Upload folder using huggingface_hub

indel_model/README.md

@@ -0,0 +1,88 @@
+---
+library_name: transformers
+tags:
+- generated_from_trainer
+datasets:
+- crispr_data
+model-index:
+- name: SX_spcas9_Lindel_indel
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+# SX_spcas9_Lindel_indel
+
+This model is a fine-tuned version of [](https://huggingface.co/) on the crispr_data dataset.
+It achieves the following results on the evaluation set:
+- Loss: 47.0798
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 0.001
+- train_batch_size: 100
+- eval_batch_size: 100
+- seed: 63036
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_ratio: 0.05
+- num_epochs: 30.0
+
+### Training results
+
+| Training Loss | Epoch | Step | Validation Loss |
+|:-------------:|:-----:|:----:|:---------------:|
+| 1022.9014     | 1.0   | 322  | 880.6261        |
+| 710.5549      | 2.0   | 644  | 540.0052        |
+| 420.3118      | 3.0   | 966  | 316.6205        |
+| 244.2501      | 4.0   | 1288 | 185.2687        |
+| 144.4954      | 5.0   | 1610 | 113.2889        |
+| 92.2284       | 6.0   | 1932 | 77.3025         |
+| 66.9501       | 7.0   | 2254 | 60.5766         |
+| 55.4648       | 8.0   | 2576 | 53.1402         |
+| 50.3219       | 9.0   | 2898 | 49.8257         |
+| 48.0616       | 10.0  | 3220 | 48.4087         |
+| 47.0631       | 11.0  | 3542 | 47.6446         |
+| 46.6081       | 12.0  | 3864 | 47.4013         |
+| 46.3857       | 13.0  | 4186 | 47.2288         |
+| 46.2889       | 14.0  | 4508 | 47.1596         |
+| 46.2447       | 15.0  | 4830 | 47.0971         |
+| 46.2195       | 16.0  | 5152 | 47.1162         |
+| 46.2015       | 17.0  | 5474 | 47.1085         |
+| 46.1924       | 18.0  | 5796 | 47.0946         |
+| 46.1853       | 19.0  | 6118 | 47.1015         |
+| 46.1784       | 20.0  | 6440 | 47.0497         |
+| 46.1757       | 21.0  | 6762 | 47.0853         |
+| 46.1676       | 22.0  | 7084 | 47.0706         |
+| 46.1614       | 23.0  | 7406 | 47.1253         |
+| 46.155        | 24.0  | 7728 | 47.0727         |
+| 46.1499       | 25.0  | 8050 | 47.0992         |
+| 46.1442       | 26.0  | 8372 | 47.0982         |
+| 46.1369       | 27.0  | 8694 | 47.0905         |
+| 46.1309       | 28.0  | 9016 | 47.0622         |
+| 46.1231       | 29.0  | 9338 | 47.0786         |
+| 46.117        | 30.0  | 9660 | 47.0798         |
+
+
+### Framework versions
+
+- Transformers 4.44.2
+- Pytorch 2.4.0+cu124
+- Datasets 2.21.0
+- Tokenizers 0.19.1

indel_model/config.json

@@ -1,5 +1,4 @@
 {
-  "_name_or_path": "/home/ljw/sdc1/CRISPR_results/Lindel/SX_spcas9_Lindel_indel",
   "architectures": [
     "LindelModel"
   ],

indel_model/model.py

@@ -0,0 +1,73 @@
+from transformers import PretrainedConfig, PreTrainedModel
+import torch.nn as nn
+import torch
+import torch.nn.functional as F
+
+class LindelConfig(PretrainedConfig):
+    model_type = "Lindel"
+    label_names = ["count"]
+
+    def __init__(
+        self,
+        dlen = 30, # the upper limit of deletion length (strictly less than dlen)
+        mh_len = 4, # the upper limit of micro-homology length
+        model = "indel", # the actual model, should be "indel", "del", or "ins"
+        reg_mode = "l2", # regularization method, should be "l2" or "l1"
+        reg_const = 0.01, # regularization coefficient
+        seed = 63036, # random seed for intialization
+        **kwargs,
+    ):
+        self.dlen = dlen
+        self.mh_len = mh_len
+        self.model = model
+        self.reg_mode = reg_mode
+        self.reg_const = reg_const
+        self.seed = seed
+        super().__init__(**kwargs)
+
+class LindelModel(PreTrainedModel):
+    config_class = LindelConfig
+
+    def __init__(self, config) -> None:
+        super().__init__(config)
+        # In more recent versions of PyTorch, you no longer need to explicitly register_parameter, it's enough to set a member of your nn.Module with nn.Parameter to "notify" pytorch that this variable should be treated as a trainable parameter (https://stackoverflow.com/questions/59234238/how-to-add-parameters-in-module-class-in-pytorch-custom-model).
+        self.generator = torch.Generator().manual_seed(config.seed)
+        self.reg_mode = config.reg_mode
+        self.reg_const = config.reg_const
+        if config.model == "indel":
+            # onehotencoder(ref[cut-17:cut+3])
+            feature_dim = 20 * 4 + 19 * 16
+            class_dim = 2
+        elif config.model == "ins":
+            # onehotencoder(ref[cut-3:cut+3])
+            feature_dim = 6 * 4 + 5 * 16
+            class_dim = 21
+        elif config.model == "del":
+            class_dim = (4 + 1 + 4 + config.dlen - 1) * (config.dlen - 1) // 2
+            # concatenate get_feature and onehotencoder(ref[cut-17:cut+3])
+            feature_dim = class_dim * (config.mh_len + 1) + 20 * 4 + 19 * 16
+        self.linear = nn.Linear(in_features=feature_dim, out_features=class_dim)
+        self.initialize_weights()
+
+    def initialize_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, mean=0, std=1, generator=self.generator)
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+
+    def forward(self, input, count=None) -> torch.Tensor:
+        logit = self.linear(input)
+        if count is not None:
+            return {
+                "logit": logit,
+                "loss": self.cross_entropy_reg(logit, count)
+            }
+        return {"logit": logit}
+
+    def cross_entropy_reg(self, logit, count):
+        if self.reg_mode == "l2":
+            reg_term = (self.linear.weight ** 2).sum()
+        elif self.reg_mode == "l1":
+            reg_term = abs(self.linear.weight).sum()
+        return -(F.log_softmax(logit, dim=1) * F.normalize(count.to(torch.float32), p=1.0, dim=1)).sum() + logit.shape[0] * self.reg_const * reg_term

indel_model/runs/Oct28_11-56-11_ljw-System-Product-Name/events.out.tfevents.1730087772.ljw-System-Product-Name.78325.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cce6bf7593c89b8019d0a4db0ff9ecb169fc9fd13b63e1bd01359dd095293d20
+size 19407

indel_model/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5bd99de9d33243b395d23a54763b3dea4ad04cb4c5dcd76a64c123da53be6ba8
+size 5304