alvp commited on
Commit
7d4020a
1 Parent(s): 57ad4f6

Upload stanzas_eval.py

Browse files
Files changed (1) hide show
  1. stanzas_eval.py +706 -0
stanzas_eval.py ADDED
@@ -0,0 +1,706 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -*- coding: utf-8 -*-
2
+ """Evaluating models for BibleBERT
3
+ Copyright 2021 © Javier de la Rosa
4
+ """
5
+
6
+ # Dependencies
7
+ # !pip install -qU transformers sacrebleu scikit-learn datasets seqeval conllu pyarrow nltk
8
+
9
+ # Dependencies and helper functions
10
+ import argparse
11
+ import logging
12
+ import os
13
+ import random
14
+ import sys
15
+ from dataclasses import dataclass
16
+ from dataclasses import field
17
+ from pathlib import Path
18
+ from typing import Optional
19
+
20
+ import datasets
21
+ import numpy as np
22
+ import pandas as pd
23
+ # from datasets import ClassLabel
24
+ from datasets import load_dataset
25
+ from nltk.tokenize import word_tokenize
26
+ from nltk.tokenize.treebank import TreebankWordDetokenizer
27
+ from seqeval.metrics.sequence_labeling import accuracy_score as seq_accuracy_score
28
+ from seqeval.metrics.sequence_labeling import f1_score as seq_f1_score
29
+ from seqeval.metrics.sequence_labeling import precision_score as seq_precision_score
30
+ from seqeval.metrics.sequence_labeling import recall_score as seq_recall_score
31
+ from seqeval.metrics.sequence_labeling import classification_report as seq_classification_report
32
+ from sklearn.metrics import accuracy_score as sk_accuracy_score
33
+ from sklearn.metrics import f1_score as sk_f1_score
34
+ from sklearn.metrics import precision_score as sk_precision_score
35
+ from sklearn.metrics import recall_score as sk_recall_score
36
+ from sklearn.metrics import classification_report as sk_classification_report
37
+ # from sklearn.preprocessing import MultiLabelBinarizer
38
+ from tqdm import tqdm
39
+ from transformers import (
40
+ AutoConfig,
41
+ AutoModelForTokenClassification,
42
+ AutoModelForSequenceClassification,
43
+ AutoTokenizer,
44
+ RobertaTokenizer,
45
+ RobertaTokenizerFast,
46
+ DataCollatorForTokenClassification,
47
+ DataCollatorWithPadding,
48
+ PreTrainedTokenizerFast,
49
+ Trainer,
50
+ TrainingArguments,
51
+ pipeline,
52
+ set_seed,
53
+ )
54
+ # from transformers.training_args import TrainingArguments
55
+ import wandb
56
+
57
+ BIBLES_BASE_URI = "https://huggingface.co/datasets/linhd-postdata/stanzas/resolve/main"
58
+ BIBLES = {
59
+ "validation": f"{BIBLES_BASE_URI}/eval.csv",
60
+ "test": f"{BIBLES_BASE_URI}/test.csv",
61
+ "train": f"{BIBLES_BASE_URI}/train.csv"
62
+ }
63
+
64
+
65
+ # Helper Funtions
66
+ def printm(string):
67
+ print(str(string))
68
+
69
+
70
+ # Tokenize all texts and align the labels with them.
71
+ def tokenize_and_align_labels(
72
+ tokenizer, examples, text_column_name, max_length, padding,
73
+ label_column_name, label_to_id, label_all_tokens
74
+ ):
75
+ tokenized_inputs = tokenizer(
76
+ examples[text_column_name],
77
+ max_length=max_length,
78
+ padding=padding,
79
+ truncation=True,
80
+ # We use this argument because the texts in our dataset are lists of words (with a label for each word).
81
+ is_split_into_words=True,
82
+ )
83
+ labels = []
84
+ for i, label in enumerate(examples[label_column_name]):
85
+ word_ids = tokenized_inputs.word_ids(batch_index=i)
86
+ previous_word_idx = None
87
+ label_ids = []
88
+ for word_idx in word_ids:
89
+ # Special tokens have a word id that is None. We set the label to -100 so they are automatically
90
+ # ignored in the loss function.
91
+ if word_idx is None:
92
+ label_ids.append(-100)
93
+ # We set the label for the first token of each word.
94
+ elif word_idx != previous_word_idx:
95
+ label_ids.append(label_to_id[label[word_idx]])
96
+ # For the other tokens in a word, we set the label to either the current label or -100, depending on
97
+ # the label_all_tokens flag.
98
+ else:
99
+ label_ids.append(label_to_id[label[word_idx]] if label_all_tokens else -100)
100
+ previous_word_idx = word_idx
101
+
102
+ labels.append(label_ids)
103
+ tokenized_inputs["labels"] = labels
104
+ return tokenized_inputs
105
+
106
+
107
+ # Metrics
108
+ def token_compute_metrics(pairs, label_list):
109
+ """Token metrics based on seqeval"""
110
+ raw_predictions, labels = pairs
111
+ predictions = np.argmax(raw_predictions, axis=2)
112
+
113
+ # Remove ignored index (special tokens)
114
+ true_predictions = [
115
+ [label_list[p] for (p, l) in zip(prediction, label) if l != -100]
116
+ for prediction, label in zip(predictions, labels)
117
+ ]
118
+ true_probas = [
119
+ [label_list[p] for (p, l) in zip(prediction, label) if l != -100]
120
+ for prediction, label in zip(predictions, labels)
121
+ ]
122
+ true_labels = [
123
+ [label_list[l] for (p, l) in zip(prediction, label) if l != -100]
124
+ for prediction, label in zip(predictions, labels)
125
+ ]
126
+ raw_scores = (
127
+ np.exp(raw_predictions) / np.exp(raw_predictions).sum(-1, keepdims=True)
128
+ )
129
+ scores = raw_scores.max(axis=2)
130
+ true_scores = [
131
+ [(s, l) for (s, l) in zip(score, label) if l != -100]
132
+ for score, label in zip(scores, labels)
133
+ ]
134
+
135
+ # mlb = MultiLabelBinarizer() # sparse_output=True
136
+ # true_predictions = mlb.fit_transform(true_predictions)
137
+ # mlb = MultiLabelBinarizer() # sparse_output=True
138
+ # true_labels = mlb.fit_transform(true_labels)
139
+ # wandb.log({
140
+ # "roc" : wandb.plot.roc_curve(
141
+ # labels,
142
+ # predictions,
143
+ # labels=label_list
144
+ # )})
145
+ metrics = {
146
+ "accuracy": seq_accuracy_score(true_labels, true_predictions),
147
+ "precision_micro": seq_precision_score(true_labels, true_predictions, average="micro"),
148
+ "recall_micro": seq_recall_score(true_labels, true_predictions, average="micro"),
149
+ "f1_micro": seq_f1_score(true_labels, true_predictions, average="micro"),
150
+ "precision_macro": seq_precision_score(true_labels, true_predictions, average="macro"),
151
+ "recall_macro": seq_recall_score(true_labels, true_predictions, average="macro"),
152
+ "f1_macro": seq_f1_score(true_labels, true_predictions, average="macro"),
153
+ # "report": seq_classification_report(true_labels, true_predictions, digits=4)
154
+ }
155
+ reports = seq_classification_report(
156
+ true_labels, true_predictions, output_dict=True, zero_division=0,
157
+ )
158
+ for label, report in reports.items():
159
+ for metric_key, metric_value in report.items():
160
+ metric_title = metric_key.replace(" avg", "_avg", 1)
161
+ metrics.update({
162
+ f"label_{label}_{metric_title}": metric_value,
163
+ })
164
+ # labels_to_plot = label_list.copy()
165
+ # if "O" in labels_to_plot:
166
+ # labels_to_plot.remove("O")
167
+ flat_true_labels = sum(true_labels, [])
168
+ flat_true_predictions = sum(true_predictions, [])
169
+ wandb.log({
170
+ # "roc": wandb.plot.roc_curve(
171
+ # labels.reshape(-1),
172
+ # raw_scores.reshape(-1, raw_predictions.shape[-1]),
173
+ # labels=label_list,
174
+ # classes_to_plot=labels_to_plot,
175
+ # ),
176
+ "matrix": wandb.sklearn.plot_confusion_matrix(
177
+ flat_true_labels, flat_true_predictions, label_list
178
+ )
179
+ })
180
+ return metrics
181
+
182
+
183
+ def sequence_compute_metrics(pairs, label_list):
184
+ """Sequence metrics based on sklearn"""
185
+ raw_predictions, labels = pairs
186
+ predictions = np.argmax(raw_predictions, axis=1)
187
+ metrics = {
188
+ "accuracy": sk_accuracy_score(labels, predictions),
189
+ "precision_micro": sk_precision_score(labels, predictions, average="micro"),
190
+ "recall_micro": sk_recall_score(labels, predictions, average="micro"),
191
+ "f1_micro": sk_f1_score(labels, predictions, average="micro"),
192
+ "precision_macro": sk_precision_score(labels, predictions, average="macro"),
193
+ "recall_macro": sk_recall_score(labels, predictions, average="macro"),
194
+ "f1_macro": sk_f1_score(labels, predictions, average="macro"),
195
+ # "report": sk_classification_report(labels, predictions, digits=4)
196
+ }
197
+ reports = sk_classification_report(
198
+ labels, predictions, target_names=label_list, output_dict=True,
199
+ )
200
+ for label, report in reports.items():
201
+ if not isinstance(report, dict):
202
+ report = {"": report}
203
+ for metric_key, metric_value in report.items():
204
+ metric_title = metric_key.replace(" avg", "_avg", 1)
205
+ metrics.update({
206
+ f"label_{label}_{metric_title}": metric_value,
207
+ })
208
+ wandb.log({
209
+ "roc": wandb.plot.roc_curve(
210
+ labels, raw_predictions, labels=label_list
211
+ ),
212
+ "matrix": wandb.sklearn.plot_confusion_matrix(
213
+ labels, predictions, label_list
214
+ )
215
+ })
216
+ return metrics
217
+
218
+
219
+ def write_file(kind, metrics, output_dir, save_artifact=False):
220
+ output_file = output_dir / f"{kind}_results.txt"
221
+ headers = []
222
+ label_headers = []
223
+ data = []
224
+ label_data = []
225
+ with open(output_file, "w") as writer:
226
+ printm(f"**{kind.capitalize()} results**")
227
+ for key, value in metrics.items():
228
+ printm(f"\t{key} = {value}")
229
+ writer.write(f"{key} = {value}\n")
230
+ title = key.replace("eval_", "", 1)
231
+ if title.startswith("label_"):
232
+ label_headers.append(title.replace("label_", "", 1))
233
+ label_data.append(value)
234
+ else:
235
+ headers.append(title)
236
+ data.append(value)
237
+ wandb.log({f"{kind}:{title}": value})
238
+ wandb.log({kind: wandb.Table(data=[data], columns=headers)})
239
+ if label_headers:
240
+ wandb.log({
241
+ f"{kind}:labels": wandb.Table(
242
+ data=[label_data], columns=label_headers
243
+ )
244
+ })
245
+ if save_artifact:
246
+ artifact = wandb.Artifact(kind, type="result")
247
+ artifact.add_file(str(output_file))
248
+ wandb.log_artifact(artifact)
249
+
250
+
251
+ def dataset_select(dataset, size):
252
+ dataset_len = len(dataset)
253
+ if size < 0 or size > dataset_len:
254
+ return dataset
255
+ elif size <= 1: # it's a percentage
256
+ return dataset.select(range(int(size * dataset_len)))
257
+ else: # it's a number
258
+ return dataset.select(range(int(size)))
259
+
260
+
261
+ def main(args):
262
+ # Set seed
263
+ if args.run:
264
+ seed = random.randrange(10**3)
265
+ else:
266
+ seed = args.seed
267
+ set_seed(seed)
268
+ # Run name
269
+ model_name = args.model_name
270
+ model_name = model_name[2:] if model_name.startswith("./") else model_name
271
+ model_name = model_name[1:] if model_name.startswith("/") else model_name
272
+ run_name = f"{model_name}_{args.task_name}"
273
+ run_name = f"{run_name}_{args.dataset_config or args.dataset_name}"
274
+ run_name = run_name.replace("/", "-")
275
+ run_name = f"{run_name}_l{str(args.dataset_language)}"
276
+ run_name = f"{run_name}_c{str(args.dataset_century)}"
277
+ run_name = f"{run_name}_e{str(args.num_train_epochs)}"
278
+ run_name = f"{run_name}_lr{str(args.learning_rate)}"
279
+ run_name = f"{run_name}_ws{str(args.warmup_steps)}"
280
+ run_name = f"{run_name}_wd{str(args.weight_decay)}"
281
+ run_name = f"{run_name}_s{str(seed)}"
282
+ run_name = f"{run_name}_eas{str(args.eval_accumulation_steps)}"
283
+ if args.max_length != 512:
284
+ run_name = f"{run_name}_seq{str(args.max_length)}"
285
+ if args.label_all_tokens:
286
+ run_name = f"{run_name}_labelall"
287
+ if args.run:
288
+ run_name = f"{run_name}_r{str(args.run)}"
289
+ output_dir = Path(args.output_dir) / run_name
290
+ # Tokenizer settings
291
+ padding = "longest" # args.task_name not in ("ner", "pos") # default: False @param ["False", "'max_length'"] {type: 'raw'}
292
+ max_length = args.max_length #@param {type: "number"}
293
+ # Training settings
294
+ weight_decay = args.weight_decay #@param {type: "number"}
295
+ adam_beta1 = 0.9 #@param {type: "number"}
296
+ adam_beta2 = 0.999 #@param {type: "number"}
297
+ adam_epsilon = 1e-08 #@param {type: "number"}
298
+ max_grad_norm = 1.0 #@param {type: "number"}
299
+ save_total_limit = 1 #@param {type: "integer"}
300
+ load_best_model_at_end = False #@param {type: "boolean"}
301
+ # wandb
302
+ wandb.init(name=run_name, project="postdata")
303
+ wandb.log({
304
+ "seed": int(seed),
305
+ })
306
+ # Loading Dataset
307
+ print("\n\n#####################################")
308
+ print(args.model_name)
309
+ print(args.task_name)
310
+ print(args.dataset_config)
311
+ print(args.dataset_language)
312
+ print(args.dataset_century)
313
+ train_split = args.dataset_split_train
314
+ test_split = args.dataset_split_test
315
+ validation_split = args.dataset_split_validation
316
+ if ":" in args.dataset_name:
317
+ dataset_name, dataset_config = args.dataset_name.split(":")
318
+ else:
319
+ dataset_name = args.dataset_name
320
+ dataset_config = args.dataset_config
321
+ use_auth_token = os.environ.get("AUTH_TOKEN", None)
322
+ if dataset_config is None or len(dataset_config) == 0:
323
+ dataset = load_dataset(dataset_name, use_auth_token=use_auth_token)
324
+ elif dataset_name == "csv" and dataset_config:
325
+ dataset = load_dataset(
326
+ dataset_name,
327
+ data_files={
328
+ "train": BIBLES["train"](dataset_config),
329
+ "validation": BIBLES["validation"](dataset_config),
330
+ "test": BIBLES["test"](dataset_config),
331
+ },
332
+ use_auth_token=use_auth_token)
333
+ else:
334
+ dataset = load_dataset(dataset_name, dataset_config, use_auth_token=use_auth_token)
335
+ if args.dataset_language and args.dataset_language.lower() not in ("all", "balanced"):
336
+ dataset = dataset.filter(lambda x: x["language"] == args.dataset_language)
337
+ if args.dataset_century and args.dataset_century.lower() != "all":
338
+ dataset = dataset.filter(lambda x: x["century"] in args.dataset_century)
339
+ if dataset["train"].shape[0] == 0 or dataset["test"].shape[0] == 0 or dataset["validation"].shape[0] == 0:
340
+ print(f"Not enough data for {str(args.dataset_language)} on {str(args.dataset_century)}: {str(dataset.shape)}")
341
+ return
342
+ column_names = dataset[train_split].column_names
343
+ features = dataset[train_split].features
344
+ if "tokens" in column_names:
345
+ text_column_name = "tokens"
346
+ elif "text" in column_names:
347
+ text_column_name = "text"
348
+ else:
349
+ text_column_name = column_names[0]
350
+ if f"{args.task_name}_tags" in column_names:
351
+ label_column_name = f"{args.task_name}_tags"
352
+ elif "label" in column_names:
353
+ label_column_name = "label"
354
+ else:
355
+ label_column_name = column_names[1]
356
+ if dataset_name == "csv":
357
+ label_list = list(set(dataset[train_split][label_column_name]))
358
+ elif isinstance(features[label_column_name], datasets.features.Sequence):
359
+ label_list = features[label_column_name].feature.names
360
+ else:
361
+ label_list = features[label_column_name].names
362
+ label_to_id = {i: i for i in range(len(label_list))}
363
+ num_labels = len(label_list)
364
+ print(f"Number of labels: {num_labels}")
365
+ print({label.split("-")[-1] for label in label_list})
366
+
367
+ # Training
368
+ config = AutoConfig.from_pretrained(
369
+ args.model_name,
370
+ num_labels=num_labels,
371
+ finetuning_task=args.task_name,
372
+ cache_dir=args.cache_dir,
373
+ force_download=args.force_download,
374
+ )
375
+ tokenizer = AutoTokenizer.from_pretrained(
376
+ args.model_name,
377
+ cache_dir=args.cache_dir,
378
+ use_fast=True,
379
+ force_download=args.force_download,
380
+ )
381
+ if isinstance(tokenizer, (RobertaTokenizer, RobertaTokenizerFast)):
382
+ tokenizer = AutoTokenizer.from_pretrained(
383
+ args.model_name,
384
+ cache_dir=args.cache_dir,
385
+ use_fast=True,
386
+ force_download=args.force_download,
387
+ add_prefix_space=True,
388
+ )
389
+
390
+ tokenizer_test_sentence = """
391
+ Ya que el Ángel del Señor tiene nombre propio, y su nombre es Yahveh.
392
+ """.strip()
393
+ printm("""Tokenizer test""")
394
+ printm(f"> {tokenizer_test_sentence}")
395
+ printm(tokenizer.tokenize(tokenizer_test_sentence))
396
+ printm(tokenizer(tokenizer_test_sentence).input_ids)
397
+ # STILTs
398
+ is_stilt = args.model_name in args.stilt.split(",") or args.stilt == "all"
399
+ model_config = dict(
400
+ from_tf=bool(".ckpt" in args.model_name),
401
+ config=config,
402
+ cache_dir=args.cache_dir,
403
+ force_download=args.force_download,
404
+ )
405
+ # Token tasks
406
+ if args.task_name in ("pos", "ner"):
407
+ if is_stilt:
408
+ # model = AutoModelForTokenClassification.from_config(
409
+ # config=config
410
+ # )
411
+ model_config.pop("config")
412
+ model = AutoModelForTokenClassification.from_pretrained(
413
+ args.model_name, num_labels=num_labels, ignore_mismatched_sizes=True, **model_config,
414
+ )
415
+ else:
416
+ model = AutoModelForTokenClassification.from_pretrained(
417
+ args.model_name, **model_config,
418
+ )
419
+ # Preprocessing the dataset
420
+ tokenized_datasets = dataset.map(
421
+ lambda examples: tokenize_and_align_labels(
422
+ tokenizer, examples, text_column_name, max_length, padding,
423
+ label_column_name, label_to_id, args.label_all_tokens),
424
+ batched=True,
425
+ load_from_cache_file=not args.overwrite_cache,
426
+ num_proc=os.cpu_count(),
427
+ )
428
+ # Data collator
429
+ data_collator = DataCollatorForTokenClassification(tokenizer)
430
+ compute_metrics = token_compute_metrics
431
+ # Sequence tasks
432
+ else:
433
+ if is_stilt:
434
+ # model = AutoModelForSequenceClassification.from_config(
435
+ # config=config
436
+ # )
437
+ model_config.pop("config")
438
+ model = AutoModelForSequenceClassification.from_pretrained(
439
+ args.model_name, num_labels=num_labels, ignore_mismatched_sizes=True, **model_config,
440
+ )
441
+ else:
442
+ model = AutoModelForSequenceClassification.from_pretrained(
443
+ args.model_name, **model_config,
444
+ )
445
+ # Preprocessing the dataset
446
+ tokenized_datasets = dataset.map(
447
+ lambda examples: tokenizer(
448
+ examples[text_column_name],
449
+ max_length=max_length,
450
+ padding=padding,
451
+ truncation=True,
452
+ is_split_into_words=False,
453
+ ),
454
+ batched=True,
455
+ load_from_cache_file=not args.overwrite_cache,
456
+ num_proc=os.cpu_count(),
457
+ )
458
+ # Data collator
459
+ data_collator = DataCollatorWithPadding(
460
+ tokenizer,
461
+ max_length=max_length,
462
+ padding=padding,
463
+ )
464
+ compute_metrics = sequence_compute_metrics
465
+ train_dataset = dataset_select(
466
+ tokenized_datasets[train_split], args.max_train_size
467
+ )
468
+ test_dataset = dataset_select(
469
+ tokenized_datasets[test_split], args.max_test_size
470
+ )
471
+ validation_dataset = dataset_select(
472
+ tokenized_datasets[validation_split], args.max_validation_size
473
+ )
474
+ wandb.log({
475
+ "train_size": len(train_dataset),
476
+ "test_size": len(test_dataset),
477
+ "validation_size": len(validation_dataset),
478
+ })
479
+ samples_per_batch = (
480
+ train_dataset.shape[0] / args.train_batch_size
481
+ )
482
+ total_steps = args.num_train_epochs * samples_per_batch
483
+ warmup_steps = int(args.warmup_steps * total_steps)
484
+ wandb.log({
485
+ "total_steps": int(total_steps),
486
+ "total_warmup_steps": warmup_steps
487
+ })
488
+ do_eval = args.do_eval and (validation_split in tokenized_datasets)
489
+ do_test = args.do_test and (test_split in tokenized_datasets)
490
+ do_predict = args.do_predict and (test_split in tokenized_datasets)
491
+ training_args = TrainingArguments(
492
+ output_dir=output_dir.as_posix(),
493
+ overwrite_output_dir=args.overwrite_output_dir,
494
+ do_train=args.do_train,
495
+ do_eval=do_eval,
496
+ do_predict=do_test or do_predict,
497
+ per_device_train_batch_size=int(args.train_batch_size),
498
+ per_device_eval_batch_size=int(args.eval_batch_size or args.train_batch_size),
499
+ learning_rate=float(args.learning_rate),
500
+ weight_decay=weight_decay,
501
+ adam_beta1=adam_beta1,
502
+ adam_beta2=adam_beta2,
503
+ adam_epsilon=adam_epsilon,
504
+ max_grad_norm=max_grad_norm,
505
+ num_train_epochs=args.num_train_epochs,
506
+ warmup_steps=warmup_steps,
507
+ load_best_model_at_end=load_best_model_at_end,
508
+ seed=seed,
509
+ save_total_limit=save_total_limit,
510
+ run_name=run_name,
511
+ disable_tqdm=False,
512
+ eval_steps=1000,
513
+ eval_accumulation_steps=args.eval_accumulation_steps or None, # it was not set
514
+ dataloader_num_workers=64, # it was not set
515
+ )
516
+ # Initialize our Trainer
517
+ trainer = Trainer(
518
+ model=model,
519
+ args=training_args,
520
+ train_dataset=train_dataset,
521
+ eval_dataset=validation_dataset if do_eval else None,
522
+ tokenizer=tokenizer,
523
+ data_collator=data_collator,
524
+ compute_metrics=lambda pairs: compute_metrics(pairs, label_list),
525
+ )
526
+ if args.do_train:
527
+ train_result = trainer.train()
528
+ trainer.save_model() # Saves the tokenizer too for easy upload
529
+ write_file("train", train_result.metrics, output_dir, save_artifact=args.save_artifacts)
530
+ # Need to save the state, since Trainer.save_model saves only the tokenizer with the model
531
+ trainer.state.save_to_json(output_dir / "trainer_state.json")
532
+ # Evaluation
533
+ if do_eval:
534
+ printm(f"**Evaluate**")
535
+ results = trainer.evaluate()
536
+ write_file("eval", results, output_dir, save_artifact=args.save_artifacts)
537
+ # Tesing and predicting
538
+ if do_test or do_predict:
539
+ printm("**Test**")
540
+ predictions, labels, metrics = trainer.predict(test_dataset)
541
+ if not do_predict:
542
+ write_file("test", metrics, output_dir, save_artifact=args.save_artifacts)
543
+ if args.task_name in ("ner", "pos"):
544
+ predictions = np.argmax(predictions, axis=2)
545
+ # Remove ignored index (special tokens)
546
+ true_predictions = [
547
+ [label_list[p] for (p, l) in zip(prediction, label) if l != -100]
548
+ for prediction, label in zip(predictions, labels)
549
+ ]
550
+ else:
551
+ predictions = np.argmax(predictions, axis=1)
552
+ true_predictions = [
553
+ label_list[p] for (p, l) in zip(predictions, labels) if l != -100
554
+ ]
555
+ # Save predictions
556
+ output_test_predictions_file = os.path.join(output_dir, "test_predictions.txt")
557
+ output_test_predictions = "\n".join(" ".join(map(str, p)) for p in true_predictions)
558
+ with open(output_test_predictions_file, "a+") as writer:
559
+ writer.write(output_test_predictions)
560
+ if args.save_artifacts:
561
+ artifact = wandb.Artifact("predictions", type="result")
562
+ artifact.add_file(output_test_predictions_file)
563
+ wandb.log_artifact(artifact)
564
+ # # Log the results
565
+ # logfile = output_dir / "evaluation.csv"
566
+ # # Check if logfile exist
567
+ # try:
568
+ # f = open(logfile)
569
+ # f.close()
570
+ # except FileNotFoundError:
571
+ # with open(logfile, 'a+') as f:
572
+ # f.write("model_name" + "\t" + "data_language" + "\t" + "task_name" + "\t" "learning_rate"+ "\t" + "num_epochs"+ "\t" + "warmup_steps"+ "\t" + "validation_f1" +"\t"+"test_f1"+"\n")
573
+ # with open(logfile, 'a') as f:
574
+ # print(results)
575
+ # f.write(args.model_name + "\t" + (args.dataset_config or args.dataset_name) + "\t" + args.task_name + "\t" + str(args.learning_rate) + "\t" + str(args.num_train_epochs)+ "\t" + str(warmup_steps)+ "\t" + str(results['eval_f1']) + "\t" + str(metrics['eval_f1']) + "\n")
576
+
577
+
578
+ if __name__ == "__main__":
579
+ # yesno = lambda x: str(x).lower() in {'true', 't', '1', 'yes', 'y'}
580
+ parser = argparse.ArgumentParser(description=f""
581
+ f"Evaluating BERT models for sequence classification on Bibles"""
582
+ f"", epilog=f"""Example usage:
583
+ {__file__} --task_name sequence --model_name "bert-base-multilingual-cased"
584
+ """, formatter_class=argparse.RawTextHelpFormatter)
585
+ parser.add_argument('--model_name',
586
+ metavar='model_name', help='Model name or path')
587
+ parser.add_argument('--dataset_name', default="csv",
588
+ metavar='dataset_name', help='Dataset name. It might enforce a config if added after a semicolon: "conll2002:es". This will ignore dataset_config, useful when run in grid search')
589
+ parser.add_argument('--dataset_config',
590
+ metavar='dataset_config', help='Dataset config name')
591
+ parser.add_argument('--dataset_language', default="all",
592
+ metavar='dataset_language', help='Dataset language name')
593
+ parser.add_argument('--dataset_century', default="all",
594
+ metavar='dataset_century', help='Dataset century')
595
+
596
+ parser.add_argument('--dataset_split_train', default="train",
597
+ metavar='dataset_split_train', help='Dataset train split name')
598
+ parser.add_argument('--dataset_split_test', default="test",
599
+ metavar='dataset_split_test', help='Dataset test split name')
600
+ parser.add_argument('--dataset_split_validation', default="validation",
601
+ metavar='dataset_split_validation', help='Dataset validation split name')
602
+
603
+ parser.add_argument('--max_train_size', type=float, default=-1.0,
604
+ metavar='max_train_size', help='Percentage of train dataset or number of rows to use')
605
+ parser.add_argument('--max_test_size', type=float, default=-1.0,
606
+ metavar='max_test_size', help='Percentage of test dataset or number of rows to use')
607
+ parser.add_argument('--max_validation_size', type=float, default=-1.0,
608
+ metavar='max_validation_size', help='Percentage of validation dataset or number of rows to use')
609
+
610
+ parser.add_argument('--do_train',
611
+ metavar='do_train', default=True, type=bool,
612
+ help='Run training',
613
+ )
614
+ parser.add_argument('--do_eval',
615
+ metavar='do_eval', default=True, type=bool,
616
+ help='Run evaluation on validation test',
617
+ )
618
+ parser.add_argument('--do_test',
619
+ metavar='do_test', default=True, type=bool,
620
+ help='Run evaluation on test set',
621
+ )
622
+ parser.add_argument('--do_predict',
623
+ metavar='do_predict', default=False, type=bool,
624
+ help='Run prediction only on test set',
625
+ )
626
+ parser.add_argument('--task_name',
627
+ metavar='task_name', default="ner",
628
+ help='Task name (supported in the dataset), either ner or pos',
629
+ )
630
+ parser.add_argument('--num_train_epochs',
631
+ metavar='num_train_epochs', default=4, type=float,
632
+ help='Number of training epochs',
633
+ )
634
+ parser.add_argument('--eval_accumulation_steps',
635
+ metavar='eval_accumulation_steps', default=0, type=int,
636
+ help='Number of predictions steps to accumulate the output tensors for, before moving the results to the CPU.',
637
+ )
638
+ parser.add_argument('--cache_dir',
639
+ metavar='cache_dir', default="/var/ml/cache/",
640
+ help='Cache dir for the transformer library',
641
+ )
642
+ parser.add_argument('--overwrite_cache',
643
+ metavar='overwrite_cache', default=False, type=bool,
644
+ help='Overwrite cache dir if present',
645
+ )
646
+ parser.add_argument('--output_dir',
647
+ metavar='output_dir', default="/var/ml/output/",
648
+ help='Output dir for models and logs',
649
+ )
650
+ parser.add_argument('--overwrite_output_dir',
651
+ metavar='overwrite_output_dir', default=True, type=bool,
652
+ help='Overwrite output dir if present',
653
+ )
654
+ parser.add_argument('--seed',
655
+ metavar='seed', type=int, default=2021,
656
+ help='Seed for the experiments',
657
+ )
658
+ parser.add_argument('--run',
659
+ metavar='run', type=int,
660
+ help='Control variable for doing several runs of the same experiment. It will force random seeds even across the same set of parameters fo a grid search',
661
+ )
662
+ parser.add_argument('--train_batch_size',
663
+ metavar='train_batch_size', type=int, default=8,
664
+ help='Batch size for training',
665
+ )
666
+ parser.add_argument('--eval_batch_size',
667
+ metavar='eval_batch_size', type=int,
668
+ help='Batch size for evaluation. Defaults to train_batch_size',
669
+ )
670
+ parser.add_argument('--max_length',
671
+ metavar='max_length', type=int, default=512,
672
+ help='Maximum sequence length',
673
+ )
674
+ parser.add_argument('--learning_rate',
675
+ metavar='learning_rate', type=str, default="3e-05",
676
+ help='Learning rate',
677
+ )
678
+ parser.add_argument('--warmup_steps',
679
+ metavar='warmup_steps', type=float, default=0.0,
680
+ help='Warmup steps as percentage of the total number of steps',
681
+ )
682
+ parser.add_argument('--weight_decay',
683
+ metavar='weight_decay', type=float, default=0.0,
684
+ help='Weight decay',
685
+ )
686
+ parser.add_argument('--label_all_tokens',
687
+ metavar='label_all_tokens', type=bool, default=False,
688
+ help=('Whether to put the label for one word on all tokens of '
689
+ 'generated by that word or just on the one (in which case the '
690
+ 'other tokens will have a padding index).'),
691
+ )
692
+ parser.add_argument('--force_download',
693
+ metavar='force_download', type=bool, default=False,
694
+ help='Force the download of model, tokenizer, and config',
695
+ )
696
+ parser.add_argument('--save_artifacts',
697
+ metavar='save_artifacts', type=bool, default=False,
698
+ help='Save train, eval, and test files in Weight & Biases',
699
+ )
700
+ parser.add_argument('--stilt',
701
+ metavar='stilt', type=str, default="",
702
+ help='Specify models already fine-tuned for other tasks',
703
+ )
704
+
705
+ args = parser.parse_args()
706
+ main(args)