autotransformers.dataset_config
1from dataclasses import dataclass, asdict, field 2from typing import Dict, List, Any 3 4 5@dataclass 6class DatasetConfig: 7 """ 8 Configure a dataset for use within the AutoTrainer class. 9 10 This determines how to load the dataset, 11 whether local files are needed, whether additional splits are needed (for example when the original 12 dataset only has train-test and we want also validation), and so on. 13 14 Parameters 15 ---------- 16 dataset_name: str 17 The name of the dataset. 18 alias: str 19 Alias for the dataset, for saving it. 20 task: str 21 The task of the dataset. Currenlty, only classification, ner and qa (question answering) are available. 22 fixed_training_args: Dict 23 The training arguments (to use in transformers.TrainingArguments) for every model on this dataset, in dictionary format. 24 is_multilabel: bool 25 Whether it is multilabel classification 26 multilabel_label_names: List 27 Names of the labels for multilabel training. 28 hf_load_kwargs: Dict 29 Arguments for loading the dataset from the huggingface datasets' hub. Example: {'path': 'wikiann', 'name': 'es'}. 30 If None, it is assumed that all necessary files exist locally and are passed in the files field. 31 type_load: str 32 The type of load to perform in load_dataset; for example, if your data is in csv format (d = load_dataset('csv', ...)), this should be csv. 33 files: Dict 34 Files to load the dataset from, in Huggingface's datasets format. Possible keys are train, validation and test. 35 data_field: str 36 Field to load data from in the case of jsons loading in datasets. 37 partial_split: bool 38 Wheter a partial split is needed, that is, if you only have train and test sets, this should be True so that a new validation set is created. 39 split: bool 40 This should be true when you only have one split, that is, a big train set; this creates new validation and test sets. 41 label_col: str 42 Name of the label column. 43 val_size: float 44 In case no validation split is provided, the proportion of the training data to leave for validation. 45 test_size: float 46 In case no test split is provided, the proportion of the total data to leave for testing. 47 pre_func 48 Function to perform previous transformations. For example, if your dataset lacks a field (like xquad with title field for example), you can fix it in a function provided here. 49 squad_v2: bool 50 Only useful for question answering. Whether it is squad v2 format or not. Default is false. 51 text_field: str 52 The name of the field containing the text. Useful only in case of unique-text-field datasets,like most datasets are. In case of 2-sentences datasets like xnli or paws-x this is not useful. Default is text. 53 is_2sents: bool 54 Whether it is a 2 sentence dataset. Useful for processing datasets like xnli or paws-x. 55 sentence1_field: str 56 In case this is a 2 sents dataset, the name of the first sentence field. 57 sentence2_field: str 58 In case this is a 2 sents dataset, the name of the second sentence field. 59 summary_field: str = field( 60 The name of the field with summaries (we assume the long texts are in the text_field field). Only useful for summarization tasks. Default is summary. 61 callbacks: List 62 Callbacks to use inside transformers. 63 metric_optimize: str 64 Name of the metric you want to optimize in the hyperparameter search. 65 direction_optimize : str 66 Direction of the optimization problem. Whether you want to maximize or minimize metric_optimize. 67 custom_eval_func: Any 68 In case we want a special evaluation function, we can provide it here. It must receive EvalPredictions by trainer, like any compute_metrics function in transformers. 69 seed : int 70 Seed for optuna sampler. 71 max_length_summary: int 72 Max length of the summaries, for tokenization purposes. It will be changed depending on the ModelConfig. 73 num_proc : int 74 Number of processes to preprocess data. 75 loaded_dataset: Any 76 In case you want to do weird things like concatenating datasets or things like that, you can do that here, by passing a (non-tokenized) dataset in this field. 77 additional_metrics: List 78 List of additional metrics loaded from datasets, to compute over the test part. 79 retrain_at_end: bool 80 whether to retrain with the best performing model. In most cases this should be True, except when training 1 model with 1 set of hyperparams. 81 config_num_labels: int 82 Number of labels to set for the config, if None it will be computed based on number of labels detected. 83 smoke_test: bool 84 Whether to select only top 10 rows of the dataset for smoke testing purposes. 85 augment_data: bool 86 Whether to augment_data or not. 87 data_augmentation_steps: List 88 List of data augmentation techniques to use from NLPAugPipeline. 89 pretokenized_dataset: Any 90 Pre-tokenized dataset, to avoid tokenizing inside AutoTrainer, which may cause memory issues with huge datasets. 91 92 Examples 93 -------- 94 One can easily create a DatasetConfig for dataset conll2002 just with the following: 95 96 >>> from autotransformers import DatasetConfig 97 98 >>> config={'fixed_training_args': {}, 'dataset_name': 'conll2002', 'alias': 'conll2002', 'task': 'ner', 'hf_load_kwargs': {'path': 'conll2002', 'name': 'es'}, 'label_col':'ner_tags'} 99 100 >>> config = DatasetConfig(**config) 101 """ 102 103 dataset_name: str = field(metadata={"help": "The name of the dataset"}) 104 alias: str = field(metadata={"help": "Alias for the dataset, for saving it."}) 105 task: str = field( 106 metadata={ 107 "help": "The task of the dataset. Currenlty, only classification, ner and qa (question answering) are available." 108 } 109 ) 110 fixed_training_args: Dict = field( 111 metadata={ 112 "help": "The training arguments (to use in transformers.TrainingArguments) for every model on this dataset, in dictionary format." 113 } 114 ) 115 is_multilabel: bool = field( 116 default=False, metadata={"help": "Whether it is multilabel classification"} 117 ) 118 multilabel_label_names: List = field( 119 default_factory=list, 120 metadata={"help": "Names of the labels for multilabel training."}, 121 ) 122 hf_load_kwargs: Dict = field( 123 default=None, 124 metadata={ 125 "help": ( 126 "arguments for loading the dataset from the huggingface datasets' hub. Example: {'path': 'wikiann', 'name': 'es'}." 127 "if None, it is assumed that all necessary files exist locally and are passed in the files field." 128 ) 129 }, 130 ) 131 type_load: str = field( 132 default="json", 133 metadata={ 134 "help": "The type of load to perform in load_dataset; for example, if your data is in csv format (d = load_dataset('csv', ...)), this should be csv." 135 }, 136 ) 137 files: Dict = field( 138 default=None, 139 metadata={ 140 "help": "Files to load the dataset from, in Huggingface's datasets format. Possible keys are train, validation and test" 141 }, 142 ) 143 data_field: str = field( 144 default="data", 145 metadata={ 146 "help": "Field to load data from in the case of jsons loading in datasets. " 147 }, 148 ) 149 partial_split: bool = field( 150 default=False, 151 metadata={ 152 "help": "Wheter a partial split is needed, that is, if you only have train and test sets, this should be True so that a new validation set is created." 153 }, 154 ) 155 split: bool = field( 156 default=False, 157 metadata={ 158 "help": "This should be true when you only have one split, that is, a big train set; this creates new validation and test sets." 159 }, 160 ) 161 label_col: str = field( 162 default="label_list", metadata={"help": "Name of the label column."} 163 ) 164 val_size: float = field( 165 default=0.15, 166 metadata={ 167 "help": "In case no validation split is provided, the proportion of the training data to leave for validation." 168 }, 169 ) 170 test_size: float = field( 171 default=0.15, 172 metadata={ 173 "help": "In case no test split is provided, the proportion of the total data to leave for testing." 174 }, 175 ) 176 pre_func: Any = field( 177 default=None, 178 metadata={ 179 "help": "function to perform previous transformations. For example, if your dataset lacks a field (like xquad with title field for example), you can fix it in a function provided here." 180 }, 181 ) 182 remove_fields_pre_func: bool = field( 183 default=False, 184 metadata={"help": "Whether to remove fields after pre_func is applied."}, 185 ) 186 squad_v2: bool = field( 187 default=False, 188 metadata={ 189 "help": "Only useful for question answering. Whether it is squad v2 format or not. Default is false" 190 }, 191 ) 192 text_field: str = field( 193 default="text", 194 metadata={ 195 "help": "The name of the field containing the text. Useful only in case of unique-text-field datasets,like most datasets are. In case of 2-sentences datasets like xnli or paws-x this is not useful." 196 }, 197 ) 198 is_2sents: bool = field( 199 default=False, 200 metadata={ 201 "help": "Whether it is a 2 sentence dataset. Useful for processing datasets like xnli or paws-x." 202 }, 203 ) 204 sentence1_field: str = field( 205 default=None, 206 metadata={ 207 "help": "In case this is a 2 sents dataset, the name of the first sentence field." 208 }, 209 ) 210 sentence2_field: str = field( 211 default=None, 212 metadata={ 213 "help": "In case this is a 2 sents dataset, the name of the second sentence field." 214 }, 215 ) 216 summary_field: str = field( 217 default="summary", 218 metadata={ 219 "help": "The name of the field with summaries (we assume the long texts are in the text_field field). Only useful for summarization tasks." 220 }, 221 ) 222 callbacks: List = field( 223 default_factory=list, metadata={"help": "Callbacks to use inside transformers."} 224 ) 225 metric_optimize: str = field( 226 default="eval_loss", 227 metadata={ 228 "help": "Name of the metric you want to optimize in the hyperparameter search." 229 }, 230 ) 231 direction_optimize: str = field( 232 default="minimize", 233 metadata={ 234 "help": "Direction of the optimization problem. Whether you want to maximize or minimize metric_optimize." 235 }, 236 ) 237 custom_eval_func: Any = field( 238 default=None, 239 metadata={ 240 "help": "In case we want a special evaluation function, we can provide it here. It must receive EvalPredictions by trainer, like any compute_metrics function in transformers." 241 }, 242 ) 243 seed: int = field(default=420, metadata={"help": "Seed for optuna sampler. "}) 244 max_length_summary: int = field( 245 default=120, 246 metadata={ 247 "help": "Max length of the summaries, for tokenization purposes. It will be changed depending on the ModelConfig." 248 }, 249 ) 250 num_proc: int = field( 251 default=4, metadata={"help": "Number of processes to preprocess data."} 252 ) 253 loaded_dataset: Any = field( 254 default=None, 255 metadata={ 256 "help": "In case you want to do weird things like concatenating datasets or things like that, you can do that here, by passing a (non-tokenized) dataset in this field." 257 }, 258 ) 259 additional_metrics: List = field( 260 default=None, 261 metadata={ 262 "help": "List of additional metrics loaded from datasets, to compute over the test part." 263 }, 264 ) 265 retrain_at_end: bool = field( 266 default=True, 267 metadata={ 268 "help": "whether to retrain with the best performing model. In most cases this should be True, except when you're only training 1 model with 1 set of hyperparams." 269 }, 270 ) 271 config_num_labels: int = field( 272 default=None, 273 metadata={ 274 "help": "Number of labels to set for the config, if None it will be computed based on number of labels detected." 275 }, 276 ) 277 smoke_test: bool = field( 278 default=False, 279 metadata={ 280 "help": "Whether to select only top 10 rows of the dataset for smoke testing purposes" 281 }, 282 ) 283 augment_data: bool = field( 284 default=False, metadata={"help": "Whether to augment_data or not."} 285 ) 286 data_augmentation_steps: List = field( 287 default_factory=list, 288 metadata={ 289 "help": "List of data augmentation techniques to use from NLPAugPipeline." 290 }, 291 ) 292 id_field_qa: str = field( 293 default="id", 294 metadata={ 295 "help": "Name of the field with the unique id of the examples in a question answering dataset." 296 }, 297 ) 298 pretokenized_dataset: Any = field( 299 default=None, 300 metadata={ 301 "help": "Pre-tokenized dataset, to avoid tokenizing inside AutoTrainer, which may cause memory issues with huge datasets." 302 }, 303 ) 304 model_config_problem_type: str = field( 305 default=None, 306 metadata={ 307 "help": "Problem type to set for the model's config. This depends on the dataset task." 308 }, 309 ) 310 chat_field: str = field( 311 default="messages", 312 metadata={ 313 "help": "The fieldname for the column in the dataset containing the messages between the user and the assistant." 314 }, 315 ) 316 317 def __str__( 318 self, 319 ): 320 """Representation of dataset config in str.""" 321 self_as_dict = asdict(self) 322 attrs_as_str = [f"{k}={v},\n" for k, v in sorted(self_as_dict.items())] 323 return f"{self.__class__.__name__}(\n{''.join(attrs_as_str)})"
@dataclass
class
DatasetConfig:
6@dataclass 7class DatasetConfig: 8 """ 9 Configure a dataset for use within the AutoTrainer class. 10 11 This determines how to load the dataset, 12 whether local files are needed, whether additional splits are needed (for example when the original 13 dataset only has train-test and we want also validation), and so on. 14 15 Parameters 16 ---------- 17 dataset_name: str 18 The name of the dataset. 19 alias: str 20 Alias for the dataset, for saving it. 21 task: str 22 The task of the dataset. Currenlty, only classification, ner and qa (question answering) are available. 23 fixed_training_args: Dict 24 The training arguments (to use in transformers.TrainingArguments) for every model on this dataset, in dictionary format. 25 is_multilabel: bool 26 Whether it is multilabel classification 27 multilabel_label_names: List 28 Names of the labels for multilabel training. 29 hf_load_kwargs: Dict 30 Arguments for loading the dataset from the huggingface datasets' hub. Example: {'path': 'wikiann', 'name': 'es'}. 31 If None, it is assumed that all necessary files exist locally and are passed in the files field. 32 type_load: str 33 The type of load to perform in load_dataset; for example, if your data is in csv format (d = load_dataset('csv', ...)), this should be csv. 34 files: Dict 35 Files to load the dataset from, in Huggingface's datasets format. Possible keys are train, validation and test. 36 data_field: str 37 Field to load data from in the case of jsons loading in datasets. 38 partial_split: bool 39 Wheter a partial split is needed, that is, if you only have train and test sets, this should be True so that a new validation set is created. 40 split: bool 41 This should be true when you only have one split, that is, a big train set; this creates new validation and test sets. 42 label_col: str 43 Name of the label column. 44 val_size: float 45 In case no validation split is provided, the proportion of the training data to leave for validation. 46 test_size: float 47 In case no test split is provided, the proportion of the total data to leave for testing. 48 pre_func 49 Function to perform previous transformations. For example, if your dataset lacks a field (like xquad with title field for example), you can fix it in a function provided here. 50 squad_v2: bool 51 Only useful for question answering. Whether it is squad v2 format or not. Default is false. 52 text_field: str 53 The name of the field containing the text. Useful only in case of unique-text-field datasets,like most datasets are. In case of 2-sentences datasets like xnli or paws-x this is not useful. Default is text. 54 is_2sents: bool 55 Whether it is a 2 sentence dataset. Useful for processing datasets like xnli or paws-x. 56 sentence1_field: str 57 In case this is a 2 sents dataset, the name of the first sentence field. 58 sentence2_field: str 59 In case this is a 2 sents dataset, the name of the second sentence field. 60 summary_field: str = field( 61 The name of the field with summaries (we assume the long texts are in the text_field field). Only useful for summarization tasks. Default is summary. 62 callbacks: List 63 Callbacks to use inside transformers. 64 metric_optimize: str 65 Name of the metric you want to optimize in the hyperparameter search. 66 direction_optimize : str 67 Direction of the optimization problem. Whether you want to maximize or minimize metric_optimize. 68 custom_eval_func: Any 69 In case we want a special evaluation function, we can provide it here. It must receive EvalPredictions by trainer, like any compute_metrics function in transformers. 70 seed : int 71 Seed for optuna sampler. 72 max_length_summary: int 73 Max length of the summaries, for tokenization purposes. It will be changed depending on the ModelConfig. 74 num_proc : int 75 Number of processes to preprocess data. 76 loaded_dataset: Any 77 In case you want to do weird things like concatenating datasets or things like that, you can do that here, by passing a (non-tokenized) dataset in this field. 78 additional_metrics: List 79 List of additional metrics loaded from datasets, to compute over the test part. 80 retrain_at_end: bool 81 whether to retrain with the best performing model. In most cases this should be True, except when training 1 model with 1 set of hyperparams. 82 config_num_labels: int 83 Number of labels to set for the config, if None it will be computed based on number of labels detected. 84 smoke_test: bool 85 Whether to select only top 10 rows of the dataset for smoke testing purposes. 86 augment_data: bool 87 Whether to augment_data or not. 88 data_augmentation_steps: List 89 List of data augmentation techniques to use from NLPAugPipeline. 90 pretokenized_dataset: Any 91 Pre-tokenized dataset, to avoid tokenizing inside AutoTrainer, which may cause memory issues with huge datasets. 92 93 Examples 94 -------- 95 One can easily create a DatasetConfig for dataset conll2002 just with the following: 96 97 >>> from autotransformers import DatasetConfig 98 99 >>> config={'fixed_training_args': {}, 'dataset_name': 'conll2002', 'alias': 'conll2002', 'task': 'ner', 'hf_load_kwargs': {'path': 'conll2002', 'name': 'es'}, 'label_col':'ner_tags'} 100 101 >>> config = DatasetConfig(**config) 102 """ 103 104 dataset_name: str = field(metadata={"help": "The name of the dataset"}) 105 alias: str = field(metadata={"help": "Alias for the dataset, for saving it."}) 106 task: str = field( 107 metadata={ 108 "help": "The task of the dataset. Currenlty, only classification, ner and qa (question answering) are available." 109 } 110 ) 111 fixed_training_args: Dict = field( 112 metadata={ 113 "help": "The training arguments (to use in transformers.TrainingArguments) for every model on this dataset, in dictionary format." 114 } 115 ) 116 is_multilabel: bool = field( 117 default=False, metadata={"help": "Whether it is multilabel classification"} 118 ) 119 multilabel_label_names: List = field( 120 default_factory=list, 121 metadata={"help": "Names of the labels for multilabel training."}, 122 ) 123 hf_load_kwargs: Dict = field( 124 default=None, 125 metadata={ 126 "help": ( 127 "arguments for loading the dataset from the huggingface datasets' hub. Example: {'path': 'wikiann', 'name': 'es'}." 128 "if None, it is assumed that all necessary files exist locally and are passed in the files field." 129 ) 130 }, 131 ) 132 type_load: str = field( 133 default="json", 134 metadata={ 135 "help": "The type of load to perform in load_dataset; for example, if your data is in csv format (d = load_dataset('csv', ...)), this should be csv." 136 }, 137 ) 138 files: Dict = field( 139 default=None, 140 metadata={ 141 "help": "Files to load the dataset from, in Huggingface's datasets format. Possible keys are train, validation and test" 142 }, 143 ) 144 data_field: str = field( 145 default="data", 146 metadata={ 147 "help": "Field to load data from in the case of jsons loading in datasets. " 148 }, 149 ) 150 partial_split: bool = field( 151 default=False, 152 metadata={ 153 "help": "Wheter a partial split is needed, that is, if you only have train and test sets, this should be True so that a new validation set is created." 154 }, 155 ) 156 split: bool = field( 157 default=False, 158 metadata={ 159 "help": "This should be true when you only have one split, that is, a big train set; this creates new validation and test sets." 160 }, 161 ) 162 label_col: str = field( 163 default="label_list", metadata={"help": "Name of the label column."} 164 ) 165 val_size: float = field( 166 default=0.15, 167 metadata={ 168 "help": "In case no validation split is provided, the proportion of the training data to leave for validation." 169 }, 170 ) 171 test_size: float = field( 172 default=0.15, 173 metadata={ 174 "help": "In case no test split is provided, the proportion of the total data to leave for testing." 175 }, 176 ) 177 pre_func: Any = field( 178 default=None, 179 metadata={ 180 "help": "function to perform previous transformations. For example, if your dataset lacks a field (like xquad with title field for example), you can fix it in a function provided here." 181 }, 182 ) 183 remove_fields_pre_func: bool = field( 184 default=False, 185 metadata={"help": "Whether to remove fields after pre_func is applied."}, 186 ) 187 squad_v2: bool = field( 188 default=False, 189 metadata={ 190 "help": "Only useful for question answering. Whether it is squad v2 format or not. Default is false" 191 }, 192 ) 193 text_field: str = field( 194 default="text", 195 metadata={ 196 "help": "The name of the field containing the text. Useful only in case of unique-text-field datasets,like most datasets are. In case of 2-sentences datasets like xnli or paws-x this is not useful." 197 }, 198 ) 199 is_2sents: bool = field( 200 default=False, 201 metadata={ 202 "help": "Whether it is a 2 sentence dataset. Useful for processing datasets like xnli or paws-x." 203 }, 204 ) 205 sentence1_field: str = field( 206 default=None, 207 metadata={ 208 "help": "In case this is a 2 sents dataset, the name of the first sentence field." 209 }, 210 ) 211 sentence2_field: str = field( 212 default=None, 213 metadata={ 214 "help": "In case this is a 2 sents dataset, the name of the second sentence field." 215 }, 216 ) 217 summary_field: str = field( 218 default="summary", 219 metadata={ 220 "help": "The name of the field with summaries (we assume the long texts are in the text_field field). Only useful for summarization tasks." 221 }, 222 ) 223 callbacks: List = field( 224 default_factory=list, metadata={"help": "Callbacks to use inside transformers."} 225 ) 226 metric_optimize: str = field( 227 default="eval_loss", 228 metadata={ 229 "help": "Name of the metric you want to optimize in the hyperparameter search." 230 }, 231 ) 232 direction_optimize: str = field( 233 default="minimize", 234 metadata={ 235 "help": "Direction of the optimization problem. Whether you want to maximize or minimize metric_optimize." 236 }, 237 ) 238 custom_eval_func: Any = field( 239 default=None, 240 metadata={ 241 "help": "In case we want a special evaluation function, we can provide it here. It must receive EvalPredictions by trainer, like any compute_metrics function in transformers." 242 }, 243 ) 244 seed: int = field(default=420, metadata={"help": "Seed for optuna sampler. "}) 245 max_length_summary: int = field( 246 default=120, 247 metadata={ 248 "help": "Max length of the summaries, for tokenization purposes. It will be changed depending on the ModelConfig." 249 }, 250 ) 251 num_proc: int = field( 252 default=4, metadata={"help": "Number of processes to preprocess data."} 253 ) 254 loaded_dataset: Any = field( 255 default=None, 256 metadata={ 257 "help": "In case you want to do weird things like concatenating datasets or things like that, you can do that here, by passing a (non-tokenized) dataset in this field." 258 }, 259 ) 260 additional_metrics: List = field( 261 default=None, 262 metadata={ 263 "help": "List of additional metrics loaded from datasets, to compute over the test part." 264 }, 265 ) 266 retrain_at_end: bool = field( 267 default=True, 268 metadata={ 269 "help": "whether to retrain with the best performing model. In most cases this should be True, except when you're only training 1 model with 1 set of hyperparams." 270 }, 271 ) 272 config_num_labels: int = field( 273 default=None, 274 metadata={ 275 "help": "Number of labels to set for the config, if None it will be computed based on number of labels detected." 276 }, 277 ) 278 smoke_test: bool = field( 279 default=False, 280 metadata={ 281 "help": "Whether to select only top 10 rows of the dataset for smoke testing purposes" 282 }, 283 ) 284 augment_data: bool = field( 285 default=False, metadata={"help": "Whether to augment_data or not."} 286 ) 287 data_augmentation_steps: List = field( 288 default_factory=list, 289 metadata={ 290 "help": "List of data augmentation techniques to use from NLPAugPipeline." 291 }, 292 ) 293 id_field_qa: str = field( 294 default="id", 295 metadata={ 296 "help": "Name of the field with the unique id of the examples in a question answering dataset." 297 }, 298 ) 299 pretokenized_dataset: Any = field( 300 default=None, 301 metadata={ 302 "help": "Pre-tokenized dataset, to avoid tokenizing inside AutoTrainer, which may cause memory issues with huge datasets." 303 }, 304 ) 305 model_config_problem_type: str = field( 306 default=None, 307 metadata={ 308 "help": "Problem type to set for the model's config. This depends on the dataset task." 309 }, 310 ) 311 chat_field: str = field( 312 default="messages", 313 metadata={ 314 "help": "The fieldname for the column in the dataset containing the messages between the user and the assistant." 315 }, 316 ) 317 318 def __str__( 319 self, 320 ): 321 """Representation of dataset config in str.""" 322 self_as_dict = asdict(self) 323 attrs_as_str = [f"{k}={v},\n" for k, v in sorted(self_as_dict.items())] 324 return f"{self.__class__.__name__}(\n{''.join(attrs_as_str)})"
Configure a dataset for use within the AutoTrainer class.
This determines how to load the dataset, whether local files are needed, whether additional splits are needed (for example when the original dataset only has train-test and we want also validation), and so on.
Parameters
dataset_name: str The name of the dataset. alias: str Alias for the dataset, for saving it. task: str The task of the dataset. Currenlty, only classification, ner and qa (question answering) are available. fixed_training_args: Dict The training arguments (to use in transformers.TrainingArguments) for every model on this dataset, in dictionary format. is_multilabel: bool Whether it is multilabel classification multilabel_label_names: List Names of the labels for multilabel training. hf_load_kwargs: Dict Arguments for loading the dataset from the huggingface datasets' hub. Example: {'path': 'wikiann', 'name': 'es'}. If None, it is assumed that all necessary files exist locally and are passed in the files field. type_load: str The type of load to perform in load_dataset; for example, if your data is in csv format (d = load_dataset('csv', ...)), this should be csv. files: Dict Files to load the dataset from, in Huggingface's datasets format. Possible keys are train, validation and test. data_field: str Field to load data from in the case of jsons loading in datasets. partial_split: bool Wheter a partial split is needed, that is, if you only have train and test sets, this should be True so that a new validation set is created. split: bool This should be true when you only have one split, that is, a big train set; this creates new validation and test sets. label_col: str Name of the label column. val_size: float In case no validation split is provided, the proportion of the training data to leave for validation. test_size: float In case no test split is provided, the proportion of the total data to leave for testing. pre_func Function to perform previous transformations. For example, if your dataset lacks a field (like xquad with title field for example), you can fix it in a function provided here. squad_v2: bool Only useful for question answering. Whether it is squad v2 format or not. Default is false. text_field: str The name of the field containing the text. Useful only in case of unique-text-field datasets,like most datasets are. In case of 2-sentences datasets like xnli or paws-x this is not useful. Default is text. is_2sents: bool Whether it is a 2 sentence dataset. Useful for processing datasets like xnli or paws-x. sentence1_field: str In case this is a 2 sents dataset, the name of the first sentence field. sentence2_field: str In case this is a 2 sents dataset, the name of the second sentence field. summary_field: str = field( The name of the field with summaries (we assume the long texts are in the text_field field). Only useful for summarization tasks. Default is summary. callbacks: List Callbacks to use inside transformers. metric_optimize: str Name of the metric you want to optimize in the hyperparameter search. direction_optimize : str Direction of the optimization problem. Whether you want to maximize or minimize metric_optimize. custom_eval_func: Any In case we want a special evaluation function, we can provide it here. It must receive EvalPredictions by trainer, like any compute_metrics function in transformers. seed : int Seed for optuna sampler. max_length_summary: int Max length of the summaries, for tokenization purposes. It will be changed depending on the ModelConfig. num_proc : int Number of processes to preprocess data. loaded_dataset: Any In case you want to do weird things like concatenating datasets or things like that, you can do that here, by passing a (non-tokenized) dataset in this field. additional_metrics: List List of additional metrics loaded from datasets, to compute over the test part. retrain_at_end: bool whether to retrain with the best performing model. In most cases this should be True, except when training 1 model with 1 set of hyperparams. config_num_labels: int Number of labels to set for the config, if None it will be computed based on number of labels detected. smoke_test: bool Whether to select only top 10 rows of the dataset for smoke testing purposes. augment_data: bool Whether to augment_data or not. data_augmentation_steps: List List of data augmentation techniques to use from NLPAugPipeline. pretokenized_dataset: Any Pre-tokenized dataset, to avoid tokenizing inside AutoTrainer, which may cause memory issues with huge datasets.
Examples
One can easily create a DatasetConfig for dataset conll2002 just with the following:
>>> from autotransformers import DatasetConfig
>>> config={'fixed_training_args': {}, 'dataset_name': 'conll2002', 'alias': 'conll2002', 'task': 'ner', 'hf_load_kwargs': {'path': 'conll2002', 'name': 'es'}, 'label_col':'ner_tags'}
>>> config = DatasetConfig(**config)
DatasetConfig( dataset_name: str, alias: str, task: str, fixed_training_args: Dict, is_multilabel: bool = False, multilabel_label_names: List = <factory>, hf_load_kwargs: Dict = None, type_load: str = 'json', files: Dict = None, data_field: str = 'data', partial_split: bool = False, split: bool = False, label_col: str = 'label_list', val_size: float = 0.15, test_size: float = 0.15, pre_func: Any = None, remove_fields_pre_func: bool = False, squad_v2: bool = False, text_field: str = 'text', is_2sents: bool = False, sentence1_field: str = None, sentence2_field: str = None, summary_field: str = 'summary', callbacks: List = <factory>, metric_optimize: str = 'eval_loss', direction_optimize: str = 'minimize', custom_eval_func: Any = None, seed: int = 420, max_length_summary: int = 120, num_proc: int = 4, loaded_dataset: Any = None, additional_metrics: List = None, retrain_at_end: bool = True, config_num_labels: int = None, smoke_test: bool = False, augment_data: bool = False, data_augmentation_steps: List = <factory>, id_field_qa: str = 'id', pretokenized_dataset: Any = None, model_config_problem_type: str = None, chat_field: str = 'messages')