initial release

README.md

@@ -0,0 +1,30 @@
+---
+language:
+- "ja"
+tags:
+- "japanese"
+- "token-classification"
+- "pos"
+base_model: HPLT/hplt_bert_base_ja
+datasets:
+- "universal_dependencies"
+license: "apache-2.0"
+pipeline_tag: "token-classification"
+widget:
+- text: "国境の長いトンネルを抜けると雪国であった。"
+---
+
+# ltgbert-base-japanese-upos
+
+## Model Description
+
+This is a LTG-BERT model for POS-tagging, derived from [hplt_bert_base_ja](https://huggingface.co/HPLT/hplt_bert_base_ja). Every short-unit-word is tagged by [UPOS](https://universaldependencies.org/u/pos/) (Universal Part-Of-Speech) and [FEATS](https://universaldependencies.org/u/feat/).
+
+## How to Use
+
+```py
+from transformers import pipeline
+nlp=pipeline("upos","KoichiYasuoka/ltgbert-base-japanese-upos",trust_remote_code=True,aggregation_strategy="simple")
+print(nlp("国境の長いトンネルを抜けると雪国であった。"))
+```
+

config.json

@@ -0,0 +1,159 @@
+{
+  "architectures": [
+    "LtgbertForTokenClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "auto_map": {
+    "AutoConfig": "configuration_ltgbert.LtgbertConfig",
+    "AutoModel": "modeling_ltgbert.LtgbertModel",
+    "AutoModelForMaskedLM": "modeling_ltgbert.LtgbertForMaskedLM",
+    "AutoModelForMultipleChoice": "modeling_ltgbert.LtgbertForMultipleChoice",
+    "AutoModelForQuestionAnswering": "modeling_ltgbert.LtgbertForQuestionAnswering",
+    "AutoModelForSequenceClassification": "modeling_ltgbert.LtgbertForSequenceClassification",
+    "AutoModelForTokenClassification": "modeling_ltgbert.LtgbertForTokenClassification"
+  },
+  "classifier_dropout": 0.1,
+  "custom_pipelines": {
+    "upos": {
+      "impl": "upos.BellmanFordTokenClassificationPipeline",
+      "pt": "AutoModelForTokenClassification"
+    }
+  },
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": "ADJ",
+    "1": "B-ADJ",
+    "2": "I-ADJ",
+    "3": "ADJ|Polarity=Neg",
+    "4": "B-ADJ|Polarity=Neg",
+    "5": "I-ADJ|Polarity=Neg",
+    "6": "ADP",
+    "7": "B-ADP",
+    "8": "I-ADP",
+    "9": "ADV",
+    "10": "B-ADV",
+    "11": "I-ADV",
+    "12": "AUX",
+    "13": "B-AUX",
+    "14": "I-AUX",
+    "15": "AUX|Polarity=Neg",
+    "16": "B-AUX|Polarity=Neg",
+    "17": "I-AUX|Polarity=Neg",
+    "18": "CCONJ",
+    "19": "B-CCONJ",
+    "20": "I-CCONJ",
+    "21": "DET",
+    "22": "B-DET",
+    "23": "I-DET",
+    "24": "INTJ",
+    "25": "B-INTJ",
+    "26": "I-INTJ",
+    "27": "NOUN",
+    "28": "B-NOUN",
+    "29": "I-NOUN",
+    "30": "NOUN|Polarity=Neg",
+    "31": "B-NOUN|Polarity=Neg",
+    "32": "I-NOUN|Polarity=Neg",
+    "33": "NUM",
+    "34": "B-NUM",
+    "35": "I-NUM",
+    "36": "PART",
+    "37": "B-PART",
+    "38": "I-PART",
+    "39": "PRON",
+    "40": "B-PRON",
+    "41": "I-PRON",
+    "42": "PROPN",
+    "43": "B-PROPN",
+    "44": "I-PROPN",
+    "45": "PUNCT",
+    "46": "B-PUNCT",
+    "47": "I-PUNCT",
+    "48": "SCONJ",
+    "49": "B-SCONJ",
+    "50": "I-SCONJ",
+    "51": "SYM",
+    "52": "B-SYM",
+    "53": "I-SYM",
+    "54": "VERB",
+    "55": "B-VERB",
+    "56": "I-VERB",
+    "57": "X",
+    "58": "B-X",
+    "59": "I-X"
+  },
+  "intermediate_size": 2560,
+  "label2id": {
+    "ADJ": 0,
+    "ADJ|Polarity=Neg": 3,
+    "ADP": 6,
+    "ADV": 9,
+    "AUX": 12,
+    "AUX|Polarity=Neg": 15,
+    "B-ADJ": 1,
+    "B-ADJ|Polarity=Neg": 4,
+    "B-ADP": 7,
+    "B-ADV": 10,
+    "B-AUX": 13,
+    "B-AUX|Polarity=Neg": 16,
+    "B-CCONJ": 19,
+    "B-DET": 22,
+    "B-INTJ": 25,
+    "B-NOUN": 28,
+    "B-NOUN|Polarity=Neg": 31,
+    "B-NUM": 34,
+    "B-PART": 37,
+    "B-PRON": 40,
+    "B-PROPN": 43,
+    "B-PUNCT": 46,
+    "B-SCONJ": 49,
+    "B-SYM": 52,
+    "B-VERB": 55,
+    "B-X": 58,
+    "CCONJ": 18,
+    "DET": 21,
+    "I-ADJ": 2,
+    "I-ADJ|Polarity=Neg": 5,
+    "I-ADP": 8,
+    "I-ADV": 11,
+    "I-AUX": 14,
+    "I-AUX|Polarity=Neg": 17,
+    "I-CCONJ": 20,
+    "I-DET": 23,
+    "I-INTJ": 26,
+    "I-NOUN": 29,
+    "I-NOUN|Polarity=Neg": 32,
+    "I-NUM": 35,
+    "I-PART": 38,
+    "I-PRON": 41,
+    "I-PROPN": 44,
+    "I-PUNCT": 47,
+    "I-SCONJ": 50,
+    "I-SYM": 53,
+    "I-VERB": 56,
+    "I-X": 59,
+    "INTJ": 24,
+    "NOUN": 27,
+    "NOUN|Polarity=Neg": 30,
+    "NUM": 33,
+    "PART": 36,
+    "PRON": 39,
+    "PROPN": 42,
+    "PUNCT": 45,
+    "SCONJ": 48,
+    "SYM": 51,
+    "VERB": 54,
+    "X": 57
+  },
+  "layer_norm_eps": 1e-07,
+  "max_position_embeddings": 512,
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "output_all_encoded_layers": true,
+  "position_bucket_size": 32,
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "torch_dtype": "float32",
+  "transformers_version": "4.44.2",
+  "vocab_size": 32768
+}

configuration_ltgbert.py

@@ -0,0 +1,34 @@
+from transformers.configuration_utils import PretrainedConfig
+
+
+class LtgbertConfig(PretrainedConfig):
+    """Configuration class to store the configuration of a `LtgbertModel`.
+    """
+    def __init__(
+        self,
+        vocab_size=32768,
+        attention_probs_dropout_prob=0.1,
+        hidden_dropout_prob=0.1,
+        hidden_size=768,
+        intermediate_size=2048,
+        max_position_embeddings=512,
+        position_bucket_size=32,
+        num_attention_heads=12,
+        num_hidden_layers=12,
+        layer_norm_eps=1.0e-7,
+        output_all_encoded_layers=True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.output_all_encoded_layers = output_all_encoded_layers
+        self.position_bucket_size = position_bucket_size
+        self.layer_norm_eps = layer_norm_eps

maker.py

@@ -0,0 +1,69 @@
+#! /usr/bin/python3
+src="hplt_bert_base_ja"
+tgt="KoichiYasuoka/ltgbert-base-japanese-upos"
+
+import os
+from transformers import AutoTokenizer,AutoConfig,AutoModelForTokenClassification,DataCollatorForTokenClassification,TrainingArguments,Trainer
+os.system(f"test -d {src} || ( curl -L https://data.hplt-project.org/one/models/encoder/{src}.tar.gz | tar xvzf - )")
+os.system("test -f ja_gsd_modern.conllu || curl -LO https://github.com/KoichiYasuoka/SuPar-UniDic/raw/main/suparunidic/suparmodels/ja_gsd_modern.conllu")
+
+class UPOSFileDataset(object):
+  def __init__(self,conllu,tokenizer):
+    self.conllu=open(conllu,"r",encoding="utf-8")
+    self.tokenizer=tokenizer
+    self.seeks=[0]
+    label=set(["SYM"])
+    s=self.conllu.readline()
+    while s!="":
+      if s=="\n":
+        self.seeks.append(self.conllu.tell())
+      else:
+        w=s.split("\t")
+        if len(w)==10:
+          if w[0].isdecimal():
+            label.add(w[3] if w[5]=="_" else w[3]+"|"+w[5])
+      s=self.conllu.readline()
+    lid={}
+    for i,l in enumerate(sorted(label)):
+      lid[l],lid["B-"+l],lid["I-"+l]=i*3,i*3+1,i*3+2
+    self.label2id=lid
+  def __call__(*args):
+    lid={l:i for i,l in enumerate(sorted(set(sum([list(t.label2id) for t in args],[]))))}
+    for t in args:
+      t.label2id=lid
+    return lid
+  def __del__(self):
+    self.conllu.close()
+  __len__=lambda self:len(self.seeks)-1
+  def __getitem__(self,i):
+    self.conllu.seek(self.seeks[i])
+    form,upos=[],[]
+    while self.conllu.tell()<self.seeks[i+1]:
+      w=self.conllu.readline().split("\t")
+      if len(w)==10:
+        form.append(w[1])
+        if w[0].isdecimal():
+          upos.append(w[3] if w[5]=="_" else w[3]+"|"+w[5])
+    v=self.tokenizer(form,add_special_tokens=False)
+    i,u=[],[]
+    for j,(x,y) in enumerate(zip(v["input_ids"],upos)):
+      if x!=[]:
+        i+=x
+        u+=[y] if len(x)==1 else ["B-"+y]+["I-"+y]*(len(x)-1)
+    if len(i)<self.tokenizer.model_max_length-3:
+      ids=[self.tokenizer.cls_token_id]+i+[self.tokenizer.sep_token_id]
+      upos=["SYM"]+u+["SYM"]
+    else:
+      ids=i[0:self.tokenizer.model_max_length-2]
+      upos=u[0:self.tokenizer.model_max_length-2]
+    return {"input_ids":ids,"labels":[self.label2id[t] for t in upos]}
+
+tkz=AutoTokenizer.from_pretrained(src,model_max_length=512)
+trainDS=UPOSFileDataset("ja_gsd_modern.conllu",tkz)
+lid=trainDS.label2id
+cfg=AutoConfig.from_pretrained(src,num_labels=len(lid),label2id=lid,id2label={i:l for l,i in lid.items()},ignore_mismatched_sizes=True,trust_remote_code=True)
+arg=TrainingArguments(num_train_epochs=3,per_device_train_batch_size=8,output_dir=tgt,overwrite_output_dir=True,save_total_limit=2,learning_rate=5e-05,warmup_ratio=0.1,save_safetensors=False)
+trn=Trainer(args=arg,data_collator=DataCollatorForTokenClassification(tkz),model=AutoModelForTokenClassification.from_pretrained(src,config=cfg,ignore_mismatched_sizes=True,trust_remote_code=True),train_dataset=trainDS)
+trn.train()
+trn.save_model(tgt)
+tkz.save_pretrained(tgt)

modeling_ltgbert.py

@@ -0,0 +1,639 @@
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.utils import checkpoint
+
+from .configuration_ltgbert import LtgbertConfig
+from transformers.modeling_utils import PreTrainedModel
+from transformers.activations import gelu_new
+from transformers.modeling_outputs import (
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+    BaseModelOutput
+)
+from transformers.pytorch_utils import softmax_backward_data
+
+
+class Encoder(nn.Module):
+    def __init__(self, config, activation_checkpointing=False):
+        super().__init__()
+        self.layers = nn.ModuleList([EncoderLayer(config) for _ in range(config.num_hidden_layers)])
+
+        for i, layer in enumerate(self.layers):
+            layer.mlp.mlp[1].weight.data *= math.sqrt(1.0 / (2.0 * (1 + i)))
+            layer.mlp.mlp[-2].weight.data *= math.sqrt(1.0 / (2.0 * (1 + i)))
+
+        self.activation_checkpointing = activation_checkpointing
+    
+    def forward(self, hidden_states, attention_mask, relative_embedding):
+        hidden_states, attention_probs = [hidden_states], []
+
+        for layer in self.layers:
+            if self.activation_checkpointing:
+                hidden_state, attention_p = checkpoint.checkpoint(layer, hidden_states[-1], attention_mask, relative_embedding)
+            else:
+                hidden_state, attention_p = layer(hidden_states[-1], attention_mask, relative_embedding)
+
+            hidden_states.append(hidden_state)
+            attention_probs.append(attention_p)
+
+        return hidden_states, attention_probs
+
+
+class MaskClassifier(nn.Module):
+    def __init__(self, config, subword_embedding):
+        super().__init__()
+        self.nonlinearity = nn.Sequential(
+            nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=False),
+            nn.Linear(config.hidden_size, config.hidden_size),
+            nn.GELU(),
+            nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=False),
+            nn.Dropout(config.hidden_dropout_prob),
+            nn.Linear(subword_embedding.size(1), subword_embedding.size(0))
+        )
+
+    def forward(self, x, masked_lm_labels=None):
+        if masked_lm_labels is not None:
+            x = torch.index_select(x.flatten(0, 1), 0, torch.nonzero(masked_lm_labels.flatten() != -100).squeeze())
+        x = self.nonlinearity(x)
+        return x
+
+
+class EncoderLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.attention = Attention(config)
+        self.mlp = FeedForward(config)
+
+    def forward(self, x, padding_mask, relative_embedding):
+        attention_output, attention_probs = self.attention(x, padding_mask, relative_embedding)
+        x = x + attention_output
+        x = x + self.mlp(x)
+        return x, attention_probs
+
+
+class GeGLU(nn.Module):
+    def forward(self, x):
+        x, gate = x.chunk(2, dim=-1)
+        x = x * gelu_new(gate)
+        return x
+
+
+class FeedForward(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, elementwise_affine=False),
+            nn.Linear(config.hidden_size, 2*config.intermediate_size, bias=False),
+            GeGLU(),
+            nn.LayerNorm(config.intermediate_size, eps=config.layer_norm_eps, elementwise_affine=False),
+            nn.Linear(config.intermediate_size, config.hidden_size, bias=False),
+            nn.Dropout(config.hidden_dropout_prob)
+        )
+
+    def forward(self, x):
+        return self.mlp(x)
+
+
+class MaskedSoftmax(torch.autograd.Function):
+    @staticmethod
+    def forward(self, x, mask, dim):
+        self.dim = dim
+        x.masked_fill_(mask, float('-inf'))
+        x = torch.softmax(x, self.dim)
+        x.masked_fill_(mask, 0.0)
+        self.save_for_backward(x)
+        return x
+
+    @staticmethod
+    def backward(self, grad_output):
+        output, = self.saved_tensors
+        input_grad = softmax_backward_data(self, grad_output, output, self.dim, output)
+        return input_grad, None, None
+
+
+class Attention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+
+        self.config = config
+
+        if config.hidden_size % config.num_attention_heads != 0:
+            raise ValueError(f"The hidden size {config.hidden_size} is not a multiple of the number of attention heads {config.num_attention_heads}")
+
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_size = config.hidden_size // config.num_attention_heads
+
+        self.in_proj_qk = nn.Linear(config.hidden_size, 2*config.hidden_size, bias=True)
+        self.in_proj_v = nn.Linear(config.hidden_size, config.hidden_size, bias=True)
+        self.out_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=True)
+
+        self.pre_layer_norm = nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=False)
+        self.post_layer_norm = nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=True)
+
+        position_indices = torch.arange(config.max_position_embeddings, dtype=torch.long).unsqueeze(1) \
+            - torch.arange(config.max_position_embeddings, dtype=torch.long).unsqueeze(0)
+        position_indices = self.make_log_bucket_position(position_indices, config.position_bucket_size, config.max_position_embeddings)
+        position_indices = config.position_bucket_size - 1 + position_indices
+        self.register_buffer("position_indices", position_indices, persistent=True)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.scale = 1.0 / math.sqrt(3 * self.head_size)
+
+    def make_log_bucket_position(self, relative_pos, bucket_size, max_position):
+        sign = torch.sign(relative_pos)
+        mid = bucket_size // 2
+        abs_pos = torch.where((relative_pos < mid) & (relative_pos > -mid), mid - 1, torch.abs(relative_pos).clamp(max=max_position - 1))
+        log_pos = torch.ceil(torch.log(abs_pos / mid) / math.log((max_position-1) / mid) * (mid - 1)).int() + mid
+        bucket_pos = torch.where(abs_pos <= mid, relative_pos, log_pos * sign).long()
+        return bucket_pos
+
+    def compute_attention_scores(self, hidden_states, relative_embedding):
+        key_len, batch_size, _ = hidden_states.size()
+        query_len = key_len
+
+        if self.position_indices.size(0) < query_len:
+            position_indices = torch.arange(query_len, dtype=torch.long).unsqueeze(1) \
+                - torch.arange(query_len, dtype=torch.long).unsqueeze(0)
+            position_indices = self.make_log_bucket_position(position_indices, self.config.position_bucket_size, 512)
+            position_indices = self.config.position_bucket_size - 1 + position_indices
+            self.position_indices = position_indices.to(hidden_states.device)
+
+        hidden_states = self.pre_layer_norm(hidden_states)
+
+        query, key = self.in_proj_qk(hidden_states).chunk(2, dim=2)  # shape: [T, B, D]
+        value = self.in_proj_v(hidden_states)  # shape: [T, B, D]
+
+        query = query.reshape(query_len, batch_size * self.num_heads, self.head_size).transpose(0, 1)
+        key = key.reshape(key_len, batch_size * self.num_heads, self.head_size).transpose(0, 1)
+        value = value.view(key_len, batch_size * self.num_heads, self.head_size).transpose(0, 1)
+
+        attention_scores = torch.bmm(query, key.transpose(1, 2) * self.scale)
+
+        query_pos, key_pos = self.in_proj_qk(self.dropout(relative_embedding)).chunk(2, dim=-1)  # shape: [2T-1, D]
+        query_pos = query_pos.view(-1, self.num_heads, self.head_size)  # shape: [2T-1, H, D]
+        key_pos = key_pos.view(-1, self.num_heads, self.head_size)  # shape: [2T-1, H, D]
+
+        query = query.view(batch_size, self.num_heads, query_len, self.head_size)
+        key = key.view(batch_size, self.num_heads, query_len, self.head_size)
+
+        attention_c_p = torch.einsum("bhqd,khd->bhqk", query, key_pos.squeeze(1) * self.scale)
+        attention_p_c = torch.einsum("bhkd,qhd->bhqk", key * self.scale, query_pos.squeeze(1))
+
+        position_indices = self.position_indices[:query_len, :key_len].expand(batch_size, self.num_heads, -1, -1)
+        attention_c_p = attention_c_p.gather(3, position_indices)
+        attention_p_c = attention_p_c.gather(2, position_indices)
+
+        attention_scores = attention_scores.view(batch_size, self.num_heads, query_len, key_len)
+        attention_scores.add_(attention_c_p)
+        attention_scores.add_(attention_p_c)
+
+        return attention_scores, value
+
+    def compute_output(self, attention_probs, value):
+        attention_probs = self.dropout(attention_probs)
+        context = torch.bmm(attention_probs.flatten(0, 1), value)  # shape: [B*H, Q, D]
+        context = context.transpose(0, 1).reshape(context.size(1), -1, self.hidden_size)  # shape: [Q, B, H*D]
+        context = self.out_proj(context)
+        context = self.post_layer_norm(context)
+        context = self.dropout(context)
+        return context
+
+    def forward(self, hidden_states, attention_mask, relative_embedding):
+        attention_scores, value = self.compute_attention_scores(hidden_states, relative_embedding)
+        attention_probs = MaskedSoftmax.apply(attention_scores, attention_mask, -1)
+        return self.compute_output(attention_probs, value), attention_probs.detach()
+
+
+class Embedding(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.word_embedding = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.word_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, elementwise_affine=False)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        self.relative_embedding = nn.Parameter(torch.empty(2 * config.position_bucket_size - 1, config.hidden_size))
+        self.relative_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, input_ids):
+        word_embedding = self.dropout(self.word_layer_norm(self.word_embedding(input_ids)))
+        relative_embeddings = self.relative_layer_norm(self.relative_embedding)
+        return word_embedding, relative_embeddings
+
+
+#
+# HuggingFace wrappers
+#
+
+class LtgbertPreTrainedModel(PreTrainedModel):
+    config_class = LtgbertConfig
+    supports_gradient_checkpointing = True
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, Encoder):
+            module.activation_checkpointing = value
+
+    def _init_weights(self, module):
+        std = math.sqrt(2.0 / (5.0 * self.hidden_size))
+
+        if isinstance(module, nn.Linear):
+            nn.init.trunc_normal_(module.weight.data, mean=0.0, std=std, a=-2*std, b=2*std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            nn.init.trunc_normal_(module.weight.data, mean=0.0, std=std, a=-2*std, b=2*std)
+        elif isinstance(module, nn.LayerNorm):
+            if module.bias is not None:
+                module.bias.data.zero_()
+            if module.weight is not None:
+                module.weight.data.fill_(1.0)
+
+
+class LtgbertModel(LtgbertPreTrainedModel):
+    def __init__(self, config, add_mlm_layer=False, gradient_checkpointing=False, **kwargs):
+        super().__init__(config, **kwargs)
+        self.config = config
+        self.hidden_size = config.hidden_size
+
+        self.embedding = Embedding(config)
+        self.transformer = Encoder(config, activation_checkpointing=gradient_checkpointing)
+        self.classifier = MaskClassifier(config, self.embedding.word_embedding.weight) if add_mlm_layer else None
+
+
+    def get_input_embeddings(self):
+        return self.embedding.word_embedding
+
+    def set_input_embeddings(self, value):
+        self.embedding.word_embedding = value
+
+    def get_contextualized_embeddings(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None
+    ) -> List[torch.Tensor]:
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            raise ValueError("You have to specify input_ids")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device
+
+        if attention_mask is None:
+            attention_mask = torch.zeros(batch_size, seq_length, dtype=torch.bool, device=device)
+        else:
+            attention_mask = ~attention_mask.bool()
+        attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+ 
+        static_embeddings, relative_embedding = self.embedding(input_ids.t())
+        contextualized_embeddings, attention_probs = self.transformer(static_embeddings, attention_mask, relative_embedding)
+        contextualized_embeddings = [e.transpose(0, 1) for e in contextualized_embeddings]
+        last_layer = contextualized_embeddings[-1]
+        contextualized_embeddings = [contextualized_embeddings[0]] + [
+            contextualized_embeddings[i] - contextualized_embeddings[i - 1]
+            for i in range(1, len(contextualized_embeddings))
+        ]
+        return last_layer, contextualized_embeddings, attention_probs
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutput]:
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        sequence_output, contextualized_embeddings, attention_probs = self.get_contextualized_embeddings(input_ids, attention_mask)
+
+        if not return_dict:
+            return (
+                sequence_output,
+                *([contextualized_embeddings] if output_hidden_states else []),
+                *([attention_probs] if output_attentions else [])
+            )
+
+        return BaseModelOutput(
+            last_hidden_state=sequence_output,
+            hidden_states=contextualized_embeddings if output_hidden_states else None,
+            attentions=attention_probs if output_attentions else None
+        )
+
+
+class LtgbertForMaskedLM(LtgbertModel):
+    _keys_to_ignore_on_load_unexpected = ["head"]
+
+    def __init__(self, config, **kwargs):
+        super().__init__(config, add_mlm_layer=True, **kwargs)
+
+    def get_output_embeddings(self):
+        return self.classifier.nonlinearity[-1].weight
+
+    def set_output_embeddings(self, new_embeddings):
+        self.classifier.nonlinearity[-1].weight = new_embeddings
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: Optional[torch.LongTensor] = None,
+        **kwargs
+    ) -> Union[Tuple[torch.Tensor], MaskedLMOutput]:
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        sequence_output, contextualized_embeddings, attention_probs = self.get_contextualized_embeddings(input_ids, attention_mask)
+        subword_prediction = self.classifier(sequence_output)
+        subword_prediction[:, :, :106+1] = float("-inf")
+
+        masked_lm_loss = None
+        if labels is not None:
+            masked_lm_loss = F.cross_entropy(subword_prediction.flatten(0, 1), labels.flatten())
+
+        if not return_dict:
+            output = (
+                subword_prediction,
+                *([contextualized_embeddings] if output_hidden_states else []),
+                *([attention_probs] if output_attentions else [])
+            )
+            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=subword_prediction,
+            hidden_states=contextualized_embeddings if output_hidden_states else None,
+            attentions=attention_probs if output_attentions else None
+        )
+
+
+class Classifier(nn.Module):
+    def __init__(self, config, num_labels: int):
+        super().__init__()
+
+        drop_out = getattr(config, "cls_dropout", None)
+        drop_out = config.hidden_dropout_prob if drop_out is None else drop_out
+
+        self.nonlinearity = nn.Sequential(
+            nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=False),
+            nn.Linear(config.hidden_size, config.hidden_size),
+            nn.GELU(),
+            nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=False),
+            nn.Dropout(drop_out),
+            nn.Linear(config.hidden_size, num_labels)
+        )
+
+    def forward(self, x):
+        x = self.nonlinearity(x)
+        return x
+
+
+class LtgbertForSequenceClassification(LtgbertModel):
+    _keys_to_ignore_on_load_unexpected = ["classifier"]
+    _keys_to_ignore_on_load_missing = ["head"]
+
+    def __init__(self, config, **kwargs):
+        super().__init__(config, add_mlm_layer=False, **kwargs)
+
+        self.num_labels = config.num_labels
+        self.head = Classifier(config, self.num_labels)
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: Optional[torch.LongTensor] = None,
+        **kwargs
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        sequence_output, contextualized_embeddings, attention_probs = self.get_contextualized_embeddings(input_ids, attention_mask)
+        logits = self.head(sequence_output[:, 0, :])
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = nn.MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = nn.CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = nn.BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (
+                logits,
+                *([contextualized_embeddings] if output_hidden_states else []),
+                *([attention_probs] if output_attentions else [])
+            )
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=contextualized_embeddings if output_hidden_states else None,
+            attentions=attention_probs if output_attentions else None
+        )
+
+
+class LtgbertForTokenClassification(LtgbertModel):
+    _keys_to_ignore_on_load_unexpected = ["classifier"]
+    _keys_to_ignore_on_load_missing = ["head"]
+
+    def __init__(self, config, **kwargs):
+        super().__init__(config, add_mlm_layer=False, **kwargs)
+
+        self.num_labels = config.num_labels
+        self.head = Classifier(config, self.num_labels)
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: Optional[torch.LongTensor] = None,
+        **kwargs
+    ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        sequence_output, contextualized_embeddings, attention_probs = self.get_contextualized_embeddings(input_ids, attention_mask)
+        logits = self.head(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (
+                logits,
+                *([contextualized_embeddings] if output_hidden_states else []),
+                *([attention_probs] if output_attentions else [])
+            )
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=contextualized_embeddings if output_hidden_states else None,
+            attentions=attention_probs if output_attentions else None
+        )
+
+
+class LtgbertForQuestionAnswering(LtgbertModel):
+    _keys_to_ignore_on_load_unexpected = ["classifier"]
+    _keys_to_ignore_on_load_missing = ["head"]
+
+    def __init__(self, config, **kwargs):
+        super().__init__(config, add_mlm_layer=False, **kwargs)
+
+        self.num_labels = config.num_labels
+        self.head = Classifier(config, self.num_labels)
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        start_positions: Optional[torch.Tensor] = None,
+        end_positions: Optional[torch.Tensor] = None,
+        **kwargs
+    ) -> Union[Tuple[torch.Tensor], QuestionAnsweringModelOutput]:
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        sequence_output, contextualized_embeddings, attention_probs = self.get_contextualized_embeddings(input_ids, attention_mask)
+        logits = self.head(sequence_output)
+
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
+
+            loss_fct = nn.CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (
+                start_logits,
+                end_logits,
+                *([contextualized_embeddings] if output_hidden_states else []),
+                *([attention_probs] if output_attentions else [])
+            )
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=contextualized_embeddings if output_hidden_states else None,
+            attentions=attention_probs if output_attentions else None
+        )
+
+
+class LtgbertForMultipleChoice(LtgbertModel):
+    _keys_to_ignore_on_load_unexpected = ["classifier"]
+    _keys_to_ignore_on_load_missing = ["head"]
+
+    def __init__(self, config, **kwargs):
+        super().__init__(config, add_mlm_layer=False, **kwargs)
+
+        self.num_labels = getattr(config, "num_labels", 2)
+        self.head = Classifier(config, self.num_labels)
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs
+    ) -> Union[Tuple[torch.Tensor], MultipleChoiceModelOutput]:
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        num_choices = input_ids.shape[1]
+
+        flat_input_ids = input_ids.view(-1, input_ids.size(-1))
+        flat_attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+
+        sequence_output, contextualized_embeddings, attention_probs = self.get_contextualized_embeddings(flat_input_ids, flat_attention_mask)
+        logits = self.head(sequence_output)
+        reshaped_logits = logits.view(-1, num_choices)
+
+        loss = None
+        if labels is not None:
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(reshaped_logits, labels)
+
+        if not return_dict:
+            output = (
+                reshaped_logits,
+                *([contextualized_embeddings] if output_hidden_states else []),
+                *([attention_probs] if output_attentions else [])
+            )
+            return ((loss,) + output) if loss is not None else output
+
+        return MultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=contextualized_embeddings if output_hidden_states else None,
+            attentions=attention_probs if output_attentions else None
+        )

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ba06cd5f37e676787d5d83261aa3beca1c4cb3f919efe730d13499b5bd016d7
+size 525206650

special_tokens_map.json

@@ -0,0 +1,39 @@
+{
+  "bos_token": "[BOS]",
+  "cls_token": {
+    "content": "[CLS]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": "[EOS]",
+  "mask_token": {
+    "content": "[MASK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "[PAD]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "content": "[SEP]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "[UNK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer_config.json

@@ -0,0 +1,870 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "[CLS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "[SEP]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "4": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "5": {
+      "content": "[MASK_1]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "6": {
+      "content": "[MASK_2]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "7": {
+      "content": "[MASK_3]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "8": {
+      "content": "[MASK_4]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "9": {
+      "content": "[MASK_5]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "10": {
+      "content": "[MASK_6]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "11": {
+      "content": "[MASK_7]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "12": {
+      "content": "[MASK_8]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "13": {
+      "content": "[MASK_9]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "14": {
+      "content": "[MASK_10]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "15": {
+      "content": "[MASK_11]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "16": {
+      "content": "[MASK_12]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "17": {
+      "content": "[MASK_13]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "18": {
+      "content": "[MASK_14]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "19": {
+      "content": "[MASK_15]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "20": {
+      "content": "[MASK_16]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "21": {
+      "content": "[MASK_17]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "22": {
+      "content": "[MASK_18]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "23": {
+      "content": "[MASK_19]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "24": {
+      "content": "[MASK_20]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "25": {
+      "content": "[MASK_21]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "26": {
+      "content": "[MASK_22]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "27": {
+      "content": "[MASK_23]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "28": {
+      "content": "[MASK_24]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "29": {
+      "content": "[MASK_25]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "30": {
+      "content": "[MASK_26]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "31": {
+      "content": "[MASK_27]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32": {
+      "content": "[MASK_28]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "33": {
+      "content": "[MASK_29]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "34": {
+      "content": "[MASK_30]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "35": {
+      "content": "[MASK_31]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "36": {
+      "content": "[MASK_32]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "37": {
+      "content": "[MASK_33]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "38": {
+      "content": "[MASK_34]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "39": {
+      "content": "[MASK_35]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "40": {
+      "content": "[MASK_36]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "41": {
+      "content": "[MASK_37]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "42": {
+      "content": "[MASK_38]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "43": {
+      "content": "[MASK_39]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "44": {
+      "content": "[MASK_40]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "45": {
+      "content": "[MASK_41]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "46": {
+      "content": "[MASK_42]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "47": {
+      "content": "[MASK_43]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "48": {
+      "content": "[MASK_44]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49": {
+      "content": "[MASK_45]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50": {
+      "content": "[MASK_46]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "51": {
+      "content": "[MASK_47]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "52": {
+      "content": "[MASK_48]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "53": {
+      "content": "[MASK_49]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "54": {
+      "content": "[MASK_50]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "55": {
+      "content": "[MASK_51]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "56": {
+      "content": "[MASK_52]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "57": {
+      "content": "[MASK_53]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "58": {
+      "content": "[MASK_54]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "59": {
+      "content": "[MASK_55]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "60": {
+      "content": "[MASK_56]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "61": {
+      "content": "[MASK_57]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "62": {
+      "content": "[MASK_58]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "63": {
+      "content": "[MASK_59]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "64": {
+      "content": "[MASK_60]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65": {
+      "content": "[MASK_61]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "66": {
+      "content": "[MASK_62]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "67": {
+      "content": "[MASK_63]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "68": {
+      "content": "[MASK_64]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "69": {
+      "content": "[MASK_65]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "70": {
+      "content": "[MASK_66]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "71": {
+      "content": "[MASK_67]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "72": {
+      "content": "[MASK_68]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "73": {
+      "content": "[MASK_69]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "74": {
+      "content": "[MASK_70]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75": {
+      "content": "[MASK_71]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "76": {
+      "content": "[MASK_72]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "77": {
+      "content": "[MASK_73]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "78": {
+      "content": "[MASK_74]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "79": {
+      "content": "[MASK_75]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "80": {
+      "content": "[MASK_76]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "81": {
+      "content": "[MASK_77]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "82": {
+      "content": "[MASK_78]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "83": {
+      "content": "[MASK_79]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "84": {
+      "content": "[MASK_80]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "85": {
+      "content": "[MASK_81]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "86": {
+      "content": "[MASK_82]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "87": {
+      "content": "[MASK_83]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "88": {
+      "content": "[MASK_84]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "89": {
+      "content": "[MASK_85]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "90": {
+      "content": "[MASK_86]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "91": {
+      "content": "[MASK_87]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "92": {
+      "content": "[MASK_88]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "93": {
+      "content": "[MASK_89]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "94": {
+      "content": "[MASK_90]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "95": {
+      "content": "[MASK_91]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "96": {
+      "content": "[MASK_92]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "97": {
+      "content": "[MASK_93]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "98": {
+      "content": "[MASK_94]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "99": {
+      "content": "[MASK_95]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "100": {
+      "content": "[MASK_96]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "101": {
+      "content": "[MASK_97]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "102": {
+      "content": "[MASK_98]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "103": {
+      "content": "[MASK_99]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "104": {
+      "content": "█",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32768": {
+      "content": "[BOS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32769": {
+      "content": "[EOS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "[BOS]",
+  "clean_up_tokenization_spaces": true,
+  "cls_token": "[CLS]",
+  "eos_token": "[EOS]",
+  "mask_token": "[MASK]",
+  "model_max_length": 512,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "unk_token": "[UNK]"
+}

upos.py

@@ -0,0 +1,41 @@
+from transformers import TokenClassificationPipeline
+
+class BellmanFordTokenClassificationPipeline(TokenClassificationPipeline):
+  def __init__(self,**kwargs):
+    import numpy
+    super().__init__(**kwargs)
+    x=self.model.config.label2id
+    y=[k for k in x if not k.startswith("I-")]
+    self.transition=numpy.full((len(x),len(x)),numpy.nan)
+    for k,v in x.items():
+      for j in ["I-"+k[2:]] if k.startswith("B-") else [k]+y if k.startswith("I-") else y:
+        self.transition[v,x[j]]=0
+  def check_model_type(self,supported_models):
+    pass
+  def postprocess(self,model_outputs,**kwargs):
+    import numpy
+    if "logits" not in model_outputs:
+      return self.postprocess(model_outputs[0],**kwargs)
+    m=model_outputs["logits"][0].numpy()
+    e=numpy.exp(m-numpy.max(m,axis=-1,keepdims=True))
+    z=e/e.sum(axis=-1,keepdims=True)
+    for i in range(m.shape[0]-1,0,-1):
+      m[i-1]+=numpy.nanmax(m[i]+self.transition,axis=1)
+    k=[numpy.nanargmax(m[0]+self.transition[0])]
+    for i in range(1,m.shape[0]):
+      k.append(numpy.nanargmax(m[i]+self.transition[k[-1]]))
+    w=[{"entity":self.model.config.id2label[j],"start":s,"end":e,"score":z[i,j]} for i,((s,e),j) in enumerate(zip(model_outputs["offset_mapping"][0].tolist(),k)) if s<e]
+    if "aggregation_strategy" in kwargs and kwargs["aggregation_strategy"]!="none":
+      for i,t in reversed(list(enumerate(w))):
+        p=t.pop("entity")
+        if p.startswith("I-"):
+          w[i-1]["score"]=min(w[i-1]["score"],t["score"])
+          w[i-1]["end"]=w.pop(i)["end"]
+        elif p.startswith("B-"):
+          t["entity_group"]=p[2:]
+        else:
+          t["entity_group"]=p
+    for t in w:
+      t["text"]=model_outputs["sentence"][t["start"]:t["end"]]
+    return w
+