Pretrained Chromoformer weights

Pretrained weights for Chromoformer models. Models were individually trained for 11 cell types from Roadmap Epigenomics Project. For each cell type, 18955 protein coding genes were divided into four non-overlapping cross-validation (CV) sets, and trained weights for each cross-validation fold is provided in this dataset.<br>The directory is organized firstly by Roadmap Epigenomics cell type (using corresponding EID), and four pytorch checkpoint files (a .pt file for each CV-fold) are placed in each directory.<br>'*.pt' files are basically dictionaries containing pretrained weights as well as some training details associated with the following keys:<br>- 'net' : Pretrained weights- 'optimizer': Status of pytorch optimizer at the time of model snapshot.- 'epoch': Epoch at the time of model snapshot.- 'last_val_loss': Validation loss at the time of model snapshot.- 'last_val_auc': Validation AUC at the time of model snapshot.- 'val_score': Output probabilities for genes in validation set.- 'val_label': True labels for genes in validation set.

Chromoformer模型的预训练权重。本数据集针对表观基因组路线图项目（Roadmap Epigenomics Project）的11种细胞类型分别完成了模型训练。针对每种细胞类型，本数据集将18955个蛋白质编码基因划分为4组无重叠的交叉验证（CV）集，并提供了每个交叉验证折对应的训练权重。
本数据集的目录首先以表观基因组路线图项目的细胞类型（采用对应的EID标识）进行层级组织，每个细胞类型对应的目录下均存放4个PyTorch模型检查点文件——每个CV折对应一个.pt格式文件。
所有*.pt后缀文件本质上均为Python字典结构，包含预训练权重以及如下训练细节信息，其对应的键名如下：
- "net"：预训练权重
- "optimizer"：模型快照时刻的PyTorch优化器状态
- "epoch"：模型快照时刻的训练轮次
- "last_val_loss"：模型快照时刻的验证损失
- "last_val_auc"：模型快照时刻的验证AUC
- "val_score"：验证集基因的输出概率值
- "val_label"：验证集基因的真实标签