Upload RoFormerForMaskedLM
Browse files- config.json +34 -0
- modeling_nt_roformer.py +1587 -0
- pytorch_model.bin +3 -0
config.json
ADDED
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "/fsx/home-zanussbaum/nucleotide-transformer/nt_transformer/saved_models/roformer/mamalian_cds_2k_window_step84999",
|
3 |
+
"architectures": [
|
4 |
+
"RoFormerForMaskedLM"
|
5 |
+
],
|
6 |
+
"attention_probs_dropout_prob": 0.1,
|
7 |
+
"auto_map": {
|
8 |
+
"AutoModelForMaskedLM": "modeling_nt_roformer.RoFormerForMaskedLM"
|
9 |
+
},
|
10 |
+
"classifier_dropout": "None",
|
11 |
+
"embedding_size": 768,
|
12 |
+
"hidden_act": "gelu",
|
13 |
+
"hidden_dropout_prob": 0.1,
|
14 |
+
"hidden_size": 768,
|
15 |
+
"initializer_range": 0.02,
|
16 |
+
"intermediate_size": 3072,
|
17 |
+
"layer_norm_eps": 1e-12,
|
18 |
+
"max_position_embeddings": 512,
|
19 |
+
"model_type": "roformer",
|
20 |
+
"num_attention_heads": 12,
|
21 |
+
"num_hidden_layers": 12,
|
22 |
+
"pad_token_id": 0,
|
23 |
+
"position_embedding_type": "absolute",
|
24 |
+
"rotary_value": true,
|
25 |
+
"summary_activation": "gelu",
|
26 |
+
"summary_last_dropout": 0.1,
|
27 |
+
"summary_type": "first",
|
28 |
+
"summary_use_proj": true,
|
29 |
+
"torch_dtype": "float32",
|
30 |
+
"transformers_version": "4.25.1",
|
31 |
+
"type_vocab_size": 2,
|
32 |
+
"use_cache": true,
|
33 |
+
"vocab_size": 9
|
34 |
+
}
|
modeling_nt_roformer.py
ADDED
@@ -0,0 +1,1587 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
""" PyTorch RoFormer model."""
|
16 |
+
|
17 |
+
|
18 |
+
import math
|
19 |
+
import os
|
20 |
+
from typing import Optional, Tuple, Union
|
21 |
+
|
22 |
+
import numpy as np
|
23 |
+
import torch
|
24 |
+
import torch.utils.checkpoint
|
25 |
+
from torch import nn
|
26 |
+
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
27 |
+
|
28 |
+
from transformers.activations import ACT2FN
|
29 |
+
from transformers.modeling_outputs import (
|
30 |
+
BaseModelOutputWithPastAndCrossAttentions,
|
31 |
+
CausalLMOutputWithCrossAttentions,
|
32 |
+
MaskedLMOutput,
|
33 |
+
MultipleChoiceModelOutput,
|
34 |
+
QuestionAnsweringModelOutput,
|
35 |
+
SequenceClassifierOutput,
|
36 |
+
TokenClassifierOutput,
|
37 |
+
)
|
38 |
+
from transformers.modeling_utils import PreTrainedModel, SequenceSummary
|
39 |
+
from transformers.pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
|
40 |
+
from transformers.utils import (
|
41 |
+
add_code_sample_docstrings,
|
42 |
+
add_start_docstrings,
|
43 |
+
add_start_docstrings_to_model_forward,
|
44 |
+
logging,
|
45 |
+
replace_return_docstrings,
|
46 |
+
)
|
47 |
+
from transformers import RoFormerConfig
|
48 |
+
|
49 |
+
|
50 |
+
logger = logging.get_logger(__name__)
|
51 |
+
|
52 |
+
_CHECKPOINT_FOR_DOC = "junnyu/roformer_chinese_base"
|
53 |
+
_CONFIG_FOR_DOC = "RoFormerConfig"
|
54 |
+
_TOKENIZER_FOR_DOC = "RoFormerTokenizer"
|
55 |
+
|
56 |
+
ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = [
|
57 |
+
"junnyu/roformer_chinese_small",
|
58 |
+
"junnyu/roformer_chinese_base",
|
59 |
+
"junnyu/roformer_chinese_char_small",
|
60 |
+
"junnyu/roformer_chinese_char_base",
|
61 |
+
"junnyu/roformer_small_discriminator",
|
62 |
+
"junnyu/roformer_small_generator"
|
63 |
+
# See all RoFormer models at https://huggingface.co/models?filter=roformer
|
64 |
+
]
|
65 |
+
|
66 |
+
|
67 |
+
# Copied from transformers.models.marian.modeling_marian.MarianSinusoidalPositionalEmbedding with Marian->RoFormer
|
68 |
+
class RoFormerSinusoidalPositionalEmbedding(nn.Embedding):
|
69 |
+
"""This module produces sinusoidal positional embeddings of any length."""
|
70 |
+
|
71 |
+
def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None) -> None:
|
72 |
+
super().__init__(num_positions, embedding_dim)
|
73 |
+
self.weight = self._init_weight(self.weight)
|
74 |
+
|
75 |
+
@staticmethod
|
76 |
+
def _init_weight(out: nn.Parameter) -> nn.Parameter:
|
77 |
+
"""
|
78 |
+
Identical to the XLM create_sinusoidal_embeddings except features are not interleaved. The cos features are in
|
79 |
+
the 2nd half of the vector. [dim // 2:]
|
80 |
+
"""
|
81 |
+
n_pos, dim = out.shape
|
82 |
+
position_enc = np.array(
|
83 |
+
[[pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)] for pos in range(n_pos)]
|
84 |
+
)
|
85 |
+
out.requires_grad = False # set early to avoid an error in pytorch-1.8+
|
86 |
+
sentinel = dim // 2 if dim % 2 == 0 else (dim // 2) + 1
|
87 |
+
out[:, 0:sentinel] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))
|
88 |
+
out[:, sentinel:] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
|
89 |
+
out.detach_()
|
90 |
+
return out
|
91 |
+
|
92 |
+
@torch.no_grad()
|
93 |
+
def forward(self, input_ids_shape: torch.Size, past_key_values_length: int = 0) -> torch.Tensor:
|
94 |
+
"""`input_ids_shape` is expected to be [bsz x seqlen]."""
|
95 |
+
bsz, seq_len = input_ids_shape[:2]
|
96 |
+
positions = torch.arange(
|
97 |
+
past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device
|
98 |
+
)
|
99 |
+
return super().forward(positions)
|
100 |
+
|
101 |
+
|
102 |
+
def load_tf_weights_in_roformer(model, config, tf_checkpoint_path):
|
103 |
+
"""Load tf checkpoints in a pytorch model."""
|
104 |
+
try:
|
105 |
+
import re
|
106 |
+
|
107 |
+
import numpy as np
|
108 |
+
import tensorflow as tf
|
109 |
+
except ImportError:
|
110 |
+
logger.error(
|
111 |
+
"Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
|
112 |
+
"https://www.tensorflow.org/install/ for installation instructions."
|
113 |
+
)
|
114 |
+
raise
|
115 |
+
tf_path = os.path.abspath(tf_checkpoint_path)
|
116 |
+
logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
|
117 |
+
# Load weights from TF model
|
118 |
+
init_vars = tf.train.list_variables(tf_path)
|
119 |
+
names = []
|
120 |
+
arrays = []
|
121 |
+
for name, shape in init_vars:
|
122 |
+
logger.info(f"Loading TF weight {name} with shape {shape}")
|
123 |
+
array = tf.train.load_variable(tf_path, name)
|
124 |
+
names.append(name.replace("bert", "roformer"))
|
125 |
+
arrays.append(array)
|
126 |
+
|
127 |
+
for name, array in zip(names, arrays):
|
128 |
+
name = name.split("/")
|
129 |
+
# adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
|
130 |
+
# which are not required for using pretrained model
|
131 |
+
if any(
|
132 |
+
n in ["adam_v", "adam_m", "AdamWeightDecayOptimizer", "AdamWeightDecayOptimizer_1", "global_step"]
|
133 |
+
for n in name
|
134 |
+
):
|
135 |
+
logger.info(f"Skipping {'/'.join(name)}")
|
136 |
+
continue
|
137 |
+
pointer = model
|
138 |
+
for m_name in name:
|
139 |
+
if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
|
140 |
+
scope_names = re.split(r"_(\d+)", m_name)
|
141 |
+
else:
|
142 |
+
scope_names = [m_name]
|
143 |
+
if scope_names[0] == "kernel" or scope_names[0] == "gamma":
|
144 |
+
pointer = getattr(pointer, "weight")
|
145 |
+
elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
|
146 |
+
pointer = getattr(pointer, "bias")
|
147 |
+
elif scope_names[0] == "output_weights":
|
148 |
+
pointer = getattr(pointer, "weight")
|
149 |
+
elif scope_names[0] == "squad":
|
150 |
+
pointer = getattr(pointer, "classifier")
|
151 |
+
else:
|
152 |
+
try:
|
153 |
+
pointer = getattr(pointer, scope_names[0])
|
154 |
+
except AttributeError:
|
155 |
+
logger.info(f"Skipping {'/'.join(name)}")
|
156 |
+
continue
|
157 |
+
if len(scope_names) >= 2:
|
158 |
+
num = int(scope_names[1])
|
159 |
+
pointer = pointer[num]
|
160 |
+
if m_name[-11:] == "_embeddings":
|
161 |
+
pointer = getattr(pointer, "weight")
|
162 |
+
elif m_name == "kernel":
|
163 |
+
array = np.transpose(array)
|
164 |
+
try:
|
165 |
+
if not pointer.shape == array.shape:
|
166 |
+
raise ValueError(f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched")
|
167 |
+
except AssertionError as e:
|
168 |
+
e.args += (pointer.shape, array.shape)
|
169 |
+
raise
|
170 |
+
logger.info(f"Initialize PyTorch weight {name}")
|
171 |
+
pointer.data = torch.from_numpy(array)
|
172 |
+
return model
|
173 |
+
|
174 |
+
|
175 |
+
class RoFormerEmbeddings(nn.Module):
|
176 |
+
"""Construct the embeddings from word and token_type embeddings."""
|
177 |
+
|
178 |
+
def __init__(self, config):
|
179 |
+
super().__init__()
|
180 |
+
self.word_embeddings = nn.Embedding(config.vocab_size, config.embedding_size, padding_idx=config.pad_token_id)
|
181 |
+
self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.embedding_size)
|
182 |
+
|
183 |
+
# self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
|
184 |
+
# any TensorFlow checkpoint file
|
185 |
+
self.LayerNorm = nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
|
186 |
+
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
187 |
+
|
188 |
+
def forward(self, input_ids=None, token_type_ids=None, inputs_embeds=None):
|
189 |
+
if input_ids is not None:
|
190 |
+
input_shape = input_ids.size()
|
191 |
+
else:
|
192 |
+
input_shape = inputs_embeds.size()[:-1]
|
193 |
+
|
194 |
+
if inputs_embeds is None:
|
195 |
+
inputs_embeds = self.word_embeddings(input_ids)
|
196 |
+
|
197 |
+
if token_type_ids is None:
|
198 |
+
token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=inputs_embeds.device)
|
199 |
+
|
200 |
+
token_type_embeddings = self.token_type_embeddings(token_type_ids)
|
201 |
+
|
202 |
+
embeddings = inputs_embeds + token_type_embeddings
|
203 |
+
|
204 |
+
embeddings = self.LayerNorm(embeddings)
|
205 |
+
embeddings = self.dropout(embeddings)
|
206 |
+
return embeddings
|
207 |
+
|
208 |
+
|
209 |
+
class RoFormerSelfAttention(nn.Module):
|
210 |
+
def __init__(self, config):
|
211 |
+
super().__init__()
|
212 |
+
if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
|
213 |
+
raise ValueError(
|
214 |
+
f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
|
215 |
+
f"heads ({config.num_attention_heads})"
|
216 |
+
)
|
217 |
+
|
218 |
+
self.num_attention_heads = config.num_attention_heads
|
219 |
+
self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
|
220 |
+
self.all_head_size = self.num_attention_heads * self.attention_head_size
|
221 |
+
|
222 |
+
self.query = nn.Linear(config.hidden_size, self.all_head_size)
|
223 |
+
self.key = nn.Linear(config.hidden_size, self.all_head_size)
|
224 |
+
self.value = nn.Linear(config.hidden_size, self.all_head_size)
|
225 |
+
|
226 |
+
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
|
227 |
+
|
228 |
+
self.is_decoder = config.is_decoder
|
229 |
+
self.rotary_value = config.rotary_value
|
230 |
+
|
231 |
+
def transpose_for_scores(self, x):
|
232 |
+
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
|
233 |
+
x = x.view(*new_x_shape)
|
234 |
+
return x.permute(0, 2, 1, 3)
|
235 |
+
|
236 |
+
def forward(
|
237 |
+
self,
|
238 |
+
hidden_states,
|
239 |
+
attention_mask=None,
|
240 |
+
sinusoidal_pos=None,
|
241 |
+
head_mask=None,
|
242 |
+
encoder_hidden_states=None,
|
243 |
+
encoder_attention_mask=None,
|
244 |
+
past_key_value=None,
|
245 |
+
output_attentions=False,
|
246 |
+
):
|
247 |
+
mixed_query_layer = self.query(hidden_states)
|
248 |
+
query_layer = self.transpose_for_scores(mixed_query_layer)
|
249 |
+
# If this is instantiated as a cross-attention module, the keys
|
250 |
+
# and values come from an encoder; the attention mask needs to be
|
251 |
+
# such that the encoder's padding tokens are not attended to.
|
252 |
+
is_cross_attention = encoder_hidden_states is not None
|
253 |
+
|
254 |
+
if is_cross_attention and past_key_value is not None:
|
255 |
+
# reuse k,v, cross_attentions
|
256 |
+
key_layer = past_key_value[0]
|
257 |
+
value_layer = past_key_value[1]
|
258 |
+
attention_mask = encoder_attention_mask
|
259 |
+
elif is_cross_attention:
|
260 |
+
key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
|
261 |
+
value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
|
262 |
+
attention_mask = encoder_attention_mask
|
263 |
+
elif past_key_value is not None:
|
264 |
+
key_layer = self.transpose_for_scores(self.key(hidden_states))
|
265 |
+
value_layer = self.transpose_for_scores(self.value(hidden_states))
|
266 |
+
key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
|
267 |
+
value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
|
268 |
+
else:
|
269 |
+
key_layer = self.transpose_for_scores(self.key(hidden_states))
|
270 |
+
value_layer = self.transpose_for_scores(self.value(hidden_states))
|
271 |
+
if sinusoidal_pos is not None:
|
272 |
+
if self.rotary_value:
|
273 |
+
query_layer, key_layer, value_layer = self.apply_rotary_position_embeddings(
|
274 |
+
sinusoidal_pos, query_layer, key_layer, value_layer
|
275 |
+
)
|
276 |
+
else:
|
277 |
+
query_layer, key_layer = self.apply_rotary_position_embeddings(
|
278 |
+
sinusoidal_pos, query_layer, key_layer
|
279 |
+
)
|
280 |
+
if self.is_decoder:
|
281 |
+
# if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
|
282 |
+
# Further calls to cross_attention layer can then reuse all cross-attention
|
283 |
+
# key/value_states (first "if" case)
|
284 |
+
# if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
|
285 |
+
# all previous decoder key/value_states. Further calls to uni-directional self-attention
|
286 |
+
# can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
|
287 |
+
# if encoder bi-directional self-attention `past_key_value` is always `None`
|
288 |
+
past_key_value = (key_layer, value_layer)
|
289 |
+
|
290 |
+
# Take the dot product between "query" and "key" to get the raw attention scores.
|
291 |
+
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
|
292 |
+
|
293 |
+
attention_scores = attention_scores / math.sqrt(self.attention_head_size)
|
294 |
+
if attention_mask is not None:
|
295 |
+
# Apply the attention mask is (precomputed for all layers in RoFormerModel forward() function)
|
296 |
+
attention_scores = attention_scores + attention_mask
|
297 |
+
|
298 |
+
# Normalize the attention scores to probabilities.
|
299 |
+
attention_probs = nn.functional.softmax(attention_scores, dim=-1)
|
300 |
+
|
301 |
+
# This is actually dropping out entire tokens to attend to, which might
|
302 |
+
# seem a bit unusual, but is taken from the original Transformer paper.
|
303 |
+
attention_probs = self.dropout(attention_probs)
|
304 |
+
|
305 |
+
# Mask heads if we want to
|
306 |
+
if head_mask is not None:
|
307 |
+
attention_probs = attention_probs * head_mask
|
308 |
+
|
309 |
+
context_layer = torch.matmul(attention_probs, value_layer)
|
310 |
+
|
311 |
+
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
|
312 |
+
new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
|
313 |
+
context_layer = context_layer.view(*new_context_layer_shape)
|
314 |
+
|
315 |
+
outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
|
316 |
+
|
317 |
+
if self.is_decoder:
|
318 |
+
outputs = outputs + (past_key_value,)
|
319 |
+
return outputs
|
320 |
+
|
321 |
+
@staticmethod
|
322 |
+
def apply_rotary_position_embeddings(sinusoidal_pos, query_layer, key_layer, value_layer=None):
|
323 |
+
# https://kexue.fm/archives/8265
|
324 |
+
# sin [batch_size, num_heads, sequence_length, embed_size_per_head//2]
|
325 |
+
# cos [batch_size, num_heads, sequence_length, embed_size_per_head//2]
|
326 |
+
sin, cos = sinusoidal_pos.chunk(2, dim=-1)
|
327 |
+
# sin [θ0,θ1,θ2......θd/2-1] -> sin_pos [θ0,θ0,θ1,θ1,θ2,θ2......θd/2-1,θd/2-1]
|
328 |
+
sin_pos = torch.stack([sin, sin], dim=-1).reshape_as(sinusoidal_pos)
|
329 |
+
# cos [θ0,θ1,θ2......θd/2-1] -> cos_pos [θ0,θ0,θ1,θ1,θ2,θ2......θd/2-1,θd/2-1]
|
330 |
+
cos_pos = torch.stack([cos, cos], dim=-1).reshape_as(sinusoidal_pos)
|
331 |
+
# rotate_half_query_layer [-q1,q0,-q3,q2......,-qd-1,qd-2]
|
332 |
+
rotate_half_query_layer = torch.stack([-query_layer[..., 1::2], query_layer[..., ::2]], dim=-1).reshape_as(
|
333 |
+
query_layer
|
334 |
+
)
|
335 |
+
query_layer = query_layer * cos_pos + rotate_half_query_layer * sin_pos
|
336 |
+
# rotate_half_key_layer [-k1,k0,-k3,k2......,-kd-1,kd-2]
|
337 |
+
rotate_half_key_layer = torch.stack([-key_layer[..., 1::2], key_layer[..., ::2]], dim=-1).reshape_as(key_layer)
|
338 |
+
key_layer = key_layer * cos_pos + rotate_half_key_layer * sin_pos
|
339 |
+
if value_layer is not None:
|
340 |
+
# rotate_half_value_layer [-v1,v0,-v3,v2......,-vd-1,vd-2]
|
341 |
+
rotate_half_value_layer = torch.stack([-value_layer[..., 1::2], value_layer[..., ::2]], dim=-1).reshape_as(
|
342 |
+
value_layer
|
343 |
+
)
|
344 |
+
value_layer = value_layer * cos_pos + rotate_half_value_layer * sin_pos
|
345 |
+
return query_layer, key_layer, value_layer
|
346 |
+
return query_layer, key_layer
|
347 |
+
|
348 |
+
|
349 |
+
# Copied from transformers.models.bert.modeling_bert.BertSelfOutput with Bert->RoFormer
|
350 |
+
class RoFormerSelfOutput(nn.Module):
|
351 |
+
def __init__(self, config):
|
352 |
+
super().__init__()
|
353 |
+
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
|
354 |
+
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
355 |
+
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
356 |
+
|
357 |
+
def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
|
358 |
+
hidden_states = self.dense(hidden_states)
|
359 |
+
hidden_states = self.dropout(hidden_states)
|
360 |
+
hidden_states = self.LayerNorm(hidden_states + input_tensor)
|
361 |
+
return hidden_states
|
362 |
+
|
363 |
+
|
364 |
+
class RoFormerAttention(nn.Module):
|
365 |
+
def __init__(self, config):
|
366 |
+
super().__init__()
|
367 |
+
self.self = RoFormerSelfAttention(config)
|
368 |
+
self.output = RoFormerSelfOutput(config)
|
369 |
+
self.pruned_heads = set()
|
370 |
+
|
371 |
+
# Copied from transformers.models.bert.modeling_bert.BertAttention.prune_heads
|
372 |
+
def prune_heads(self, heads):
|
373 |
+
if len(heads) == 0:
|
374 |
+
return
|
375 |
+
heads, index = find_pruneable_heads_and_indices(
|
376 |
+
heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
|
377 |
+
)
|
378 |
+
|
379 |
+
# Prune linear layers
|
380 |
+
self.self.query = prune_linear_layer(self.self.query, index)
|
381 |
+
self.self.key = prune_linear_layer(self.self.key, index)
|
382 |
+
self.self.value = prune_linear_layer(self.self.value, index)
|
383 |
+
self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
|
384 |
+
|
385 |
+
# Update hyper params and store pruned heads
|
386 |
+
self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
|
387 |
+
self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
|
388 |
+
self.pruned_heads = self.pruned_heads.union(heads)
|
389 |
+
|
390 |
+
# End Copy
|
391 |
+
def forward(
|
392 |
+
self,
|
393 |
+
hidden_states,
|
394 |
+
attention_mask=None,
|
395 |
+
sinusoidal_pos=None,
|
396 |
+
head_mask=None,
|
397 |
+
encoder_hidden_states=None,
|
398 |
+
encoder_attention_mask=None,
|
399 |
+
past_key_value=None,
|
400 |
+
output_attentions=False,
|
401 |
+
):
|
402 |
+
self_outputs = self.self(
|
403 |
+
hidden_states,
|
404 |
+
attention_mask,
|
405 |
+
sinusoidal_pos,
|
406 |
+
head_mask,
|
407 |
+
encoder_hidden_states,
|
408 |
+
encoder_attention_mask,
|
409 |
+
past_key_value,
|
410 |
+
output_attentions,
|
411 |
+
)
|
412 |
+
attention_output = self.output(self_outputs[0], hidden_states)
|
413 |
+
outputs = (attention_output,) + self_outputs[1:] # add attentions if we output them
|
414 |
+
return outputs
|
415 |
+
|
416 |
+
|
417 |
+
# Copied from transformers.models.bert.modeling_bert.BertIntermediate with Bert->RoFormer
|
418 |
+
class RoFormerIntermediate(nn.Module):
|
419 |
+
def __init__(self, config):
|
420 |
+
super().__init__()
|
421 |
+
self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
|
422 |
+
if isinstance(config.hidden_act, str):
|
423 |
+
self.intermediate_act_fn = ACT2FN[config.hidden_act]
|
424 |
+
else:
|
425 |
+
self.intermediate_act_fn = config.hidden_act
|
426 |
+
|
427 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
428 |
+
hidden_states = self.dense(hidden_states)
|
429 |
+
hidden_states = self.intermediate_act_fn(hidden_states)
|
430 |
+
return hidden_states
|
431 |
+
|
432 |
+
|
433 |
+
# Copied from transformers.models.bert.modeling_bert.BertOutput with Bert->RoFormer
|
434 |
+
class RoFormerOutput(nn.Module):
|
435 |
+
def __init__(self, config):
|
436 |
+
super().__init__()
|
437 |
+
self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
|
438 |
+
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
439 |
+
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
440 |
+
|
441 |
+
def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
|
442 |
+
hidden_states = self.dense(hidden_states)
|
443 |
+
hidden_states = self.dropout(hidden_states)
|
444 |
+
hidden_states = self.LayerNorm(hidden_states + input_tensor)
|
445 |
+
return hidden_states
|
446 |
+
|
447 |
+
|
448 |
+
class RoFormerLayer(nn.Module):
|
449 |
+
def __init__(self, config):
|
450 |
+
super().__init__()
|
451 |
+
self.chunk_size_feed_forward = config.chunk_size_feed_forward
|
452 |
+
self.seq_len_dim = 1
|
453 |
+
self.attention = RoFormerAttention(config)
|
454 |
+
self.is_decoder = config.is_decoder
|
455 |
+
self.add_cross_attention = config.add_cross_attention
|
456 |
+
if self.add_cross_attention:
|
457 |
+
if not self.is_decoder:
|
458 |
+
raise ValueError(f"{self} should be used as a decoder model if cross attention is added")
|
459 |
+
self.crossattention = RoFormerAttention(config)
|
460 |
+
self.intermediate = RoFormerIntermediate(config)
|
461 |
+
self.output = RoFormerOutput(config)
|
462 |
+
|
463 |
+
def forward(
|
464 |
+
self,
|
465 |
+
hidden_states,
|
466 |
+
attention_mask=None,
|
467 |
+
sinusoidal_pos=None,
|
468 |
+
head_mask=None,
|
469 |
+
encoder_hidden_states=None,
|
470 |
+
encoder_attention_mask=None,
|
471 |
+
past_key_value=None,
|
472 |
+
output_attentions=False,
|
473 |
+
):
|
474 |
+
# decoder uni-directional self-attention cached key/values tuple is at positions 1,2
|
475 |
+
self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
|
476 |
+
self_attention_outputs = self.attention(
|
477 |
+
hidden_states,
|
478 |
+
attention_mask,
|
479 |
+
sinusoidal_pos,
|
480 |
+
head_mask,
|
481 |
+
output_attentions=output_attentions,
|
482 |
+
past_key_value=self_attn_past_key_value,
|
483 |
+
)
|
484 |
+
attention_output = self_attention_outputs[0]
|
485 |
+
|
486 |
+
# if decoder, the last output is tuple of self-attn cache
|
487 |
+
if self.is_decoder:
|
488 |
+
outputs = self_attention_outputs[1:-1]
|
489 |
+
present_key_value = self_attention_outputs[-1]
|
490 |
+
else:
|
491 |
+
outputs = self_attention_outputs[1:] # add self attentions if we output attention weights
|
492 |
+
|
493 |
+
cross_attn_present_key_value = None
|
494 |
+
if self.is_decoder and encoder_hidden_states is not None:
|
495 |
+
if not hasattr(self, "crossattention"):
|
496 |
+
raise ValueError(
|
497 |
+
f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention "
|
498 |
+
"layers by setting `config.add_cross_attention=True`"
|
499 |
+
)
|
500 |
+
|
501 |
+
# cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
|
502 |
+
cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
|
503 |
+
cross_attention_outputs = self.crossattention(
|
504 |
+
attention_output,
|
505 |
+
attention_mask,
|
506 |
+
sinusoidal_pos,
|
507 |
+
head_mask,
|
508 |
+
encoder_hidden_states,
|
509 |
+
encoder_attention_mask,
|
510 |
+
cross_attn_past_key_value,
|
511 |
+
output_attentions,
|
512 |
+
)
|
513 |
+
attention_output = cross_attention_outputs[0]
|
514 |
+
outputs = outputs + cross_attention_outputs[1:-1] # add cross attentions if we output attention weights
|
515 |
+
|
516 |
+
# add cross-attn cache to positions 3,4 of present_key_value tuple
|
517 |
+
cross_attn_present_key_value = cross_attention_outputs[-1]
|
518 |
+
present_key_value = present_key_value + cross_attn_present_key_value
|
519 |
+
|
520 |
+
layer_output = apply_chunking_to_forward(
|
521 |
+
self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
|
522 |
+
)
|
523 |
+
outputs = (layer_output,) + outputs
|
524 |
+
|
525 |
+
# if decoder, return the attn key/values as the last output
|
526 |
+
if self.is_decoder:
|
527 |
+
outputs = outputs + (present_key_value,)
|
528 |
+
|
529 |
+
return outputs
|
530 |
+
|
531 |
+
def feed_forward_chunk(self, attention_output):
|
532 |
+
intermediate_output = self.intermediate(attention_output)
|
533 |
+
layer_output = self.output(intermediate_output, attention_output)
|
534 |
+
return layer_output
|
535 |
+
|
536 |
+
|
537 |
+
class RoFormerEncoder(nn.Module):
|
538 |
+
def __init__(self, config):
|
539 |
+
super().__init__()
|
540 |
+
self.config = config
|
541 |
+
self.embed_positions = RoFormerSinusoidalPositionalEmbedding(
|
542 |
+
config.max_position_embeddings, config.hidden_size // config.num_attention_heads
|
543 |
+
)
|
544 |
+
self.layer = nn.ModuleList([RoFormerLayer(config) for _ in range(config.num_hidden_layers)])
|
545 |
+
self.gradient_checkpointing = False
|
546 |
+
|
547 |
+
def forward(
|
548 |
+
self,
|
549 |
+
hidden_states,
|
550 |
+
attention_mask=None,
|
551 |
+
head_mask=None,
|
552 |
+
encoder_hidden_states=None,
|
553 |
+
encoder_attention_mask=None,
|
554 |
+
past_key_values=None,
|
555 |
+
use_cache=None,
|
556 |
+
output_attentions=False,
|
557 |
+
output_hidden_states=False,
|
558 |
+
return_dict=True,
|
559 |
+
):
|
560 |
+
all_hidden_states = () if output_hidden_states else None
|
561 |
+
all_self_attentions = () if output_attentions else None
|
562 |
+
all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
|
563 |
+
|
564 |
+
# [sequence_length, embed_size_per_head] -> [batch_size, num_heads, sequence_length, embed_size_per_head]
|
565 |
+
sinusoidal_pos = self.embed_positions(hidden_states.shape[:-1])[None, None, :, :]
|
566 |
+
|
567 |
+
next_decoder_cache = () if use_cache else None
|
568 |
+
for i, layer_module in enumerate(self.layer):
|
569 |
+
if output_hidden_states:
|
570 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
571 |
+
|
572 |
+
layer_head_mask = head_mask[i] if head_mask is not None else None
|
573 |
+
past_key_value = past_key_values[i] if past_key_values is not None else None
|
574 |
+
|
575 |
+
if self.gradient_checkpointing and self.training:
|
576 |
+
|
577 |
+
if use_cache:
|
578 |
+
logger.warning(
|
579 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
580 |
+
)
|
581 |
+
use_cache = False
|
582 |
+
|
583 |
+
def create_custom_forward(module):
|
584 |
+
def custom_forward(*inputs):
|
585 |
+
return module(*inputs, past_key_value, output_attentions)
|
586 |
+
|
587 |
+
return custom_forward
|
588 |
+
|
589 |
+
layer_outputs = torch.utils.checkpoint.checkpoint(
|
590 |
+
create_custom_forward(layer_module),
|
591 |
+
hidden_states,
|
592 |
+
attention_mask,
|
593 |
+
sinusoidal_pos,
|
594 |
+
layer_head_mask,
|
595 |
+
encoder_hidden_states,
|
596 |
+
encoder_attention_mask,
|
597 |
+
)
|
598 |
+
else:
|
599 |
+
layer_outputs = layer_module(
|
600 |
+
hidden_states,
|
601 |
+
attention_mask,
|
602 |
+
sinusoidal_pos,
|
603 |
+
layer_head_mask,
|
604 |
+
encoder_hidden_states,
|
605 |
+
encoder_attention_mask,
|
606 |
+
past_key_value,
|
607 |
+
output_attentions,
|
608 |
+
)
|
609 |
+
|
610 |
+
hidden_states = layer_outputs[0]
|
611 |
+
if use_cache:
|
612 |
+
next_decoder_cache += (layer_outputs[-1],)
|
613 |
+
if output_attentions:
|
614 |
+
all_self_attentions = all_self_attentions + (layer_outputs[1],)
|
615 |
+
if self.config.add_cross_attention:
|
616 |
+
all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
|
617 |
+
|
618 |
+
if output_hidden_states:
|
619 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
620 |
+
|
621 |
+
if not return_dict:
|
622 |
+
return tuple(
|
623 |
+
v
|
624 |
+
for v in [
|
625 |
+
hidden_states,
|
626 |
+
next_decoder_cache,
|
627 |
+
all_hidden_states,
|
628 |
+
all_self_attentions,
|
629 |
+
all_cross_attentions,
|
630 |
+
]
|
631 |
+
if v is not None
|
632 |
+
)
|
633 |
+
return BaseModelOutputWithPastAndCrossAttentions(
|
634 |
+
last_hidden_state=hidden_states,
|
635 |
+
past_key_values=next_decoder_cache,
|
636 |
+
hidden_states=all_hidden_states,
|
637 |
+
attentions=all_self_attentions,
|
638 |
+
cross_attentions=all_cross_attentions,
|
639 |
+
)
|
640 |
+
|
641 |
+
|
642 |
+
class RoFormerPredictionHeadTransform(nn.Module):
|
643 |
+
def __init__(self, config):
|
644 |
+
super().__init__()
|
645 |
+
self.dense = nn.Linear(config.hidden_size, config.embedding_size)
|
646 |
+
if isinstance(config.hidden_act, str):
|
647 |
+
self.transform_act_fn = ACT2FN[config.hidden_act]
|
648 |
+
else:
|
649 |
+
self.transform_act_fn = config.hidden_act
|
650 |
+
self.LayerNorm = nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
|
651 |
+
|
652 |
+
def forward(self, hidden_states):
|
653 |
+
hidden_states = self.dense(hidden_states)
|
654 |
+
hidden_states = self.transform_act_fn(hidden_states)
|
655 |
+
hidden_states = self.LayerNorm(hidden_states)
|
656 |
+
return hidden_states
|
657 |
+
|
658 |
+
|
659 |
+
class RoFormerLMPredictionHead(nn.Module):
|
660 |
+
def __init__(self, config):
|
661 |
+
super().__init__()
|
662 |
+
self.transform = RoFormerPredictionHeadTransform(config)
|
663 |
+
|
664 |
+
# The output weights are the same as the input embeddings, but there is
|
665 |
+
# an output-only bias for each token.
|
666 |
+
self.decoder = nn.Linear(config.embedding_size, config.vocab_size, bias=False)
|
667 |
+
|
668 |
+
self.bias = nn.Parameter(torch.zeros(config.vocab_size))
|
669 |
+
|
670 |
+
# Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
|
671 |
+
self.decoder.bias = self.bias
|
672 |
+
|
673 |
+
def forward(self, hidden_states):
|
674 |
+
hidden_states = self.transform(hidden_states)
|
675 |
+
hidden_states = self.decoder(hidden_states)
|
676 |
+
return hidden_states
|
677 |
+
|
678 |
+
|
679 |
+
# Copied from transformers.models.bert.modeling_bert.BertOnlyMLMHead with Bert->RoFormer
|
680 |
+
class RoFormerOnlyMLMHead(nn.Module):
|
681 |
+
def __init__(self, config):
|
682 |
+
super().__init__()
|
683 |
+
self.predictions = RoFormerLMPredictionHead(config)
|
684 |
+
|
685 |
+
def forward(self, sequence_output: torch.Tensor) -> torch.Tensor:
|
686 |
+
prediction_scores = self.predictions(sequence_output)
|
687 |
+
return prediction_scores
|
688 |
+
|
689 |
+
|
690 |
+
class RoFormerPreTrainedModel(PreTrainedModel):
|
691 |
+
"""
|
692 |
+
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
693 |
+
models.
|
694 |
+
"""
|
695 |
+
|
696 |
+
config_class = RoFormerConfig
|
697 |
+
load_tf_weights = load_tf_weights_in_roformer
|
698 |
+
base_model_prefix = "roformer"
|
699 |
+
supports_gradient_checkpointing = True
|
700 |
+
_keys_to_ignore_on_load_missing = []
|
701 |
+
_keys_to_ignore_on_load_unexpected = [
|
702 |
+
r"roformer.embeddings_project.weight",
|
703 |
+
r"roformer.embeddings_project.bias",
|
704 |
+
]
|
705 |
+
|
706 |
+
def _init_weights(self, module):
|
707 |
+
"""Initialize the weights"""
|
708 |
+
if isinstance(module, nn.Linear):
|
709 |
+
# Slightly different from the TF version which uses truncated_normal for initialization
|
710 |
+
# cf https://github.com/pytorch/pytorch/pull/5617
|
711 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
712 |
+
if module.bias is not None:
|
713 |
+
module.bias.data.zero_()
|
714 |
+
elif isinstance(module, RoFormerSinusoidalPositionalEmbedding):
|
715 |
+
pass
|
716 |
+
elif isinstance(module, nn.Embedding):
|
717 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
718 |
+
if module.padding_idx is not None:
|
719 |
+
module.weight.data[module.padding_idx].zero_()
|
720 |
+
elif isinstance(module, nn.LayerNorm):
|
721 |
+
module.bias.data.zero_()
|
722 |
+
module.weight.data.fill_(1.0)
|
723 |
+
|
724 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
725 |
+
if isinstance(module, RoFormerEncoder):
|
726 |
+
module.gradient_checkpointing = value
|
727 |
+
|
728 |
+
|
729 |
+
ROFORMER_START_DOCSTRING = r"""
|
730 |
+
This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. Use
|
731 |
+
it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
|
732 |
+
behavior.
|
733 |
+
|
734 |
+
Parameters:
|
735 |
+
config ([`RoFormerConfig`]): Model configuration class with all the parameters of the model.
|
736 |
+
Initializing with a config file does not load the weights associated with the model, only the
|
737 |
+
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
738 |
+
"""
|
739 |
+
|
740 |
+
ROFORMER_INPUTS_DOCSTRING = r"""
|
741 |
+
Args:
|
742 |
+
input_ids (`torch.LongTensor` of shape `({0})`):
|
743 |
+
Indices of input sequence tokens in the vocabulary.
|
744 |
+
|
745 |
+
Indices can be obtained using [`RoFormerTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
746 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
747 |
+
|
748 |
+
[What are input IDs?](../glossary#input-ids)
|
749 |
+
attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
|
750 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
751 |
+
|
752 |
+
- 1 for tokens that are **not masked**,
|
753 |
+
- 0 for tokens that are **masked**.
|
754 |
+
|
755 |
+
[What are attention masks?](../glossary#attention-mask)
|
756 |
+
token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
|
757 |
+
Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
|
758 |
+
1]`:
|
759 |
+
|
760 |
+
- 0 corresponds to a *sentence A* token,
|
761 |
+
- 1 corresponds to a *sentence B* token.
|
762 |
+
|
763 |
+
[What are token type IDs?](../glossary#token-type-ids)
|
764 |
+
head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
|
765 |
+
Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
|
766 |
+
|
767 |
+
- 1 indicates the head is **not masked**,
|
768 |
+
- 0 indicates the head is **masked**.
|
769 |
+
|
770 |
+
inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
|
771 |
+
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
772 |
+
is useful if you want more control over how to convert *input_ids* indices into associated vectors than the
|
773 |
+
model's internal embedding lookup matrix.
|
774 |
+
output_attentions (`bool`, *optional*):
|
775 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
776 |
+
tensors for more detail.
|
777 |
+
output_hidden_states (`bool`, *optional*):
|
778 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
779 |
+
more detail.
|
780 |
+
return_dict (`bool`, *optional*):
|
781 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
782 |
+
"""
|
783 |
+
|
784 |
+
|
785 |
+
@add_start_docstrings(
|
786 |
+
"The bare RoFormer Model transformer outputting raw hidden-states without any specific head on top.",
|
787 |
+
ROFORMER_START_DOCSTRING,
|
788 |
+
)
|
789 |
+
class RoFormerModel(RoFormerPreTrainedModel):
|
790 |
+
"""
|
791 |
+
|
792 |
+
The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
|
793 |
+
cross-attention is added between the self-attention layers, following the architecture described in [Attention is
|
794 |
+
all you need](https://arxiv.org/abs/1706.03762) by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
|
795 |
+
Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
|
796 |
+
|
797 |
+
To behave as an decoder the model needs to be initialized with the `is_decoder` argument of the configuration set
|
798 |
+
to `True`. To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder` argument and
|
799 |
+
`add_cross_attention` set to `True`; an `encoder_hidden_states` is then expected as an input to the forward pass.
|
800 |
+
"""
|
801 |
+
|
802 |
+
def __init__(self, config):
|
803 |
+
super().__init__(config)
|
804 |
+
self.config = config
|
805 |
+
self.embeddings = RoFormerEmbeddings(config)
|
806 |
+
|
807 |
+
if config.embedding_size != config.hidden_size:
|
808 |
+
self.embeddings_project = nn.Linear(config.embedding_size, config.hidden_size)
|
809 |
+
|
810 |
+
self.encoder = RoFormerEncoder(config)
|
811 |
+
|
812 |
+
# Initialize weights and apply final processing
|
813 |
+
self.post_init()
|
814 |
+
|
815 |
+
def get_input_embeddings(self):
|
816 |
+
return self.embeddings.word_embeddings
|
817 |
+
|
818 |
+
def set_input_embeddings(self, value):
|
819 |
+
self.embeddings.word_embeddings = value
|
820 |
+
|
821 |
+
def _prune_heads(self, heads_to_prune):
|
822 |
+
"""
|
823 |
+
Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
|
824 |
+
class PreTrainedModel
|
825 |
+
"""
|
826 |
+
for layer, heads in heads_to_prune.items():
|
827 |
+
self.encoder.layer[layer].attention.prune_heads(heads)
|
828 |
+
|
829 |
+
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
|
830 |
+
@add_code_sample_docstrings(
|
831 |
+
processor_class=_TOKENIZER_FOR_DOC,
|
832 |
+
checkpoint=_CHECKPOINT_FOR_DOC,
|
833 |
+
output_type=BaseModelOutputWithPastAndCrossAttentions,
|
834 |
+
config_class=_CONFIG_FOR_DOC,
|
835 |
+
)
|
836 |
+
def forward(
|
837 |
+
self,
|
838 |
+
input_ids: Optional[torch.LongTensor] = None,
|
839 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
840 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
841 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
842 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
843 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
844 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
845 |
+
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
846 |
+
use_cache: Optional[bool] = None,
|
847 |
+
output_attentions: Optional[bool] = None,
|
848 |
+
output_hidden_states: Optional[bool] = None,
|
849 |
+
return_dict: Optional[bool] = None,
|
850 |
+
) -> Union[BaseModelOutputWithPastAndCrossAttentions, Tuple[torch.Tensor]]:
|
851 |
+
r"""
|
852 |
+
encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
853 |
+
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
|
854 |
+
the model is configured as a decoder.
|
855 |
+
encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
856 |
+
Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
|
857 |
+
the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
|
858 |
+
|
859 |
+
- 1 for tokens that are **not masked**,
|
860 |
+
- 0 for tokens that are **masked**.
|
861 |
+
past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
|
862 |
+
Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
|
863 |
+
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
|
864 |
+
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
|
865 |
+
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
|
866 |
+
use_cache (`bool`, *optional*):
|
867 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
|
868 |
+
`past_key_values`).
|
869 |
+
"""
|
870 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
871 |
+
output_hidden_states = (
|
872 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
873 |
+
)
|
874 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
875 |
+
|
876 |
+
if self.config.is_decoder:
|
877 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
878 |
+
else:
|
879 |
+
use_cache = False
|
880 |
+
|
881 |
+
if input_ids is not None and inputs_embeds is not None:
|
882 |
+
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
|
883 |
+
elif input_ids is not None:
|
884 |
+
input_shape = input_ids.size()
|
885 |
+
elif inputs_embeds is not None:
|
886 |
+
input_shape = inputs_embeds.size()[:-1]
|
887 |
+
else:
|
888 |
+
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
889 |
+
|
890 |
+
batch_size, seq_length = input_shape
|
891 |
+
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
892 |
+
|
893 |
+
# past_key_values_length
|
894 |
+
past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
|
895 |
+
|
896 |
+
if attention_mask is None:
|
897 |
+
attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
|
898 |
+
if token_type_ids is None:
|
899 |
+
token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
|
900 |
+
|
901 |
+
# We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
|
902 |
+
# ourselves in which case we just need to make it broadcastable to all heads.
|
903 |
+
extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape)
|
904 |
+
|
905 |
+
# If a 2D or 3D attention mask is provided for the cross-attention
|
906 |
+
# we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
|
907 |
+
if self.config.is_decoder and encoder_hidden_states is not None:
|
908 |
+
encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
|
909 |
+
encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
|
910 |
+
if encoder_attention_mask is None:
|
911 |
+
encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
|
912 |
+
encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
|
913 |
+
else:
|
914 |
+
encoder_extended_attention_mask = None
|
915 |
+
|
916 |
+
# Prepare head mask if needed
|
917 |
+
# 1.0 in head_mask indicate we keep the head
|
918 |
+
# attention_probs has shape bsz x n_heads x N x N
|
919 |
+
# input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
|
920 |
+
# and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
|
921 |
+
head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
|
922 |
+
|
923 |
+
embedding_output = self.embeddings(
|
924 |
+
input_ids=input_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
|
925 |
+
)
|
926 |
+
if hasattr(self, "embeddings_project"):
|
927 |
+
embedding_output = self.embeddings_project(embedding_output)
|
928 |
+
|
929 |
+
encoder_outputs = self.encoder(
|
930 |
+
embedding_output,
|
931 |
+
attention_mask=extended_attention_mask,
|
932 |
+
head_mask=head_mask,
|
933 |
+
encoder_hidden_states=encoder_hidden_states,
|
934 |
+
encoder_attention_mask=encoder_extended_attention_mask,
|
935 |
+
past_key_values=past_key_values,
|
936 |
+
use_cache=use_cache,
|
937 |
+
output_attentions=output_attentions,
|
938 |
+
output_hidden_states=output_hidden_states,
|
939 |
+
return_dict=return_dict,
|
940 |
+
)
|
941 |
+
sequence_output = encoder_outputs[0]
|
942 |
+
|
943 |
+
if not return_dict:
|
944 |
+
return (sequence_output,) + encoder_outputs[1:]
|
945 |
+
|
946 |
+
return BaseModelOutputWithPastAndCrossAttentions(
|
947 |
+
last_hidden_state=sequence_output,
|
948 |
+
past_key_values=encoder_outputs.past_key_values,
|
949 |
+
hidden_states=encoder_outputs.hidden_states,
|
950 |
+
attentions=encoder_outputs.attentions,
|
951 |
+
cross_attentions=encoder_outputs.cross_attentions,
|
952 |
+
)
|
953 |
+
|
954 |
+
|
955 |
+
@add_start_docstrings("""RoFormer Model with a `language modeling` head on top.""", ROFORMER_START_DOCSTRING)
|
956 |
+
class RoFormerForMaskedLM(RoFormerPreTrainedModel):
|
957 |
+
_keys_to_ignore_on_load_missing = ["cls.predictions.decoder.bias", "cls.predictions.decoder.weight"]
|
958 |
+
|
959 |
+
def __init__(self, config):
|
960 |
+
super().__init__(config)
|
961 |
+
|
962 |
+
if config.is_decoder:
|
963 |
+
logger.warning(
|
964 |
+
"If you want to use `RoFormerForMaskedLM` make sure `config.is_decoder=False` for "
|
965 |
+
"bi-directional self-attention."
|
966 |
+
)
|
967 |
+
|
968 |
+
self.roformer = RoFormerModel(config)
|
969 |
+
self.cls = RoFormerOnlyMLMHead(config)
|
970 |
+
|
971 |
+
# Initialize weights and apply final processing
|
972 |
+
self.post_init()
|
973 |
+
|
974 |
+
def get_output_embeddings(self):
|
975 |
+
return self.cls.predictions.decoder
|
976 |
+
|
977 |
+
def set_output_embeddings(self, new_embeddings):
|
978 |
+
self.cls.predictions.decoder = new_embeddings
|
979 |
+
|
980 |
+
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
|
981 |
+
@add_code_sample_docstrings(
|
982 |
+
processor_class=_TOKENIZER_FOR_DOC,
|
983 |
+
checkpoint=_CHECKPOINT_FOR_DOC,
|
984 |
+
output_type=MaskedLMOutput,
|
985 |
+
config_class=_CONFIG_FOR_DOC,
|
986 |
+
)
|
987 |
+
def forward(
|
988 |
+
self,
|
989 |
+
input_ids: Optional[torch.LongTensor] = None,
|
990 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
991 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
992 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
993 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
994 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
995 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
996 |
+
labels: Optional[torch.LongTensor] = None,
|
997 |
+
output_attentions: Optional[bool] = None,
|
998 |
+
output_hidden_states: Optional[bool] = None,
|
999 |
+
return_dict: Optional[bool] = None,
|
1000 |
+
loss_weight: Optional[torch.FloatTensor] = None,
|
1001 |
+
) -> Union[MaskedLMOutput, Tuple[torch.Tensor]]:
|
1002 |
+
r"""
|
1003 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
1004 |
+
Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
|
1005 |
+
config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
|
1006 |
+
loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
1007 |
+
"""
|
1008 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1009 |
+
|
1010 |
+
outputs = self.roformer(
|
1011 |
+
input_ids,
|
1012 |
+
attention_mask=attention_mask,
|
1013 |
+
token_type_ids=token_type_ids,
|
1014 |
+
head_mask=head_mask,
|
1015 |
+
inputs_embeds=inputs_embeds,
|
1016 |
+
encoder_hidden_states=encoder_hidden_states,
|
1017 |
+
encoder_attention_mask=encoder_attention_mask,
|
1018 |
+
output_attentions=output_attentions,
|
1019 |
+
output_hidden_states=output_hidden_states,
|
1020 |
+
return_dict=return_dict,
|
1021 |
+
)
|
1022 |
+
|
1023 |
+
sequence_output = outputs[0]
|
1024 |
+
prediction_scores = self.cls(sequence_output)
|
1025 |
+
|
1026 |
+
masked_lm_loss = None
|
1027 |
+
if labels is not None:
|
1028 |
+
loss_fct = CrossEntropyLoss(reduction="none") # -100 index = padding token
|
1029 |
+
labels = labels.view(-1)
|
1030 |
+
masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels)
|
1031 |
+
loss_weight = loss_weight.view(-1)
|
1032 |
+
loss_weight[labels==-100] = 0.0
|
1033 |
+
masked_lm_loss = (masked_lm_loss * loss_weight / loss_weight.sum()).sum()
|
1034 |
+
|
1035 |
+
if not return_dict:
|
1036 |
+
output = (prediction_scores,) + outputs[1:]
|
1037 |
+
return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
|
1038 |
+
|
1039 |
+
return MaskedLMOutput(
|
1040 |
+
loss=masked_lm_loss,
|
1041 |
+
logits=prediction_scores,
|
1042 |
+
hidden_states=outputs.hidden_states,
|
1043 |
+
attentions=outputs.attentions,
|
1044 |
+
)
|
1045 |
+
|
1046 |
+
def prepare_inputs_for_generation(self, input_ids, attention_mask=None, **model_kwargs):
|
1047 |
+
input_shape = input_ids.shape
|
1048 |
+
effective_batch_size = input_shape[0]
|
1049 |
+
|
1050 |
+
# add a dummy token
|
1051 |
+
assert self.config.pad_token_id is not None, "The PAD token should be defined for generation"
|
1052 |
+
attention_mask = torch.cat([attention_mask, attention_mask.new_zeros((attention_mask.shape[0], 1))], dim=-1)
|
1053 |
+
dummy_token = torch.full(
|
1054 |
+
(effective_batch_size, 1), self.config.pad_token_id, dtype=torch.long, device=input_ids.device
|
1055 |
+
)
|
1056 |
+
input_ids = torch.cat([input_ids, dummy_token], dim=1)
|
1057 |
+
|
1058 |
+
return {"input_ids": input_ids, "attention_mask": attention_mask}
|
1059 |
+
|
1060 |
+
|
1061 |
+
@add_start_docstrings(
|
1062 |
+
"""RoFormer Model with a `language modeling` head on top for CLM fine-tuning.""", ROFORMER_START_DOCSTRING
|
1063 |
+
)
|
1064 |
+
class RoFormerForCausalLM(RoFormerPreTrainedModel):
|
1065 |
+
_keys_to_ignore_on_load_missing = ["cls.predictions.decoder.bias", "cls.predictions.decoder.weight"]
|
1066 |
+
|
1067 |
+
def __init__(self, config):
|
1068 |
+
super().__init__(config)
|
1069 |
+
|
1070 |
+
if not config.is_decoder:
|
1071 |
+
logger.warning("If you want to use `RoFormerForCausalLM` as a standalone, add `is_decoder=True.`")
|
1072 |
+
|
1073 |
+
self.roformer = RoFormerModel(config)
|
1074 |
+
self.cls = RoFormerOnlyMLMHead(config)
|
1075 |
+
|
1076 |
+
# Initialize weights and apply final processing
|
1077 |
+
self.post_init()
|
1078 |
+
|
1079 |
+
def get_output_embeddings(self):
|
1080 |
+
return self.cls.predictions.decoder
|
1081 |
+
|
1082 |
+
def set_output_embeddings(self, new_embeddings):
|
1083 |
+
self.cls.predictions.decoder = new_embeddings
|
1084 |
+
|
1085 |
+
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
|
1086 |
+
@replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)
|
1087 |
+
def forward(
|
1088 |
+
self,
|
1089 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1090 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1091 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1092 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1093 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
1094 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
1095 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1096 |
+
cross_attn_head_mask: Optional[torch.Tensor] = None,
|
1097 |
+
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
1098 |
+
labels: Optional[torch.LongTensor] = None,
|
1099 |
+
use_cache: Optional[bool] = None,
|
1100 |
+
output_attentions: Optional[bool] = None,
|
1101 |
+
output_hidden_states: Optional[bool] = None,
|
1102 |
+
return_dict: Optional[bool] = None,
|
1103 |
+
) -> Union[CausalLMOutputWithCrossAttentions, Tuple[torch.Tensor]]:
|
1104 |
+
r"""
|
1105 |
+
encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
1106 |
+
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
|
1107 |
+
the model is configured as a decoder.
|
1108 |
+
encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
1109 |
+
Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
|
1110 |
+
the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
|
1111 |
+
|
1112 |
+
- 1 for tokens that are **not masked**,
|
1113 |
+
- 0 for tokens that are **masked**.
|
1114 |
+
past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
|
1115 |
+
Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
|
1116 |
+
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
|
1117 |
+
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
|
1118 |
+
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
|
1119 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
1120 |
+
Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
|
1121 |
+
`[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are
|
1122 |
+
ignored (masked), the loss is only computed for the tokens with labels n `[0, ..., config.vocab_size]`.
|
1123 |
+
use_cache (`bool`, *optional*):
|
1124 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
|
1125 |
+
`past_key_values`).
|
1126 |
+
|
1127 |
+
Returns:
|
1128 |
+
|
1129 |
+
Example:
|
1130 |
+
|
1131 |
+
```python
|
1132 |
+
>>> from transformers import RoFormerTokenizer, RoFormerForCausalLM, RoFormerConfig
|
1133 |
+
>>> import torch
|
1134 |
+
|
1135 |
+
>>> tokenizer = RoFormerTokenizer.from_pretrained("junnyu/roformer_chinese_base")
|
1136 |
+
>>> config = RoFormerConfig.from_pretrained("junnyu/roformer_chinese_base")
|
1137 |
+
>>> config.is_decoder = True
|
1138 |
+
>>> model = RoFormerForCausalLM.from_pretrained("junnyu/roformer_chinese_base", config=config)
|
1139 |
+
|
1140 |
+
>>> inputs = tokenizer("今天天气非常好。", return_tensors="pt")
|
1141 |
+
>>> outputs = model(**inputs)
|
1142 |
+
|
1143 |
+
>>> prediction_logits = outputs.logits
|
1144 |
+
```"""
|
1145 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1146 |
+
|
1147 |
+
outputs = self.roformer(
|
1148 |
+
input_ids,
|
1149 |
+
attention_mask=attention_mask,
|
1150 |
+
token_type_ids=token_type_ids,
|
1151 |
+
head_mask=head_mask,
|
1152 |
+
inputs_embeds=inputs_embeds,
|
1153 |
+
encoder_hidden_states=encoder_hidden_states,
|
1154 |
+
encoder_attention_mask=encoder_attention_mask,
|
1155 |
+
past_key_values=past_key_values,
|
1156 |
+
use_cache=use_cache,
|
1157 |
+
output_attentions=output_attentions,
|
1158 |
+
output_hidden_states=output_hidden_states,
|
1159 |
+
return_dict=return_dict,
|
1160 |
+
)
|
1161 |
+
|
1162 |
+
sequence_output = outputs[0]
|
1163 |
+
prediction_scores = self.cls(sequence_output)
|
1164 |
+
|
1165 |
+
lm_loss = None
|
1166 |
+
if labels is not None:
|
1167 |
+
# we are doing next-token prediction; shift prediction scores and input ids by one
|
1168 |
+
shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
|
1169 |
+
labels = labels[:, 1:].contiguous()
|
1170 |
+
loss_fct = CrossEntropyLoss()
|
1171 |
+
lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
|
1172 |
+
|
1173 |
+
if not return_dict:
|
1174 |
+
output = (prediction_scores,) + outputs[1:]
|
1175 |
+
return ((lm_loss,) + output) if lm_loss is not None else output
|
1176 |
+
|
1177 |
+
return CausalLMOutputWithCrossAttentions(
|
1178 |
+
loss=lm_loss,
|
1179 |
+
logits=prediction_scores,
|
1180 |
+
past_key_values=outputs.past_key_values,
|
1181 |
+
hidden_states=outputs.hidden_states,
|
1182 |
+
attentions=outputs.attentions,
|
1183 |
+
cross_attentions=outputs.cross_attentions,
|
1184 |
+
)
|
1185 |
+
|
1186 |
+
def prepare_inputs_for_generation(self, input_ids, past=None, attention_mask=None, **model_kwargs):
|
1187 |
+
input_shape = input_ids.shape
|
1188 |
+
|
1189 |
+
# if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
|
1190 |
+
if attention_mask is None:
|
1191 |
+
attention_mask = input_ids.new_ones(input_shape)
|
1192 |
+
|
1193 |
+
# cut decoder_input_ids if past is used
|
1194 |
+
if past is not None:
|
1195 |
+
input_ids = input_ids[:, -1:]
|
1196 |
+
|
1197 |
+
return {"input_ids": input_ids, "attention_mask": attention_mask, "past_key_values": past}
|
1198 |
+
|
1199 |
+
def _reorder_cache(self, past, beam_idx):
|
1200 |
+
reordered_past = ()
|
1201 |
+
for layer_past in past:
|
1202 |
+
reordered_past += (
|
1203 |
+
tuple(past_state.index_select(0, beam_idx) for past_state in layer_past[:2]) + layer_past[2:],
|
1204 |
+
)
|
1205 |
+
return reordered_past
|
1206 |
+
|
1207 |
+
|
1208 |
+
class RoFormerClassificationHead(nn.Module):
|
1209 |
+
"""Head for sentence-level classification tasks."""
|
1210 |
+
|
1211 |
+
def __init__(self, config):
|
1212 |
+
super().__init__()
|
1213 |
+
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
|
1214 |
+
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
1215 |
+
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
|
1216 |
+
|
1217 |
+
self.config = config
|
1218 |
+
|
1219 |
+
def forward(self, features, **kwargs):
|
1220 |
+
x = features[:, 0, :] # take <s> token (equiv. to [CLS])
|
1221 |
+
x = self.dropout(x)
|
1222 |
+
x = self.dense(x)
|
1223 |
+
x = ACT2FN[self.config.hidden_act](x)
|
1224 |
+
x = self.dropout(x)
|
1225 |
+
x = self.out_proj(x)
|
1226 |
+
return x
|
1227 |
+
|
1228 |
+
|
1229 |
+
@add_start_docstrings(
|
1230 |
+
"""
|
1231 |
+
RoFormer Model transformer with a sequence classification/regression head on top (a linear layer on top of the
|
1232 |
+
pooled output) e.g. for GLUE tasks.
|
1233 |
+
""",
|
1234 |
+
ROFORMER_START_DOCSTRING,
|
1235 |
+
)
|
1236 |
+
class RoFormerForSequenceClassification(RoFormerPreTrainedModel):
|
1237 |
+
def __init__(self, config):
|
1238 |
+
super().__init__(config)
|
1239 |
+
self.num_labels = config.num_labels
|
1240 |
+
self.roformer = RoFormerModel(config)
|
1241 |
+
self.classifier = RoFormerClassificationHead(config)
|
1242 |
+
|
1243 |
+
# Initialize weights and apply final processing
|
1244 |
+
self.post_init()
|
1245 |
+
|
1246 |
+
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
|
1247 |
+
@add_code_sample_docstrings(
|
1248 |
+
processor_class=_TOKENIZER_FOR_DOC,
|
1249 |
+
checkpoint=_CHECKPOINT_FOR_DOC,
|
1250 |
+
output_type=SequenceClassifierOutput,
|
1251 |
+
config_class=_CONFIG_FOR_DOC,
|
1252 |
+
)
|
1253 |
+
def forward(
|
1254 |
+
self,
|
1255 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1256 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1257 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1258 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1259 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1260 |
+
labels: Optional[torch.LongTensor] = None,
|
1261 |
+
output_attentions: Optional[bool] = None,
|
1262 |
+
output_hidden_states: Optional[bool] = None,
|
1263 |
+
return_dict: Optional[bool] = None,
|
1264 |
+
) -> Union[SequenceClassifierOutput, Tuple[torch.Tensor]]:
|
1265 |
+
r"""
|
1266 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
1267 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
1268 |
+
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
1269 |
+
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
1270 |
+
"""
|
1271 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1272 |
+
|
1273 |
+
outputs = self.roformer(
|
1274 |
+
input_ids,
|
1275 |
+
attention_mask=attention_mask,
|
1276 |
+
token_type_ids=token_type_ids,
|
1277 |
+
head_mask=head_mask,
|
1278 |
+
inputs_embeds=inputs_embeds,
|
1279 |
+
output_attentions=output_attentions,
|
1280 |
+
output_hidden_states=output_hidden_states,
|
1281 |
+
return_dict=return_dict,
|
1282 |
+
)
|
1283 |
+
|
1284 |
+
sequence_output = outputs[0]
|
1285 |
+
logits = self.classifier(sequence_output)
|
1286 |
+
|
1287 |
+
loss = None
|
1288 |
+
if labels is not None:
|
1289 |
+
if self.config.problem_type is None:
|
1290 |
+
if self.num_labels == 1:
|
1291 |
+
self.config.problem_type = "regression"
|
1292 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
1293 |
+
self.config.problem_type = "single_label_classification"
|
1294 |
+
else:
|
1295 |
+
self.config.problem_type = "multi_label_classification"
|
1296 |
+
|
1297 |
+
if self.config.problem_type == "regression":
|
1298 |
+
loss_fct = MSELoss()
|
1299 |
+
if self.num_labels == 1:
|
1300 |
+
loss = loss_fct(logits.squeeze(), labels.squeeze())
|
1301 |
+
else:
|
1302 |
+
loss = loss_fct(logits, labels)
|
1303 |
+
elif self.config.problem_type == "single_label_classification":
|
1304 |
+
loss_fct = CrossEntropyLoss()
|
1305 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
1306 |
+
elif self.config.problem_type == "multi_label_classification":
|
1307 |
+
loss_fct = BCEWithLogitsLoss()
|
1308 |
+
loss = loss_fct(logits, labels)
|
1309 |
+
if not return_dict:
|
1310 |
+
output = (logits,) + outputs[1:]
|
1311 |
+
return ((loss,) + output) if loss is not None else output
|
1312 |
+
|
1313 |
+
return SequenceClassifierOutput(
|
1314 |
+
loss=loss,
|
1315 |
+
logits=logits,
|
1316 |
+
hidden_states=outputs.hidden_states,
|
1317 |
+
attentions=outputs.attentions,
|
1318 |
+
)
|
1319 |
+
|
1320 |
+
|
1321 |
+
@add_start_docstrings(
|
1322 |
+
"""
|
1323 |
+
RoFormer Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
|
1324 |
+
softmax) e.g. for RocStories/SWAG tasks.
|
1325 |
+
""",
|
1326 |
+
ROFORMER_START_DOCSTRING,
|
1327 |
+
)
|
1328 |
+
class RoFormerForMultipleChoice(RoFormerPreTrainedModel):
|
1329 |
+
def __init__(self, config):
|
1330 |
+
super().__init__(config)
|
1331 |
+
|
1332 |
+
self.roformer = RoFormerModel(config)
|
1333 |
+
self.sequence_summary = SequenceSummary(config)
|
1334 |
+
self.classifier = nn.Linear(config.hidden_size, 1)
|
1335 |
+
|
1336 |
+
# Initialize weights and apply final processing
|
1337 |
+
self.post_init()
|
1338 |
+
|
1339 |
+
@add_start_docstrings_to_model_forward(
|
1340 |
+
ROFORMER_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
|
1341 |
+
)
|
1342 |
+
@add_code_sample_docstrings(
|
1343 |
+
processor_class=_TOKENIZER_FOR_DOC,
|
1344 |
+
checkpoint=_CHECKPOINT_FOR_DOC,
|
1345 |
+
output_type=MultipleChoiceModelOutput,
|
1346 |
+
config_class=_CONFIG_FOR_DOC,
|
1347 |
+
)
|
1348 |
+
def forward(
|
1349 |
+
self,
|
1350 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1351 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1352 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1353 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1354 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1355 |
+
labels: Optional[torch.LongTensor] = None,
|
1356 |
+
output_attentions: Optional[bool] = None,
|
1357 |
+
output_hidden_states: Optional[bool] = None,
|
1358 |
+
return_dict: Optional[bool] = None,
|
1359 |
+
) -> Union[MultipleChoiceModelOutput, Tuple[torch.Tensor]]:
|
1360 |
+
r"""
|
1361 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
1362 |
+
Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
|
1363 |
+
num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
|
1364 |
+
`input_ids` above)
|
1365 |
+
"""
|
1366 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1367 |
+
num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
|
1368 |
+
|
1369 |
+
input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
|
1370 |
+
attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
|
1371 |
+
token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
|
1372 |
+
|
1373 |
+
inputs_embeds = (
|
1374 |
+
inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
|
1375 |
+
if inputs_embeds is not None
|
1376 |
+
else None
|
1377 |
+
)
|
1378 |
+
|
1379 |
+
outputs = self.roformer(
|
1380 |
+
input_ids,
|
1381 |
+
attention_mask=attention_mask,
|
1382 |
+
token_type_ids=token_type_ids,
|
1383 |
+
head_mask=head_mask,
|
1384 |
+
inputs_embeds=inputs_embeds,
|
1385 |
+
output_attentions=output_attentions,
|
1386 |
+
output_hidden_states=output_hidden_states,
|
1387 |
+
return_dict=return_dict,
|
1388 |
+
)
|
1389 |
+
|
1390 |
+
sequence_output = outputs[0]
|
1391 |
+
|
1392 |
+
pooled_output = self.sequence_summary(sequence_output)
|
1393 |
+
logits = self.classifier(pooled_output)
|
1394 |
+
reshaped_logits = logits.view(-1, num_choices)
|
1395 |
+
|
1396 |
+
loss = None
|
1397 |
+
if labels is not None:
|
1398 |
+
loss_fct = CrossEntropyLoss()
|
1399 |
+
loss = loss_fct(reshaped_logits, labels)
|
1400 |
+
|
1401 |
+
if not return_dict:
|
1402 |
+
output = (reshaped_logits,) + outputs[1:]
|
1403 |
+
return ((loss,) + output) if loss is not None else output
|
1404 |
+
|
1405 |
+
return MultipleChoiceModelOutput(
|
1406 |
+
loss=loss,
|
1407 |
+
logits=reshaped_logits,
|
1408 |
+
hidden_states=outputs.hidden_states,
|
1409 |
+
attentions=outputs.attentions,
|
1410 |
+
)
|
1411 |
+
|
1412 |
+
|
1413 |
+
@add_start_docstrings(
|
1414 |
+
"""
|
1415 |
+
RoFormer Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
|
1416 |
+
Named-Entity-Recognition (NER) tasks.
|
1417 |
+
""",
|
1418 |
+
ROFORMER_START_DOCSTRING,
|
1419 |
+
)
|
1420 |
+
class RoFormerForTokenClassification(RoFormerPreTrainedModel):
|
1421 |
+
def __init__(self, config):
|
1422 |
+
super().__init__(config)
|
1423 |
+
self.num_labels = config.num_labels
|
1424 |
+
|
1425 |
+
self.roformer = RoFormerModel(config)
|
1426 |
+
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
1427 |
+
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
|
1428 |
+
|
1429 |
+
# Initialize weights and apply final processing
|
1430 |
+
self.post_init()
|
1431 |
+
|
1432 |
+
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
|
1433 |
+
@add_code_sample_docstrings(
|
1434 |
+
processor_class=_TOKENIZER_FOR_DOC,
|
1435 |
+
checkpoint=_CHECKPOINT_FOR_DOC,
|
1436 |
+
output_type=TokenClassifierOutput,
|
1437 |
+
config_class=_CONFIG_FOR_DOC,
|
1438 |
+
)
|
1439 |
+
def forward(
|
1440 |
+
self,
|
1441 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1442 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1443 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1444 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1445 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1446 |
+
labels: Optional[torch.LongTensor] = None,
|
1447 |
+
output_attentions: Optional[bool] = None,
|
1448 |
+
output_hidden_states: Optional[bool] = None,
|
1449 |
+
return_dict: Optional[bool] = None,
|
1450 |
+
) -> Union[TokenClassifierOutput, Tuple[torch.Tensor]]:
|
1451 |
+
r"""
|
1452 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
1453 |
+
Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
|
1454 |
+
"""
|
1455 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1456 |
+
|
1457 |
+
outputs = self.roformer(
|
1458 |
+
input_ids,
|
1459 |
+
attention_mask=attention_mask,
|
1460 |
+
token_type_ids=token_type_ids,
|
1461 |
+
head_mask=head_mask,
|
1462 |
+
inputs_embeds=inputs_embeds,
|
1463 |
+
output_attentions=output_attentions,
|
1464 |
+
output_hidden_states=output_hidden_states,
|
1465 |
+
return_dict=return_dict,
|
1466 |
+
)
|
1467 |
+
|
1468 |
+
sequence_output = outputs[0]
|
1469 |
+
|
1470 |
+
sequence_output = self.dropout(sequence_output)
|
1471 |
+
logits = self.classifier(sequence_output)
|
1472 |
+
|
1473 |
+
loss = None
|
1474 |
+
if labels is not None:
|
1475 |
+
loss_fct = CrossEntropyLoss()
|
1476 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
1477 |
+
|
1478 |
+
if not return_dict:
|
1479 |
+
output = (logits,) + outputs[1:]
|
1480 |
+
return ((loss,) + output) if loss is not None else output
|
1481 |
+
|
1482 |
+
return TokenClassifierOutput(
|
1483 |
+
loss=loss,
|
1484 |
+
logits=logits,
|
1485 |
+
hidden_states=outputs.hidden_states,
|
1486 |
+
attentions=outputs.attentions,
|
1487 |
+
)
|
1488 |
+
|
1489 |
+
|
1490 |
+
@add_start_docstrings(
|
1491 |
+
"""
|
1492 |
+
RoFormer Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
|
1493 |
+
layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
|
1494 |
+
""",
|
1495 |
+
ROFORMER_START_DOCSTRING,
|
1496 |
+
)
|
1497 |
+
class RoFormerForQuestionAnswering(RoFormerPreTrainedModel):
|
1498 |
+
def __init__(self, config):
|
1499 |
+
super().__init__(config)
|
1500 |
+
|
1501 |
+
config.num_labels = 2
|
1502 |
+
self.num_labels = config.num_labels
|
1503 |
+
|
1504 |
+
self.roformer = RoFormerModel(config)
|
1505 |
+
self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
|
1506 |
+
|
1507 |
+
# Initialize weights and apply final processing
|
1508 |
+
self.post_init()
|
1509 |
+
|
1510 |
+
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
|
1511 |
+
@add_code_sample_docstrings(
|
1512 |
+
processor_class=_TOKENIZER_FOR_DOC,
|
1513 |
+
checkpoint=_CHECKPOINT_FOR_DOC,
|
1514 |
+
output_type=QuestionAnsweringModelOutput,
|
1515 |
+
config_class=_CONFIG_FOR_DOC,
|
1516 |
+
)
|
1517 |
+
def forward(
|
1518 |
+
self,
|
1519 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1520 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1521 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1522 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1523 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1524 |
+
start_positions: Optional[torch.LongTensor] = None,
|
1525 |
+
end_positions: Optional[torch.LongTensor] = None,
|
1526 |
+
output_attentions: Optional[bool] = None,
|
1527 |
+
output_hidden_states: Optional[bool] = None,
|
1528 |
+
return_dict: Optional[bool] = None,
|
1529 |
+
) -> Union[QuestionAnsweringModelOutput, Tuple[torch.Tensor]]:
|
1530 |
+
r"""
|
1531 |
+
start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
1532 |
+
Labels for position (index) of the start of the labelled span for computing the token classification loss.
|
1533 |
+
Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
|
1534 |
+
are not taken into account for computing the loss.
|
1535 |
+
end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
1536 |
+
Labels for position (index) of the end of the labelled span for computing the token classification loss.
|
1537 |
+
Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
|
1538 |
+
are not taken into account for computing the loss.
|
1539 |
+
"""
|
1540 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1541 |
+
|
1542 |
+
outputs = self.roformer(
|
1543 |
+
input_ids,
|
1544 |
+
attention_mask=attention_mask,
|
1545 |
+
token_type_ids=token_type_ids,
|
1546 |
+
head_mask=head_mask,
|
1547 |
+
inputs_embeds=inputs_embeds,
|
1548 |
+
output_attentions=output_attentions,
|
1549 |
+
output_hidden_states=output_hidden_states,
|
1550 |
+
return_dict=return_dict,
|
1551 |
+
)
|
1552 |
+
|
1553 |
+
sequence_output = outputs[0]
|
1554 |
+
|
1555 |
+
logits = self.qa_outputs(sequence_output)
|
1556 |
+
start_logits, end_logits = logits.split(1, dim=-1)
|
1557 |
+
start_logits = start_logits.squeeze(-1)
|
1558 |
+
end_logits = end_logits.squeeze(-1)
|
1559 |
+
|
1560 |
+
total_loss = None
|
1561 |
+
if start_positions is not None and end_positions is not None:
|
1562 |
+
# If we are on multi-GPU, split add a dimension
|
1563 |
+
if len(start_positions.size()) > 1:
|
1564 |
+
start_positions = start_positions.squeeze(-1)
|
1565 |
+
if len(end_positions.size()) > 1:
|
1566 |
+
end_positions = end_positions.squeeze(-1)
|
1567 |
+
# sometimes the start/end positions are outside our model inputs, we ignore these terms
|
1568 |
+
ignored_index = start_logits.size(1)
|
1569 |
+
start_positions = start_positions.clamp(0, ignored_index)
|
1570 |
+
end_positions = end_positions.clamp(0, ignored_index)
|
1571 |
+
|
1572 |
+
loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
|
1573 |
+
start_loss = loss_fct(start_logits, start_positions)
|
1574 |
+
end_loss = loss_fct(end_logits, end_positions)
|
1575 |
+
total_loss = (start_loss + end_loss) / 2
|
1576 |
+
|
1577 |
+
if not return_dict:
|
1578 |
+
output = (start_logits, end_logits) + outputs[1:]
|
1579 |
+
return ((total_loss,) + output) if total_loss is not None else output
|
1580 |
+
|
1581 |
+
return QuestionAnsweringModelOutput(
|
1582 |
+
loss=total_loss,
|
1583 |
+
start_logits=start_logits,
|
1584 |
+
end_logits=end_logits,
|
1585 |
+
hidden_states=outputs.hidden_states,
|
1586 |
+
attentions=outputs.attentions,
|
1587 |
+
)
|
pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:ae885a57f9cc3d9491bb414c0471026a7695381cb83c719a76f8f0094ad7b44a
|
3 |
+
size 342824309
|