File size: 12,316 Bytes
e7d3e35
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" GPT Neo model configuration"""
import os
from typing import Tuple, Union

from transformers import AutoConfig
from transformers.configuration_utils import PretrainedConfig
from transformers.utils import logging


logger = logging.get_logger(__name__)

GPT_NEO_PRETRAINED_CONFIG_ARCHIVE_MAP = {
    "EleutherAI/gpt-neo-125M": "https://huggingface.co/EleutherAI/gpt-neo-125M/resolve/main/config.json",
    "EleutherAI/gpt-neo-1.3B": "https://huggingface.co/EleutherAI/gpt-neo-1.3B/resolve/main/config.json",
    # See all GPTNeo models at https://huggingface.co/models?filter=gpt_neo
}


class VGPTNeoConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`GPTNeoModel`]. It is used to instantiate a GPT
    Neo model according to the specified arguments, defining the model architecture. Instantiating a configuration with
    the defaults will yield a similar configuration to that of the GPTNeo
    [EleutherAI/gpt-neo-1.3B](https://huggingface.co/EleutherAI/gpt-neo-1.3B) architecture.

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.

    TODO: this doc is completely out of sync with the actual args

    Args:
        vocab_size (`int`, *optional*, defaults to 50257):
            Vocabulary size of the GPT Neo model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`GPTNeoModel`]. Vocabulary size of the model. Defines the different
            tokens that can be represented by the *inputs_ids* passed to the forward method of [`GPTNeoModel`].
        additional_vocab_size (`int`, *optional`, defaults to 0):
            Additional vocabulary size of the model, typically for the special "<img>" token. Additional vocab tokens
            are always trainable whereas regular vocab tokens can be frozen or not.
        attention_types (`List`, *optional*, defaults to `[[["global", "local"], 12]]`):
            The type of attention for each layer in a `List` of the following format `[[["attention_type"],
            num_layerss]]` e.g. for a 24 layer model `[[["global"], 24]]` or `[[["global", "local"], 12]]` Choose the
            value of `attention_type` from `["global", "local"]`
        hidden_size (`int`, *optional*, defaults to 2048):
            Dimensionality of the encoder layers and the pooler layer.
        num_layers (`int`, *optional*, defaults to 24):
            Number of hidden layers in the Transformer encoder.
        num_heads (`int`, *optional*, defaults to 16):
            Number of attention heads for each attention layer in the Transformer encoder.
        intermediate_size (`int`, *optional*, defaults to 8192):
            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
        activation_function (`str` or `function`, *optional*, defaults to `"gelu_new"`):
            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
            `"relu"`, `"selu"` and `"gelu_new"` are supported.
        embed_dropout (`float`, *optional*, defaults to 0.0):
            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        max_position_embeddings (`int`, *optional*, defaults to 2048):
            The maximum sequence length that this model might ever be used with. Typically set this to something large
            just in case (e.g., 512 or 1024 or 2048).
        type_vocab_size (`int`, *optional*, defaults to 2):
            The vocabulary size of the `token_type_ids` passed when calling [`GPTNeoModel`].
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        alpha_initializer (`str`, *optional*, defaults to `"ones"`):
            Initialization type for the alphas.
        alphas_initializer_range (`float`, *optional*, defaults to 0.0):
            The standard deviation of the truncated_normal_initializer for initializing the alphas in the Gated Cross Attention.
        alpha_type (`str`, *optional*, defaults to `"vector"`):
            Whether the gating alphas should be vectors or single floats.
        layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
            The epsilon used by the layer normalization layers.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models). Only
            relevant if `config.is_decoder=True`.
        cross_layer_interval (`int`, *optional*, default to 1)
            Interval for cross attention (from text to image) layers.
    Example:
    ```python
    >>> from transformers import GPTNeoConfig, GPTNeoModel
    >>> # Initializing a GPTNeo EleutherAI/gpt-neo-1.3B style configuration
    >>> configuration = GPTNeoConfig()
    >>> # Initializing a model (with random weights) from the EleutherAI/gpt-neo-1.3B style configuration
    >>> model = GPTNeoModel(configuration)
    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""
    model_type = "vgpt_neo"
    keys_to_ignore_at_inference = ["past_key_values"]
    attribute_map = {"num_attention_heads": "num_heads", "num_hidden_layers": "num_layers"}

    def __init__(
        self,
        vocab_size=50257,
        additional_vocab_size=0,
        max_position_embeddings=2048,
        hidden_size=2048,
        num_layers=24,
        attention_types=[[["global", "local"], 12]],
        num_heads=16,
        intermediate_size=None,
        window_size=256,
        activation_function="gelu_new",
        resid_dropout=0.0,
        embed_dropout=0.0,
        attention_dropout=0.0,
        layer_norm_epsilon=1e-5,
        initializer_range=0.02,
        alpha_initializer="ones",
        alphas_initializer_range=0.0,
        alpha_type="vector",
        summary_type="cls_index",
        summary_use_proj=True,
        summary_activation=None,
        summary_proj_to_labels=True,
        summary_first_dropout=0.1,
        use_cache=True,
        bos_token_id=50256,
        eos_token_id=50256,
        cross_layer_interval=1,
        tie_word_embeddings=False,
        freeze_text_layers=True,
        freeze_lm_head=False,
        freeze_vision_layers=True,
        vision_model_name="google/vit-base-patch16-224",
        vision_model_params="{}",
        vision_embed_dim=768,
        vision_image_size=224,
        image_token_index=50257,
        use_resampler=False,
        resampler_n_latents=64,
        resampler_depth=6,
        resampler_n_heads=16,
        resampler_head_dim=96,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.additional_vocab_size = additional_vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.num_layers = num_layers
        self.num_heads = num_heads
        self.intermediate_size = intermediate_size
        self.window_size = window_size
        self.activation_function = activation_function
        self.resid_dropout = resid_dropout
        self.embed_dropout = embed_dropout
        self.attention_dropout = attention_dropout
        self.layer_norm_epsilon = layer_norm_epsilon
        self.initializer_range = initializer_range
        self.alpha_initializer = alpha_initializer
        self.alphas_initializer_range = alphas_initializer_range
        self.alpha_type = alpha_type
        self.summary_type = summary_type
        self.summary_use_proj = summary_use_proj
        self.summary_activation = summary_activation
        self.summary_first_dropout = summary_first_dropout
        self.summary_proj_to_labels = summary_proj_to_labels
        self.use_cache = use_cache

        self.bos_token_id = bos_token_id
        self.eos_token_id = eos_token_id
        self.cross_layer_interval = cross_layer_interval
        self.freeze_vision_layers = freeze_vision_layers
        self.vision_model_name = vision_model_name
        self.vision_model_params = vision_model_params

        self.tie_word_embeddings = tie_word_embeddings
        self.freeze_text_layers = freeze_text_layers
        self.freeze_lm_head = freeze_lm_head
        self.image_token_index = image_token_index
        self.attention_types = attention_types
        self.attention_layers = self.expand_attention_types_params(attention_types)

        self.vision_embed_dim = vision_embed_dim
        self.vision_image_size = vision_image_size

        # Resampler params
        self.use_resampler = use_resampler
        self.resampler_n_latents = resampler_n_latents
        self.resampler_depth = resampler_depth
        self.resampler_n_heads = resampler_n_heads
        self.resampler_head_dim = resampler_head_dim

        # IMPORTANT: Do not do any __init__ args-based checks in the constructor, since
        # PretrainedConfig.from_dict first instantiates the class with the config dict and only then
        # updates the config object with `kwargs` from from_pretrained, so during the instantiation
        # of this object many attributes have default values and haven't yet been overridden.
        # Do any required checks inside `from_pretrained` once the superclass' `from_pretrained` was run.

        super().__init__(
            bos_token_id=bos_token_id, eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs
        )

    def check_compatibilities(self):
        if self.tie_word_embeddings and (self.freeze_text_layers != self.freeze_lm_head):
            raise ValueError(
                "if `tie_word_embeddings` is True, then `freeze_lm_head` and `freeze_text_layers` must be equal."
            )

        vision_model_params = eval(self.vision_model_params)
        config = AutoConfig.from_pretrained(self.vision_model_name, **vision_model_params)
        if hasattr(config, "vision_config"):
            vision_config = config.vision_config
        else:
            vision_config = config
        vision_embed_dim = vision_config.hidden_size
        if self.vision_embed_dim != vision_embed_dim:
            raise ValueError(
                f"vision_embed_dim ({self.vision_embed_dim}) must match the hidden size of the vision model"
                f" ({vision_embed_dim})"
            )
        vision_image_size = vision_config.image_size
        if self.vision_image_size != vision_image_size:
            raise ValueError(
                f"vision_image_size ({self.vision_image_size}) must match the hidden size of the vision model"
                f" ({vision_image_size})"
            )

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
        outputs = super(VGPTNeoConfig, cls).from_pretrained(pretrained_model_name_or_path, **kwargs)
        if isinstance(outputs, Tuple):
            # When called with return_unused_kwargs=True, the first item will be the config
            outputs[0].check_compatibilities()
        else:
            outputs.check_compatibilities()
        return outputs

    @staticmethod
    def expand_attention_types_params(attention_types):
        attentions = []
        for item in attention_types:
            for _ in range(item[1]):
                attentions.extend(item[0])
        return attentions