SinaTools 0.1.1__py2.py3-none-any.whl

nlptools/arabiner/data/datasets.py

@@ -0,0 +1,146 @@
+import logging
+import torch
+from torch.utils.data import Dataset
+from torch.nn.utils.rnn import pad_sequence
+from nlptools.arabiner.data.transforms import (
+    BertSeqTransform,
+    NestedTagsTransform
+)
+
+logger = logging.getLogger(__name__)
+
+
+class Token:
+    def __init__(self, text=None, pred_tag=None, gold_tag=None):
+        """
+        Token object to hold token attributes
+        :param text: str
+        :param pred_tag: str
+        :param gold_tag: str
+        """
+        self.text = text
+        self.gold_tag = gold_tag
+        self.pred_tag = pred_tag
+        self.subwords = None
+
+    @property
+    def subwords(self):
+        return self._subwords
+
+    @subwords.setter
+    def subwords(self, value):
+        self._subwords = value
+
+    def __str__(self):
+        """
+        Token text representation
+        :return: str
+        """
+        gold_tags = "|".join(self.gold_tag)
+        pred_tags = "|".join([pred_tag["tag"] for pred_tag in self.pred_tag])
+
+        if self.gold_tag:
+            r = f"{self.text}\t{gold_tags}\t{pred_tags}"
+        else:
+            r = f"{self.text}\t{pred_tags}"
+
+        return r
+
+
+class DefaultDataset(Dataset):
+    def __init__(
+        self,
+        examples=None,
+        vocab=None,
+        bert_model="aubmindlab/bert-base-arabertv2",
+        max_seq_len=512,
+    ):
+        """
+        The dataset that used to transform the segments into training data
+        :param examples: list[[tuple]] - [[(token, tag), (token, tag), ...], [(token, tag), ...]]
+                         You can get generate examples from -- arabiner.data.dataset.parse_conll_files
+        :param vocab: vocab object containing indexed tags and tokens
+        :param bert_model: str - BERT model
+        :param: int - maximum sequence length
+        """
+        self.transform = BertSeqTransform(bert_model, vocab, max_seq_len=max_seq_len)
+        self.examples = examples
+        self.vocab = vocab
+
+    def __len__(self):
+        return len(self.examples)
+
+    def __getitem__(self, item):
+        subwords, tags, tokens, valid_len = self.transform(self.examples[item])
+        return subwords, tags, tokens, valid_len
+
+    def collate_fn(self, batch):
+        """
+        Collate function that is called when the batch is called by the trainer
+        :param batch: Dataloader batch
+        :return: Same output as the __getitem__ function
+        """
+        subwords, tags, tokens, valid_len = zip(*batch)
+
+        # Pad sequences in this batch
+        # subwords and tokens are padded with zeros
+        # tags are padding with the index of the O tag
+        subwords = pad_sequence(subwords, batch_first=True, padding_value=0)
+        tags = pad_sequence(
+            tags, batch_first=True, padding_value=self.vocab.tags[0].get_stoi()["O"]
+        )
+        return subwords, tags, tokens, valid_len
+
+
+class NestedTagsDataset(Dataset):
+    def __init__(
+        self,
+        examples=None,
+        vocab=None,
+        bert_model="aubmindlab/bert-base-arabertv2",
+        max_seq_len=512,
+    ):
+        """
+        The dataset that used to transform the segments into training data
+        :param examples: list[[tuple]] - [[(token, tag), (token, tag), ...], [(token, tag), ...]]
+                         You can get generate examples from -- arabiner.data.dataset.parse_conll_files
+        :param vocab: vocab object containing indexed tags and tokens
+        :param bert_model: str - BERT model
+        :param: int - maximum sequence length
+        """
+        self.transform = NestedTagsTransform(
+            bert_model, vocab, max_seq_len=max_seq_len
+        )
+        self.examples = examples
+        self.vocab = vocab
+
+    def __len__(self):
+        return len(self.examples)
+
+    def __getitem__(self, item):
+        subwords, tags, tokens, masks, valid_len = self.transform(self.examples[item])
+        return subwords, tags, tokens, masks, valid_len
+
+    def collate_fn(self, batch):
+        """
+        Collate function that is called when the batch is called by the trainer
+        :param batch: Dataloader batch
+        :return: Same output as the __getitem__ function
+        """
+        subwords, tags, tokens, masks, valid_len = zip(*batch)
+
+        # Pad sequences in this batch
+        # subwords and tokens are padded with zeros
+        # tags are padding with the index of the O tag
+        subwords = pad_sequence(subwords, batch_first=True, padding_value=0)
+
+        masks = [torch.nn.ConstantPad1d((0, subwords.shape[-1] - tag.shape[-1]), 0)(mask)
+                 for tag, mask in zip(tags, masks)]
+        masks = torch.cat(masks)
+
+        # Pad the tags, do the padding for each tag type
+        tags = [torch.nn.ConstantPad1d((0, subwords.shape[-1] - tag.shape[-1]), vocab.get_stoi()["<pad>"])(tag)
+                for tag, vocab in zip(tags, self.vocab.tags[1:])]
+        tags = torch.cat(tags)
+
+        return subwords, tags, tokens, masks, valid_len

nlptools/arabiner/data/transforms.py

@@ -0,0 +1,118 @@
+import torch
+from transformers import BertTokenizer
+from functools import partial
+import re
+import itertools
+from nlptools.arabiner.data import datasets
+class BertSeqTransform:
+    def __init__(self, bert_model, vocab, max_seq_len=512):
+        self.tokenizer = BertTokenizer.from_pretrained(bert_model)
+        self.encoder = partial(
+            self.tokenizer.encode,
+            max_length=max_seq_len,
+            truncation=True,
+        )
+        self.max_seq_len = max_seq_len
+        self.vocab = vocab
+
+    def __call__(self, segment):
+        subwords, tags, tokens = list(), list(), list()
+        unk_token = datasets.Token(text="UNK")
+
+        for token in segment:
+            token_subwords = self.encoder(token.text)[1:-1]
+            subwords += token_subwords
+            tags += [self.vocab.tags[0].get_stoi()[token.gold_tag[0]]] + [self.vocab.tags[0].get_stoi()["O"]] * (len(token_subwords) - 1)
+            tokens += [token] + [unk_token] * (len(token_subwords) - 1)
+
+        # Truncate to max_seq_len
+        if len(subwords) > self.max_seq_len - 2:
+            text = " ".join([t.text for t in tokens if t.text != "UNK"])
+          
+            subwords = subwords[:self.max_seq_len - 2]
+            tags = tags[:self.max_seq_len - 2]
+            tokens = tokens[:self.max_seq_len - 2]
+
+        subwords.insert(0, self.tokenizer.cls_token_id)
+        subwords.append(self.tokenizer.sep_token_id)
+
+        tags.insert(0, self.vocab.tags[0].get_stoi()["O"])
+        tags.append(self.vocab.tags[0].get_stoi()["O"])
+
+        tokens.insert(0, unk_token)
+        tokens.append(unk_token)
+
+        return torch.LongTensor(subwords), torch.LongTensor(tags), tokens, len(tokens)
+
+
+class NestedTagsTransform:
+    def __init__(self, bert_model, vocab, max_seq_len=512):
+        self.tokenizer = BertTokenizer.from_pretrained(bert_model)
+        self.encoder = partial(
+            self.tokenizer.encode,
+            max_length=max_seq_len,
+            truncation=True,
+        )
+        self.max_seq_len = max_seq_len
+        self.vocab = vocab
+
+    def __call__(self, segment):
+        tags, tokens, subwords = list(), list(), list()
+        unk_token = datasets.Token(text="UNK")
+
+        # Encode each token and get its subwords and IDs
+        for token in segment:
+            token.subwords = self.encoder(token.text)[1:-1]
+            subwords += token.subwords
+            tokens += [token] + [unk_token] * (len(token.subwords ) - 1)
+
+        # Construct the labels for each tag type
+        # The sequence will have a list of tags for each type
+        # The final tags for a sequence is a matrix NUM_TAG_TYPES x SEQ_LEN
+        # Example:
+        #   [
+        #       [O,     O,     B-PERS, I-PERS, O, O, O]
+        #       [B-ORG, I-ORG, O,      O,      O, O, O]
+        #       [O,     O,     O,      O,      O, O, B-GPE]
+        #   ]
+        for vocab in self.vocab.tags[1:]:
+            vocab_tags = "|".join([t for t in vocab.get_itos() if "-" in t])
+            r = re.compile(vocab_tags)
+
+            # This is really messy
+            # For a given token we find a matching tag_name, BUT we might find
+            # multiple matches (i.e. a token can be labeled B-ORG and I-ORG) in this
+            # case we get only the first tag as we do not have overlapping of same type
+            single_type_tags = [[(list(filter(r.match, token.gold_tag))
+                                or ["O"])[0]] + ["O"] * (len(token.subwords) - 1)
+                                for token in segment]
+            single_type_tags = list(itertools.chain(*single_type_tags))
+            tags.append([vocab.get_stoi()[tag] for tag in single_type_tags])
+
+        # Truncate to max_seq_len
+        if len(subwords) > self.max_seq_len - 2:
+            text = " ".join([t.text for t in tokens if t.text != "UNK"])
+          
+            subwords = subwords[:self.max_seq_len - 2]
+            tags = [t[:self.max_seq_len - 2] for t in tags]
+            tokens = tokens[:self.max_seq_len - 2]
+
+        # Add dummy token at the start end of sequence
+        tokens.insert(0, unk_token)
+        tokens.append(unk_token)
+
+        # Add CLS and SEP at start end of subwords
+        subwords.insert(0, self.tokenizer.cls_token_id)
+        subwords.append(self.tokenizer.sep_token_id)
+        subwords = torch.LongTensor(subwords)
+
+        # Add "O" tags for the first and last subwords
+        tags = torch.Tensor(tags)
+        tags = torch.column_stack((
+            torch.Tensor([vocab.get_stoi()["O"] for vocab in self.vocab.tags[1:]]),
+            tags,
+            torch.Tensor([vocab.get_stoi()["O"] for vocab in self.vocab.tags[1:]]),
+        )).unsqueeze(0)
+
+        mask = torch.ones_like(tags)
+        return subwords, tags, tokens, mask, len(tokens)

nlptools/arabiner/nn/BaseModel.py

@@ -0,0 +1,22 @@
+from torch import nn
+from transformers import BertModel
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class BaseModel(nn.Module):
+    def __init__(self,
+                 bert_model="aubmindlab/bert-base-arabertv2",
+                 num_labels=2,
+                 dropout=0.1,
+                 num_types=0):
+        super().__init__()
+
+        self.bert_model = bert_model
+        self.num_labels = num_labels
+        self.num_types = num_types
+        self.dropout = dropout
+
+        self.bert = BertModel.from_pretrained(bert_model)
+        self.dropout = nn.Dropout(dropout)

nlptools/arabiner/nn/BertNestedTagger.py

@@ -0,0 +1,34 @@
+import torch
+import torch.nn as nn
+from nlptools.arabiner.nn import BaseModel
+
+
+class BertNestedTagger(BaseModel):
+    def __init__(self, **kwargs):
+        super(BertNestedTagger, self).__init__(**kwargs)
+
+        self.max_num_labels = max(self.num_labels)
+        classifiers = [nn.Linear(768, num_labels) for num_labels in self.num_labels]
+        self.classifiers = torch.nn.Sequential(*classifiers)
+
+    def forward(self, x):
+        y = self.bert(x)
+        y = self.dropout(y["last_hidden_state"])
+        output = list()
+
+        for i, classifier in enumerate(self.classifiers):
+            logits = classifier(y)
+
+            # Pad logits to allow Multi-GPU/DataParallel training to work
+            # We will truncate the padded dimensions when we compute the loss in the trainer
+            logits = torch.nn.ConstantPad1d((0, self.max_num_labels - logits.shape[-1]), 0)(logits)
+            output.append(logits)
+
+        # Return tensor of the shape B x T x L x C
+        # B: batch size
+        # T: sequence length
+        # L: number of tag types
+        # C: number of classes per tag type
+        output = torch.stack(output).permute((1, 2, 0, 3))
+        return output
+

nlptools/arabiner/nn/BertSeqTagger.py

@@ -0,0 +1,17 @@
+import torch.nn as nn
+from transformers import BertModel
+
+
+class BertSeqTagger(nn.Module):
+    def __init__(self, bert_model, num_labels=2, dropout=0.1):
+        super().__init__()
+
+        self.bert = BertModel.from_pretrained(bert_model)
+        self.dropout = nn.Dropout(dropout)
+        self.linear = nn.Linear(768, num_labels)
+
+    def forward(self, x):
+        y = self.bert(x)
+        y = self.dropout(y["last_hidden_state"])
+        logits = self.linear(y)
+        return logits

nlptools/arabiner/nn/__init__.py

@@ -0,0 +1,3 @@
+from nlptools.arabiner.nn.BaseModel import BaseModel
+from nlptools.arabiner.nn.BertSeqTagger import BertSeqTagger
+from nlptools.arabiner.nn.BertNestedTagger import BertNestedTagger

nlptools/arabiner/trainers/BaseTrainer.py

@@ -0,0 +1,117 @@
+import os
+import torch
+import logging
+import natsort
+import glob
+
+logger = logging.getLogger(__name__)
+
+
+class BaseTrainer:
+    def __init__(
+        self,
+        model=None,
+        max_epochs=50,
+        optimizer=None,
+        scheduler=None,
+        loss=None,
+        train_dataloader=None,
+        val_dataloader=None,
+        test_dataloader=None,
+        log_interval=10,
+        summary_writer=None,
+        output_path=None,
+        clip=5,
+        patience=5
+    ):
+        self.model = model
+        self.max_epochs = max_epochs
+        self.train_dataloader = train_dataloader
+        self.val_dataloader = val_dataloader
+        self.test_dataloader = test_dataloader
+        self.optimizer = optimizer
+        self.scheduler = scheduler
+        self.loss = loss
+        self.log_interval = log_interval
+        self.summary_writer = summary_writer
+        self.output_path = output_path
+        self.current_timestep = 0
+        self.current_epoch = 0
+        self.clip = clip
+        self.patience = patience
+
+    def tag(self, dataloader, is_train=True):
+        """
+        Given a dataloader containing segments, predict the tags
+        :param dataloader: torch.utils.data.DataLoader
+        :param is_train: boolean - True for training model, False for evaluation
+        :return: Iterator
+                    subwords (B x T x NUM_LABELS)- torch.Tensor - BERT subword ID
+                    gold_tags (B x T x NUM_LABELS) - torch.Tensor - ground truth tags IDs
+                    tokens - List[arabiner.data.dataset.Token] - list of tokens
+                    valid_len (B x 1) - int - valiud length of each sequence
+                    logits (B x T x NUM_LABELS) - logits for each token and each tag
+        """
+        for subwords, gold_tags, tokens, valid_len in dataloader:
+            self.model.train(is_train)
+
+            if torch.cuda.is_available():
+                subwords = subwords.cuda()
+                gold_tags = gold_tags.cuda()
+
+            if is_train:
+                self.optimizer.zero_grad()
+                logits = self.model(subwords)
+            else:
+                with torch.no_grad():
+                    logits = self.model(subwords)
+
+            yield subwords, gold_tags, tokens, valid_len, logits
+
+    def segments_to_file(self, segments, filename):
+        """
+        Write segments to file
+        :param segments: [List[arabiner.data.dataset.Token]] - list of list of tokens
+        :param filename: str - output filename
+        :return: None
+        """
+        with open(filename, "w") as fh:
+            results = "\n\n".join(["\n".join([t.__str__() for t in segment]) for segment in segments])
+            fh.write("Token\tGold Tag\tPredicted Tag\n")
+            fh.write(results)
+            logging.info("Predictions written to %s", filename)
+
+    def save(self):
+        """
+        Save model checkpoint
+        :return:
+        """
+        filename = os.path.join(
+            self.output_path,
+            "checkpoints",
+            "checkpoint_{}.pt".format(self.current_epoch),
+        )
+
+        checkpoint = {
+            "model": self.model.state_dict(),
+            "optimizer": self.optimizer.state_dict(),
+            "epoch": self.current_epoch
+        }
+
+        logger.info("Saving checkpoint to %s", filename)
+        torch.save(checkpoint, filename)
+
+    def load(self, checkpoint_path):
+        """
+        Load model checkpoint
+        :param checkpoint_path: str - path/to/checkpoints
+        :return: None
+        """
+        checkpoint_path = natsort.natsorted(glob.glob(f"{checkpoint_path}/checkpoint_*.pt"))
+        checkpoint_path = checkpoint_path[-1]
+
+        logger.info("Loading checkpoint %s", checkpoint_path)
+
+        device = None if torch.cuda.is_available() else torch.device('cpu')
+        checkpoint = torch.load(checkpoint_path, map_location=device)
+        self.model.load_state_dict(checkpoint["model"])

nlptools/arabiner/trainers/BertNestedTrainer.py

@@ -0,0 +1,203 @@
+import os
+import logging
+import torch
+import numpy as np
+from nlptools.arabiner.trainers import BaseTrainer
+from nlptools.arabiner.utils.metrics import compute_nested_metrics
+
+logger = logging.getLogger(__name__)
+
+
+class BertNestedTrainer(BaseTrainer):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+    def train(self):
+        best_val_loss, test_loss = np.inf, np.inf
+        num_train_batch = len(self.train_dataloader)
+        num_labels = [len(v) for v in self.train_dataloader.dataset.vocab.tags[1:]]
+        patience = self.patience
+
+        for epoch_index in range(self.max_epochs):
+            self.current_epoch = epoch_index
+            train_loss = 0
+
+            for batch_index, (subwords, gold_tags, tokens, valid_len, logits) in enumerate(self.tag(
+                self.train_dataloader, is_train=True
+            ), 1):
+                self.current_timestep += 1
+
+                # Compute loses for each output
+                # logits = B x T x L x C
+                losses = [self.loss(logits[:, :, i, 0:l].view(-1, logits[:, :, i, 0:l].shape[-1]),
+                                    torch.reshape(gold_tags[:, i, :], (-1,)).long())
+                          for i, l in enumerate(num_labels)]
+
+                torch.autograd.backward(losses)
+
+                # Avoid exploding gradient by doing gradient clipping
+                torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.clip)
+
+                self.optimizer.step()
+                self.scheduler.step()
+                batch_loss = sum(l.item() for l in losses)
+                train_loss += batch_loss
+
+                if self.current_timestep % self.log_interval == 0:
+                    logger.info(
+                        "Epoch %d | Batch %d/%d | Timestep %d | LR %.10f | Loss %f",
+                        epoch_index,
+                        batch_index,
+                        num_train_batch,
+                        self.current_timestep,
+                        self.optimizer.param_groups[0]['lr'],
+                        batch_loss
+                    )
+
+            train_loss /= num_train_batch
+
+            logger.info("** Evaluating on validation dataset **")
+            val_preds, segments, valid_len, val_loss = self.eval(self.val_dataloader)
+            val_metrics = compute_nested_metrics(segments, self.val_dataloader.dataset.transform.vocab.tags[1:])
+
+            epoch_summary_loss = {
+                "train_loss": train_loss,
+                "val_loss": val_loss
+            }
+            epoch_summary_metrics = {
+                "val_micro_f1": val_metrics.micro_f1,
+                "val_precision": val_metrics.precision,
+                "val_recall": val_metrics.recall
+            }
+
+            logger.info(
+                "Epoch %d | Timestep %d | Train Loss %f | Val Loss %f | F1 %f",
+                epoch_index,
+                self.current_timestep,
+                train_loss,
+                val_loss,
+                val_metrics.micro_f1
+            )
+
+            if val_loss < best_val_loss:
+                patience = self.patience
+                best_val_loss = val_loss
+                logger.info("** Validation improved, evaluating test data **")
+                test_preds, segments, valid_len, test_loss = self.eval(self.test_dataloader)
+                self.segments_to_file(segments, os.path.join(self.output_path, "predictions.txt"))
+                test_metrics = compute_nested_metrics(segments, self.test_dataloader.dataset.transform.vocab.tags[1:])
+
+                epoch_summary_loss["test_loss"] = test_loss
+                epoch_summary_metrics["test_micro_f1"] = test_metrics.micro_f1
+                epoch_summary_metrics["test_precision"] = test_metrics.precision
+                epoch_summary_metrics["test_recall"] = test_metrics.recall
+
+                logger.info(
+                    f"Epoch %d | Timestep %d | Test Loss %f | F1 %f",
+                    epoch_index,
+                    self.current_timestep,
+                    test_loss,
+                    test_metrics.micro_f1
+                )
+
+                self.save()
+            else:
+                patience -= 1
+
+            # No improvements, terminating early
+            if patience == 0:
+                logger.info("Early termination triggered")
+                break
+
+            self.summary_writer.add_scalars("Loss", epoch_summary_loss, global_step=self.current_timestep)
+            self.summary_writer.add_scalars("Metrics", epoch_summary_metrics, global_step=self.current_timestep)
+
+    def tag(self, dataloader, is_train=True):
+        """
+        Given a dataloader containing segments, predict the tags
+        :param dataloader: torch.utils.data.DataLoader
+        :param is_train: boolean - True for training model, False for evaluation
+        :return: Iterator
+                    subwords (B x T x NUM_LABELS)- torch.Tensor - BERT subword ID
+                    gold_tags (B x T x NUM_LABELS) - torch.Tensor - ground truth tags IDs
+                    tokens - List[arabiner.data.dataset.Token] - list of tokens
+                    valid_len (B x 1) - int - valiud length of each sequence
+                    logits (B x T x NUM_LABELS) - logits for each token and each tag
+        """
+        for subwords, gold_tags, tokens, mask, valid_len in dataloader:
+            self.model.train(is_train)
+
+            if torch.cuda.is_available():
+                subwords = subwords.cuda()
+                gold_tags = gold_tags.cuda()
+
+            if is_train:
+                self.optimizer.zero_grad()
+                logits = self.model(subwords)
+            else:
+                with torch.no_grad():
+                    logits = self.model(subwords)
+
+            yield subwords, gold_tags, tokens, valid_len, logits
+
+    def eval(self, dataloader):
+        golds, preds, segments, valid_lens = list(), list(), list(), list()
+        num_labels = [len(v) for v in dataloader.dataset.vocab.tags[1:]]
+        loss = 0
+
+        for _, gold_tags, tokens, valid_len, logits in self.tag(
+            dataloader, is_train=False
+        ):
+            losses = [self.loss(logits[:, :, i, 0:l].view(-1, logits[:, :, i, 0:l].shape[-1]),
+                                torch.reshape(gold_tags[:, i, :], (-1,)).long())
+                      for i, l in enumerate(num_labels)]
+            loss += sum(losses)
+            preds += torch.argmax(logits, dim=3)
+            segments += tokens
+            valid_lens += list(valid_len)
+
+        loss /= len(dataloader)
+
+        # Update segments, attach predicted tags to each token
+        segments = self.to_segments(segments, preds, valid_lens, dataloader.dataset.vocab)
+
+        return preds, segments, valid_lens, loss
+
+    def infer(self, dataloader):
+        golds, preds, segments, valid_lens = list(), list(), list(), list()
+
+        for _, gold_tags, tokens, valid_len, logits in self.tag(
+            dataloader, is_train=False
+        ):
+            preds += torch.argmax(logits, dim=3)
+            segments += tokens
+            valid_lens += list(valid_len)
+
+        segments = self.to_segments(segments, preds, valid_lens, dataloader.dataset.vocab)
+        return segments
+
+    def to_segments(self, segments, preds, valid_lens, vocab):
+        if vocab is None:
+            vocab = self.vocab
+
+        tagged_segments = list()
+        tokens_stoi = vocab.tokens.get_stoi()
+        unk_id = tokens_stoi["UNK"]
+
+        for segment, pred, valid_len in zip(segments, preds, valid_lens):
+            # First, the token at 0th index [CLS] and token at nth index [SEP]
+            # Combine the tokens with their corresponding predictions
+            segment_pred = zip(segment[1:valid_len-1], pred[1:valid_len-1])
+
+            # Ignore the sub-tokens/subwords, which are identified with text being UNK
+            segment_pred = list(filter(lambda t: tokens_stoi[t[0].text] != unk_id, segment_pred))
+
+            # Attach the predicted tags to each token
+            list(map(lambda t: setattr(t[0], 'pred_tag', [{"tag": vocab.get_itos()[tag_id]}
+                                                     for tag_id, vocab in zip(t[1].int().tolist(), vocab.tags[1:])]), segment_pred))
+
+            # We are only interested in the tagged tokens, we do no longer need raw model predictions
+            tagged_segment = [t for t, _ in segment_pred]
+            tagged_segments.append(tagged_segment)
+
+        return tagged_segments