[WIP] add option to use MultipleNegativesRankingLoss

src/setfit/data.py

@@ -1,3 +1,5 @@
+import math
+import random
 from typing import TYPE_CHECKING, Dict, List, Tuple
 
 import pandas as pd
@@ -280,3 +282,42 @@ def collate_fn(batch):
         labels = torch.Tensor(labels).long()
 
         return features, labels
+
+
+class NoDuplicateClassesDataLoader:
+    def __init__(self, train_examples, batch_size):
+        self.batch_size = batch_size
+        self.collate_fn = None
+        self.train_examples = train_examples
+
+        # TODO: add assert batch_size <= num_classes
+
+    def __iter__(self):
+        label_class_dict = {}
+        random.shuffle(self.train_examples)
+        for example in self.train_examples:
+            example_label_list = label_class_dict.get(example.label, [])
+            example_label_list.append(example)
+            label_class_dict[example.label] = example_label_list
+
+        for _ in range(self.__len__()):
+            batch = []
+            classes_in_batch = set()
+
+            while len(batch) < self.batch_size:
+                class_to_add = random.choice(label_class_dict.keys())
+                if class_to_add not in classes_in_batch:
+                    example = label_class_dict[class_to_add].pop(0)
+                    batch.append(example)
+
+                    # list of examples for this class is empty and needs to be refilled
+                    if len(label_class_dict[class_to_add]) == 0:
+                        random.shuffle(self.train_examples)
+                        for example in self.train_examples:
+                            if example.label == class_to_add:
+                                label_class_dict[class_to_add].append(example)
+
+            yield self.collate_fn(batch) if self.collate_fn is not None else batch
+
+    def __len__(self):
+        return math.floor(len(self.train_examples) / self.batch_size)