "settings.conf")
def main(): def _compute_class_weights_lengths(gen, preprocessor, MC_weighting=False): # determine the actual size of the available dataset and adjust the sample weights correspondingly H1_data = gen.H1_collection.get_data(Config.branches, 0.0, 1.0) H0_data = gen.H0_collection.get_data(Config.branches, 0.0, 1.0) H1_length = len(preprocessor.process(H1_data).values()[0]) H1_indices = preprocessor.get_last_indices() H0_length = len(preprocessor.process(H0_data).values()[0]) H0_indices = preprocessor.get_last_indices() print "H1_length = " + str(H1_length) print "H0_length = " + str(H0_length) # if per-sample weighting is enabled, also set up the normalization of the event weights if MC_weighting: H1_weight_sum = np.sum( np.maximum(np.array(H1_data["training_weight"][H1_indices]), 0.0)) H0_weight_sum = np.sum( np.maximum(np.array(H0_data["training_weight"][H0_indices]), 0.0)) H1_class_weight = float(H0_length) / H1_weight_sum H0_class_weight = float(H1_length) / H0_weight_sum else: # H1_class_weight = 1.0 # H0_class_weight = float(H1_length) / float(H0_length) H1_class_weight = 1.0 + float(H0_length) / float(H1_length) H0_class_weight = 1.0 + float(H1_length) / float(H0_length) return H1_class_weight, H0_class_weight, H1_length, H0_length # this computes low-level performance metrics for a model collection, i.e. the mean-quare error # computed on the validation dataset for each discriminant. Since the validation datasets will be held constant, # this is an easy way to directly compare different models setting_dir = sys.argv[1] training_dir = sys.argv[2] out_dir = sys.argv[3] # first, need to read in the trained ModelCollection: mconfhandler = ModelCollectionConfigFileHandler() mconfhandler.load_configuration(setting_dir + "settings.conf") mcolls = mconfhandler.GetModelCollection(weightpath=training_dir) confhandler = ConfigFileHandler() out_path = out_dir + "model_benchmark.txt" # for the evaluation, need to proceed in the same way as for training, but evaluate the models on the validation # data instead of training them on the training data for mcoll in mcolls: models, preprocessors, settings = mcoll.get_models( ), mcoll.get_preprocessors(), mcoll.get_settings() for cur_model, cur_preprocessor, cur_settings in zip( models, preprocessors, settings): val_gen = Generator(mcoll.H1_stream, mcoll.H0_stream, Config.branches, preprocessor=cur_preprocessor, chunks=1, MC_weighting=False) val_gen.setup_validation_data() val_H1_classweight, val_H0_classweight, H1_length, H0_length = _compute_class_weights_lengths( val_gen, cur_preprocessor, False) print val_H1_classweight print val_H0_classweight print H1_length print H0_length val_gen.set_H1_weight(val_H1_classweight) val_gen.set_H0_weight(val_H0_classweight) val_gen.set_minimum_length(0) cur_model.get_keras_model().compile(optimizer=optimizers.Adam(), loss="mean_squared_error", metrics=["binary_accuracy"]) res = cur_model.get_keras_model().evaluate_generator( val_gen.preprocessed_generator(), steps=1) print "statistics for model " + cur_model.name print res print cur_model.get_keras_model().metrics_names confhandler.new_section(cur_model.name) confhandler.set_field(cur_model.name, 'H0_val_length', str(H0_length)) confhandler.set_field(cur_model.name, 'H1_val_length', str(H1_length)) confhandler.set_field(cur_model.name, 'val_loss', str(res[0])) confhandler.save_configuration(out_path)
def main(): # runs to check for (good) models (the first one passed is taken as reference run from which the available models # are taken - it is expected that all others runs also follow this structure): input_runs = [] print "===================================================================" print "looking for models in the following runs:" for campaign_dir in sys.argv[1:-2]: for run_dir in next(os.walk(campaign_dir))[1]: if not "bin" in run_dir: run_path = os.path.join(campaign_dir, run_dir) print run_path input_runs.append(run_path) print "===================================================================" # output training campaign, this will consist of a combination of the models found in the campaigns listed above, in such a way that the overall performance is optimized output_run = os.path.join(sys.argv[-1], "optimized") # where the configuration file for the hyperparameter settings should be stored hyperparam_output = os.path.join(output_run, "../hyperparameters.conf") os.makedirs(output_run) # load the available model names reference_run = input_runs[0] available_mcolls = os.walk(os.path.join(reference_run, "training")).next()[1] mcolls_winning = [] for mcoll in available_mcolls: models = os.walk(os.path.join(reference_run, "training", mcoll)).next()[1] # load a representative version of the current model collection... mconfhandler = ModelCollectionConfigFileHandler() mconfhandler.load_configuration( os.path.join(reference_run, "settings_training", mcoll, "settings.conf")) mcoll_template = mconfhandler.GetModelCollection()[0] # ... but strip away all the actual model components mcoll_template.model_dict = {} mcoll_template.preprocessor_dict = {} mcoll_template.settings_dict = {} for model in models: # compare this model across the different runs losses = [get_loss(run, mcoll, model) for run in input_runs] winner = np.argmin(losses) winning_run = input_runs[winner] # copy the winning model into the output run shutil.copytree( os.path.join(winning_run, "training", mcoll, model), os.path.join(output_run, "training", mcoll, model)) print "--------------------------------------------" print " take " + model + " from " + winning_run print "--------------------------------------------" # load the winning model to keep track of its settings mconfhandler = ModelCollectionConfigFileHandler() mconfhandler.load_configuration( os.path.join(winning_run, "settings_training", mcoll, "settings.conf")) mcoll_winning = mconfhandler.GetModelCollection()[0] # then pull the winning model over into the template winning_model = mcoll_winning.model_dict[model] winning_preprocessor = mcoll_winning.preprocessor_dict[model] winning_settings = mcoll_winning.settings_dict[model] mcoll_template.add_model(winning_preprocessor, winning_model, winning_settings) mcolls_winning.append(mcoll_template) # now save the put-together config file also into the output run mconfhandler = ModelCollectionConfigFileHandler() mconfhandler.ToConfiguration(mcolls_winning) mconfhandler.save_configuration(os.path.join(output_run, "settings.conf")) # now distriute again the training settings, as usual: distribute_training_settings(output_run + '/') # now create the hyperparameter config file for each model, taken from the winners hp_confhandler = ConfigFileHandler() for mcoll in mcolls_winning: for model_name, model in mcoll.model_dict.iteritems(): hp_confhandler.new_section(model_name) hp_confhandler.set_field( model_name, "hyperparameters", ConfigFileUtils.serialize_dict(model.hyperparameters, lambda x: str(x))) hp_confhandler.save_configuration(hyperparam_output) print "===================================================================" print "hyperparameter configuration file written to " + hyperparam_output print "==================================================================="