Reproduction results are slightly lower than reported in the paper

[cfrfree]

Hi authors,

First of all, thank you very much for your outstanding work and for open-sourcing the TACMT code! This paper is very insightful.

I have recently been trying to reproduce the results reported in the paper, specifically for the SARVG dataset (e.g., Table 3 or Table 4 ).

Expected Results (from the paper)

According to Table 3 (Val Set), the performance metrics for TACMT (ours) are:

• Pr@0.5: 88.57
• mIoU: 81.59

According to Table 4 (Test Set), the performance metrics for TACMT (ours) are:

• Pr@0.5: 89.38
• mIoU: 82.81

Actual Results (My Reproduction)

In my experiments, the best result I obtained (evaluated on the val_split) is:

• Pr@0.5: 0.8647
• mIoU: 0.7847

The best result evaluated on the test_split is:

• Pr@0.5: 0.8676
• mIoU: 0.7939

As you can see, my results are about 2 percentage points lower than those reported in the paper.

---

Reproduction Steps

I have strictly followed the steps in the paper and the README.md:

1. Environment Installation:

   • torch==1.9.1+cu111
   • torchvision==0.10.1+cu111
   • pytorch-pretrained-bert==0.6.2
   • rasterio==1.3.11

2. Configuration File:

   • I used configs/SARVG_R50.py.
   • I downloaded the load_weights_path specified in the config file.
   • I have modified the data_root and split_root in the config file to my local paths.

3. Training Command:

   • I followed the hyperparameters described in Section 4.2 of the paper (e.g., 90 epochs total, lr_drop=60, freeze_epochs=5, L1 loss-coef=5, GIoU loss-coef=2) .
   • My training launch command is as follows:

   Bash

   # Started with 2 GPUs
   python -m torch.distributed.launch --nproc_per_node=2 --use_env train.py --config configs/SARVG_R50.py --world_size 2 --checkpoint_best --enable_batch_accum --batch_size 10 --freeze_epochs 5
   

My Environment

• PyTorch Version: 1.9.1
• CUDA Version: 11.1
• GPU Model: RTX3090 * 2
• Operating System: CentOS Linux release 7.9.2009

---

Attachments

I have attached my full training log (from epoch 0 to 90) to this issue so you can review the detailed loss and evaluation metrics.

sarvg_train.log

Question

Could you please help me check if I missed any critical settings? Or is this performance variation an expected fluctuation, possibly due to minor differences in random seeds?

Thank you very much for any help!