Error on training Toronto3d On RandlaNet

[7Vyshak7]

I am training the toronto3d dataset on the Randlanet model.But on training I am getting this error

RuntimeError: weight tensor should be defined either for all 8 classes or no classes but got weight tensor of shape: [1, 8]

I have not changed anything from the code and config file. What Could be the problem I will attach the config file and training code

Config

dataset: name: Toronto3D cache_dir: ./logs/cache class_weights: [41697357, 1745448, 6572572, 19136493, 674897, 897825, 4634634, 374721] ignored_label_inds:

• 0 num_classes: 8 num_points: 65536 test_files:
• L002.ply test_result_folder: ./test train_files:
• L001.ply
• L003.ply
• L004.ply use_cache: true val_files:
• L002.ply steps_per_epoch_train: 100 steps_per_epoch_valid: 10 model: name: RandLANet batcher: DefaultBatcher ckpt_path: ./logs/randlanet_toronto3d_202201071330utc.pth num_neighbors: 16 num_layers: 5 num_points: 65536 num_classes: 8 ignored_label_inds: [0] sub_sampling_ratio: [4, 4, 4, 4, 2] in_channels: 6 dim_features: 8 dim_output: [16, 64, 128, 256, 512] grid_size: 0.05 augment: recenter: dim: [0, 1, 2] normalize: points: method: linear pipeline: name: SemanticSegmentation optimizer: lr: 0.001 batch_size: 2 main_log_dir: ./logs max_epoch: 200 save_ckpt_freq: 5 scheduler_gamma: 0.99 test_batch_size: 1 train_sum_dir: train_log val_batch_size: 2 summary: record_for: [] max_pts: use_reference: false max_outputs: 1

Training Code

import os
import open3d.ml as _ml3d
import open3d.ml.torch as ml3d

cfg_file = "Config/randlanet_toronto3d.yml"
cfg = _ml3d.utils.Config.load_from_file(cfg_file)

dataset = ml3d.datasets.Toronto3D(dataset_path='./TrainingData/Toronto3d/',**cfg.dataset)
model = ml3d.models.RandLANet(**cfg.model)

pipeline = ml3d.pipelines.SemanticSegmentation(model=model, dataset=dataset,**cfg.pipeline)
pipeline.run_train()

[msanov]

Hi, you are encountering this problem because there is a bug in the original code located in ./ml3d/torch/modules/losses/semseg_loss.py. The issue arises because the code expects a tensor with 8 values (num of valid classes), but you are passing a list of tensors instead. You should use .squeeze() to fix this. Here is my corrected code (line 40):" `

def __init__(self, pipeline, model, dataset, device):
    super(SemSegLoss, self).__init__()
    # weighted_CrossEntropyLoss
    if 'class_weights' in dataset.cfg.keys() and len(
            dataset.cfg.class_weights) != 0:
        class_wt = DataProcessing.get_class_weights(
            dataset.cfg.class_weights)
        weights = torch.tensor(class_wt, dtype=torch.float, device=device)
        weights = weights.squeeze()
        self.weighted_CrossEntropyLoss = nn.CrossEntropyLoss(weight=weights)
    else:
        self.weighted_CrossEntropyLoss = nn.CrossEntropyLoss()

`