add dnalm-bert-base model

README.md

@@ -0,0 +1,44 @@
+# DNALM
+
+DNALM is a transformer masked language model trained on human DNA sequence.
+
+Differences between DNALM and DNABERT:
+- BPE tokenization instead of k-mers;
+- input sequence size is about 3000 nucleotides (512 BPE tokens) compared to 510 nucleotides of DNABERT
+- pre-training on T2T vs. GRCh38.p13 human genome assembly.
+
+Source code and data: https://github.com/AIRI-Institute/dna-lm
+
+## Examples
+### How to load the model to fine-tune it on classification task
+```python
+from src.dnalm.modeling_bert import BertForSequenceClassification
+from transformers import AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained('AIRI-Institute/dnalm-bert-base')
+model = BertForSequenceClassification.from_pretrained('AIRI-Institute/dnalm-bert-base')
+```
+
+## Model description
+DNALM model is trained in a masked language model (MLM) fashion, following the methods proposed in the BigBird paper by masking 85% of tokens. Model config for `dnalm-bert-base` is similar to the bert-base:
+
+- 512 Maximum sequence length
+- 12 Layers, 12 Attention heads
+- 768 Hidden size
+- 32k Vocabulary size
+
+We pre-trained dnalm-bert-base using the latest T2T human genome assembly (https://www.ncbi.nlm.nih.gov/assembly/GCA_009914755.3/). Pre-training was performed for 500,000 iterations with the same parameters as in BigBird, except sequence length was equal to 512 tokens and we used pre-layer normalization in Transformer.
+
+## Downstream tasks
+Currently, dnalm-bert-base model has been finetuned  and tested on promoter prediction task.  Its' performance is comparable to previous SOTA results. We plan to fine-tune and make available models for other downstream tasks in the near future.
+
+### Fine-tuning DNALM on our data and scoring
+After fine-tuning dnalm-bert-base on promoter prediction dataset, following results were achieved: 
+
+| model                  | seq_len (bp) | F1    |
+|------------------------|--------------|-------|
+| DeePromoter            | 300          | 95.60 |
+| DNALM bert_base (ours) | 2000         | 95.72 |
+| BigBird                | 16000        | 99.90 |
+
+We can conclude that our model achieves comparable performance to the previously published results for promoter prediction task.

config.json

@@ -0,0 +1,24 @@
+{
+    "architectures": [
+      "BertForPretraining"
+    ],
+    "attention_probs_dropout_prob": 0.1,
+    "gradient_checkpointing": false,
+    "hidden_act": "gelu",
+    "hidden_dropout_prob": 0.1,
+    "hidden_size": 768,
+    "initializer_range": 0.02,
+    "intermediate_size": 3072,
+    "layer_norm_eps": 1e-12,
+    "max_position_embeddings": 512,
+    "model_type": "bert",
+    "num_attention_heads": 12,
+    "num_hidden_layers": 12,
+    "pad_token_id": 3,
+    "pre_layer_norm": true,
+    "position_embedding_type": "absolute",
+    "transformers_version": "4.6.0.dev0",
+    "type_vocab_size": 2,
+    "use_cache": true,
+    "vocab_size": 32000
+  }

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4310578f475a716f6343c7da90fba6230f2c917e80444726a481a611c1faf054
+size 543494664

special_tokens_map.json

@@ -0,0 +1 @@
+{"unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]"}

tokenizer_config.json

@@ -0,0 +1 @@
+{"tokenizer_class": "PreTrainedTokenizerFast"}