Initial commit

README.md

@@ -1,3 +1,113 @@
 ---
 license: cc-by-nc-sa-4.0
+pipeline_tag: fill-mask
+language: en
+tags:
+- long_documents
+- hierarchical_transformers
+datasets:
+- wikipedia
+model-index:
+- name: kiddothe2b/hat-mini-1024-I2
+  results: []
 ---
+
+# Hierarchical Attention Transformer (HAT) / hat-mini-1024-I2
+
+## Model description
+
+This is a Hierarchical Attention Transformer (HAT) model as presented in [An Exploration of Hierarchical Attention Transformers for Efficient Long Document Classification (Chalkidis et al., 2022)](https://arxiv.org/abs/xxx). 
+
+The model has been warm-started re-using the weights of miniature BERT [(Turc et al., 2019)](https://arxiv.org/abs/1908.08962), and continued pre-trained for MLM following the paradigm of Longformer released by [Beltagy et al. (2020)](](https://arxiv.org/abs/1908.08962)). It supports sequences of length up to 1,024.
+
+HAT use a hierarchical attention, which is a combination of segment-wise and cross-segment attention operations. You can think segments as paragraphs or sentences.
+
+## Intended uses & limitations
+
+You can use the raw model for masked language modeling, but it's mostly intended to be fine-tuned on a downstream task.
+See the [model hub](https://huggingface.co/models?other=hierarchical-transformer) to look for fine-tuned versions on a task that interests you.
+
+Note that this model is primarily aimed at being fine-tuned on tasks that use the whole document to make decisions, such as document classification, sequential sentence classification or question answering.
+
+## How to use
+
+You can use this model directly with a pipeline for masked language modeling:
+
+```python
+from transformers import pipeline
+mlm_model = pipeline('fill-mask', model='kiddothe2b/hat-mini-1024-I1', trust_remote_code=True)
+mlm_model("Hello I'm a <mask> model.")
+```
+
+You can also fine-tun it for SequenceClassification, SequentialSentenceClassification, and MultipleChoice down-stream tasks:
+
+```python
+from transformers import AutoTokenizer, AutoModelforSequenceClassification
+tokenizer = AutoTokenizer.from_pretrained("kiddothe2b/hat-mini-1024-I1", trust_remote_code=True)
+doc_classifier = AutoModelforSequenceClassification(model='kiddothe2b/hat-base-4096', trust_remote_code=True)
+```
+
+## Limitations and bias
+
+The training data used for this model contains a lot of unfiltered content from the internet, which is far from
+neutral. Therefore, the model can have biased predictions.
+
+
+## Training procedure
+
+### Training and evaluation data
+
+The model has been warm-started from [google/bert_uncased_L-6_H-256_A-4](https://huggingface.co/google/bert_uncased_L-6_H-256_A-4) checkpoint and has been continued pre-trained for additional 50k steps on English [Wikipedia](https://huggingface.co/datasets/wikipedia).
+
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 0.0001
+- train_batch_size: 4
+- eval_batch_size: 4
+- seed: 42
+- distributed_type: tpu
+- num_devices: 8
+- gradient_accumulation_steps: 4
+- total_train_batch_size: 128
+- total_eval_batch_size: 32
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_ratio: 0.1
+- training_steps: 50000
+
+
+### Training results
+
+| Training Loss | Epoch | Step  | Validation Loss |
+|:-------------:|:-----:|:-----:|:---------------:|
+| 2.7353        | 0.2   | 10000 | 2.5067          |
+| 2.6081        | 0.4   | 20000 | 2.3966          |
+| 2.5552        | 0.6   | 30000 | 2.3446          |
+| 2.5105        | 0.8   | 40000 | 2.3117          |
+| 2.4978        | 1.14  | 50000 | 2.2954          |
+
+
+### Framework versions
+
+- Transformers 4.19.0.dev0
+- Pytorch 1.11.0+cu102
+- Datasets 2.0.0
+- Tokenizers 0.11.6
+
+
+##Citing
+If you use HAT in your research, please cite [An Exploration of Hierarchical Attention Transformers for Efficient Long Document Classification](https://arxiv.org/abs/xxx)
+
+```
+@misc{chalkidis-etal-2022-hat,
+  url = {https://arxiv.org/abs/xxx},
+  author = {Chalkidis, Ilias and Dai, Xiang and Fergadiotis, Manos and Malakasiotis, Prodromos and Elliott, Desmond},
+  title = {An Exploration of Hierarchical Attention Transformers for Efficient Long Document Classification},
+  publisher = {arXiv},
+  year = {2022},
+}
+```
+
+

all_results.json

@@ -0,0 +1,12 @@
+{
+    "epoch": 1.14,
+    "eval_loss": 2.2956442832946777,
+    "eval_runtime": 3135.1617,
+    "eval_samples_per_second": 159.481,
+    "eval_steps_per_second": 4.984,
+    "perplexity": 9.930832220881081,
+    "train_loss": 2.669982685546875,
+    "train_runtime": 65951.2344,
+    "train_samples_per_second": 97.041,
+    "train_steps_per_second": 0.758
+}

config.json

@@ -0,0 +1,79 @@
+{
+  "_name_or_path": "kiddothe2b/hat-mini-1024-I2",
+  "architectures": [
+    "HATForMaskedLM"
+  ],
+  "auto_map": {
+     "AutoConfig": "configuration_hat.HATConfig",
+     "AutoTokenizer": "tokenization_hat.HATTokenizer",
+     "AutoModel": "modelling_hat.HATModel",
+     "AutoModelForMaskedLM": "modelling_hat.HATForMaskedLM",
+     "AutoModelForMultipleChoice": "modelling_hat.HATForMultipleChoice",
+     "AutoModelForQuestionAnswering": "modelling_hat.HATForQuestionAnswering",
+     "AutoModelForSequenceClassification": "modelling_hat.HATForSequenceClassification",
+     "AutoModelForTokenClassification": "modelling_hat.HATForTokenClassification"
+   },
+  "attention_probs_dropout_prob": 0.1,
+  "classifier_dropout": null,
+  "encoder_layout": {
+    "0": {
+      "document_encoder": false,
+      "sentence_encoder": true
+    },
+    "1": {
+      "document_encoder": true,
+      "sentence_encoder": true
+    },
+    "2": {
+      "document_encoder": true,
+      "sentence_encoder": false
+    },
+    "3": {
+      "document_encoder": false,
+      "sentence_encoder": true
+    },
+    "4": {
+      "document_encoder": true,
+      "sentence_encoder": true
+    },
+    "5": {
+      "document_encoder": true,
+      "sentence_encoder": false
+    },
+    "6": {
+      "document_encoder": false,
+      "sentence_encoder": true
+    },
+    "7": {
+      "document_encoder": true,
+      "sentence_encoder": true
+    },
+    "8": {
+      "document_encoder": true,
+      "sentence_encoder": false
+    }
+  },
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 256,
+  "initializer_range": 0.02,
+  "intermediate_size": 1024,
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 128,
+  "max_sentence_length": 128,
+  "max_sentence_size": 128,
+  "max_sentences": 8,
+  "model_max_length": 1024,
+  "model_type": "hierarchical-transformer",
+  "num_attention_heads": 4,
+  "num_hidden_layers": 9,
+  "output_past": true,
+  "pad_token_id": 0,
+  "parameters": 136350720,
+  "position_embedding_type": "absolute",
+  "torch_dtype": "float32",
+  "transformers_version": "4.19.0.dev0",
+  "type_vocab_size": 2,
+  "use_cache": true,
+  "vocab_size": 30522
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9846499976ce5546ddcadf89828a6d5199eb731aba25a56fba46bd702958e5a7
+size 101198747

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 @@
+{"do_lower_case": true, "unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]", "tokenize_chinese_chars": true, "strip_accents": null, "model_max_length": 1024, "special_tokens_map_file": null, "name_or_path": "data/PLMs/hi-transformer-s2-grouped", "do_basic_tokenize": true, "never_split": null, "tokenizer_class": "BertTokenizer", "auto_map":  {"AutoTokenizer": ["tokenization_hat.HATTokenizer", "tokenization_hat.HATTokenizer"]}}