how to get 69.27

[jinluo12345]

hello,author! I follow your thumos14 dataset and use the data and code to train a model,but the final result is only 68.88 and i wonder how to get 69.27 really thanks,its really helpful to me

[jinluo12345]

dataset_name: thumos train_split: [ 'validation' ] val_split: [ 'test' ] dataset: { json_file: ./data/thumos/annotations/thumos14.json, feat_folder: ./data/thumos/i3d_features, file_prefix: ~, file_ext: .npy, num_classes: 20, input_dim: 2048, feat_stride: 4, num_frames: 16, trunc_thresh: 0.5, crop_ratio: [ 0.9, 1.0 ], max_seq_len: 2304, } model: { fpn_type: identity, max_buffer_len_factor: 6.0, backbone_arch: [ 2, 2, 5 ], n_sgp_win_size: 1, regression_range: [ [ 0, 4 ], [ 4, 8 ], [ 8, 16 ], [ 16, 32 ], [ 32, 64 ], [ 64, 10000 ] ], num_bins: 16, k: 5, iou_weight_power: 0.2, use_trident_head: True, sgp_mlp_dim: 768, input_noise: 0.0005 } opt: { learning_rate: 0.0001, warmup_epochs: 20, epochs: 20, weight_decay: 0.025, } loader: { batch_size: 2, } train_cfg: { init_loss_norm: 100, clip_grad_l2norm: 1.0, cls_prior_prob: 0.01, center_sample: radius, center_sample_radius: 1.5, }

test_cfg: { voting_thresh: 0.7, pre_nms_topk: 2000, max_seg_num: 2000, min_score: 0.001, multiclass_nms: True, } output_folder: ./ckpt/

this is the parameters i used ,not changing,but can not reproduce the result qvq

[dingfengshi]

Hi, we use exactly the same setting and data for training. It seems that there are fluctuations in the results when using different devices and environments, and THUMOS14 appears to be more sensitive to these variations.

[jinluo12345]

really thanks for your reply! and may i ask the version of your cuda?thanks!

[dingfengshi]

really thanks for your reply! and may i ask the version of your cuda?thanks!

Hi, we conduct experiments with a single GPU in the DGX-A100-320G station. The version of CUDA is 11.3 and the version of Pytorch is 1.11.0 and for other requirements, please see the requirements.txt file. I hope this information can be helpful to you.