Mlm and llrd

notebooks/training/bert_masked_lm.py

@@ -0,0 +1,112 @@
+
+import argparse
+import json
+import numpy as np
+import torch as th
+
+from huggingface_hub import login
+from transformers import (AutoTokenizer, AutoModelForMaskedLM,
+                          Trainer, TrainingArguments)
+
+from utils.generic_utils import preprocess_tweet
+
+
+LABEL_NAMES = [
+    'Alakasiz', 'Barinma', 'Elektronik',
+    'Giysi', 'Kurtarma', 'Lojistik', 'Saglik',
+    'Su', 'Guvenlik', 'Yemek']
+
+
+class DepremTweetUnlabeledDataset(th.utils.data.Dataset):
+
+    def __init__(self, tweets, tokenizer):
+        self.tweets = tweets
+        self.tokenizer = tokenizer
+
+    def __len__(self):
+        return len(self.tweets)
+
+    def __getitem__(self, idx):
+        text = self.tweets[idx]
+        encoding = self.tokenizer(text, max_length=64, padding="max_length", truncation=True)
+        encoding = {key: th.tensor(val) for key, val in encoding.items()}
+        encoding["labels"] = encoding["input_ids"]
+
+        inp = encoding["input_ids"]
+        tokens = range(len(inp))
+        # We need to select 15% random tokens from the given list
+        num_of_token_to_mask = round(len(tokens) * 0.15)
+        token_to_mask = np.random.choice(np.array(tokens),
+                                         size=num_of_token_to_mask,
+                                         replace=False).tolist()
+        # Now we have the indices where we need to mask the tokens
+        inp[token_to_mask] = self.tokenizer.mask_token_id
+        encoding["input_ids"] = inp
+
+        return dict(encoding)
+
+
+def prepare_datasets(json_path, tokenizer):
+    with open(json_path, 'r') as f:
+        json_ = json.load(f)
+
+    tweets = [preprocess_tweet(tweet['full_text'], remove_hashtags=False, remove_handles=False)
+              for tweet in json_]
+    tweets = list(set(tweets))  # Remove duplicates
+    print("Number of tweets: {}".format(len(tweets)))
+    print("Sample tweets:")
+    print(np.random.choice(tweets, 10))
+    print()
+
+    n_train = int(len(tweets) * 0.85)
+    train_ds = DepremTweetUnlabeledDataset(tweets[:n_train], tokenizer)
+    val_ds = DepremTweetUnlabeledDataset(tweets[n_train:], tokenizer)
+
+    return train_ds, val_ds
+
+
+def clean_dataset(dataset):
+    raise NotImplementedError
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_name", type=str, default="dbmdz/bert-base-turkish-uncased")
+    parser.add_argument("--output_dir", type=str, default="./output-intent")
+    parser.add_argument("--hf_token", type=str, required=True)
+    args = parser.parse_args()
+
+    login(token=args.hf_token)
+
+    tokenizer = AutoTokenizer.from_pretrained(args.model_name)
+    train_ds, val_ds = prepare_datasets(
+        "postgres_public_feeds_entry.json", tokenizer)
+    # Note: above line of code could be replaced with downloading the dataset
+    # from HF and preprocessing it (see next two lines for example)
+    # train_ds = datasets.load_dataset("deprem-private/deprem_tweet_unlabeled", "plain_text")
+    # train_ds = clean_dataset(train_ds)
+
+    model = AutoModelForMaskedLM.from_pretrained(args.model_name)
+
+    training_args = TrainingArguments(
+        output_dir=args.output_dir,
+        evaluation_strategy="steps",
+        save_strategy="steps",
+        per_device_train_batch_size=32,
+        per_device_eval_batch_size=32,
+        weight_decay=0.01,
+        num_train_epochs=1,
+        eval_steps=1000,
+        logging_steps=1000
+    )
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        train_dataset=train_ds,
+        eval_dataset=val_ds
+    )
+    trainer.train()
+
+
+if __name__ == '__main__':
+    main()

notebooks/training/intent_bert_llrd.py

@@ -0,0 +1,175 @@
+
+import argparse
+import numpy as np
+import os
+import torch as th
+
+from collections import OrderedDict
+from huggingface_hub import login
+from sklearn.metrics import classification_report
+from transformers import (AdamW, AutoTokenizer, AutoModelForSequenceClassification,
+                          DataCollatorWithPadding, Trainer, TrainingArguments, EarlyStoppingCallback,
+                          get_cosine_schedule_with_warmup)
+
+from utils.generic_utils import set_seed_everywhere, select_thresholds, compute_f1
+from utils.dataset_utils import prep_datasets
+from utils.training_utils import ImbalancedTrainer, compute_class_weights
+
+
+LABEL_IDX2NAME = OrderedDict([
+        (0, 'Lojistik'),
+        (1, 'Elektrik Kaynagi'),
+        (2, 'Arama Ekipmani'),
+        (3, 'Cenaze'),
+        (4, 'Giysi'),
+        (5, 'Enkaz Kaldirma'),
+        (6, 'Isinma'),
+        (7, 'Barınma'),
+        (8, 'Tuvalet'),
+        (9, 'Su'),
+        (10, 'Yemek'),
+        (11, 'Saglik'),
+        (12, 'Alakasiz')])
+
+os.environ["WANDB_DISABLED"] = "true"
+os.environ['TRANSFORMERS_NO_ADVISORY_WARNINGS'] = 'true'
+
+
+def get_optimizer_grouped_parameters(
+    model, model_type,
+    learning_rate, weight_decay,
+    layerwise_learning_rate_decay
+):
+    no_decay = ["bias", "LayerNorm.weight"]
+    # initialize lr for task specific layer
+    optimizer_grouped_parameters = [
+        {
+            "params": [p for n, p in model.named_parameters() if "classifier" in n or "pooler" in n],
+            "weight_decay": 0.0,
+            "lr": learning_rate,
+        },
+    ]
+    # initialize lrs for every layer
+    num_layers = model.config.num_hidden_layers
+    layers = [getattr(model, model_type).embeddings] + list(getattr(model, model_type).encoder.layer)
+    layers.reverse()
+    lr = learning_rate
+    for layer in layers:
+        lr *= layerwise_learning_rate_decay
+        optimizer_grouped_parameters += [
+            {
+                "params": [p for n, p in layer.named_parameters() if not any(nd in n for nd in no_decay)],
+                "weight_decay": weight_decay,
+                "lr": lr,
+            },
+            {
+                "params": [p for n, p in layer.named_parameters() if any(nd in n for nd in no_decay)],
+                "weight_decay": 0.0,
+                "lr": lr,
+            },
+        ]
+    return optimizer_grouped_parameters
+
+
+def get_llrd_optimizer_scheduler(model, learning_rate=1e-5, weight_decay=0.01, layerwise_learning_rate_decay=0.95):
+    grouped_optimizer_params = get_optimizer_grouped_parameters(
+        model, 'bert',
+        learning_rate, weight_decay,
+        layerwise_learning_rate_decay
+    )
+    optimizer = AdamW(
+        grouped_optimizer_params,
+        lr=learning_rate,
+        eps=1e-6,
+        correct_bias=True
+    )
+    scheduler = get_cosine_schedule_with_warmup(
+        optimizer,
+        num_warmup_steps=0,
+        num_training_steps=15
+    )
+    # Note: linear schedule fails to converge for unknown reasons.
+
+    return optimizer, scheduler
+
+
+def main():
+    # Define argpars for training parameters
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--n_seeds", type=int, default=1, help="Number of trials to run with different seeds")
+    parser.add_argument("--model_name", type=str, default="dbmdz/bert-base-turkish-uncased",
+                        help="Name or path of the model to use. For example, could be"
+                             "<path-to-BERT-finetuned-for-MLM-on-unlabelled-tweets>")
+    parser.add_argument("--output_dir", type=str, default="./output-intent")
+    parser.add_argument("--batch_size", type=int, default=32)
+    parser.add_argument("--hf_token", type=str, required=True)
+    parser.add_argument("--layerwise_LR_decay_rate", type=float, default=0.8)
+    args = parser.parse_args()
+
+    login(token=args.hf_token)
+    tokenizer = AutoTokenizer.from_pretrained("dbmdz/bert-base-turkish-uncased")
+    data_collator = DataCollatorWithPadding(tokenizer, padding="max_length", max_length=100)
+    train_ds, val_ds, test_ds, mlb_labels = prep_datasets(
+        tokenizer,
+        labelidx2name=LABEL_IDX2NAME,
+        path="deprem-private/intent-v13")
+
+    f1s = []
+    for i in range(args.n_seeds):
+        set_seed_everywhere(i)
+        model = AutoModelForSequenceClassification.from_pretrained(
+            args.model_name, num_labels=len(LABEL_IDX2NAME), problem_type="multi_label_classification")
+
+        training_args = TrainingArguments(
+            output_dir=args.output_dir,
+            evaluation_strategy="epoch",
+            save_strategy="epoch",
+            per_device_train_batch_size=args.batch_size,
+            per_device_eval_batch_size=args.batch_size * 2,
+            report_to=None,
+            num_train_epochs=15,
+            metric_for_best_model="macro f1",
+            load_best_model_at_end=True,
+            group_by_length=True
+        )
+        optimizer, scheduler = get_llrd_optimizer_scheduler(
+            model,
+            learning_rate=5e-5,
+            weight_decay=0.01,  # Weight decay defined here instead of training_args
+            layerwise_learning_rate_decay=args.layerwise_LR_decay_rate)
+
+        trainer = ImbalancedTrainer(
+            class_weights=compute_class_weights(mlb_labels),
+            model=model,
+            args=training_args,
+            train_dataset=train_ds,
+            eval_dataset=val_ds,
+            data_collator=data_collator,
+            compute_metrics=compute_f1,
+            callbacks=[EarlyStoppingCallback(early_stopping_patience=3)],
+            optimizers=(optimizer, scheduler)
+        )
+        trainer.train()
+
+        # Choose the best thresholds per label using train+val data
+        train_preds = trainer.predict(train_ds)
+        val_preds = trainer.predict(val_ds)
+        thresholds = select_thresholds(
+            np.concatenate([train_preds.label_ids, val_preds.label_ids]),
+            np.concatenate([train_preds.predictions, val_preds.predictions])
+        )
+        # Evaluate on test data
+        test_preds = trainer.predict(test_ds)
+        f1 = compute_f1((test_preds.predictions, test_preds.label_ids), thresholds=thresholds)
+        f1s.append(f1["macro f1"])
+        report = classification_report(
+            test_preds.label_ids.astype(int),
+            (th.sigmoid(th.from_numpy(test_preds.predictions)).numpy() > thresholds).astype(int),
+            target_names=LABEL_IDX2NAME.values(), digits=3)
+        print(report)
+
+    print("Mean F1: {:.2f}, Std F1: {:.2f}".format(np.mean(f1s) * 100, np.std(f1s) * 100))
+
+
+if __name__ == '__main__':
+    main()

utils/dataset_utils.py

@@ -0,0 +1,82 @@
+
+import numpy as np
+import pandas as pd
+import torch as th
+
+from datasets import load_dataset
+from iterstrat.ml_stratifiers import MultilabelStratifiedKFold
+from sklearn.preprocessing import MultiLabelBinarizer
+from unicode_tr import unicode_tr
+
+
+def prep_datasets(tokenizer, labelidx2name, path, label_col="label", text_col="image_url"):
+    intent = load_dataset(path, use_auth_token=True)
+    print(intent["train"], intent["test"])
+
+    for instance in intent["train"]:
+        print(unicode_tr(instance["image_url"]).lower())
+        break
+
+    df_train = pd.DataFrame().from_records(list(intent["train"]))
+    df_test = pd.DataFrame().from_records(list(intent["test"]))
+
+    df_train[text_col] = df_train[text_col].apply(lambda x: unicode_tr(x).lower())
+    df_test[text_col] = df_test[text_col].apply(lambda x: unicode_tr(x).lower())
+
+    # Next, we remove the rows that have no labels
+    df_train = df_train[df_train[label_col].notnull()].reset_index(drop=True)
+    df_test = df_test[df_test[label_col].notnull()].reset_index(drop=True)
+
+    # df_train.labels.apply(lambda x: len(x))
+    #
+    # labels = set()
+    # for label in df_train.labels.values:
+    #     labels.update({l for l in label})
+    #
+    # name2ix = {v: k for k, v in labelidx2name.items()}
+    # labels = name2ix.keys()
+
+    mlb = MultiLabelBinarizer(classes=list(labelidx2name.values()))
+    mlb_labels = mlb.fit_transform(df_train.label.tolist())
+
+    cv = MultilabelStratifiedKFold(n_splits=5, shuffle=True, random_state=42)
+
+    for fold, (train_idx, val_idx) in enumerate(cv.split(df_train.index.tolist(), mlb_labels)):
+        df_train.loc[val_idx, 'kfold'] = int(fold)
+
+    df_train, df_val = df_train[df_train['kfold'] != 0], df_train[df_train['kfold'] == 0]
+
+    train_ds = IntentDataset(df_train, tokenizer, labelidx2name, label_col, text_col)
+    val_ds = IntentDataset(df_val, tokenizer, labelidx2name, label_col, text_col)
+    test_ds = IntentDataset(df_test, tokenizer, labelidx2name, label_col, text_col)
+
+    return train_ds, val_ds, test_ds, mlb_labels
+
+
+class IntentDataset(th.utils.data.Dataset):
+
+    def __init__(self, df, tokenizer, labelidx2name, label_col="label", text_col="image_url"):
+        self.df = df
+        self.tokenizer = tokenizer
+        self.labelidx2name = labelidx2name
+        self.name2ix = {v: k for k, v in labelidx2name.items()}
+        self.num_classes = len(labelidx2name)
+        self.label_col = label_col
+        self.text_col = text_col
+
+    def __len__(self):
+        return len(self.df)
+
+    def __getitem__(self, idx):
+        row = self.df.iloc[idx]
+        text, label = row.image_url, self._encode_label(row[self.label_col])
+        encoding = self.tokenizer(text, max_length=100, padding="max_length", truncation=True)
+        encoding = {key: th.tensor(val, dtype=th.int64) for key, val in encoding.items()}
+        encoding[self.label_col] = th.tensor(label, dtype=th.float32)
+        return dict(encoding)
+
+    def _encode_label(self, labels):
+        encoded_labels = np.zeros(self.num_classes)
+        for label in labels:
+            encoded_labels[self.name2ix[label]] = 1.0
+        return encoded_labels