top1-5 accuracy did not achieve the expected effect（Mosi/Finetuned on UCF101/HMDB51 dataset） #18
Comments
TJQdoIt9527
changed the title
TJQdoIt9527
changed the title
|
Hi, thanks for the interest. I will look into this problem ASAP. |
|
The result was generated on the first split for both UCF and HMDB. We did not average the performance on all three splits. |
Oh, okay, that's even more confusing for me. Here's val_ 10clipsx1crops.log output (including the results of three splits): [09/01 10:46:56][INFO] tadaconv.utils.checkpoint: 511: Load from given checkpoint file. The last result is the first val split The configuration content in val.log is as follows： |
|
It is important to note that our model is trained on the first split, so it is only valid when it is evaluated on the test set of the first split. For the problem, we suspect that we might have unmatched test lists, and we provide ours here for your information. |
The accuracy (in the test file you provided) is still unsatisfactory, and the results are as follows Checkpoint file path: /home/lzh/2022/tjq/TAdaConv/checkpoint/r2p1d_pt_hmdb_ft_hmdb_5183_public.pyth There is a fact that I have to mention, when decord=0.4.1, my program reported the following error Error at decoding. 7/10. Vid index: 1130, Vid path: /data/hmdb51/videos/sit/TheBoondockSaints_sit_u_cm_np1_fr_med_29.avi Based on this（https://github.com/dmlc/decord/issues/124） link, I feel that there is a problem with the decord version。 The conda list is as follows: Is it due to the decord version? Is there anything else I haven't noticed? |
|
This is caused by the change of cropping function that we use. Set the cfg.DATA.TEST_SCALE to 112 and the problem would be solved. |
After adding this parameter, testing in the test file you provided has indeed improved the effect, but it is still about 2 points short of your best result (using backbone r2p1d/r2d3s on hmdb51). The results are as follows: [09/07 13:49:12][INFO] tadaconv.utils.checkpoint: 511: Load from given checkpoint file. [09/07 13:49:39][INFO] tadaconv.datasets.base.base_dataset: 197: Dataset HMDB51 split test loaded. Length 15300. The configuration file is as follows： Are there any other parameters that I need to pay attention to? For example, AUTOAUGMENT? |
|
I do not see a problem in your current config. Autoaugment only affects the training process. |
|
On our side, we have reproduced a result similar to yours using decord==0.6.0. We used decord==0.4.0 for producing 51.83 on HMDB51. |
Yes, I am using the test list you provided. You are right. The results I obtained above were at decord=0.6.0. When I converted decord=0.4.0, my results were as follows: It is very close to yours result. This makes me extremely excited, and I am also very grateful for your patient guidance! However, I attempted to test using the r2d3ds backbone on the hmdb51 dataset, but the results still differ by two points from yours. (I added TEST SCALE: 112 and decord=0.4.0 and the test list you provided), and the results are as follows: Checkpoint file path: /home/lwd/aim_tada/TAdaConv/checkpoint/r2d3ds_pt_hmdb_ft_hmdb_4693_public.pyth May I ask if this result is normal, or what additional parameters do I need to add for this backbone network |
|
Could you try using the code of this version to reproduce the result and see whether it is normal? |
Thanks for your great work!
It is worth affirming that your work has great research value. However, when I loaded the fine-tuned checkpoint in the MOSI project (with R (2+1) D-10 as the backbone), I ran r2p1d_test.yaml, the accuracy of top1-5 is approximately 72.77% and 91.49%, respectively. I don't know where I made the mistake. I haven't made any changes to the configuration file. Do I need to add some additional configuration parameter? Or can you provide the configuration in val.log
Looking forward to your reply, thank you very much!
The text was updated successfully, but these errors were encountered: