Added speed task

src/seb/cli/table.py

@@ -5,10 +5,11 @@
 from rich.table import Table
 
 import seb
+from seb.types import Language
 
 
-def get_main_score(task: seb.TaskResult, langs: Optional[list[str]]) -> float:
-    if langs is None:
+def get_main_score(task: seb.TaskResult, langs: Optional[list[Language]]) -> float:
+    if langs is None:  # noqa
         _langs = task.languages
     else:
         _langs = set(langs) & set(task.languages)

src/seb/full_benchmark.py

@@ -3,14 +3,18 @@
 """
 
 
+import logging
 from pathlib import Path
 from typing import Optional
 
 from .benchmark import Benchmark
 from .interfaces.model import EmbeddingModel
+from .registered_tasks.speed import CPUSpeedTask, GPUSpeedTask
 from .registries import get_all_models
 from .result_dataclasses import BenchmarkResults
 
+logger = logging.getLogger(__name__)
+
 BENCHMARKS = {
     "Mainland Scandinavian": ["da", "sv", "nn", "nb"],
     "Danish": ["da"],
@@ -44,3 +48,36 @@ def run_benchmark(
         results[subset] = bm_results
 
     return results
+
+
+def run_speed_benchmark(
+    use_cache: bool = True,
+    run_models: bool = True,
+    raise_errors: bool = True,
+    cache_dir: Optional[Path] = None,
+) -> dict[str, list[BenchmarkResults]]:
+    """
+    Run the speed benchmark.
+    """
+    models: list[EmbeddingModel] = get_all_models()
+    tasks = [CPUSpeedTask(), GPUSpeedTask()]  # type: ignore
+
+    if use_cache:
+        logger.warn(
+            "Running the speed benchmark with use_cache=True will load speed results from the cache, this might lead to incomparable results."
+        )
+
+    results = {}
+    for subset, langs in BENCHMARKS.items():
+        benchmark = Benchmark(languages=langs)
+        bm_results = benchmark.evaluate_models(
+            models=models,
+            use_cache=use_cache,
+            run_model=run_models,
+            raise_errors=raise_errors,
+            cache_dir=cache_dir,
+        )
+
+        results[subset] = bm_results
+
+    return results

src/seb/interfaces/model.py

@@ -25,6 +25,7 @@ def encode(
         **kwargs: Any,
     ) -> ArrayLike:
         """Returns a list of embeddings for the given sentences.
+
         Args:
             sentences: List of sentences to encode
             task: The task to encode for. This allows the model to encode differently for different tasks. Will always be given but does not need

src/seb/interfaces/mteb_task.py

@@ -1,13 +1,13 @@
 from datetime import datetime
-from typing import Any
+from typing import Any, Union
 
 import numpy as np
 from datasets import DatasetDict, concatenate_datasets
 from mteb import AbsTask
 from mteb import __version__ as mteb_version
 
 from ..result_dataclasses import TaskResult
-from ..types import ArrayLike
+from ..types import ArrayLike, Language
 from .model import Encoder
 from .task import DescriptiveDatasetStats, Task
 
@@ -88,7 +88,7 @@ def evaluate(self, model: Encoder) -> TaskResult:
         scores = scores.get(split, scores)
         score_is_nested = isinstance(scores[next(iter(scores.keys()))], dict)
         if not score_is_nested:
-            _scores = {lang: scores for lang in self.languages}
+            _scores: dict[str, dict[str, Union[float, str]]] = {lang: scores for lang in self.languages}
             scores = _scores
 
         task_result = TaskResult(

src/seb/interfaces/task.py

@@ -1,5 +1,7 @@
 from typing import Protocol, runtime_checkable
 
+from attr import dataclass
+
 from ..result_dataclasses import TaskResult
 from ..types import DescriptiveDatasetStats, Domain, Language, TaskType
 from .model import Encoder

src/seb/registered_tasks/speed.py

@@ -0,0 +1,103 @@
+import logging
+import platform
+import time
+from datetime import datetime
+from pathlib import Path
+from typing import Optional, Union
+
+import numpy as np
+import psutil
+import torch
+
+from seb.interfaces.model import EmbeddingModel
+from seb.interfaces.task import Task
+from seb.result_dataclasses import TaskResult
+from seb.types import DescriptiveDatasetStats, languages_in_seb
+
+logger = logging.getLogger(__name__)
+
+
+class CPUSpeedTask(Task):
+    reference = "NA"
+    version = "0.0.1"
+    task_type = "Speed"
+    languages = languages_in_seb
+    main_score = "Inference speed (seconds)"
+    domain = ["fiction"]  # noqa
+    name = "Speed (CPU)"
+    description = "Time taken to encode the text 'The Ugly Duckling' split by paragraphs on a CPU."
+    device = "cpu"
+    _dataset: Optional[list[str]] = None
+
+    def load_dataset(self) -> list[str]:
+        file_path = Path(__file__).parent / "the_ugly_duckling.txt"
+        with file_path.open("r") as f:
+            text = f.read()
+        return text.split("\n\n")
+
+    @property
+    def dataset(self) -> list[str]:
+        if self._dataset is None:
+            self._dataset = self.load_dataset()
+        return self._dataset
+
+    def get_descriptive_stats(self) -> DescriptiveDatasetStats:
+        dataset = self.load_dataset()
+        lengths = np.array([len(x) for x in dataset])
+        return DescriptiveDatasetStats(
+            mean_document_length=float(np.mean(lengths)), std_document_length=float(np.std(lengths)), num_documents=len(dataset)
+        )
+
+    def get_time_taken(self, model: EmbeddingModel) -> float:
+        dataset = self.load_dataset()
+        start = time.time()
+        with torch.no_grad():
+            model.encode(dataset, batch_size=1, device=self.device, task=self)
+        time_taken = time.time() - start
+        return time_taken
+
+    def evaluate(self, model: EmbeddingModel) -> TaskResult:
+        model.loader()  # ensure model is loaded
+
+        has_to_method = callable(getattr(model._model, "to", None))
+        if has_to_method:
+            model = model.to(self.device)  # type: ignore
+
+        run_inference = not (self.device == "cuda" and not has_to_method)
+        if run_inference:
+            time_taken = self.get_time_taken(model)
+        else:
+            logger.warn(f"Could not run inference on {model.meta.name} on {self.device} as it does not have a 'to' method. Skipping")
+            time_taken = np.nan
+
+        scores: dict[str, Union[str, float]] = {self.main_score: time_taken, **self.get_system_info()}
+
+        return TaskResult(
+            task_name=self.name,
+            task_description=self.description,
+            task_version=self.version,
+            scores={Language: scores for Language in self.languages},
+            time_of_run=datetime.now(),
+            main_score=self.main_score,
+        )
+
+    def get_system_info(self) -> dict[str, str]:
+        """
+        Returns a dictionary with system information.
+        """
+        info = {}
+        info["platform"] = platform.system()
+        info["platform-release"] = platform.release()
+        info["platform-version"] = platform.version()
+        info["architecture"] = platform.machine()
+        info["processor"] = platform.processor()
+        info["ram"] = str(round(psutil.virtual_memory().total / (1024.0**3))) + " GB"
+        info["Physical cores"] = psutil.cpu_count(logical=False)
+        info["Total cores"] = psutil.cpu_count(logical=True)
+        return info
+
+
+class GPUSpeedTask(CPUSpeedTask):
+    name = "Speed (GPU)"
+    description = "Time taken to encode the text 'The Ugly Duckling' split by paragraphs on a GPU."
+    device: str = "cuda"