Small fixes in EasySet and AbstractMetaLearner

* Fix best validation accuracy update

* Sort data instances for each class in EasySet

* Move switch to train mode inside fit_on_task()

* Add citation file

* Make AbstractMetaLearner.fit() return average loss

* Add EasySet.number_of_classes()

* Fix python version in CI linter

* Fix linter version

* Bump version: 0.2.1 → 0.2.2

Author: ebennequin

.bumpversion.cfg

@@ -1,5 +1,5 @@
 [bumpversion]
-current_version = 0.2.1
+current_version = 0.2.2
 commit = True
 tag = False
 

.circleci/config.yml

@@ -33,7 +33,7 @@ jobs:
   lint:
     working_directory: ~/project
     docker:
-      - image: cimg/python:3.7
+      - image: cimg/python:3.7.10
     steps:
       - run:
           name: Set python 3.7.10

CITATION.cff

@@ -0,0 +1,19 @@
+# This CITATION.cff file was generated with cffinit.
+# Visit https://bit.ly/cffinit to generate yours today!
+
+cff-version: 1.2.0
+title: easyfsl
+message: >-
+  If you use easyfsl in your research, please cite it
+  using these metadata.
+type: software
+authors:
+  - given-names: Etienne
+    family-names: Bennequin
+    email: [email protected]
+    affiliation: Université Paris-Saclay
+repository-code: 'https://github.com/sicara/easy-few-shot-learning'
+abstract: >-
+  Ready-to-use code and tutorial notebooks to boost
+  your way into few-shot image classification. 
+license: MIT

dev_requirements.txt

@@ -2,9 +2,9 @@ black>=20.8b1
 loguru>=0.5.3
 matplotlib>=3.3.4
 pandas>=1.1.0
-pylint>=2.7.0
-pytest>=6.2.2
-pytest-mock>=3.6.1
+pylint==2.8.2
+pytest==6.2.2
+pytest-mock==3.6.1
 torch>=1.7.1
 torchvision>=0.8.2
 tqdm>=4.56.0

easyfsl/__init__.py

@@ -5,4 +5,4 @@
 for few-shot learning experiences.
 """
 
-__version__ = "0.2.1"
+__version__ = "0.2.2"

easyfsl/data_tools/easy_set.py

@@ -61,7 +61,8 @@ def load_specs(specs_file: Path) -> dict:
         if specs_file.suffix != ".json":
             raise ValueError("EasySet requires specs in a JSON file.")
 
-        specs = json.load(open(specs_file, "r"))
+        with open(specs_file, "r") as file:
+            specs = json.load(file)
 
         if "class_names" not in specs.keys() or "class_roots" not in specs.keys():
             raise ValueError(
@@ -124,7 +125,7 @@ def list_data_instances(class_roots: List[str]) -> (List[str], List[int]):
         for class_id, class_root in enumerate(class_roots):
             class_images = [
                 str(image_path)
-                for image_path in Path(class_root).glob("*")
+                for image_path in sorted(Path(class_root).glob("*"))
                 if image_path.is_file()
             ]
             images += class_images
@@ -155,3 +156,6 @@ def __getitem__(self, item: int):
 
     def __len__(self) -> int:
         return len(self.labels)
+
+    def number_of_classes(self):
+        return len(self.class_names)

easyfsl/methods/abstract_meta_learner.py

@@ -1,5 +1,6 @@
 from abc import abstractmethod
 from pathlib import Path
+from statistics import mean
 from typing import Union
 
 import torch
@@ -162,6 +163,7 @@ def fit_on_task(
         Returns:
             the value of the classification loss (for reporting purposes)
         """
+        self.train()
         optimizer.zero_grad()
         self.process_support_set(support_images.cuda(), support_labels.cuda())
         classification_scores = self(query_images.cuda())
@@ -178,7 +180,7 @@ def fit(
         optimizer: optim.Optimizer,
         val_loader: DataLoader = None,
         validation_frequency: int = 1000,
-    ):
+    ) -> float:
         """
         Train the model on few-shot classification tasks.
         Args:
@@ -187,11 +189,12 @@ def fit(
             val_loader: loads data from the validation set in the shape of few-shot classification
                 tasks
             validation_frequency: number of training episodes between two validations
+        Returns:
+            average loss
         """
         log_update_frequency = 10
 
         all_loss = []
-        self.train()
         with tqdm(
             enumerate(train_loader), total=len(train_loader), desc="Meta-Training"
         ) as tqdm_train:
@@ -222,6 +225,8 @@ def fit(
                     if (episode_index + 1) % validation_frequency == 0:
                         self.validate(val_loader)
 
+        return mean(all_loss)
+
     def validate(self, val_loader: DataLoader) -> float:
         """
         Validate the model on the validation set.
@@ -237,6 +242,7 @@ def validate(self, val_loader: DataLoader) -> float:
         if validation_accuracy > self.best_validation_accuracy:
             print("Best validation accuracy so far!")
             self.best_model_state = self.state_dict()
+            self.best_validation_accuracy = validation_accuracy
 
         return validation_accuracy
 

easyfsl/methods/matching_networks.py

@@ -4,7 +4,7 @@
 """
 
 import torch
-import torch.nn as nn
+from torch import nn
 from easyfsl.methods import AbstractMetaLearner
 
 

easyfsl/methods/relation_networks.py

@@ -4,7 +4,7 @@
 """
 
 import torch
-import torch.nn as nn
+from torch import nn
 from easyfsl.methods import AbstractMetaLearner
 from easyfsl.utils import compute_prototypes
 

easyfsl/tests/methods/abstract_meta_learner_test.py

@@ -82,6 +82,16 @@ def test_validate_updates_best_model_state_if_it_has_best_validation_accuracy(
         meta_learner.validate(None)
         assert meta_learner.best_model_state is not None
 
+    @staticmethod
+    def test_validate_updates_best_accuracy_if_it_has_best_validation_accuracy(
+        mocker,
+    ):
+        mocker.patch("easyfsl.methods.AbstractMetaLearner.evaluate", return_value=0.5)
+        meta_learner = AbstractMetaLearner(resnet18())
+        meta_learner.best_validation_accuracy = 0.1
+        meta_learner.validate(None)
+        assert meta_learner.best_validation_accuracy == 0.5
+
     @staticmethod
     @pytest.mark.parametrize(
         "accuracy",