File size: 8,768 Bytes
14c75cc |
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 |
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team.
#
# 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.
"""
Processor class for MiniCPMV.
"""
from typing import List, Optional, Union, Dict, Any
import torch
import re
from transformers.image_processing_utils import BatchFeature
from transformers.image_utils import ImageInput
from transformers.processing_utils import ProcessorMixin
from transformers.tokenization_utils_base import PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
from transformers.utils import TensorType, requires_backends, is_torch_dtype, is_torch_device
from .image_processing_minicpmv import MiniCPMVBatchFeature
class MiniCPMVProcessor(ProcessorMixin):
r"""
Constructs a MiniCPMV processor which wraps a MiniCPMV image processor and a MiniCPMV tokenizer into a single processor.
[`MiniCPMVProcessor`] offers all the functionalities of [`MiniCPMVImageProcessor`] and [`LlamaTokenizerWrapper`]. See the
[`~MiniCPMVProcessor.__call__`] and [`~MiniCPMVProcessor.decode`] for more information.
Args:
image_processor ([`MiniCPMVImageProcessor`], *optional*):
The image processor is a required input.
tokenizer ([`LlamaTokenizerWrapper`], *optional*):
The tokenizer is a required input.
"""
attributes = ["image_processor", "tokenizer"]
image_processor_class = "AutoImageProcessor"
tokenizer_class = "AutoTokenizer"
def __init__(self, image_processor=None, tokenizer=None):
super().__init__(image_processor, tokenizer)
self.version = image_processor.version
def __call__(
self,
images: ImageInput,
max_length: Optional[int] = None,
do_pad: Optional[bool] = True,
max_slice_nums: int = None,
use_image_id: bool = None,
return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
**kwargs
) -> MiniCPMVBatchFeature:
if images is not None:
image_inputs = self.image_processor(images, do_pad=do_pad, max_slice_nums=max_slice_nums, return_tensors=return_tensors)
return self._convert_images_texts_to_inputs(image_inputs, max_slice_nums=max_slice_nums, use_image_id=use_image_id, max_length=max_length, **kwargs)
# Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->Llama
def batch_decode(self, *args, **kwargs):
"""
This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
refer to the docstring of this method for more information.
"""
output_ids = args[0]
result_text = []
for result in output_ids:
result = result[result != 0]
if result[0] == self.tokenizer.bos_id:
result = result[1:]
if result[-1] == self.tokenizer.eos_id:
result = result[:-1]
result_text.append(self.tokenizer.decode(result, *args[1:], **kwargs).strip())
return result_text
# return self.tokenizer.batch_decode(*args, **kwargs)
# Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode with CLIP->Llama
def decode(self, *args, **kwargs):
"""
This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
the docstring of this method for more information.
"""
result = args[0]
result = result[result != 0]
if result[0] == self.tokenizer.bos_id:
result = result[1:]
if result[-1] == self.tokenizer.eos_id or (hasattr(self.tokenizer, "eot_id") and result[-1] == self.tokenizer.eot_id):
result = result[:-1]
return self.tokenizer.decode(result, *args[1:], **kwargs).strip()
def _convert(
self, input_str, max_inp_length: Optional[int] = None
):
if self.version > 2.5 or not getattr(self.tokenizer, "add_bos_token", False):
input_ids = self.tokenizer.encode(input_str)
else:
input_ids = [self.tokenizer.bos_id] + self.tokenizer.encode(input_str)
if max_inp_length is not None:
input_ids = input_ids[:max_inp_length]
input_ids = torch.tensor(input_ids, dtype=torch.int32)
start_cond = (input_ids == self.tokenizer.im_start_id) | (input_ids == self.tokenizer.slice_start_id)
end_cond = (input_ids == self.tokenizer.im_end_id) | (input_ids == self.tokenizer.slice_end_id)
image_start_tokens = torch.where(start_cond)[0]
image_start_tokens += 1
image_end_tokens = torch.where(end_cond)[0]
valid_image_nums = max(len(image_start_tokens), len(image_end_tokens))
image_bounds = torch.hstack(
[
image_start_tokens[:valid_image_nums].unsqueeze(-1),
image_end_tokens[:valid_image_nums].unsqueeze(-1),
]
)
return input_ids, image_bounds
def _convert_images_texts_to_inputs(
self,
images,
truncation=None,
max_length=None,
max_slice_nums=None,
use_image_id=None,
return_tensors=None,
**kwargs
):
pattern = "(<image>./</image>)"
images, image_sizes, tgt_sizes = images["pixel_values"], images["image_sizes"], images["tgt_sizes"]
input_ids_list = []
image_bounds_list = []
padded_input_ids, padding_lengths = self.pad(
input_ids_list,
padding_side="left"
)
for i, length in enumerate(padding_lengths):
image_bounds_list[i] = image_bounds_list[i] + length
attention_mask = padded_input_ids.ne(0)
return MiniCPMVBatchFeature(data={
"input_ids": padded_input_ids,
"attention_mask": attention_mask,
"pixel_values": images,
"image_sizes": image_sizes,
"image_bound": image_bounds_list,
"tgt_sizes": tgt_sizes
})
@property
# Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names
def model_input_names(self):
tokenizer_input_names = self.tokenizer.model_input_names
image_processor_input_names = self.image_processor.model_input_names
return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
def pad(self, inputs, max_length=None, padding_value=0, padding_side="left"):
items = []
if isinstance(inputs[0], list):
assert isinstance(inputs[0][0], torch.Tensor)
for it in inputs:
for tr in it:
items.append(tr)
else:
assert isinstance(inputs[0], torch.Tensor)
items = inputs
batch_size = len(items)
shape = items[0].shape
dim = len(shape)
assert dim <= 2
if max_length is None:
max_length = 0
max_length = max(max_length, max(item.shape[-1] for item in items))
min_length = min(item.shape[-1] for item in items)
dtype = items[0].dtype
if dim == 0:
return torch.stack([item for item in items], dim=0), [0]
elif dim == 1:
if max_length == min_length:
return torch.stack([item for item in items], dim=0), [0] * batch_size
tensor = torch.zeros((batch_size, max_length), dtype=dtype) + padding_value
else:
tensor = (
torch.zeros((batch_size, max_length, shape[-1]), dtype=dtype)
+ padding_value
)
padding_length = []
for i, item in enumerate(items):
if dim == 1:
if padding_side == "left":
tensor[i, -len(item) :] = item.clone()
else:
tensor[i, : len(item)] = item.clone()
elif dim == 2:
if padding_side == "left":
tensor[i, -len(item) :, :] = item.clone()
else:
tensor[i, : len(item), :] = item.clone()
padding_length.append(tensor.shape[-1] - len(item))
return tensor, padding_length
|