ImageNet pretrained model about lite-mono HOT 7 CLOSED

done, similar performance, and I notice that you uploaded the training code, close issue.

from lite-mono.

Hi, I will upload the pre-trained weights soon and let you know.

from lite-mono.

Thanks.

from lite-mono.

Hi @FangGet , I have uploaded the pre-trained weights. A tip for training, you need to set the weight_decay to 1e-2 and drop_path to 0.2 to mitigate overfitting. For the latest Lite-Mono-8M, I use a larger drop path rate, which is 0.4.
Please let me know if you have more questions.

from lite-mono.

Thanks @noahzn .
With AdamW optimizer, weight_decay=1e-2, drop_path=0.2 learning_rate=1e-4, pretrained Imagenet, I got the results:

Gap still exists, waiting for your training detail(maybe different learning rate for depth and pose network?).

from lite-mono.

Hi! Glad to see that you trained the network by yourself! Some comments:

1. I used the initial learning rate 1e-4 for both the DepthNet and the PoseNet of Lite-Mono. Then, the learning rates decrease to 5e-6 and 1e-5, respectively.
2. The best results are usually not achieved in the last epoch due to training fluctuations.
3. I used the cosine annealing scheduler to reduce learning rates.
4. The drop path rate introduces uncertainty into the training. The same results cannot be got unless the randomness of the training is fixed, by setting fixed seeds.
5. I have just done several rounds of training after reading your comment, and found that some training achieved better results in 2 or 3 metrics than the results reported in the paper. But also one round of training couldn't converge.
6. You can use Tensorboard to check the training process. tensorboard --logdir ./tmp/your_saved_model_folder.

from lite-mono.

Thanks.

1. useful, I will try that.
2. I also get best results not from last epoch.
3. Cosine annealing has also been tried, but with a much smaller eta lr. I will try it with tips. 1.
4. 4.5.6. understand.

thanks again.

from lite-mono.