point_encoder: type: pointnet feature_dim: 64 pointnet: 0: npoint: 256 radius: 0.01 nsample: 32 in_channel: 0 mlp: - 64 - 64 - 128 group_all: false 1: npoint: 64 radius: 0.02 nsample: 64 in_channel: 128 mlp: - 128 - 128 - 256 group_all: false 2: npoint: 16 radius: 0.04 nsample: 128 in_channel: 256 mlp: - 256 - 256 - 512 group_all: true experiment_name: concat_tools_mixed_focal_0.9_0.75_force_mag_dir_ada_weight_hist5_smoothpc_pcd256+512 use_env_pointcloud: true split: test model: type: tactile_pointnet_concat use_pose_encoding: false reconstruct_tactile: false tactile_channels: 3 contact_prob_init_bias_enabled: true contact_prob_init_bias: 0.02 joint_processing: enabled: true label_channel: true data: max_memory_samples: 10000000 use_partial_pointcloud: false include_env_pointcloud: ${use_env_pointcloud} knn_k: 5 knn_range: 0.01 dist_lambda: 100.0 contact_depth_threshold: -0.005 force_propagation: weight_method: inv_square dist_lambda: 50.0 force_clip_percentile: 98.0 normalize_per_point: true smooth_sigma: 0.0 contact_L: -0.005 contact_alpha: 0.5 contact_k: 1.7 tactile_shape: - 7 - 9 - 3 task: tools point_cloud_downsampling: enabled: true method: random object_points: 256 env_points: 512 precontact_smoothing: enabled: true postcontact_enabled: true method: linear history: enabled: true length: 5 tactile_history: true pose_history: true temporal_dropout: 0.1 temporal_attention_heads: 4 reference_tactile: enabled: true n_steps: 5 method: median use_difference: true episode_filtering: filter_initial_contact: true initial_contact_threshold: 0.1 filter_low_tactile: true low_tactile_threshold: 1.0e-07 low_tactile_ratio: 0.5 train_data_path: /users/kevinma/tacsl/collected_data/tools_mixed/train val_data_path: /users/kevinma/tacsl/collected_data/tools_mixed/val test_data_path: /users/kevinma/tacsl/collected_data/tools_mixed/test validate_grasping: false grasp_threshold: 0.1 augmentation: enabled: true rotation_range: 30 translation_range: 0.1 noise_std: 0.001 point_cloud: position_noise_std: 0.01 orientation_noise_std: 0.05 patch_removal: enabled: true num_patches: 3 patch_size_ratio: 0.05 removal_probability: 0.3 tactile: noise_std: 0.001 pose: noise_std: 0.005 velocity_noise_std: 0.01 delay_steps: 1 delay_probability: 0.2 tactile_filtering: enabled: true spatial: enabled: true method: gaussian sigma: 0.25 temporal: enabled: true window_length: 7 polyorder: 1 tactile_normalization: enabled: true method: scale scale_factor: 1000.0 clip_range: - -10.0 - 10.0 per_channel: true training: batch_size: 320 num_epochs: 400 learning_rate: 0.0001 weight_decay: 0.0001 accelerator: gpu devices: - 0 - 1 - 2 - 3 - 4 - 5 - 6 - 7 strategy: auto precision: 32 gradient_clip_val: 1.0 accumulate_grad_batches: 1 scheduler: type: plateau mode: min factor: 0.7 patience: 20 threshold: 1.0e-05 min_lr: 1.0e-07 verbose: true loss: contact_prob_weight: 1.0 contact_force_weight: 2.0 force_magnitude_weight: 0.0 tactile_reconstruction_weight: 0.5 contact_prob_loss_type: focal focal_loss: alpha: 0.9 gamma: 0.75 use_hard_labels: false hard_label_cutoff: 0.5 bce_pos_weight: 1.0 contact_force_loss_type: magnitude_direction use_magnitude_loss: false magnitude_direction: magnitude_loss_type: mse magnitude_weight: 1.5 direction_loss_type: cosine direction_weight: 1.0 magnitude_threshold: 0.005 use_adaptive_weighting: true adaptive_weighting: weighting_scheme: log_magnitude min_weight: 1.0 max_weight: 3.0 temperature: 1.0 early_stopping: patience: 100 min_delta: 1.0e-05 val_frequency: 1 save_frequency: 20 optimizer: type: AdamW lr: 0.0001 weight_decay: 0.0001 betas: - 0.9 - 0.999 device: cuda num_workers: 0 pin_memory: true logging: log_dir: logs experiment_name: ${experiment_name} log_frequency: 100 use_timestamp: true wandb: project: contact-field-estimation entity: null tags: - contact_field - tactile_pointnet_concat - tactile_as_pointcloud - no_pose_encoding - no_reconstruction - tools_mixed notes: 'Tactile-as-pointcloud architecture: tactile converted to point cloud with force channels, no explicit pose encoding, spatial reasoning only' offline: false checkpoint: save_dir: checkpoints use_timestamp: true save_best: true save_latest: true resume_from: null resume_from: null evaluation: metrics: - accuracy - precision - recall - f1_score - force_mse visualize_predictions: true save_predictions: true