yuq-1s / BARTNER

Towards parameter-efficient tuning of large models for NER

Currently, this repo targets T5 for NER. Current results on CoNLL 2003:

model	#param	p	r	f
bart-large (paper)	406M	92.61	93.87	93.24
t5-large	738M	91.4	92.05	91.73
t5-11b	11.3B	92.46	92.33	92.40

TODO

• Add MLP, set it as trainable
• Debug T5: SequenceGenerator always 0 f1
• Debug T5: >200 loss on initialization
• Finetune t5-base and achieve comparable performance with bart-large
• Finetune t5-11b and achieve better performance than bart-large
• Add LM adaption for T5
• Tuning only MLP after encoder_out and embed_tokens on t5-base
• Tuning only MLP after encoder_out and a soft prompt

Install

To run this version with model parallel, some modifications to the following libraries are needed.

• fastNLP==0.6.0

--- /tmp/trainer_orig.py        2021-07-10 03:57:06.348091151 +0000
+++ .venv/lib/python3.9/site-packages/fastNLP/core/trainer.py       2021-07-10 03:58:34.512642014 +0000
@@ -500,7 +500,11 @@
             raise TypeError("train_data type {} not support".format(type(train_data)))
 
         model.train()
-        self.model = _move_model_to_device(model, device=device)
+        some_param = next(iter(model.parameters()))
+        if some_param.device.type == 'cpu':
+            self.model = _move_model_to_device(model, device=device)
+        else:
+            self.model = model
         if _model_contains_inner_module(self.model):
             self._forward_func = self.model.module.forward
         else:
@@ -789,7 +793,7 @@
         """
         return self.losser(predict, truth)
 
-    def _save_model(self, model, model_name, only_param=False):
+    def _save_model(self, model, model_name, only_param=True):
         r""" 存储不含有显卡信息的state_dict或model
         :param model:
         :param model_name:
--- /tmp/tester_orig.py 2021-07-11 05:47:02.723080056 +0000
+++ .venv/lib/python3.9/site-packages/fastNLP/core/tester.py  2021-07-11 05:47:07.457005462 +0000
@@ -103,7 +103,11 @@
         self.metrics = _prepare_metrics(metrics)
         
         self.data = data
-        self._model = _move_model_to_device(model, device=device)
+        some_param = next(iter(model.parameters()))
+        if some_param.device.type == 'cpu':
+            self._model = _move_model_to_device(model, device=device)
+        else:
+            self._model = model
         self.batch_size = batch_size
         self.verbose = verbose
         self.use_tqdm = use_tqdm

Otherwise fastNLP will load all parameters into cuda:0, invalidating model parallel.

• transformers==4.7.0

--- /tmp/modeling_t5_orig.py    2021-07-10 10:32:13.078207891 +0800
+++ .venv/lib/python3.9/site-packages/transformers/models/t5/modeling_t5.py        2021-07-09 16:46:14.836178715 +0800
@@ -244,6 +244,8 @@
         # convert into float16 if necessary
         if self.weight.dtype == torch.float16:
             hidden_states = hidden_states.to(torch.float16)
+        if hidden_states.device != self.weight.device:
+            self.to(hidden_states.device)
         return self.weight * hidden_states
@@ -993,6 +996,8 @@
                     None,  # past_key_value is always None with gradient checkpointing
                 )
             else:
+                if self.model_parallel:
+                    layer_module.to(hidden_states.device)
                 layer_outputs = layer_module(
                     hidden_states,
                     attention_mask=extended_attention_mask,

Otherwise torch complains incompatible cuda devices with model parallel.

Original README

This is the code for ACL-ICJNLP2021 paper A Unified Generative Framework for Various NER Subtasks.

Install the package in the requirements.txt, then use the following commands to install two other packages

pip install git+https://github.com/fastnlp/fastNLP@dev
pip install git+https://github.com/fastnlp/fitlog

You need to put your data in the parallel folder of this repo

    - BARTNER/
        - train.py
        ...
    - data/
        - conll2003
            - train.txt
            - text.txt
            - dev.txt
        - en-ontonotes
            - ...
        - Share_2013
        - Share_2014
        - CADEC
        - en_ace04
        - en_ace05
        - genia

For the conll2003 and en-ontonotes you data in each split should like (The first column is words, the second column is tags. We assume the tag is the BIO-tagging)

LONDON B-LOC
1996-08-30 O

West B-MISC
Indian I-MISC
all-rounder O
Phil B-PER

For nested dataset en_ace04, en_ace05 and genia, the data should like (each line is a jsonline, contains ners and sentences keys.)

{"ners": [[[16, 16, "DNA"], [4, 8, "DNA"], [24, 26, "DNA"], [19, 20, "DNA"]], [[31, 31, "DNA"], [2, 2, "DNA"], [4, 4, "DNA"], [30, 31, "DNA"]], [[23, 24, "RNA"], [14, 15, "cell_type"], [1, 2, "RNA"]], [[2, 2, "DNA"]], [], [[0, 0, "DNA"], [9, 9, "cell_type"]]], "sentences": [["There", "is", "a", "single", "methionine", "codon-initiated", "open", "reading", "frame", "of", "1,458", "nt", "in", "frame", "with", "a", "homeobox", "and", "a", "CAX", "repeat", ",", "and", "the", "open", "reading", "frame", "is", "predicted", "to", "encode", "a", "protein", "of", "51,659", "daltons."], ["When", "the", "homeodomain", "from", "HB24", "was", "compared", "to", "known", "mammalian", "and", "Drosophila", "homeodomains", "it", "was", "found", "to", "be", "only", "moderately", "conserved,", "but", "when", "it", "was", "compared", "to", "a", "highly", "diverged", "Drosophila", "homeodomain", ",", "H2.0,", "it", "was", "found", "to", "be", "80%", "identical."], ["The", "HB24", "mRNA", "was", "absent", "or", "present", "at", "low", "levels", "in", "normal", "B", "and", "T", "lymphocytes", ";", "however,", "with", "the", "appropriate", "activation", "signal", "HB24", "mRNA", "was", "induced", "within", "several", "hours", "even", "in", "the", "presence", "of", "cycloheximide", "."], ["Characterization", "of", "HB24", "expression", "in", "lymphoid", "and", "select", "developing", "tissues", "was", "performed", "by", "in", "situ", "hybridization", "."], ["Positive", "hybridization", "was", "found", "in", "thymus", ",", "tonsil", ",", "bone", "marrow", ",", "developing", "vessels", ",", "and", "in", "fetal", "brain", "."], ["HB24", "is", "likely", "to", "have", "an", "important", "role", "in", "lymphocytes", "as", "well", "as", "in", "certain", "developing", "tissues", "."]]}
{"ners": [[[16, 16, "DNA"], [4, 8, "DNA"], [24, 26, "DNA"], [19, 20, "DNA"]], [[31, 31, "DNA"], [2, 2, "DNA"], [4, 4, "DNA"], [30, 31, "DNA"]], [[23, 24, "RNA"], [14, 15, "cell_type"], [1, 2, "RNA"]], [[2, 2, "DNA"]], [], [[0, 0, "DNA"], [9, 9, "cell_type"]]], "sentences": [["There", "is", "a", "single", "methionine", "codon-initiated", "open", "reading", "frame", "of", "1,458", "nt", "in", "frame", "with", "a", "homeobox", "and", "a", "CAX", "repeat", ",", "and", "the", "open", "reading", "frame", "is", "predicted", "to", "encode", "a", "protein", "of", "51,659", "daltons."], ["When", "the", "homeodomain", "from", "HB24", "was", "compared", "to", "known", "mammalian", "and", "Drosophila", "homeodomains", "it", "was", "found", "to", "be", "only", "moderately", "conserved,", "but", "when", "it", "was", "compared", "to", "a", "highly", "diverged", "Drosophila", "homeodomain", ",", "H2.0,", "it", "was", "found", "to", "be", "80%", "identical."], ["The", "HB24", "mRNA", "was", "absent", "or", "present", "at", "low", "levels", "in", "normal", "B", "and", "T", "lymphocytes", ";", "however,", "with", "the", "appropriate", "activation", "signal", "HB24", "mRNA", "was", "induced", "within", "several", "hours", "even", "in", "the", "presence", "of", "cycloheximide", "."], ["Characterization", "of", "HB24", "expression", "in", "lymphoid", "and", "select", "developing", "tissues", "was", "performed", "by", "in", "situ", "hybridization", "."], ["Positive", "hybridization", "was", "found", "in", "thymus", ",", "tonsil", ",", "bone", "marrow", ",", "developing", "vessels", ",", "and", "in", "fetal", "brain", "."], ["HB24", "is", "likely", "to", "have", "an", "important", "role", "in", "lymphocytes", "as", "well", "as", "in", "certain", "developing", "tissues", "."]]}
...

For discontinuous dataset Share_2013, Share_2014 and CADEC, the data should like ( each sample has two lines, if the second line is empty means there is not entity. )

Abdominal cramps , flatulence , gas , bloating .
0,1 ADR|3,3 ADR|7,7 ADR|5,5 ADR

Cramps would start within 15 minutes of taking pill , even during meals .
0,0 ADR

...

We use code from https://github.com/daixiangau/acl2020-transition-discontinuous-ner to pre-process the data.

You can run the code by directly using

python train.py

The following output should be achieved

Save cache to caches/data_facebook/bart-large_conll2003_word.pt.                                                                                                        
max_len_a:0.6, max_len:10
In total 3 datasets:
        test has 3453 instances.
        train has 14041 instances.
        dev has 3250 instances.

The number of tokens in tokenizer  50265
50269 50274
input fields after batch(if batch size is 2):
        tgt_tokens: (1)type:torch.Tensor (2)dtype:torch.int64, (3)shape:torch.Size([2, 8]) 
        src_tokens: (1)type:torch.Tensor (2)dtype:torch.int64, (3)shape:torch.Size([2, 11]) 
        first: (1)type:torch.Tensor (2)dtype:torch.int64, (3)shape:torch.Size([2, 11]) 
        src_seq_len: (1)type:torch.Tensor (2)dtype:torch.int64, (3)shape:torch.Size([2]) 
        tgt_seq_len: (1)type:torch.Tensor (2)dtype:torch.int64, (3)shape:torch.Size([2]) 
target fields after batch(if batch size is 2):
        entities: (1)type:numpy.ndarray (2)dtype:object, (3)shape:(2,) 
        tgt_tokens: (1)type:torch.Tensor (2)dtype:torch.int64, (3)shape:torch.Size([2, 8]) 
        target_span: (1)type:numpy.ndarray (2)dtype:object, (3)shape:(2,) 
        tgt_seq_len: (1)type:torch.Tensor (2)dtype:torch.int64, (3)shape:torch.Size([2]) 

training epochs started 2021-06-02-11-49-26-964889
Epoch 1/30:   0%|                                                         | 15/32430 [00:06<3:12:37,  2.80it/s, loss:6.96158

Some important python files are listed below

- BartNER
  - data
     - pipe.py # load and process data
  - model
     - bart.py # the model file
  - train.py  # the training file

The different Loaders in the data/pipe.py is meant to load data, and the data.BartNERPipe class is to process data, the loader should load data into a DataBundle object, you can mock the provided Loader to write your own loader, as long as your dataset has the following four fields, the BartNERPipe should be able to process it

- raw_words  # List[str]
    # ['AL-AIN', ',', 'United', 'Arab', 'Emirates', '1996-12-06']
- entities  # List[List[str]]
    # [['AL-AIN'], ['United', 'Arab', 'Emirates']]
- entity_tags  # List[str], the same length as entities
    # ['loc', 'loc']
- entity_spans # List[List[int]], the inner list must have an even number of ints, means the start(inclusive，开区间) and end(exclusive，开区间) of an entity segment
    # [[0, 1], [2, 5]] or for discontinous NER [[0, 1, 5, 7], [2, 3, 5, 7],...]

In order to help you reproduce the results, we have hardcoded the hyper-parameters for each dataset in the code, you can change them based on your need. We conduct all experiments in NVIDIA-3090(24G memory). Some known difficulties about the reproduction of this code: (1) Some datasets (nested and discontinous) will drop to 0 or near 0 F1 during training, please drop these results; (2) randomness will cause large performance variance for some datasets, please try to run multiple times.

We deeply understand how frustrating it can be if the results are hard to reproduce, we tried our best to make sure the results were at least reproducible in our equipment (Usually take average from at least five runs).