pere commited on
Commit
c539a6e
1 Parent(s): 801f7b3

moved model from pere/nb-bert-roberta-base-scandinavian-long

Browse files
README.md ADDED
@@ -0,0 +1,68 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # This is just a Test Model. Do NOT use for anything!
2
+
3
+ Continued pretrained from the nb-roberta-base.
4
+
5
+ The domain specific pretraining is done on the 102GB (Scandinavian corpus)[https://huggingface.co/datasets/NbAiLab/scandinavian].
6
+
7
+ ## Train for 180k steps for 128 sequences:
8
+ ```bash
9
+ ./run_mlm_flax_stream.py \
10
+ --output_dir="./" \
11
+ --model_type="roberta" \
12
+ --config_name="./" \
13
+ --tokenizer_name="./" \
14
+ --model_name_or_path="./" \
15
+ --dataset_name="NbAiLab/scandinavian" \
16
+ --max_seq_length="128" \
17
+ --weight_decay="0.01" \
18
+ --per_device_train_batch_size="128" \
19
+ --per_device_eval_batch_size="128" \
20
+ --learning_rate="6e-5" \
21
+ --warmup_steps="5000" \
22
+ --overwrite_output_dir \
23
+ --cache_dir /mnt/disks/flaxdisk/cache/ \
24
+ --num_train_steps="180000" \
25
+ --adam_beta1="0.9" \
26
+ --adam_beta2="0.98" \
27
+ --logging_steps="10000" \
28
+ --save_steps="10000" \
29
+ --eval_steps="10000" \
30
+ --preprocessing_num_workers 96 \
31
+ --auth_token True \
32
+ --adafactor \
33
+ --push_to_hub
34
+ ```
35
+ ## Train for 20k steps for 512 sequences:
36
+ ```bash
37
+ ./run_mlm_flax_stream.py \
38
+ --output_dir="./" \
39
+ --model_type="roberta" \
40
+ --config_name="./" \
41
+ --tokenizer_name="./" \
42
+ --model_name_or_path="./" \
43
+ --dataset_name="NbAiLab/scandinavian" \
44
+ --max_seq_length="512" \
45
+ --weight_decay="0.01" \
46
+ --per_device_train_batch_size="48" \
47
+ --per_device_eval_batch_size="48" \
48
+ --learning_rate="3e-5" \
49
+ --warmup_steps="5000" \
50
+ --overwrite_output_dir \
51
+ --cache_dir /mnt/disks/flaxdisk/cache/ \
52
+ --num_train_steps="20000" \
53
+ --adam_beta1="0.9" \
54
+ --adam_beta2="0.98" \
55
+ --logging_steps="20000" \
56
+ --save_steps="10000" \
57
+ --eval_steps="10000" \
58
+ --preprocessing_num_workers 96 \
59
+ --auth_token True \
60
+ --adafactor \
61
+ --push_to_hub
62
+ ```
63
+
64
+
65
+
66
+ Approximate additional training time: 1 week.
67
+
68
+
config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": ".",
3
+ "architectures": [
4
+ "RobertaForMaskedLM"
5
+ ],
6
+ "attention_probs_dropout_prob": 0.1,
7
+ "bos_token_id": 0,
8
+ "classifier_dropout": null,
9
+ "eos_token_id": 2,
10
+ "gradient_checkpointing": false,
11
+ "hidden_act": "gelu",
12
+ "hidden_dropout_prob": 0.1,
13
+ "hidden_size": 768,
14
+ "initializer_range": 0.02,
15
+ "intermediate_size": 3072,
16
+ "layer_norm_eps": 1e-05,
17
+ "max_position_embeddings": 514,
18
+ "model_type": "roberta",
19
+ "num_attention_heads": 12,
20
+ "num_hidden_layers": 12,
21
+ "pad_token_id": 1,
22
+ "position_embedding_type": "absolute",
23
+ "torch_dtype": "float32",
24
+ "transformers_version": "4.13.0.dev0",
25
+ "type_vocab_size": 1,
26
+ "use_cache": true,
27
+ "vocab_size": 50265
28
+ }
create_config.py ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ from transformers import RobertaConfig
2
+
3
+ model_dir = "./" # ${MODEL_DIR}
4
+
5
+ config = RobertaConfig.from_pretrained("roberta-base")
6
+ config.save_pretrained(model_dir)
events.out.tfevents.1637101724.t1v-n-358ff5d1-w-0.857283.3.v2 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:44b63ae973927ce95868a7d7123f56e8e33d3e9dad9a5cabe627c84d796f9396
3
+ size 5928006
events.out.tfevents.1637157195.t1v-n-358ff5d1-w-0.908301.3.v2 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d5a20f58ee9305c56b75e5ee15c4c982270549bc390b0fc331fd3605cb7b4b5a
3
+ size 105772210
events.out.tfevents.1637704340.t1v-n-358ff5d1-w-0.1392746.3.v2 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:af735f4f4a2ddde9d5811f927ea089b226643a4a66a2491fcf160a17a8a785e3
3
+ size 8908221
flax_model.msgpack ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:64f7ff612ca7eb3abfe9ae1d2c0088df78d9a5ab1af89b2368ceabadf34e46d4
3
+ size 498796983
generate_pytorch_model.py ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ # This script overwrites any existing PyTorch model. Generates a new one with an LM head from the pretrained Flax model.
2
+ from transformers import RobertaForMaskedLM, AutoTokenizer
3
+ model = RobertaForMaskedLM.from_pretrained(".",from_flax=True)
4
+ tokenizer = AutoTokenizer.from_pretrained(".")
5
+ tokenizer.save_pretrained(".")
6
+ model.save_pretrained(".")
7
+
merges.txt ADDED
The diff for this file is too large to render. See raw diff
 
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:95e26332a282f00db4b34330bceaec2b4bc6c2a6166aeef8692dcf9ee7a5881a
3
+ size 498872555
run_mlm_flax.py ADDED
@@ -0,0 +1,680 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ # coding=utf-8
3
+ # Copyright 2021 The HuggingFace Team All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+ # limitations under the License.
16
+ """
17
+ Fine-tuning the library models for masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
18
+ text file or a dataset.
19
+
20
+ Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
21
+ https://huggingface.co/models?filter=masked-lm
22
+ """
23
+ import logging
24
+ import os
25
+ import sys
26
+ import time
27
+ from dataclasses import dataclass, field
28
+
29
+ # You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
30
+ from pathlib import Path
31
+ from typing import Dict, List, Optional, Tuple
32
+
33
+ import numpy as np
34
+ from datasets import load_dataset
35
+ from tqdm import tqdm
36
+
37
+ import flax
38
+ import jax
39
+ import jax.numpy as jnp
40
+ import optax
41
+ from flax import jax_utils, traverse_util
42
+ from flax.training import train_state
43
+ from flax.training.common_utils import get_metrics, onehot, shard
44
+ from transformers import (
45
+ CONFIG_MAPPING,
46
+ FLAX_MODEL_FOR_MASKED_LM_MAPPING,
47
+ AutoConfig,
48
+ AutoTokenizer,
49
+ FlaxAutoModelForMaskedLM,
50
+ HfArgumentParser,
51
+ PreTrainedTokenizerBase,
52
+ TensorType,
53
+ TrainingArguments,
54
+ is_tensorboard_available,
55
+ set_seed,
56
+ )
57
+
58
+
59
+ MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_MASKED_LM_MAPPING.keys())
60
+ MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
61
+
62
+
63
+ @dataclass
64
+ class ModelArguments:
65
+ """
66
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
67
+ """
68
+
69
+ model_name_or_path: Optional[str] = field(
70
+ default=None,
71
+ metadata={
72
+ "help": "The model checkpoint for weights initialization."
73
+ "Don't set if you want to train a model from scratch."
74
+ },
75
+ )
76
+ model_type: Optional[str] = field(
77
+ default=None,
78
+ metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
79
+ )
80
+ config_name: Optional[str] = field(
81
+ default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
82
+ )
83
+ tokenizer_name: Optional[str] = field(
84
+ default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
85
+ )
86
+ cache_dir: Optional[str] = field(
87
+ default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
88
+ )
89
+ use_fast_tokenizer: bool = field(
90
+ default=True,
91
+ metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
92
+ )
93
+ dtype: Optional[str] = field(
94
+ default="float32",
95
+ metadata={
96
+ "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
97
+ },
98
+ )
99
+
100
+
101
+ @dataclass
102
+ class DataTrainingArguments:
103
+ """
104
+ Arguments pertaining to what data we are going to input our model for training and eval.
105
+ """
106
+
107
+ dataset_name: Optional[str] = field(
108
+ default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
109
+ )
110
+ dataset_config_name: Optional[str] = field(
111
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
112
+ )
113
+ train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
114
+ validation_file: Optional[str] = field(
115
+ default=None,
116
+ metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
117
+ )
118
+ train_ref_file: Optional[str] = field(
119
+ default=None,
120
+ metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
121
+ )
122
+ validation_ref_file: Optional[str] = field(
123
+ default=None,
124
+ metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
125
+ )
126
+ overwrite_cache: bool = field(
127
+ default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
128
+ )
129
+ validation_split_percentage: Optional[int] = field(
130
+ default=5,
131
+ metadata={
132
+ "help": "The percentage of the train set used as validation set in case there's no validation split"
133
+ },
134
+ )
135
+ max_seq_length: Optional[int] = field(
136
+ default=None,
137
+ metadata={
138
+ "help": "The maximum total input sequence length after tokenization. Sequences longer "
139
+ "than this will be truncated. Default to the max input length of the model."
140
+ },
141
+ )
142
+ preprocessing_num_workers: Optional[int] = field(
143
+ default=None,
144
+ metadata={"help": "The number of processes to use for the preprocessing."},
145
+ )
146
+ mlm_probability: float = field(
147
+ default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
148
+ )
149
+ pad_to_max_length: bool = field(
150
+ default=False,
151
+ metadata={
152
+ "help": "Whether to pad all samples to `max_seq_length`. "
153
+ "If False, will pad the samples dynamically when batching to the maximum length in the batch."
154
+ },
155
+ )
156
+ line_by_line: bool = field(
157
+ default=False,
158
+ metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
159
+ )
160
+
161
+ def __post_init__(self):
162
+ if self.dataset_name is None and self.train_file is None and self.validation_file is None:
163
+ raise ValueError("Need either a dataset name or a training/validation file.")
164
+ else:
165
+ if self.train_file is not None:
166
+ extension = self.train_file.split(".")[-1]
167
+ assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
168
+ if self.validation_file is not None:
169
+ extension = self.validation_file.split(".")[-1]
170
+ assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
171
+
172
+
173
+ @flax.struct.dataclass
174
+ class FlaxDataCollatorForLanguageModeling:
175
+ """
176
+ Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
177
+ are not all of the same length.
178
+
179
+ Args:
180
+ tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
181
+ The tokenizer used for encoding the data.
182
+ mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
183
+ The probability with which to (randomly) mask tokens in the input.
184
+
185
+ .. note::
186
+
187
+ For best performance, this data collator should be used with a dataset having items that are dictionaries or
188
+ BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
189
+ :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
190
+ argument :obj:`return_special_tokens_mask=True`.
191
+ """
192
+
193
+ tokenizer: PreTrainedTokenizerBase
194
+ mlm_probability: float = 0.15
195
+
196
+ def __post_init__(self):
197
+ if self.tokenizer.mask_token is None:
198
+ raise ValueError(
199
+ "This tokenizer does not have a mask token which is necessary for masked language modeling. "
200
+ "You should pass `mlm=False` to train on causal language modeling instead."
201
+ )
202
+
203
+ def __call__(self, examples: List[Dict[str, np.ndarray]], pad_to_multiple_of: int) -> Dict[str, np.ndarray]:
204
+ # Handle dict or lists with proper padding and conversion to tensor.
205
+ batch = self.tokenizer.pad(examples, pad_to_multiple_of=pad_to_multiple_of, return_tensors=TensorType.NUMPY)
206
+
207
+ # If special token mask has been preprocessed, pop it from the dict.
208
+ special_tokens_mask = batch.pop("special_tokens_mask", None)
209
+
210
+ batch["input_ids"], batch["labels"] = self.mask_tokens(
211
+ batch["input_ids"], special_tokens_mask=special_tokens_mask
212
+ )
213
+ return batch
214
+
215
+ def mask_tokens(
216
+ self, inputs: np.ndarray, special_tokens_mask: Optional[np.ndarray]
217
+ ) -> Tuple[jnp.ndarray, jnp.ndarray]:
218
+ """
219
+ Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
220
+ """
221
+ labels = inputs.copy()
222
+ # We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
223
+ probability_matrix = np.full(labels.shape, self.mlm_probability)
224
+ special_tokens_mask = special_tokens_mask.astype("bool")
225
+
226
+ probability_matrix[special_tokens_mask] = 0.0
227
+ masked_indices = np.random.binomial(1, probability_matrix).astype("bool")
228
+ labels[~masked_indices] = -100 # We only compute loss on masked tokens
229
+
230
+ # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
231
+ indices_replaced = np.random.binomial(1, np.full(labels.shape, 0.8)).astype("bool") & masked_indices
232
+ inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
233
+
234
+ # 10% of the time, we replace masked input tokens with random word
235
+ indices_random = np.random.binomial(1, np.full(labels.shape, 0.5)).astype("bool")
236
+ indices_random &= masked_indices & ~indices_replaced
237
+
238
+ random_words = np.random.randint(self.tokenizer.vocab_size, size=labels.shape, dtype="i4")
239
+ inputs[indices_random] = random_words[indices_random]
240
+
241
+ # The rest of the time (10% of the time) we keep the masked input tokens unchanged
242
+ return inputs, labels
243
+
244
+
245
+ def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndarray:
246
+ num_samples = len(samples_idx)
247
+ samples_to_remove = num_samples % batch_size
248
+
249
+ if samples_to_remove != 0:
250
+ samples_idx = samples_idx[:-samples_to_remove]
251
+ sections_split = num_samples // batch_size
252
+ batch_idx = np.split(samples_idx, sections_split)
253
+ return batch_idx
254
+
255
+
256
+ def write_train_metric(summary_writer, train_metrics, train_time, step):
257
+ summary_writer.scalar("train_time", train_time, step)
258
+
259
+ train_metrics = get_metrics(train_metrics)
260
+ for key, vals in train_metrics.items():
261
+ tag = f"train_{key}"
262
+ for i, val in enumerate(vals):
263
+ summary_writer.scalar(tag, val, step - len(vals) + i + 1)
264
+
265
+
266
+ def write_eval_metric(summary_writer, eval_metrics, step):
267
+ for metric_name, value in eval_metrics.items():
268
+ summary_writer.scalar(f"eval_{metric_name}", value, step)
269
+
270
+
271
+ if __name__ == "__main__":
272
+ # See all possible arguments in src/transformers/training_args.py
273
+ # or by passing the --help flag to this script.
274
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
275
+
276
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
277
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
278
+ # If we pass only one argument to the script and it's the path to a json file,
279
+ # let's parse it to get our arguments.
280
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
281
+ else:
282
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
283
+
284
+ if (
285
+ os.path.exists(training_args.output_dir)
286
+ and os.listdir(training_args.output_dir)
287
+ and training_args.do_train
288
+ and not training_args.overwrite_output_dir
289
+ ):
290
+ raise ValueError(
291
+ f"Output directory ({training_args.output_dir}) already exists and is not empty."
292
+ "Use --overwrite_output_dir to overcome."
293
+ )
294
+
295
+ # Setup logging
296
+ logging.basicConfig(
297
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
298
+ level="NOTSET",
299
+ datefmt="[%X]",
300
+ )
301
+
302
+ # Log on each process the small summary:
303
+ logger = logging.getLogger(__name__)
304
+
305
+ # Set the verbosity to info of the Transformers logger (on main process only):
306
+ logger.info(f"Training/evaluation parameters {training_args}")
307
+
308
+ # Set seed before initializing model.
309
+ set_seed(training_args.seed)
310
+
311
+ # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
312
+ # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
313
+ # (the dataset will be downloaded automatically from the datasets Hub).
314
+ #
315
+ # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
316
+ # 'text' is found. You can easily tweak this behavior (see below).
317
+ #
318
+ # In distributed training, the load_dataset function guarantees that only one local process can concurrently
319
+ # download the dataset.
320
+ chunksize = 10<<20
321
+ if data_args.dataset_name is not None:
322
+ # Downloading and loading a dataset from the hub.
323
+ datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir, chunksize=chunksize)
324
+
325
+ if "validation" not in datasets.keys():
326
+ datasets["validation"] = load_dataset(
327
+ data_args.dataset_name,
328
+ data_args.dataset_config_name,
329
+ split=f"train[:{data_args.validation_split_percentage}%]",
330
+ cache_dir=model_args.cache_dir,
331
+ chunksize=chunksize
332
+ )
333
+ datasets["train"] = load_dataset(
334
+ data_args.dataset_name,
335
+ data_args.dataset_config_name,
336
+ split=f"train[{data_args.validation_split_percentage}%:]",
337
+ cache_dir=model_args.cache_dir,
338
+ chunksize=chunksize
339
+ )
340
+ else:
341
+ data_files = {}
342
+ if data_args.train_file is not None:
343
+ data_files["train"] = data_args.train_file
344
+ if data_args.validation_file is not None:
345
+ data_files["validation"] = data_args.validation_file
346
+ extension = data_args.train_file.split(".")[-1]
347
+ if extension == "txt":
348
+ extension = "text"
349
+ datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir, chunksize=chunksize)
350
+ # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
351
+ # https://huggingface.co/docs/datasets/loading_datasets.html.
352
+
353
+ # Load pretrained model and tokenizer
354
+
355
+ # Distributed training:
356
+ # The .from_pretrained methods guarantee that only one local process can concurrently
357
+ # download model & vocab.
358
+ if model_args.config_name:
359
+ config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
360
+ elif model_args.model_name_or_path:
361
+ config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
362
+ else:
363
+ config = CONFIG_MAPPING[model_args.model_type]()
364
+ logger.warning("You are instantiating a new config instance from scratch.")
365
+
366
+ if model_args.tokenizer_name:
367
+ tokenizer = AutoTokenizer.from_pretrained(
368
+ model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
369
+ )
370
+ elif model_args.model_name_or_path:
371
+ tokenizer = AutoTokenizer.from_pretrained(
372
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
373
+ )
374
+ else:
375
+ raise ValueError(
376
+ "You are instantiating a new tokenizer from scratch. This is not supported by this script."
377
+ "You can do it from another script, save it, and load it from here, using --tokenizer_name."
378
+ )
379
+
380
+ # Preprocessing the datasets.
381
+ # First we tokenize all the texts.
382
+ if training_args.do_train:
383
+ column_names = datasets["train"].column_names
384
+ else:
385
+ column_names = datasets["validation"].column_names
386
+ text_column_name = "text" if "text" in column_names else column_names[0]
387
+
388
+ max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
389
+
390
+ if data_args.line_by_line:
391
+ # When using line_by_line, we just tokenize each nonempty line.
392
+ padding = "max_length" if data_args.pad_to_max_length else False
393
+
394
+ def tokenize_function(examples):
395
+ # Remove empty lines
396
+ examples = [line for line in examples if len(line) > 0 and not line.isspace()]
397
+ return tokenizer(
398
+ examples,
399
+ return_special_tokens_mask=True,
400
+ padding=padding,
401
+ truncation=True,
402
+ max_length=max_seq_length,
403
+ )
404
+
405
+ tokenized_datasets = datasets.map(
406
+ tokenize_function,
407
+ input_columns=[text_column_name],
408
+ batched=True,
409
+ num_proc=data_args.preprocessing_num_workers,
410
+ remove_columns=column_names,
411
+ load_from_cache_file=not data_args.overwrite_cache,
412
+ )
413
+
414
+ else:
415
+ # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
416
+ # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
417
+ # efficient when it receives the `special_tokens_mask`.
418
+ def tokenize_function(examples):
419
+ return tokenizer(examples[text_column_name], return_special_tokens_mask=True)
420
+
421
+ tokenized_datasets = datasets.map(
422
+ tokenize_function,
423
+ batched=True,
424
+ num_proc=data_args.preprocessing_num_workers,
425
+ remove_columns=column_names,
426
+ load_from_cache_file=not data_args.overwrite_cache,
427
+ )
428
+
429
+ # Main data processing function that will concatenate all texts from our dataset and generate chunks of
430
+ # max_seq_length.
431
+ def group_texts(examples):
432
+ # Concatenate all texts.
433
+ concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
434
+ total_length = len(concatenated_examples[list(examples.keys())[0]])
435
+ # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
436
+ # customize this part to your needs.
437
+ if total_length >= max_seq_length:
438
+ total_length = (total_length // max_seq_length) * max_seq_length
439
+ # Split by chunks of max_len.
440
+ result = {
441
+ k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
442
+ for k, t in concatenated_examples.items()
443
+ }
444
+ return result
445
+
446
+ # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
447
+ # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
448
+ # might be slower to preprocess.
449
+ #
450
+ # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
451
+ # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
452
+ tokenized_datasets = tokenized_datasets.map(
453
+ group_texts,
454
+ batched=True,
455
+ num_proc=data_args.preprocessing_num_workers,
456
+ load_from_cache_file=not data_args.overwrite_cache,
457
+ )
458
+
459
+ # Enable tensorboard only on the master node
460
+ has_tensorboard = is_tensorboard_available()
461
+ if has_tensorboard and jax.process_index() == 0:
462
+ try:
463
+ from flax.metrics.tensorboard import SummaryWriter
464
+
465
+ summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
466
+ except ImportError as ie:
467
+ has_tensorboard = False
468
+ logger.warning(
469
+ f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
470
+ )
471
+ else:
472
+ logger.warning(
473
+ "Unable to display metrics through TensorBoard because the package is not installed: "
474
+ "Please run pip install tensorboard to enable."
475
+ )
476
+
477
+ # Data collator
478
+ # This one will take care of randomly masking the tokens.
479
+ data_collator = FlaxDataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
480
+
481
+ # Initialize our training
482
+ rng = jax.random.PRNGKey(training_args.seed)
483
+ dropout_rngs = jax.random.split(rng, jax.local_device_count())
484
+
485
+ if model_args.model_name_or_path:
486
+ model = FlaxAutoModelForMaskedLM.from_pretrained(
487
+ model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
488
+ )
489
+ else:
490
+ model = FlaxAutoModelForMaskedLM.from_config(
491
+ config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
492
+ )
493
+
494
+ # Store some constant
495
+ num_epochs = int(training_args.num_train_epochs)
496
+ train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
497
+ eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
498
+
499
+ num_train_steps = len(tokenized_datasets["train"]) // train_batch_size * num_epochs
500
+
501
+ # Create learning rate schedule
502
+ warmup_fn = optax.linear_schedule(
503
+ init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps
504
+ )
505
+ decay_fn = optax.linear_schedule(
506
+ init_value=training_args.learning_rate,
507
+ end_value=0,
508
+ transition_steps=num_train_steps - training_args.warmup_steps,
509
+ )
510
+ linear_decay_lr_schedule_fn = optax.join_schedules(
511
+ schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]
512
+ )
513
+
514
+ # We use Optax's "masking" functionality to not apply weight decay
515
+ # to bias and LayerNorm scale parameters. decay_mask_fn returns a
516
+ # mask boolean with the same structure as the parameters.
517
+ # The mask is True for parameters that should be decayed.
518
+ # Note that this mask is specifically adapted for FlaxBERT-like models.
519
+ # For other models, one should correct the layer norm parameter naming
520
+ # accordingly.
521
+ def decay_mask_fn(params):
522
+ flat_params = traverse_util.flatten_dict(params)
523
+ flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params}
524
+ return traverse_util.unflatten_dict(flat_mask)
525
+
526
+ # create adam optimizer
527
+ if training_args.adafactor:
528
+ # We use the default parameters here to initialize adafactor,
529
+ # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
530
+ optimizer = optax.adafactor(
531
+ learning_rate=linear_decay_lr_schedule_fn,
532
+ )
533
+ else:
534
+ optimizer = optax.adamw(
535
+ learning_rate=linear_decay_lr_schedule_fn,
536
+ b1=training_args.adam_beta1,
537
+ b2=training_args.adam_beta2,
538
+ eps=training_args.adam_epsilon,
539
+ weight_decay=training_args.weight_decay,
540
+ mask=decay_mask_fn,
541
+ )
542
+
543
+ # Setup train state
544
+ state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)
545
+
546
+ # Define gradient update step fn
547
+ def train_step(state, batch, dropout_rng):
548
+ dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
549
+
550
+ def loss_fn(params):
551
+ labels = batch.pop("labels")
552
+
553
+ logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
554
+
555
+ # compute loss, ignore padded input tokens
556
+ label_mask = jnp.where(labels > 0, 1.0, 0.0)
557
+ loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
558
+
559
+ # take average
560
+ loss = loss.sum() / label_mask.sum()
561
+
562
+ return loss
563
+
564
+ grad_fn = jax.value_and_grad(loss_fn)
565
+ loss, grad = grad_fn(state.params)
566
+ grad = jax.lax.pmean(grad, "batch")
567
+ new_state = state.apply_gradients(grads=grad)
568
+
569
+ metrics = jax.lax.pmean(
570
+ {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch"
571
+ )
572
+
573
+ return new_state, metrics, new_dropout_rng
574
+
575
+ # Create parallel version of the train step
576
+ p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
577
+
578
+ # Define eval fn
579
+ def eval_step(params, batch):
580
+ labels = batch.pop("labels")
581
+
582
+ logits = model(**batch, params=params, train=False)[0]
583
+
584
+ # compute loss, ignore padded input tokens
585
+ label_mask = jnp.where(labels > 0, 1.0, 0.0)
586
+ loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
587
+
588
+ # compute accuracy
589
+ accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels) * label_mask
590
+
591
+ # summarize metrics
592
+ metrics = {"loss": loss.sum(), "accuracy": accuracy.sum(), "normalizer": label_mask.sum()}
593
+ metrics = jax.lax.psum(metrics, axis_name="batch")
594
+
595
+ return metrics
596
+
597
+ p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
598
+
599
+ # Replicate the train state on each device
600
+ state = jax_utils.replicate(state)
601
+
602
+ train_time = 0
603
+ epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
604
+ for epoch in epochs:
605
+ # ======================== Training ================================
606
+ train_start = time.time()
607
+ train_metrics = []
608
+
609
+ # Create sampling rng
610
+ rng, input_rng = jax.random.split(rng)
611
+
612
+ # Generate an epoch by shuffling sampling indices from the train dataset
613
+ num_train_samples = len(tokenized_datasets["train"])
614
+ train_samples_idx = jax.random.permutation(input_rng, jnp.arange(num_train_samples))
615
+ train_batch_idx = generate_batch_splits(train_samples_idx, train_batch_size)
616
+
617
+ # Gather the indexes for creating the batch and do a training step
618
+ for step, batch_idx in enumerate(tqdm(train_batch_idx, desc="Training...", position=1)):
619
+ samples = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
620
+ model_inputs = data_collator(samples, pad_to_multiple_of=16)
621
+
622
+ # Model forward
623
+ model_inputs = shard(model_inputs.data)
624
+ state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
625
+ train_metrics.append(train_metric)
626
+
627
+ cur_step = epoch * (num_train_samples // train_batch_size) + step
628
+
629
+ if cur_step % training_args.logging_steps == 0 and cur_step > 0:
630
+ # Save metrics
631
+ train_metric = jax_utils.unreplicate(train_metric)
632
+ train_time += time.time() - train_start
633
+ if has_tensorboard and jax.process_index() == 0:
634
+ write_train_metric(summary_writer, train_metrics, train_time, cur_step)
635
+
636
+ epochs.write(
637
+ f"Step... ({cur_step} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
638
+ )
639
+
640
+ train_metrics = []
641
+
642
+ if cur_step % training_args.eval_steps == 0 and cur_step > 0:
643
+ # ======================== Evaluating ==============================
644
+ num_eval_samples = len(tokenized_datasets["validation"])
645
+ eval_samples_idx = jnp.arange(num_eval_samples)
646
+ eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
647
+
648
+ eval_metrics = []
649
+ for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
650
+ samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
651
+ model_inputs = data_collator(samples, pad_to_multiple_of=16)
652
+
653
+ # Model forward
654
+ model_inputs = shard(model_inputs.data)
655
+ metrics = p_eval_step(state.params, model_inputs)
656
+ eval_metrics.append(metrics)
657
+
658
+ # normalize eval metrics
659
+ eval_metrics = get_metrics(eval_metrics)
660
+ eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
661
+ eval_normalizer = eval_metrics.pop("normalizer")
662
+ eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
663
+
664
+ # Update progress bar
665
+ epochs.desc = f"Step... ({cur_step} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
666
+
667
+ # Save metrics
668
+ if has_tensorboard and jax.process_index() == 0:
669
+ write_eval_metric(summary_writer, eval_metrics, cur_step)
670
+
671
+ if cur_step % training_args.save_steps == 0 and cur_step > 0:
672
+ # save checkpoint after each epoch and push checkpoint to the hub
673
+ if jax.process_index() == 0:
674
+ params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
675
+ model.save_pretrained(
676
+ training_args.output_dir,
677
+ params=params,
678
+ push_to_hub=training_args.push_to_hub,
679
+ commit_message=f"Saving weights and logs of step {cur_step}",
680
+ )
run_mlm_flax_stream.py ADDED
@@ -0,0 +1,638 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ # coding=utf-8
3
+ # Copyright 2021 The HuggingFace Team All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+ # limitations under the License.
16
+ """
17
+ Fine-tuning the library models for masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
18
+ text file or a dataset.
19
+
20
+ Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
21
+ https://huggingface.co/models?filter=masked-lm
22
+ """
23
+ import logging
24
+ import os
25
+ import sys
26
+ import time
27
+ from collections import defaultdict
28
+ from dataclasses import dataclass, field
29
+
30
+ # You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
31
+ from pathlib import Path
32
+ from typing import Dict, List, Optional, Tuple
33
+
34
+ import datasets
35
+ import numpy as np
36
+ from datasets import load_dataset
37
+ from tqdm import tqdm
38
+
39
+ import flax
40
+ import jax
41
+ import jax.numpy as jnp
42
+ import optax
43
+ from flax import jax_utils, traverse_util
44
+ from flax.training import train_state
45
+ from flax.training.common_utils import get_metrics, onehot, shard
46
+ from transformers import (
47
+ CONFIG_MAPPING,
48
+ FLAX_MODEL_FOR_MASKED_LM_MAPPING,
49
+ AutoConfig,
50
+ AutoTokenizer,
51
+ FlaxAutoModelForMaskedLM,
52
+ HfArgumentParser,
53
+ PreTrainedTokenizerBase,
54
+ TensorType,
55
+ TrainingArguments,
56
+ is_tensorboard_available,
57
+ set_seed,
58
+ )
59
+
60
+
61
+ #if datasets.__version__ <= "1.8.0":
62
+ # raise ValueError("Make sure to upgrade `datasets` to a version >= 1.9.0 to use dataset streaming")
63
+
64
+
65
+ MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_MASKED_LM_MAPPING.keys())
66
+ MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
67
+
68
+
69
+ @dataclass
70
+ class ModelArguments:
71
+ """
72
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
73
+ """
74
+
75
+ model_name_or_path: Optional[str] = field(
76
+ default=None,
77
+ metadata={
78
+ "help": "The model checkpoint for weights initialization."
79
+ "Don't set if you want to train a model from scratch."
80
+ },
81
+ )
82
+ model_type: Optional[str] = field(
83
+ default=None,
84
+ metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
85
+ )
86
+ config_name: Optional[str] = field(
87
+ default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
88
+ )
89
+ tokenizer_name: Optional[str] = field(
90
+ default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
91
+ )
92
+ cache_dir: Optional[str] = field(
93
+ default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
94
+ )
95
+ use_fast_tokenizer: bool = field(
96
+ default=True,
97
+ metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
98
+ )
99
+ dtype: Optional[str] = field(
100
+ default="float32",
101
+ metadata={
102
+ "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
103
+ },
104
+ )
105
+
106
+
107
+ @dataclass
108
+ class DataTrainingArguments:
109
+ """
110
+ Arguments pertaining to what data we are going to input our model for training and eval.
111
+ """
112
+
113
+ dataset_name: Optional[str] = field(
114
+ default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
115
+ )
116
+ dataset_config_name: Optional[str] = field(
117
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
118
+ )
119
+ train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
120
+ validation_file: Optional[str] = field(
121
+ default=None,
122
+ metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
123
+ )
124
+ train_ref_file: Optional[str] = field(
125
+ default=None,
126
+ metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
127
+ )
128
+ validation_ref_file: Optional[str] = field(
129
+ default=None,
130
+ metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
131
+ )
132
+ auth_token: bool = field(
133
+ default=False, metadata={"help": "Use authorisation token"}
134
+ )
135
+ overwrite_cache: bool = field(
136
+ default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
137
+ )
138
+ validation_split_percentage: Optional[int] = field(
139
+ default=5,
140
+ metadata={
141
+ "help": "The percentage of the train set used as validation set in case there's no validation split"
142
+ },
143
+ )
144
+ max_seq_length: Optional[int] = field(
145
+ default=None,
146
+ metadata={
147
+ "help": "The maximum total input sequence length after tokenization. Sequences longer "
148
+ "than this will be truncated. Default to the max input length of the model."
149
+ },
150
+ )
151
+ preprocessing_num_workers: Optional[int] = field(
152
+ default=None,
153
+ metadata={"help": "The number of processes to use for the preprocessing."},
154
+ )
155
+ mlm_probability: float = field(
156
+ default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
157
+ )
158
+ pad_to_max_length: bool = field(
159
+ default=False,
160
+ metadata={
161
+ "help": "Whether to pad all samples to `max_seq_length`. "
162
+ "If False, will pad the samples dynamically when batching to the maximum length in the batch."
163
+ },
164
+ )
165
+ line_by_line: bool = field(
166
+ default=False,
167
+ metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
168
+ )
169
+ text_column_name: str = field(
170
+ default="text", metadata={"help": "The name of the column to retrieve the training text."}
171
+ )
172
+ shuffle_buffer_size: int = field(
173
+ default=10000, metadata={"help": "The number of examples to pre-load for shuffling."}
174
+ )
175
+ num_train_steps: int = field(default=50000, metadata={"help": "The number of training steps."})
176
+ num_eval_samples: int = field(default=50000, metadata={"help": "The number of samples to be used for evaluation"})
177
+
178
+ def __post_init__(self):
179
+ if self.dataset_name is None and self.train_file is None and self.validation_file is None:
180
+ raise ValueError("Need either a dataset name or a training/validation file.")
181
+ else:
182
+ if self.train_file is not None:
183
+ extension = self.train_file.split(".")[-1]
184
+ assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
185
+ if self.validation_file is not None:
186
+ extension = self.validation_file.split(".")[-1]
187
+ assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
188
+
189
+
190
+ @flax.struct.dataclass
191
+ class FlaxDataCollatorForLanguageModeling:
192
+ """
193
+ Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
194
+ are not all of the same length.
195
+
196
+ Args:
197
+ tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
198
+ The tokenizer used for encoding the data.
199
+ mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
200
+ The probability with which to (randomly) mask tokens in the input.
201
+
202
+ .. note::
203
+
204
+ For best performance, this data collator should be used with a dataset having items that are dictionaries or
205
+ BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
206
+ :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
207
+ argument :obj:`return_special_tokens_mask=True`.
208
+ """
209
+
210
+ tokenizer: PreTrainedTokenizerBase
211
+ mlm_probability: float = 0.15
212
+
213
+ def __post_init__(self):
214
+ if self.tokenizer.mask_token is None:
215
+ raise ValueError(
216
+ "This tokenizer does not have a mask token which is necessary for masked language modeling. "
217
+ "You should pass `mlm=False` to train on causal language modeling instead."
218
+ )
219
+
220
+ def __call__(self, examples: List[Dict[str, np.ndarray]]) -> Dict[str, np.ndarray]:
221
+ # Handle dict or lists with proper padding and conversion to tensor.
222
+ batch = self.tokenizer.pad(examples, return_tensors=TensorType.NUMPY)
223
+
224
+ # If special token mask has been preprocessed, pop it from the dict.
225
+ special_tokens_mask = batch.pop("special_tokens_mask", None)
226
+
227
+ batch["input_ids"], batch["labels"] = self.mask_tokens(
228
+ batch["input_ids"], special_tokens_mask=special_tokens_mask
229
+ )
230
+ return batch
231
+
232
+ def mask_tokens(
233
+ self, inputs: np.ndarray, special_tokens_mask: Optional[np.ndarray]
234
+ ) -> Tuple[jnp.ndarray, jnp.ndarray]:
235
+ """
236
+ Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
237
+ """
238
+ labels = inputs.copy()
239
+ # We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
240
+ probability_matrix = np.full(labels.shape, self.mlm_probability)
241
+ special_tokens_mask = special_tokens_mask.astype("bool")
242
+
243
+ probability_matrix[special_tokens_mask] = 0.0
244
+ masked_indices = np.random.binomial(1, probability_matrix).astype("bool")
245
+ labels[~masked_indices] = -100 # We only compute loss on masked tokens
246
+
247
+ # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
248
+ indices_replaced = np.random.binomial(1, np.full(labels.shape, 0.8)).astype("bool") & masked_indices
249
+ inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
250
+
251
+ # 10% of the time, we replace masked input tokens with random word
252
+ indices_random = np.random.binomial(1, np.full(labels.shape, 0.5)).astype("bool")
253
+ indices_random &= masked_indices & ~indices_replaced
254
+
255
+ random_words = np.random.randint(self.tokenizer.vocab_size, size=labels.shape, dtype="i4")
256
+ inputs[indices_random] = random_words[indices_random]
257
+
258
+ # The rest of the time (10% of the time) we keep the masked input tokens unchanged
259
+ return inputs, labels
260
+
261
+
262
+ def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndarray:
263
+ num_samples = len(samples_idx)
264
+ samples_to_remove = num_samples % batch_size
265
+
266
+ if samples_to_remove != 0:
267
+ samples_idx = samples_idx[:-samples_to_remove]
268
+ sections_split = num_samples // batch_size
269
+ batch_idx = np.split(samples_idx, sections_split)
270
+ return batch_idx
271
+
272
+
273
+ def advance_iter_and_group_samples(train_iterator, num_samples, max_seq_length):
274
+ """
275
+ The training iterator is advanced so that after groupifying the samples,
276
+ `num_samples` of length `max_seq_length` are returned.
277
+ """
278
+ num_total_tokens = max_seq_length * num_samples
279
+ samples = defaultdict(list)
280
+
281
+ i = 0
282
+ while i < num_total_tokens:
283
+ tokenized_samples = next(train_iterator)
284
+ i += len(tokenized_samples["input_ids"])
285
+
286
+ # concatenate tokenized samples to list
287
+ samples = {k: samples[k] + tokenized_samples[k] for k in tokenized_samples.keys()}
288
+
289
+ # Concatenated tokens are split to lists of length `max_seq_length`.
290
+ # Note that remainedr of % max_seq_length are thrown away.
291
+ def group_texts(examples):
292
+ result = {
293
+ k: [t[i : i + max_seq_length] for i in range(0, num_total_tokens, max_seq_length)]
294
+ for k, t in examples.items()
295
+ }
296
+ return result
297
+
298
+ grouped_samples = group_texts(samples)
299
+ return grouped_samples
300
+
301
+
302
+ def write_train_metric(summary_writer, train_metrics, train_time, step):
303
+ summary_writer.scalar("train_time", train_time, step)
304
+
305
+ train_metrics = get_metrics(train_metrics)
306
+ for key, vals in train_metrics.items():
307
+ tag = f"train_{key}"
308
+ for i, val in enumerate(vals):
309
+ summary_writer.scalar(tag, val, step - len(vals) + i + 1)
310
+
311
+
312
+ def write_eval_metric(summary_writer, eval_metrics, step):
313
+ for metric_name, value in eval_metrics.items():
314
+ summary_writer.scalar(f"eval_{metric_name}", value, step)
315
+
316
+
317
+ if __name__ == "__main__":
318
+ # See all possible arguments in src/transformers/training_args.py
319
+ # or by passing the --help flag to this script.
320
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
321
+
322
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
323
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
324
+ # If we pass only one argument to the script and it's the path to a json file,
325
+ # let's parse it to get our arguments.
326
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
327
+ else:
328
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
329
+
330
+ if (
331
+ os.path.exists(training_args.output_dir)
332
+ and os.listdir(training_args.output_dir)
333
+ and training_args.do_train
334
+ and not training_args.overwrite_output_dir
335
+ ):
336
+ raise ValueError(
337
+ f"Output directory ({training_args.output_dir}) already exists and is not empty."
338
+ "Use --overwrite_output_dir to overcome."
339
+ )
340
+
341
+ # Setup logging
342
+ logging.basicConfig(
343
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
344
+ level="INFO",
345
+ datefmt="[%X]",
346
+ )
347
+
348
+ # Log on each process the small summary:
349
+ logger = logging.getLogger(__name__)
350
+ #logger.warning(
351
+ # f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
352
+ # + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
353
+ #)
354
+
355
+ # Set the verbosity to info of the Transformers logger (on main process only):
356
+ logger.info(f"Training/evaluation parameters {training_args}")
357
+
358
+ # Set seed before initializing model.
359
+ set_seed(training_args.seed)
360
+
361
+ # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
362
+ # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
363
+ # (the dataset will be downloaded automatically from the datasets Hub).
364
+ #
365
+ # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
366
+ # 'text' is found. You can easily tweak this behavior (see below).
367
+
368
+ if data_args.dataset_name is not None:
369
+ # Downloading and loading a dataset from the hub.
370
+ dataset = load_dataset(
371
+ data_args.dataset_name,
372
+ data_args.dataset_config_name,
373
+ cache_dir=model_args.cache_dir,
374
+ streaming=True,
375
+ use_auth_token=data_args.auth_token,
376
+ split="train",
377
+ )
378
+
379
+ if model_args.config_name:
380
+ config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
381
+ elif model_args.model_name_or_path:
382
+ config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
383
+ else:
384
+ config = CONFIG_MAPPING[model_args.model_type]()
385
+ logger.warning("You are instantiating a new config instance from scratch.")
386
+
387
+ if model_args.tokenizer_name:
388
+ tokenizer = AutoTokenizer.from_pretrained(
389
+ model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
390
+ )
391
+ elif model_args.model_name_or_path:
392
+ tokenizer = AutoTokenizer.from_pretrained(
393
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
394
+ )
395
+ else:
396
+ raise ValueError(
397
+ "You are instantiating a new tokenizer from scratch. This is not supported by this script."
398
+ "You can do it from another script, save it, and load it from here, using --tokenizer_name."
399
+ )
400
+
401
+ # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
402
+ # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
403
+ # efficient when it receives the `special_tokens_mask`.
404
+ def tokenize_function(examples):
405
+ return tokenizer(examples[data_args.text_column_name], return_special_tokens_mask=True)
406
+
407
+ tokenized_datasets = dataset.map(
408
+ tokenize_function,
409
+ batched=True,
410
+ )
411
+
412
+ shuffle_seed = training_args.seed
413
+ tokenized_datasets = tokenized_datasets.shuffle(buffer_size=data_args.shuffle_buffer_size, seed=shuffle_seed)
414
+
415
+ has_tensorboard = is_tensorboard_available()
416
+ if has_tensorboard and jax.process_index() == 0:
417
+ try:
418
+ from flax.metrics.tensorboard import SummaryWriter
419
+ except ImportError as ie:
420
+ has_tensorboard = False
421
+ logger.warning(
422
+ f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
423
+ )
424
+
425
+ summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
426
+
427
+ # Data collator
428
+ # This one will take care of randomly masking the tokens.
429
+ data_collator = FlaxDataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
430
+
431
+ # Initialize our training
432
+ rng = jax.random.PRNGKey(training_args.seed)
433
+ dropout_rngs = jax.random.split(rng, jax.local_device_count())
434
+
435
+ if model_args.model_name_or_path:
436
+ model = FlaxAutoModelForMaskedLM.from_pretrained(
437
+ model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
438
+ )
439
+ else:
440
+ model = FlaxAutoModelForMaskedLM.from_config(
441
+ config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
442
+ )
443
+
444
+ # Store some constant
445
+ num_epochs = int(training_args.num_train_epochs)
446
+ train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
447
+ eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
448
+
449
+ # define number steps per stream epoch
450
+ num_train_steps = data_args.num_train_steps
451
+
452
+ # Create learning rate schedule
453
+ warmup_fn = optax.linear_schedule(
454
+ init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps
455
+ )
456
+ decay_fn = optax.linear_schedule(
457
+ init_value=training_args.learning_rate,
458
+ end_value=0,
459
+ transition_steps=num_train_steps - training_args.warmup_steps,
460
+ )
461
+ linear_decay_lr_schedule_fn = optax.join_schedules(
462
+ schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]
463
+ )
464
+
465
+ # We use Optax's "masking" functionality to not apply weight decay
466
+ # to bias and LayerNorm scale parameters. decay_mask_fn returns a
467
+ # mask boolean with the same structure as the parameters.
468
+ # The mask is True for parameters that should be decayed.
469
+ # Note that this mask is specifically adapted for FlaxBERT-like models.
470
+ # For other models, one should correct the layer norm parameter naming
471
+ # accordingly.
472
+ def decay_mask_fn(params):
473
+ flat_params = traverse_util.flatten_dict(params)
474
+ flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params}
475
+ return traverse_util.unflatten_dict(flat_mask)
476
+
477
+
478
+
479
+ #te adam optimizer
480
+ if training_args.adafactor:
481
+ # We use the default parameters here to initialize adafactor,
482
+ # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
483
+ optimizer = optax.adafactor(
484
+ learning_rate=linear_decay_lr_schedule_fn,
485
+ )
486
+ else:
487
+ optimizer = optax.adamw(
488
+ learning_rate=linear_decay_lr_schedule_fn,
489
+ b1=training_args.adam_beta1,
490
+ b2=training_args.adam_beta2,
491
+ eps=training_args.adam_epsilon,
492
+ weight_decay=training_args.weight_decay,
493
+ mask=decay_mask_fn,
494
+ )
495
+
496
+ # Setup train state
497
+ state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)
498
+
499
+ # Define gradient update step fn
500
+ def train_step(state, batch, dropout_rng):
501
+ dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
502
+
503
+ def loss_fn(params):
504
+ labels = batch.pop("labels")
505
+
506
+ logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
507
+
508
+ # compute loss, ignore padded input tokens
509
+ label_mask = jnp.where(labels > 0, 1.0, 0.0)
510
+ loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
511
+
512
+ # take average
513
+ loss = loss.sum() / label_mask.sum()
514
+
515
+ return loss
516
+
517
+ grad_fn = jax.value_and_grad(loss_fn)
518
+ loss, grad = grad_fn(state.params)
519
+ grad = jax.lax.pmean(grad, "batch")
520
+ new_state = state.apply_gradients(grads=grad)
521
+
522
+ metrics = jax.lax.pmean(
523
+ {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch"
524
+ )
525
+
526
+ return new_state, metrics, new_dropout_rng
527
+
528
+ # Create parallel version of the train step
529
+ p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
530
+
531
+ # Define eval fn
532
+ def eval_step(params, batch):
533
+ labels = batch.pop("labels")
534
+
535
+ logits = model(**batch, params=params, train=False)[0]
536
+
537
+ # compute loss, ignore padded input tokens
538
+ label_mask = jnp.where(labels > 0, 1.0, 0.0)
539
+ loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
540
+
541
+ # compute accuracy
542
+ accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels) * label_mask
543
+
544
+ # summarize metrics
545
+ metrics = {"loss": loss.sum(), "accuracy": accuracy.sum(), "normalizer": label_mask.sum()}
546
+ metrics = jax.lax.psum(metrics, axis_name="batch")
547
+
548
+ return metrics
549
+
550
+ p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
551
+
552
+ # Replicate the train state on each device
553
+ state = jax_utils.replicate(state)
554
+
555
+ train_time = 0
556
+ train_start = time.time()
557
+ train_metrics = []
558
+ eval_metrics = []
559
+
560
+ training_iter = iter(tokenized_datasets)
561
+
562
+ max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
563
+ eval_samples = advance_iter_and_group_samples(training_iter, data_args.num_eval_samples, max_seq_length)
564
+
565
+ steps = tqdm(range(num_train_steps), desc="Training...", position=0)
566
+ for step in range(num_train_steps):
567
+ # ======================== Training ================================
568
+ try:
569
+ samples = advance_iter_and_group_samples(training_iter, train_batch_size, max_seq_length)
570
+ except StopIteration:
571
+ # Once the end of the dataset stream is reached, the training iterator
572
+ # is reinitialized and reshuffled and a new eval dataset is randomely chosen.
573
+ shuffle_seed += 1
574
+ tokenized_datasets.set_epoch(shuffle_seed)
575
+
576
+ training_iter = iter(tokenized_datasets)
577
+
578
+ eval_dataset = advance_iter_and_group_samples(training_iter, data_args.num_eval_samples, max_seq_length)
579
+ samples = advance_iter_and_group_samples(training_iter, train_batch_size, max_seq_length)
580
+
581
+ # process input samples
582
+ model_inputs = data_collator(samples)
583
+
584
+ # Model forward
585
+ model_inputs = shard(model_inputs.data)
586
+ state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
587
+
588
+ train_metrics.append(train_metric)
589
+
590
+ if step % training_args.logging_steps == 0 and step > 0:
591
+ steps.write(
592
+ f"Step... ({step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
593
+ )
594
+ train_time += time.time() - train_start
595
+ if has_tensorboard and jax.process_index() == 0:
596
+ write_train_metric(summary_writer, train_metrics, train_time, step)
597
+ train_metrics = []
598
+
599
+ # ======================== Evaluating ==============================
600
+ if step % training_args.eval_steps == 0 and step > 0:
601
+ eval_samples_idx = jnp.arange(data_args.num_eval_samples)
602
+ eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
603
+
604
+ for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=1)):
605
+ # process input samples
606
+ batch_eval_samples = {k: [v[idx] for idx in batch_idx] for k, v in eval_samples.items()}
607
+ model_inputs = data_collator(batch_eval_samples)
608
+
609
+ # Model forward
610
+ model_inputs = shard(model_inputs.data)
611
+ metrics = p_eval_step(state.params, model_inputs)
612
+ eval_metrics.append(metrics)
613
+
614
+ # normalize eval metrics
615
+ eval_metrics = get_metrics(eval_metrics)
616
+ eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
617
+ eval_normalizer = eval_metrics.pop("normalizer")
618
+ eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
619
+
620
+ # Update progress bar
621
+ steps.desc = f"Step... ({step + 1}/{num_train_steps} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
622
+
623
+ if has_tensorboard and jax.process_index() == 0:
624
+ write_eval_metric(summary_writer, eval_metrics, step)
625
+ eval_metrics = []
626
+
627
+ # save checkpoint after each epoch and push checkpoint to the hub
628
+ if jax.process_index() == 0:
629
+ params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
630
+ model.save_pretrained(
631
+ training_args.output_dir,
632
+ params=params,
633
+ push_to_hub=training_args.push_to_hub,
634
+ commit_message=f"Saving weights and logs of step {step+1}",
635
+ )
636
+
637
+ # update tqdm bar
638
+ steps.update(1)
run_stream_128.sh ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ./run_mlm_flax_stream.py \
2
+ --output_dir="./" \
3
+ --model_type="roberta" \
4
+ --config_name="./" \
5
+ --tokenizer_name="./" \
6
+ --model_name_or_path="./" \
7
+ --dataset_name="NbAiLab/scandinavian" \
8
+ --max_seq_length="128" \
9
+ --weight_decay="0.01" \
10
+ --per_device_train_batch_size="128" \
11
+ --per_device_eval_batch_size="128" \
12
+ --learning_rate="6e-5" \
13
+ --warmup_steps="5000" \
14
+ --overwrite_output_dir \
15
+ --cache_dir /mnt/disks/flaxdisk/cache/ \
16
+ --num_train_steps="720000" \
17
+ --adam_beta1="0.9" \
18
+ --adam_beta2="0.98" \
19
+ --logging_steps="10000" \
20
+ --save_steps="10000" \
21
+ --eval_steps="10000" \
22
+ --preprocessing_num_workers 96 \
23
+ --auth_token True \
24
+ --adafactor \
25
+ --push_to_hub
run_stream_512.sh ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ./run_mlm_flax_stream.py \
2
+ --output_dir="./" \
3
+ --model_type="roberta" \
4
+ --config_name="./" \
5
+ --tokenizer_name="./" \
6
+ --model_name_or_path="./" \
7
+ --dataset_name="NbAiLab/scandinavian" \
8
+ --max_seq_length="512" \
9
+ --weight_decay="0.01" \
10
+ --per_device_train_batch_size="48" \
11
+ --per_device_eval_batch_size="48" \
12
+ --learning_rate="3e-5" \
13
+ --warmup_steps="2000" \
14
+ --overwrite_output_dir \
15
+ --cache_dir /mnt/disks/flaxdisk/cache/ \
16
+ --num_train_steps="70000" \
17
+ --adam_beta1="0.9" \
18
+ --adam_beta2="0.98" \
19
+ --logging_steps="20000" \
20
+ --save_steps="10000" \
21
+ --eval_steps="10000" \
22
+ --preprocessing_num_workers 96 \
23
+ --auth_token True \
24
+ --adafactor \
25
+ --push_to_hub
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "sep_token": "</s>", "pad_token": "<pad>", "cls_token": "<s>", "mask_token": {"content": "<mask>", "single_word": false, "lstrip": true, "rstrip": false, "normalized": false}}
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"unk_token": "<unk>", "bos_token": "<s>", "eos_token": "</s>", "add_prefix_space": false, "errors": "replace", "sep_token": "</s>", "cls_token": "<s>", "pad_token": "<pad>", "mask_token": "<mask>", "special_tokens_map_file": null, "name_or_path": ".", "tokenizer_class": "RobertaTokenizer"}
train_tokenizer.py ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from datasets import load_dataset, concatenate_datasets
2
+ from tokenizers import trainers, Tokenizer, normalizers, ByteLevelBPETokenizer
3
+
4
+ model_dir = "./" # ${MODEL_DIR}
5
+
6
+ # load dataset
7
+ dataset = load_dataset("json", data_files=["/mnt/disks/flaxdisk/corpus/norwegian_colossal_corpus_validation.json","/mnt/disks/flaxdisk/corpus/special_chars.json"], split='train')
8
+
9
+
10
+ # Instantiate tokenizer
11
+ tokenizer = ByteLevelBPETokenizer()
12
+
13
+ def batch_iterator(batch_size=1000):
14
+ for i in range(0, len(dataset), batch_size):
15
+ yield dataset[i: i + batch_size]["text"]
16
+
17
+ # Customized training
18
+ tokenizer.train_from_iterator(batch_iterator(), vocab_size=50265, min_frequency=2, special_tokens=[
19
+ "<s>",
20
+ "<pad>",
21
+ "</s>",
22
+ "<unk>",
23
+ "<mask>",
24
+ ])
25
+
26
+
27
+ # Save files to disk
28
+ tokenizer.save(f"{model_dir}/tokenizer.json")
vocab.json ADDED
The diff for this file is too large to render. See raw diff