train_test.md

Training / Testing

Before training/testing, we hope that you have read data and finish preparing dataset.

Quick Start

For training/testing depth completion :

python3 -m depth_completion.main

Note : modify main.py for different agent and configuration if needed.

Agent

For different agent, you can see agent_list and then modify main.py.

Configuration

For different configuration corresponding to the agent, you can see configs for example configurations. Here we will take one for example :

train_config = {
    "dataset_name": "matterport",
    "model_name": ("GatedConvSkipConnectionModel", "UNet"),
    "in_channel": (9, 1),
    "device_ids": [3],
    "seed": 7122,

    "num_workers": 8,
    "mode": "train",
    "train_path": "/tmp2/tsunghan/new_matterport/v1",
    "lr": (1e-4, 1e-4),
    "batch_size": 2,
    "loss_func": {('depth(L1)', 'depth_L1_loss', 1.), ('bc(L1)', 'bc_L1_loss', 1)},
    "load_model_path": (None, None),
    "param_only": (False, False),
    "validation": True,
    "valid_path": "/tmp2/tsunghan/new_matterport/v1",
    "epoches": 100,
    "save_prefix": "official_ver",
}

Common

• model_name : Models in models you use for backbone model. (Second input for boundary consistency network)
• in_channel : input channel for the model
• device_ids : GPU-ids (support multi-gpu training)

Training

• training_path, valid_path : training and validation dataset root path. (Training list and Testing list are here)
• loss_func : dictionary for loss functions, each element is a tuple : (description, function_name, weight). You can see loss_funcs for all loss functions.
• param_only : Set False as default. If set as True, you need to pass load_model_path for re-training.

Note

• Training log and model will be saved in experiments directory.

Testing

• test_path : testing dataset path.
• batch_size : set as 1 for testing
• param_only : Set as True and pass load\_model\_path.
• output : output directory saving .npy files.

Note

• You can download pre-trained model from Google Drive if needed.
• Testing scripts only give you .npy files. If you want to evaluate the result or visualize them, please refer visualize