train_Vaihingen3D_WeakLabel.py

#
#
#      0=================================0
#      |    Kernel Point Convolutions    |
#      0=================================0
#
#
# ----------------------------------------------------------------------------------------------------------------------
#
#      Callable script to start a training with weak region-labels (WL) on Vaihingen3D dataset.
#      This is based on the multi-path region mining (mprm) approach.
#      - adapted by Johannes Ernst
#
# ----------------------------------------------------------------------------------------------------------------------
#
#      Hugues THOMAS - 06/03/2020
#
# ----------------------------------------------------------------------------------------------------------------------
#
#           Imports and global variables
#       \**********************************/
#

# Common libs
import signal
import os
import time

# Dataset
from datasets.Vaihingen3D_WeakLabel import *
from torch.utils.data import DataLoader

# Utils
from utils.config import Config
from utils.trainer_WeakLabel import ModelTrainer
from utils.tester_WeakLabel import ModelTesterWL
from models.architectures import *


# ----------------------------------------------------------------------------------------------------------------------
#
#           Config Class
#       \******************/
#

class Vaihingen3DWLConfig(Config):
    """
    Override the parameters you want to modify for this dataset
    """

    ####################
    # Dataset parameters
    ####################

    # Dataset name
    dataset = 'Vaihingen3DWL'

    # Number of classes in the dataset (This value is overwritten by dataset class when Initializating dataset).
    num_classes = None

    # Type of task performed on this dataset (also overwritten)
    dataset_task = ''

    # Number of CPU threads for the input pipeline
    input_threads = 10

    #########################
    # Architecture definition
    #########################

    # Define layers
    architecture = ['simple',
                    'resnetb',
                    'resnetb_strided',
                    'resnetb',
                    'resnetb_strided',
                    'resnetb',
                    'nearest_upsample',
                    'nearest_upsample']


    ###################
    # KPConv parameters
    ###################

    # Number of kernel points
    num_kernel_points = 15

    # Radius of the input sphere (decrease value to reduce memory cost)
    in_radius = 18

    # Radius of the subcloud for weak labels (smaller means more labels but better results)
    # Increse to reduce memory cost
    # NOTE: Delete old _anchors.pkl files in input_xy/ when changing
    sub_radius = 5

    # Size of the first subsampling grid in meter (increase value to reduce memory cost)
    first_subsampling_dl = 0.24

    # Radius of convolution in "number grid cell" (2.5 is the standard value)
    conv_radius = 2.5

    # Radius of deformable convolution in "number grid cell". Larger so that deformed kernel can spread out
    deform_radius = 1.0

    # Radius of the area of influence of each kernel point in "number grid cell" (1.0 is the standard value)
    KP_extent = 1.0

    # Behavior of convolutions in ('constant', 'linear', 'gaussian')
    KP_influence = 'linear'

    # Aggregation function of KPConv in ('closest', 'sum')
    aggregation_mode = 'sum'

    # Choice of input features (4 here means [ones  intensity absoluteHeight reducedHeight])
    first_features_dim = 64
    in_features_dim = 4

    # Can the network learn modulations
    modulated = False

    # Batch normalization parameters
    use_batch_norm = True
    batch_norm_momentum = 0.02

    # Deformable offset loss
    # 'point2point' fitting geometry by penalizing distance from deform point to input points
    # 'point2plane' fitting geometry by penalizing distance from deform point to input point triplet (not implemented)
    deform_fitting_mode = 'point2point'
    deform_fitting_power = 1.0              # Multiplier for the fitting/repulsive loss
    deform_lr_factor = 0.1                  # Multiplier for learning rate applied to the deformations
    repulse_extent = 1.2                    # Distance of repulsion for deformed kernel points

    #####################
    # Training parameters
    #####################

    # Maximal number of epochs
    max_epoch = 80

    # Learning rate management (standard value is 1e-2)
    learning_rate = 0.01
    momentum = 0.98
    lr_decays = {i: 0.98 for i in range(1, 1000)}
    grad_clip_norm = 1

    # Number of batch (or number of input spheres)
    batch_num = 3

    # Number of steps per epochs
    epoch_steps = 600

    # Number of validation examples per epoch
    validation_size = 200

    # Number of epoch between each checkpoint
    checkpoint_gap = 40

    # Augmentations
    augment_scale_anisotropic = True
    augment_symmetries = [True, True, False]
    augment_rotation = 'vertical'
    augment_scale_min = 0.8
    augment_scale_max = 1.2
    augment_noise = 0.04

    # Choose weights for class
    class_w = [1, 1, 1, 1, 1, 1, 1, 1, 1]

    # Active learning parameters (label parameters are per input file)
    # NOTE: Delete old anchor files when switching active learning on or of
    # (i.e. when switching "active_learning_iterations = 0" to > 0 or vice versa)
    active_learning_iterations = 20
    initial_labels_per_file = 600
    subsample_method = 'balanced'
    added_labels_per_epoch = 200

    # Decide whether to subsample weak labels 
    # NOTE: Must be True for active_learning_iterations > 0
    subsample_labels = True if active_learning_iterations > 0 else False

    # Other parameters
    model_name = 'KPFCNN_mprm'
    loss_type = 'region_mprm_loss'
    anchor_method = 'reduced'

    # Do we nee to save convergence
    saving = True
    saving_path = None


# ----------------------------------------------------------------------------------------------------------------------
#
#           Main Call
#       \***************/
#

if __name__ == '__main__':

    ############################
    # Initialize the environment
    ############################

    # Set which gpu is going to be used
    GPU_ID = '0'

    # Set GPU visible device
    os.environ['CUDA_VISIBLE_DEVICES'] = GPU_ID

    ###############
    # Previous chkp
    ###############

    # Choose here if you want to start training from a previous snapshot (None for new training)
    # previous_training_path = 'Log_2020-03-19_19-53-27'
    previous_training_path = ''

    # Choose index of checkpoint to start from. If None, uses the latest chkp
    chkp_idx = None
    if previous_training_path:

        # Find all snapshot in the chosen training folder
        chkp_path = os.path.join('results/WeakLabel', previous_training_path, 'checkpoints')
        chkps = [f for f in os.listdir(chkp_path) if f[:4] == 'chkp']

        # Find which snapshot to restore
        if chkp_idx is None:
            chosen_chkp = 'current_chkp.tar'
        else:
            chosen_chkp = np.sort(chkps)[chkp_idx]
        chosen_chkp = os.path.join('results/WeakLabel', previous_training_path, 'checkpoints', chosen_chkp)

    else:
        chosen_chkp = None

    ##############
    # Prepare Data
    ##############

    print()
    print('Data Preparation')
    print('****************')

    # Initialize configuration class
    config = Vaihingen3DWLConfig()
    if previous_training_path:
        config.load(os.path.join('results/WeakLabel', previous_training_path))

        # Find the current active learning iteration
        iteration_files = [f for f in os.listdir(config.saving_path) if f[:18] == 'training_iteration']
        iteration_previous = len(iteration_files)-1

        # Reset saving path
        config.saving_path = None

    # Get path from argument if given
    if len(sys.argv) > 1:
        config.saving_path = sys.argv[1]

    # Active learning loop
    for iteration in range(config.active_learning_iterations + 1):

        if previous_training_path:
            iteration += iteration_previous

        # Initialize datasets for training and validation
        training_dataset = Vaihingen3DWLDataset(config, set='training', use_potentials=True, al_iteration=iteration)
        validation_dataset = Vaihingen3DWLDataset(config, set='validation', use_potentials=True)

        # Initialize dataset for testing on the training set (test_on_train=True)
        test_dataset = Vaihingen3DWLDataset(config, set='test', use_potentials=True, test_on_train=True)

        # Initialize samplers
        training_sampler = Vaihingen3DWLSampler(training_dataset)
        validation_sampler = Vaihingen3DWLSampler(validation_dataset)
        test_sampler = Vaihingen3DWLSampler(test_dataset)

        # Initialize the dataloader
        training_loader = DataLoader(training_dataset,
                                     batch_size=1,
                                     sampler=training_sampler,
                                     collate_fn=Vaihingen3DWLCollate,
                                     num_workers=config.input_threads,
                                     pin_memory=True)
        validation_loader = DataLoader(validation_dataset,
                                       batch_size=1,
                                       sampler=validation_sampler,
                                       collate_fn=Vaihingen3DWLCollate,
                                       num_workers=config.input_threads,
                                       pin_memory=True)
        test_loader = DataLoader(test_dataset,
                                 batch_size=1,
                                 sampler=test_sampler,
                                 collate_fn=Vaihingen3DWLCollate,
                                 num_workers=config.input_threads,
                                 pin_memory=True)

        # Calibrate samplers
        training_sampler.calibration(training_loader, verbose=True)
        validation_sampler.calibration(validation_loader, verbose=True)
        test_sampler.calibration(test_loader, verbose=True)

        # Optional debug functions
        # debug_timing(training_dataset, training_loader)
        # debug_timing(validation_dataset, validation_loader)
        # debug_upsampling(training_dataset, training_loader)

        print('\nModel Preparation')
        print('*****************')

        # Define network model
        t1 = time.time()
        net = KPFCNN_mprm(config, training_dataset.label_values, training_dataset.ignored_labels)
    
        debug = False
        if debug:
            print('\n*************************************\n')
            print(net)
            print('\n*************************************\n')
            for param in net.parameters():
                if param.requires_grad:
                    print(param.shape)
            print('\n*************************************\n')
            print("Model size %i" % sum(param.numel() for param in net.parameters() if param.requires_grad))
            print('\n*************************************\n')

        # Define a trainer class
        trainer = ModelTrainer(net, config, chkp_path=chosen_chkp)
        print('Done in {:.1f}s\n'.format(time.time() - t1))

        print('\nStart training')
        print('**************')

        # Training
        trainer.train(net, training_loader, validation_loader, config, al_iteration=iteration)

        # Print amount of used weak labels
        anchor_num_init = training_dataset.config.initial_labels_per_file*len(training_dataset.cloud_names) + \
                          iteration*training_dataset.config.added_labels_per_epoch*len(training_dataset.cloud_names)
        anchor_num_over = np.sum([len(f) for f in training_dataset.anchors])
        print('\n*************************************')
        print('Initial amount of weak labels:  {:d}'.format(anchor_num_init))
        print('Amount of weak labels with overlaps:  {:d}'.format(anchor_num_over))
        print('*************************************\n')

        # Test network on training data to get probabilities for active learning
        # --> Weak label set is extended for the next iteration in cloud_segmentation_test
        if config.active_learning_iterations and not iteration == config.active_learning_iterations:
            torch.cuda.empty_cache()
            chosen_chkp = os.path.join(config.saving_path, 'checkpoints/current_chkp.tar')
            tester = ModelTesterWL(net, chkp_path=chosen_chkp)
            tester.cloud_segmentation_test(net, test_loader, config, num_votes=10, active_learning=True)
        
        # Reset the checkpoint to ensure a new training network for the next iteration
        chosen_chkp = None

        # Break active learning loop if desired iteration is reached
        if iteration == config.active_learning_iterations:
            break

    print('Forcing exit now')
    os.kill(os.getpid(), signal.SIGINT)