backend: tensorflow cache: caches/cms_gen dataset: schema: cms target_particles: gen num_input_features: 42 # NONE = 0, # TRACK = 1, # PS1 = 2, # PS2 = 3, # ECAL = 4, # HCAL = 5, # GSF = 6, # BREM = 7, # HFEM = 8, # HFHAD = 9, # SC = 10, # HO = 11, num_input_classes: 12 #(none=0, ch.had=1, n.had=2, hfem=3, hfhad=4, gamma=5, e=6, mu=7) num_output_classes: 8 padded_num_elem_size: 6400 cls_weight_by_pt: no reg_weight_by_pt: no enable_tfds_caching: no loss: classification_loss_coef: 100.0 charge_loss_coef: 1.0 pt_loss_coef: 1.0 eta_loss_coef: 1.0 sin_phi_loss_coef: 1.0 cos_phi_loss_coef: 1.0 energy_loss_coef: 1.0 cls_loss: type: SigmoidFocalCrossEntropy from_logits: yes gamma: 2.0 charge_loss: type: CategoricalCrossentropy from_logits: yes energy_loss: type: Huber pt_loss: type: Huber sin_phi_loss: type: Huber delta: 0.1 cos_phi_loss: type: Huber delta: 0.1 eta_loss: type: Huber delta: 0.1 event_loss: none #none, sliced_wasserstein, gen_jet_logcosh, gen_jet_mse, hist_2d event_loss_coef: 1.0 met_loss: none met_loss_coef: 1.0 tensorflow: eager: no setup: train: yes weights: weights_config: lr: 0.00005 num_epochs: 55 dtype: float32 trainable: lr_schedule: none # cosinedecay, exponentialdecay, onecycle, none optimizer: adam # adam, adamw, sgd horovod_enabled: no cls_output_as_logits: yes #if enabled, do not create LSH bins for small graphs (less than one bin size) #enabling results in some speedup for gun samples, but must be disabled for XLA small_graph_opt: yes use_normalizer: no batching: # if enabled, use dynamic batching instead of the fixed-size batches configured in batch_per_gpu bucket_by_sequence_length: yes bucket_batch_sizes: auto batch_multiplier: 1 optimizer: adam: amsgrad: no adamw: amsgrad: yes weight_decay: 0.001 sgd: nesterov: no momentum: 0.9 # LR Schedules exponentialdecay: decay_steps: 2000 decay_rate: 0.99 staircase: yes onecycle: mom_min: 0.85 mom_max: 0.95 warmup_ratio: 0.3 div_factor: 25.0 final_div: 100000.0 parameters: model: gnn_dense input_encoding: cms node_update_mode: additive do_node_encoding: yes node_encoding_hidden_dim: 512 combined_graph_layer: bin_size: 640 max_num_bins: 200 distance_dim: 128 layernorm: yes dropout: 0.0 dist_activation: elu ffn_dist_num_layers: 2 ffn_dist_hidden_dim: 128 # GCN kernel: type: NodePairGaussianKernel dist_mult: 0.1 clip_value_low: 0.0 dist_norm: l2 num_node_messages: 2 node_message: type: GHConvDense output_dim: 512 activation: elu #if this is enabled, it will break float16 training normalize_degrees: no activation: elu num_graph_layers_id: 3 num_graph_layers_reg: 3 output_decoding: activation: elu regression_use_classification: yes dropout: 0.1 pt_as_correction: yes id_dim_decrease: yes charge_dim_decrease: yes pt_dim_decrease: yes eta_dim_decrease: yes phi_dim_decrease: yes energy_dim_decrease: yes id_hidden_dim: 512 charge_hidden_dim: 256 pt_hidden_dim: 512 eta_hidden_dim: 256 phi_hidden_dim: 256 energy_hidden_dim: 512 id_num_layers: 3 charge_num_layers: 2 pt_num_layers: 2 eta_num_layers: 2 phi_num_layers: 2 energy_num_layers: 2 layernorm: yes mask_reg_cls0: yes skip_connection: no debug: no timing: num_ev: 100 num_iter: 3 callbacks: checkpoint: monitor: "val_loss" plot_freq: 1 tensorboard: dump_history: yes hist_freq: 1 hypertune: algorithm: hyperband # random, bayesian, hyperband random: objective: val_loss max_trials: 100 bayesian: objective: val_loss max_trials: 100 num_initial_points: 2 hyperband: objective: val_loss max_epochs: 10 factor: 3 iterations: 1 executions_per_trial: 1 raytune: local_dir: # Note: please specify an absolute path sched: asha # asha, hyperband search_alg: # bayes, bohb, hyperopt, nevergrad, scikit default_metric: "val_loss" default_mode: "min" # Tune schedule specific parameters asha: max_t: 200 reduction_factor: 4 brackets: 1 grace_period: 10 hyperband: max_t: 200 reduction_factor: 4 hyperopt: n_random_steps: 10 nevergrad: n_random_steps: 10 train_test_datasets: multiparticlegun: batch_per_gpu: 1 event_pad_size: -1 datasets: - cms_pf_multi_particle_gun physical: batch_per_gpu: 1 event_pad_size: -1 datasets: - cms_pf_ttbar - cms_pf_ztt - cms_pf_qcd - cms_pf_qcd_high_pt - cms_pf_sms_t1tttt gun: batch_per_gpu: 50 event_pad_size: -1 datasets: - cms_pf_single_electron - cms_pf_single_gamma - cms_pf_single_neutron - cms_pf_single_pi0 - cms_pf_single_pi - cms_pf_single_tau - cms_pf_single_mu - cms_pf_single_proton evaluation_datasets: cms_pf_qcd_high_pt: batch_size: 5 num_events: -1 cms_pf_single_neutron: batch_size: 100 num_events: -1 validation_dataset: cms_pf_qcd_high_pt validation_batch_size: 5 validation_num_events: 500 evaluation_jet_algo: antikt_algorithm datasets: cms_pf_ttbar: version: 1.6.0 data_dir: manual_dir: cms_pf_ztt: version: 1.6.0 data_dir: manual_dir: cms_pf_qcd: version: 1.6.0 data_dir: manual_dir: cms_pf_qcd_high_pt: version: 1.6.0 data_dir: manual_dir: cms_pf_single_electron: version: 1.6.0 data_dir: manual_dir: cms_pf_single_gamma: version: 1.6.0 data_dir: manual_dir: cms_pf_single_pi0: version: 1.6.0 data_dir: manual_dir: cms_pf_single_neutron: version: 1.6.0 data_dir: manual_dir: cms_pf_single_pi: version: 1.6.0 data_dir: manual_dir: cms_pf_single_tau: version: 1.6.0 data_dir: manual_dir: cms_pf_single_mu: version: 1.6.0 data_dir: manual_dir: cms_pf_single_proton: version: 1.6.0 data_dir: manual_dir: cms_pf_multi_particle_gun: version: 1.6.0 data_dir: manual_dir: cms_pf_sms_t1tttt: version: 1.6.0 data_dir: manual_dir: