Obtain pre-trained .pt files

[hwl26]

Excuse me,
Would it be necessary to run the process and generate a .pt file every time for different types of sample data?
Could I ask if there are any online available pre-trained .pt files that can be used ?

[ValentinDebarnot]

Hi, we are still working on the possibility of releasing a pre-trained model. This is not trivial and requires a lot of engineering. Currently, the recommended way is to generate a .pt file for every different type of sample.
On similar samples, you can specify the path to an already trained model using the option "-pretrain-path".
I hope it helps!

[hwl26]

Thank you so much for your clear reply!
I now understand the engineering challenges involved, and I’ll follow the recommended approach for now.
Looking forward to the pre-trained model release in the future. 🙏
Thank you ！

[13567222319]

hi.
If several sets of data were collected on the same lamella, the same .pt file would be necessary to run the process ?
By the way, which batch-size would you prefer to use when GPU is A40?

[ValentinDebarnot]

Hi,
There are two possibilities:

1. Precise but slower: train one model per tomogram. This will produce the best results, but can be long.
2. Fast but less precise: generate one .pt file by training on several volumes. Results are likely to not be optimal on data not used for training. Notice that we are still working on releasing a simple tutorial on how to train on multiple volumes. This is not quite optimal now.

Regarding the batch size, we would recommend taking it large, as soon as it fits in your GPU memory.

[13567222319]

Hi, There are two possibilities:

1. Precise but slower: train one model per tomogram. This will produce the best results, but can be long.
2. Fast but less precise: generate one .pt file by training on several volumes. Results are likely to not be optimal on data not used for training. Notice that we are still working on releasing a simple tutorial on how to train on multiple volumes. This is not quite optimal now.

Regarding the batch size, we would recommend taking it large, as soon as it fits in your GPU memory.

Thank you for your reply.
However, when conducting training on my computer, I noticed that the larger the batch size, the longer the training process takes. I wonder if this is normal?