Could you provide the code forgenerate the results for figure 1c and the extended figures

[caiquanyou]

Could you provide the code forgenerate the results for figure 1c and the extended figures

[caiquanyou]

I follow the tutorials and get InferenceRegionMotifDataset error:File /opt/conda/lib/python3.12/site-packages/requests/adapters.py:700, in HTTPAdapter.send(self, request, stream, timeout, verify, cert, proxies)
696 if isinstance(e.reason, _SSLError):
697 # This branch is for urllib3 v1.22 and later.
698 raise SSLError(e, request=request)
--> 700 raise ConnectionError(e, request=request)
702 except ClosedPoolError as e:
703 raise ConnectionError(e, request=request)

ConnectionError: HTTPConnectionPool(host='ftp.ebi.ac.uk', port=80): Max retries exceeded with url: /pub/databases/gencode/Gencode_human/release_40/gencode.v40.basic.annotation.gtf.gz (Caused by NameResolutionError("<urllib3.connection.HTTPConnection object at 0x7fb6d1ade090>: Failed to resolve 'ftp.ebi.ac.uk' ([Errno -3] Temporary failure in name resolution)"))

[fuxialexander]

Fig1c data and code and some extended figures are in the Zenodo link. The error you have looks like a DNS resolving issue. You can try manually download the file here: https://www.gencodegenes.org/human/release_40.html (the basic annotation one) and put it in $HOME/.gcell_data/annotations/; alternatively you might try use Google DNS (8.8.8.8) or AliDNS (223.5.5.5)

[caiquanyou]

Thanks ，it was solved。 Another question is could you provide the pretrainning code from scratch？I want to train more epochs

[fuxialexander]

that's basically the same as in the 'pretrain_pbmc'? you can remove the checkpoint specification and enlarge the number of epochs

[caiquanyou]

that's basically the same as in the 'pretrain_pbmc'? you can remove the checkpoint specification and enlarge the number of epochs

thanks，I’d like to pretrain the model from scratch instead of loading a pretrained checkpoint like in pretrain_pbmc. How should I modify the config.yaml file? Specifically, what else parameters need to be changed (e.g., remove checkpoint path, increase training epochs, etc.)? Also, how many GPUs are typically needed for pretraining from scratch?

[caiquanyou]

Or could you please provide the yml file for pretrain？The pretrain datas in get_demo directory are entire pretrain data in the article？

[fuxialexander]

You can just set cfg.finetune.checkpoint to null and set cfg.training.epochs to 800. We used 8*A100 for pretrain from scratch.

[caiquanyou]

Hello @fuxialexander , thank you for your help.

I am trying to load the provided pre-training data for model training, but I noticed that the format of the pre-training data is somewhat different from the demo data format used in the project.

I would like to ask:

Does the pre-training data need to be converted to the same format as the demo data, or can it be used directly in its current format?

If conversion is required, do you have any recommended processing steps or scripts for this?

Thank you very much for your help. I look forward to your response!

[fuxialexander]

Check out dataset section of https://github.com/GET-Foundation/get_model/blob/master/get_model/config/fetal_region.yaml this should be about to load it. Main difference is to inherit from base_region_config (which use

get_model/get_model/dataset/zarr_dataset.py

Line 139 in c25885c

class RegionDataset(Dataset):

, rather than

get_model/get_model/dataset/zarr_dataset.py

Line 996 in c25885c

class RegionMotifDataset(Dataset):

). also you want to change metadata_path to cell_type_pretrain_human_bingren_shendure_apr2023.txt (in s3://2023-get-xf2217/get_demo/pretrain_human_bingren_shendure_apr2023/data/cell_type_pretrain_human_bingren_shendure_apr2023.txt) and assembly to hg38