Lal Claude Opus 4.6 commited on
Commit
9f2aaba
·
1 Parent(s): a492bc0

Add performance metrics, training details, fix loading code

Browse files

- Fix malformed YAML frontmatter
- Add performance metrics (MSE, Pearson for val/test)
- Add training hyperparameters
- Add parameter count (6.3M)
- Add weights_only=False to loading code

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Files changed (1) hide show
  1. README.md +28 -6
README.md CHANGED
@@ -1,5 +1,4 @@
1
  ---
2
- # 1. Metadata Block
3
  license: mit
4
  library_name: pytorch-lightning
5
  pipeline_tag: tabular-regression
@@ -16,8 +15,31 @@ datasets:
16
  This model is a single-task regression model trained to take in 2114 bp genomic intervals and predict the total GM12878 DNase-seq coverage in the central 1000 bp. It is described in Lal et al. 2025 (https://www.nature.com/articles/s41592-025-02868-z).
17
 
18
  - **Architecture:** DilatedConvModel (gReLU)
19
- - **Input:** Genomic sequences (hg38)
20
- - **Output:** Total DNase-seq coverage in the central 1000 bp.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
21
 
22
  ## Repository Content
23
  1. `model.ckpt`: The trained model weights and hyperparameters (PyTorch Lightning checkpoint).
@@ -33,10 +55,10 @@ from grelu.lightning import LightningModel
33
  from huggingface_hub import hf_hub_download
34
 
35
  ckpt_path = hf_hub_download(
36
- repo_id="Genentech/GM12878_dnase-model",
37
  filename="model.ckpt"
38
  )
39
 
40
- model = LightningModel.load_from_checkpoint(ckpt_path)
41
  model.eval()
42
- ```
 
1
  ---
 
2
  license: mit
3
  library_name: pytorch-lightning
4
  pipeline_tag: tabular-regression
 
15
  This model is a single-task regression model trained to take in 2114 bp genomic intervals and predict the total GM12878 DNase-seq coverage in the central 1000 bp. It is described in Lal et al. 2025 (https://www.nature.com/articles/s41592-025-02868-z).
16
 
17
  - **Architecture:** DilatedConvModel (gReLU)
18
+ - **Input:** 2,114 bp genomic sequences (hg38)
19
+ - **Output:** Total DNase-seq coverage in the central 1000 bp
20
+ - **Parameters:** 6.3M
21
+
22
+ ## Performance
23
+
24
+ | Split | MSE | Pearson |
25
+ |-------|-----|---------|
26
+ | Validation | 0.4458 | 0.7524 |
27
+ | Test | 0.4113 | 0.8056 |
28
+
29
+ ## Training Details
30
+
31
+ | Parameter | Value |
32
+ |-----------|-------|
33
+ | Task | Regression |
34
+ | Loss | MSE |
35
+ | Optimizer | Adam |
36
+ | Learning rate | 0.0001 |
37
+ | Batch size | 512 |
38
+ | Max epochs | 15 |
39
+ | Channels | 512 |
40
+ | n_conv | 9 |
41
+ | crop_len | 557 |
42
+ | grelu version | 1.0.4.post1.dev39 |
43
 
44
  ## Repository Content
45
  1. `model.ckpt`: The trained model weights and hyperparameters (PyTorch Lightning checkpoint).
 
55
  from huggingface_hub import hf_hub_download
56
 
57
  ckpt_path = hf_hub_download(
58
+ repo_id="Genentech/GM12878_dnase-model",
59
  filename="model.ckpt"
60
  )
61
 
62
+ model = LightningModel.load_from_checkpoint(ckpt_path, weights_only=False)
63
  model.eval()
64
+ ```