genhpf 1.0.11__py3-none-any.whl

genhpf/criterions/binary_cross_entropy_with_logits.py

@@ -0,0 +1,115 @@
+import math
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Tuple
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+import genhpf.utils.utils as utils
+from genhpf.criterions import BaseCriterion, register_criterion
+from genhpf.criterions.criterion import CriterionConfig
+from genhpf.loggings import meters, metrics
+from genhpf.loggings.meters import safe_round
+
+
+@dataclass
+class BinaryCrossEntropyWithLogitsConfig(CriterionConfig):
+    threshold: float = field(default=0.5, metadata={"help": "threshold value for binary classification"})
+
+
+@register_criterion("binary_cross_entropy_with_logits", dataclass=BinaryCrossEntropyWithLogitsConfig)
+class BinaryCrossEntropyWithLogits(BaseCriterion):
+    def __init__(self, cfg: BinaryCrossEntropyWithLogitsConfig):
+        super().__init__(cfg)
+
+        if self.task_names is not None and len(self.task_names) > 1:
+            raise ValueError(
+                "binary_cross_entropy_with_logits only supports single task training."
+                " if you want to train multiple tasks, use multi_task_criterion instead."
+            )
+
+        self.threshold = cfg.threshold
+
+    def compute_loss(
+        self, logits: torch.Tensor, targets: torch.Tensor, sample=None, net_output=None, model=None
+    ) -> Tuple[torch.Tensor, List[float]]:
+        assert (
+            logits.size() == targets.size()
+        ), f"logits and targets must have the same size: {logits.size()} vs {targets.size()}"
+        targets = targets.float()
+        loss = F.binary_cross_entropy_with_logits(input=logits, target=targets, reduction="sum")
+        return loss, [loss.detach().item()]
+
+    def get_sample_size(self, sample, targets: torch.Tensor) -> int:
+        if "sample_size" in sample:
+            sample_size = sample["sample_size"]
+        else:
+            sample_size = targets.numel()
+        return sample_size
+
+    def get_logging_outputs(
+        self, logging_output, logits: torch.Tensor, targets: torch.Tensor, sample=None
+    ) -> List[Dict[str, Any]]:
+        with torch.no_grad():
+            probs = torch.sigmoid(logits)
+            outputs = probs > self.threshold
+
+            if probs.numel() == 0:
+                corr = 0
+                count = 0
+            else:
+                count = float(probs.numel())
+                corr = (outputs == targets).sum().item()
+
+            logging_output["correct"] = corr
+            logging_output["count"] = count
+
+            # report aucs only in eval mode
+            if not self.training:
+                logging_output["_y_true"] = targets.cpu().numpy()
+                logging_output["_y_score"] = probs.cpu().numpy()
+
+        return logging_output
+
+    @staticmethod
+    def reduce_metrics(logging_outputs: List[Dict[str, Any]], prefix: str = None) -> None:
+        """Aggregate logging outputs from data parallel training."""
+        if prefix is None:
+            prefix = ""
+        elif prefix is not None and not prefix.endswith("_"):
+            prefix = prefix + "_"
+
+        loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs))
+
+        sample_size = utils.item(sum(log.get("sample_size", 0) for log in logging_outputs))
+
+        metrics.log_scalar(f"{prefix}loss", loss_sum / (sample_size or 1) / math.log(2), sample_size, round=3)
+
+        if "_y_true" in logging_outputs[0] and "_y_score" in logging_outputs[0]:
+            y_true = np.concatenate([log.get("_y_true", []) for log in logging_outputs])
+            y_score = np.concatenate([log.get("_y_score", []) for log in logging_outputs])
+
+            metrics.log_custom(meters.AUCMeter, f"_{prefix}auc", y_score, y_true)
+
+        correct = sum(log.get("correct", 0) for log in logging_outputs)
+        metrics.log_scalar(f"_{prefix}correct", correct)
+
+        total = sum(log.get("count", 0) for log in logging_outputs)
+        metrics.log_scalar(f"_{prefix}total", total)
+
+        if total > 0:
+            metrics.log_derived(
+                f"{prefix}accuracy",
+                lambda meters: safe_round(meters[f"_{prefix}correct"].sum / meters[f"_{prefix}total"].sum, 5)
+                if meters[f"_{prefix}total"].sum > 0
+                else float("nan"),
+            )
+
+    def post_validate(self, stats, agg, **kwargs):
+        for key in agg.keys():
+            if key.startswith("_") and key.endswith("auc"):
+                stats[key[1:-3] + "auroc"] = agg[key].auroc
+                stats[key[1:-3] + "auprc"] = agg[key].auprc
+
+        return stats

genhpf/criterions/criterion.py

@@ -0,0 +1,87 @@
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Tuple
+
+import torch
+from torch.nn.modules.loss import _Loss
+
+from genhpf.configs import BaseConfig
+from genhpf.models.genhpf import GenHPF
+
+
+@dataclass
+class CriterionConfig(BaseConfig):
+    task_names: Optional[List[str]] = field(
+        default=None, metadata={"help": "a list of task names for multi-task learning"}
+    )
+    num_labels: Optional[List[int]] = field(
+        default=None, metadata={"help": "a list of number of labels for each task"}
+    )
+
+
+class BaseCriterion(_Loss):
+    def __init__(self, cfg: CriterionConfig):
+        super().__init__()
+        self.cfg = cfg
+
+        self.task_names = cfg.task_names
+        self.num_labels = cfg.num_labels
+
+    @classmethod
+    def build_criterion(cls, cfg: CriterionConfig):
+        """Construct a new criterion instance."""
+        return cls(cfg)
+
+    def compute_loss(
+        self, logits: torch.Tensor, targets: torch.Tensor, sample=None, net_output=None, model=None
+    ) -> Tuple[torch.Tensor, List[float]]:
+        """Compute the loss given the logits and targets from the model."""
+        raise NotImplementedError("Criterion must implement the `compute_loss` method")
+
+    def get_sample_size(self, sample, targets: torch.Tensor) -> int:
+        """Get the sample size, which is used as the denominator for the gradient."""
+        raise NotImplementedError("Criterion must implement the `get_sample_size` method")
+
+    def get_logging_outputs(
+        self, logging_output, logits: torch.Tensor, targets: torch.Tensor, sample=None
+    ) -> List[Dict[str, Any]]:
+        """
+        Get the logging output to display while training
+        """
+        raise NotImplementedError("Criterion must implement the `get_logging_outputs` method")
+
+    def forward(self, model: GenHPF, sample, return_net_output=False):
+        """Compute the loss for the given sample.
+
+        Returns a tuple with three elements:
+        1. the loss
+        2. the sample size, which is used as the denominator for the gradient
+        3. logging outputs to display while training
+        """
+        net_output = model(**sample["net_input"])
+        logits = model.get_logits(sample, net_output)
+        targets = model.get_targets(sample, net_output)
+
+        loss, losses_to_log = self.compute_loss(
+            logits, targets, sample=sample, net_output=net_output, model=model
+        )
+        sample_size = self.get_sample_size(sample, targets)
+
+        logging_output = {}
+        if len(losses_to_log) > 1:
+            logging_output["loss"] = loss.item()
+            for i, l in enumerate(losses_to_log):
+                logging_output[f"loss_{i}"] = l
+        else:
+            logging_output["loss"] = losses_to_log[0]
+        logging_output["sample_size"] = sample_size
+        logging_output = self.get_logging_outputs(logging_output, logits, targets, sample)
+
+        if return_net_output:
+            return loss, sample_size, logging_output, net_output
+        else:
+            return loss, sample_size, logging_output
+
+    @staticmethod
+    def reduce_metrics(stats: Dict[str, Any], prefix: str = None) -> None:
+        """Aggregate logging outputs from data parallel training."""
+        raise NotImplementedError

genhpf/criterions/cross_entropy.py

@@ -0,0 +1,202 @@
+import math
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Tuple
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from omegaconf import II
+
+import genhpf.utils.utils as utils
+from genhpf.criterions import BaseCriterion, register_criterion
+from genhpf.criterions.criterion import CriterionConfig
+from genhpf.loggings import meters, metrics
+from genhpf.loggings.meters import safe_round
+
+
+@dataclass
+class CrossEntropyConfig(CriterionConfig):
+    report_auc: bool = field(
+        default=False,
+        metadata={
+            "help": "whether to report auc. note that this is only available in eval mode and "
+            "can cause memory and performance issues if enabled."
+        },
+    )
+    ignore_index: int = II("dataset.ignore_index")
+
+
+@register_criterion("cross_entropy", dataclass=CrossEntropyConfig)
+class CrossEntropy(BaseCriterion):
+    def __init__(self, cfg: CrossEntropyConfig):
+        super().__init__(cfg)
+
+        if self.task_names is not None and len(self.task_names) > 1:
+            raise ValueError(
+                "cross_entropy only supports single task training. if you want to train multiple"
+                " tasks, use multi_task_criterion instead."
+            )
+
+        self.report_auc = cfg.report_auc
+        self.ignore_index = cfg.ignore_index
+
+    def compute_loss(
+        self, logits: torch.Tensor, targets: torch.Tensor, sample=None, net_output=None, model=None
+    ) -> Tuple[torch.Tensor, List[float]]:
+        """Compute the loss given the logits and targets from the model."""
+        logits = logits.view(-1, logits.size(-1))
+        targets = targets.view(-1).long()
+
+        if torch.all(targets == self.ignore_index):
+            return logits.new_tensor(0.0), [0.0]
+
+        loss = F.cross_entropy(logits, targets, reduction="sum", ignore_index=self.ignore_index)
+
+        return loss, [loss.detach().item()]
+
+    def get_sample_size(self, sample, targets: torch.Tensor) -> int:
+        if "sample_size" in sample:
+            sample_size = sample["sample_size"]
+        else:
+            sample_size = targets.numel()
+        return sample_size
+
+    def get_logging_outputs(
+        self, logging_output, logits: torch.Tensor, targets: torch.Tensor, sample=None
+    ) -> List[Dict[str, Any]]:
+        with torch.no_grad():
+            logits = logits.view(-1, logits.size(-1))
+            targets = targets.view(-1).long()
+
+            valid_indices = torch.where(targets != self.ignore_index)
+            if len(valid_indices[0]) == 0:
+                return {}
+
+            logits = logits[valid_indices]
+            targets = targets[valid_indices]
+
+            preds = logits.argmax(dim=-1)
+            count = targets.numel()
+            corr = (preds == targets).sum().item()
+
+            logging_output["correct"] = corr
+            logging_output["count"] = count
+
+            # report aucs only in eval mode
+            if self.report_auc and not self.training:
+                probs = torch.sigmoid(logits).view(-1)
+                targets = F.one_hot(targets, logits.size(-1)).float().view(-1)
+
+                logging_output["_y_true"] = targets.cpu().numpy()
+                logging_output["_y_score"] = probs.cpu().numpy()
+
+        return logging_output
+
+    # def forward(self, model, sample):
+    #     net_output = model(**sample['net_input'])
+    #     if isinstance(model, DistributedDataParallel):
+    #         logits = model.module.get_outputs(
+    #                 net_output,
+    #                 task=self.args.train_task,
+    #                 normalize=False
+    #             )
+    #         targets = model.module.get_targets(sample, net_output, self.args.train_task)
+    #     else:
+    #         logits = model.get_outputs(
+    #                 net_output,
+    #                 task=self.args.train_task,
+    #                 normalize=False
+    #             )
+    #         targets = model.get_targets(sample, net_output, self.args.train_task)
+
+    #     loss_dict = {}
+    #     logging_output = {}
+
+    #     if self.args.train_task == 'pretrain' and self.args.pretrain_task in ['mlm', 'spanmlm']:
+    #         B, S= targets['input_label'].shape
+    #         for victim in self.args.mask_list:
+    #             loss = F.cross_entropy(
+    #                 logits[victim+'_ids'].view(B*S, -1),
+    #                 targets[victim+'_label'].view(-1)
+    #             )
+    #             loss_dict[victim+'_loss'] = loss
+
+    #             with torch.no_grad():
+    #                 preds = torch.argmax(logits[victim+'_ids'], dim=-1).view(-1).detach().cpu()
+    #                 target_label = targets[victim+'_label'].view(-1).detach().cpu()
+    #                 mask_idcs = (target_label != -100) & (target_label != 0)
+    #                 total = mask_idcs.sum()
+    #                 correct = (preds[mask_idcs] == target_label[mask_idcs]).sum().float()
+
+    #                 logging_output[victim+'_correct'] = correct
+    #                 logging_output[victim+'_total'] = total
+
+    #         loss = sum(loss_dict.values())
+    #         sample_size = len(sample)
+    #         logging_output['loss'] = loss.item()
+    #         logging_output['sample_size'] = sample_size
+
+    #     elif self.args.train_task in ['finetune', 'scratch']:
+
+    #         sample_size = len(targets)
+    #         loss = F.cross_entropy(
+    #             logits, F.one_hot(
+    #                 targets.long(),
+    #                 self.multi_label_dict[self.args.pred_src][self.args.pred_target]
+    #             ).float().to(logits.device),
+    #             reduction=self.ce_reduction_mode
+    #         )
+
+    #         logging_output['loss'] = loss.item()
+    #         logging_output['sample_size'] = sample_size
+
+    #         with torch.no_grad():
+    #             probs = torch.sigmoid(logits).view(-1).detach()
+    #             targets = self.mlb.transform(np.expand_dims(targets.view(-1).cpu(), axis=1)).flatten()
+
+    #             logging_output["_y_true"] = targets
+    #             logging_output["_y_score"] = probs.cpu().numpy()
+
+    #     return loss, sample_size, logging_output
+
+    @staticmethod
+    def reduce_metrics(logging_outputs: List[Dict[str, Any]], prefix: str = None) -> None:
+        """Aggregate logging outputs from data parallel training."""
+        if prefix is None:
+            prefix = ""
+        elif prefix is not None and not prefix.endswith("_"):
+            prefix = prefix + "_"
+
+        loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs))
+
+        sample_size = utils.item(sum(log.get("sample_size", 0) for log in logging_outputs))
+
+        metrics.log_scalar(f"{prefix}loss", loss_sum / (sample_size or 1) / math.log(2), sample_size, round=3)
+
+        if "_y_true" in logging_outputs[0] and "_y_score" in logging_outputs[0]:
+            y_true = np.concatenate([log.get("_y_true", []) for log in logging_outputs])
+            y_score = np.concatenate([log.get("_y_score", []) for log in logging_outputs])
+
+            metrics.log_custom(meters.AUCMeter, f"_{prefix}auc", y_score, y_true)
+
+        correct = sum(log.get("correct", 0) for log in logging_outputs)
+        metrics.log_scalar(f"_{prefix}correct", correct)
+
+        total = sum(log.get("count", 0) for log in logging_outputs)
+        metrics.log_scalar(f"_{prefix}total", total)
+
+        if total > 0:
+            metrics.log_derived(
+                f"{prefix}accuracy",
+                lambda meters: safe_round(meters[f"_{prefix}correct"].sum / meters[f"_{prefix}total"].sum, 5)
+                if meters[f"_{prefix}total"].sum > 0
+                else float("nan"),
+            )
+
+    def post_validate(self, stats, agg, **kwargs):
+        for key in agg.keys():
+            if key.startswith("_") and key.endswith("auc"):
+                stats[key[1:-3] + "auroc"] = agg[key].auroc
+                stats[key[1:-3] + "auprc"] = agg[key].auprc
+
+        return stats

genhpf/criterions/multi_task_criterion.py

@@ -0,0 +1,177 @@
+import math
+import re
+from collections import defaultdict
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional
+
+import genhpf.utils.utils as utils
+from genhpf.criterions import BaseCriterion, register_criterion
+from genhpf.criterions.criterion import CriterionConfig
+from genhpf.loggings import metrics
+from genhpf.models.genhpf import GenHPF
+
+from . import build_criterion
+
+
+@dataclass
+class MultiTaskCriterionConfig(CriterionConfig):
+    task_loss_weights: Optional[List[float]] = field(
+        default=None,
+        metadata={
+            "help": "weights for each loss term. if given, has to be a float list of size " "n_criterions"
+        },
+    )
+    args: Any = field(
+        default=None,
+        metadata={
+            "help": "configurations for each criterion where the name of each argument should "
+            "match with the corresponding task name."
+        },
+    )
+
+
+@register_criterion("multi_task_criterion", dataclass=MultiTaskCriterionConfig)
+class MultiTaskCriterion(BaseCriterion):
+    def __init__(self, cfg: MultiTaskCriterionConfig):
+        super().__init__(cfg)
+
+        criterions = {}
+        for task_name in self.task_names:
+            criterion_cfg = getattr(cfg.args, task_name)
+            criterions[task_name] = build_criterion(criterion_cfg)
+        self.criterions = criterions
+
+        if cfg.task_loss_weights is None:
+            self.task_loss_weights = [1.0] * len(criterions)
+        else:
+            self.task_loss_weights = cfg.task_loss_weights
+
+    def forward(self, model: GenHPF, sample, return_net_output=False):
+        net_output = model(**sample["net_input"])
+        logits = model.get_logits(sample, net_output)
+        targets = model.get_targets(sample, net_output)
+
+        if not isinstance(logits, dict):
+            logits = {self.task_names[0]: logits}
+        if not isinstance(targets, dict):
+            targets = {self.task_names[0]: targets}
+
+        if len(logits) != len(self.task_names) or len(targets) != len(self.task_names):
+            raise ValueError(
+                "number of logits and targets should be equal to the number of tasks. "
+                f"got {len(logits)} logits and {len(targets)} targets for "
+                f"{len(self.task_names)} tasks"
+            )
+
+        loss = 0.0
+        logging_outputs = dict()
+        for i, task_name in enumerate(self.task_names):
+            criterion = self.criterions[task_name]
+            assert (
+                task_name in logits and task_name in targets
+            ), f"task name {task_name} not found in logits or targets"
+            task_logits = logits[task_name]
+            task_targets = targets[task_name]
+            task_loss, task_losses_to_log = criterion.compute_loss(
+                logits=task_logits, targets=task_targets, sample=sample, net_output=net_output, model=model
+            )
+            task_loss *= self.task_loss_weights[i]
+            sample_size = criterion.get_sample_size(sample, task_targets)
+
+            logging_outputs[f"<{task_name}>_criterion_cls"] = criterion.__class__
+            if len(task_losses_to_log) > 1:
+                logging_outputs[f"{task_name}_loss"] = task_loss.item()
+                for j, l in enumerate(task_losses_to_log):
+                    logging_outputs[f"<{task_name}>_loss_{j}"] = l
+            else:
+                logging_outputs[f"<{task_name}>_loss"] = task_losses_to_log[0]
+            logging_outputs[f"<{task_name}>_sample_size"] = sample_size
+
+            task_logging_output = criterion.get_logging_outputs({}, task_logits, task_targets, sample)
+            for log, value in task_logging_output.items():
+                if log.startswith("_"):
+                    log = log[1:]
+                    logging_outputs[f"_<{task_name}>_{log}"] = value
+                else:
+                    logging_outputs[f"<{task_name}>_{log}"] = value
+
+            # divide task loss by the sample size beforehand to handle different sample
+            # sizes for multiple criterions
+            loss += task_loss / logging_outputs[f"<{task_name}>_sample_size"]
+
+        # manipulate sample_size to be 1 to avoid double-dividing gradients in optimizer later
+        sample_size = 1
+
+        if return_net_output:
+            return loss, sample_size, logging_outputs, net_output
+        else:
+            return loss, sample_size, logging_outputs
+
+    @staticmethod
+    def reduce_metrics(logging_outputs: List[Dict[str, Any]]) -> None:
+        log_keys = logging_outputs[0].keys()
+
+        grouped_log_keys = defaultdict(list)
+        for lk in log_keys:
+            group = re.search(r"\<.*\>", lk)
+            offset = group.end() + 1
+            group = group.group()[1:-1]
+            key = lk[offset:]
+            if lk.startswith("_"):
+                key = "_" + key
+            grouped_log_keys[group].append(key)
+
+        total_loss = 0
+        for group, log_keys in grouped_log_keys.items():
+            criterion_cls = logging_outputs[0][f"<{group}>_criterion_cls"]
+            logging_output = []
+            for log in logging_outputs:
+                log_dict = {}
+                for log_key in set(log_keys) - {"criterion_cls"}:
+                    if log_key.startswith("_") and f"_<{group}>{log_key}" in log:
+                        log_dict[log_key] = log[f"_<{group}>{log_key}"]
+                    elif f"<{group}>_{log_key}" in log:
+                        log_dict[log_key] = log[f"<{group}>_{log_key}"]
+                logging_output.append(log_dict)
+            criterion_cls.reduce_metrics(logging_output, prefix=group)
+
+            loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_output))
+            sample_size = utils.item(sum(log.get("sample_size", 0) for log in logging_output))
+
+            total_loss += loss_sum / (sample_size or 1) / math.log(2)
+
+        metrics.log_scalar("loss", total_loss, 1, round=3)
+
+    def post_validate(self, stats, agg, **kwargs):
+        task_agg = {}
+        for key in agg:
+            for task_name in self.task_names:
+                if key.startswith(task_name) or key[1:].startswith(task_name):
+                    if task_name not in task_agg:
+                        task_agg[task_name] = {}
+                    task_agg[task_name][key] = agg[key]
+                    break
+
+        for task_name, task_agg in task_agg.items():
+            if hasattr(self.criterions[task_name], "post_validate"):
+                stats = self.criterions[task_name].post_validate(stats, task_agg, **kwargs)
+
+        for key in list(stats.keys()):
+            for task_name in self.task_names:
+                if key.startswith(task_name):
+                    stat_key = key[len(task_name) + 1 :]
+                    if f"avg_{stat_key}" not in stats:
+                        stats[f"avg_{stat_key}"] = []
+                    stats[f"avg_{stat_key}"].append(stats[key])
+                    break
+
+        for key in list(stats.keys()):
+            if key.startswith("avg_"):
+                stats[key] = sum(stats[key]) / len(stats[key])
+        return stats
+
+    def eval(self):
+        super().eval()
+        for criterion in self.criterions.values():
+            criterion.eval()
+        return self

genhpf/criterions/simclr_criterion.py

@@ -0,0 +1,84 @@
+import math
+from dataclasses import dataclass, field
+from typing import List, Tuple
+
+import torch
+import torch.nn.functional as F
+
+import genhpf.utils.utils as utils
+from genhpf.criterions import BaseCriterion, register_criterion
+from genhpf.criterions.criterion import CriterionConfig
+from genhpf.loggings import metrics
+
+
+@dataclass
+class SimCLRCriterionConfig(CriterionConfig):
+    temp: float = field(default=0.1, metadata={"help": "temperature to divide logits by"})
+
+
+@register_criterion("simclr_criterion", dataclass=SimCLRCriterionConfig)
+class SimCLRCriterion(BaseCriterion):
+    def __init__(self, cfg: SimCLRCriterionConfig):
+        super().__init__(cfg)
+
+        self.temp = cfg.temp
+
+    def compute_loss(
+        self, logits: torch.Tensor, targets: torch.Tensor = None, sample=None, net_output=None, model=None
+    ) -> Tuple[torch.Tensor, List[float]]:
+        """Compute the loss given the logits and targets from the model."""
+        logits = F.normalize(logits, dim=1)  # normalize logits
+
+        bsz = int(logits.shape[0] / 2)
+
+        mask = 1 - torch.eye(bsz * 2, dtype=torch.uint8).to(logits.device)
+        pos_ind = (
+            torch.arange(bsz * 2).to(logits.device),
+            2
+            * torch.arange(bsz, dtype=torch.long)
+            .unsqueeze(1)
+            .repeat(1, 2)
+            .view(-1, 1)
+            .squeeze()
+            .to(logits.device),
+        )
+        neg_mask = torch.ones((bsz * 2, bsz * 2 - 1), dtype=torch.uint8).to(logits.device)
+        neg_mask[pos_ind] = 0
+
+        # Cosine similarity computation
+        sim_matrix = torch.matmul(logits, logits.T)  # cosine similarity computation
+
+        # Eliminate similarity between same view
+        sim_matrix = torch.masked_select(sim_matrix, mask.bool()).view(sim_matrix.size(0), -1)
+
+        positives = sim_matrix[pos_ind].unsqueeze(1)
+        negatives = torch.masked_select(sim_matrix, neg_mask.bool()).view(sim_matrix.size(0), -1)
+
+        logits = torch.cat((positives, negatives), dim=1)
+        logits /= self.temp
+
+        target = torch.zeros((logits.size(0),), dtype=torch.long).to(logits.device)
+
+        loss = F.cross_entropy(logits, target, reduction="sum")
+
+        return loss, [loss.detach().item()]
+
+    def get_sample_size(self, sample, targets: torch.Tensor = None) -> int:
+        return sample["net_input"]["input_ids"].size(0)
+
+    def get_logging_outputs(self, logging_output, logits, target, sample=None, net_output=None):
+        return logging_output
+
+    @staticmethod
+    def reduce_metrics(logging_outputs, prefix: str = None) -> None:
+        """Aggregate logging outputs from data parallel training."""
+        if prefix is None:
+            prefix = ""
+        elif prefix is not None and not prefix.endswith("_"):
+            prefix = prefix + "_"
+
+        loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs))
+
+        sample_size = utils.item(sum(log.get("sample_size", 0) for log in logging_outputs))
+
+        metrics.log_scalar(f"{prefix}loss", loss_sum / (sample_size or 1) / math.log(2), sample_size, round=3)