Merge pull request #67 from KevinMusgrave/dev

v0.0.72

Author: KevinMusgrave

.gitignore

@@ -17,4 +17,5 @@ examples/**/officehome_resized.tar.gz
 examples/**/officehome
 examples/**/saved_models
 examples/**/lightning_logs
-zzz_pytorch_adapt_dataset_test_folder
+zzz_pytorch_adapt_dataset_test_folder
+.env

build_script.sh

@@ -1,5 +1,5 @@
 ./format_code.sh
-python -m unittest discover && \
+RUN_DATASET_TESTS=true python -m unittest discover && \
 rm -rfv build/ && \
 rm -rfv dist/ && \
 rm -rfv src/pytorch_adapt.egg-info/ && \

src/pytorch_adapt/__init__.py

@@ -1 +1 @@
-__version__ = "0.0.71"
+__version__ = "0.0.72"

src/pytorch_adapt/datasets/getters.py

@@ -24,6 +24,7 @@ def get_datasets(
     folder,
     download=False,
     return_target_with_labels=False,
+    supervised=False,
     transform_getter=None,
 ):
     def getter(domains, train, is_training):
@@ -47,14 +48,19 @@ def getter(domains, train, is_training):
         output["src_train"] = SourceDataset(getter(src_domains, True, False))
         output["src_val"] = SourceDataset(getter(src_domains, False, False))
     if target_domains:
-        output["target_train"] = TargetDataset(getter(target_domains, True, False))
-        output["target_val"] = TargetDataset(getter(target_domains, False, False))
+        output["target_train"] = TargetDataset(
+            getter(target_domains, True, False), supervised=supervised
+        )
+        output["target_val"] = TargetDataset(
+            getter(target_domains, False, False), supervised=supervised
+        )
+        # For academic setting: unsupervised learning w/ seperate target datasets that have gt lables for eval.
         if return_target_with_labels:
-            output["target_train_with_labels"] = SourceDataset(
-                getter(target_domains, True, False), domain=1
+            output["target_train_with_labels"] = TargetDataset(
+                getter(target_domains, True, False), domain=1, supervised=True
             )
-            output["target_val_with_labels"] = SourceDataset(
-                getter(target_domains, False, False), domain=1
+            output["target_val_with_labels"] = TargetDataset(
+                getter(target_domains, False, False), domain=1, supervised=True
             )
     if src_domains and target_domains:
         output["train"] = CombinedSourceAndTargetDataset(

tests/__init__.py

@@ -29,6 +29,7 @@
 TEST_FOLDER = "zzz_pytorch_adapt_test_folder"
 DATASET_FOLDER = "zzz_pytorch_adapt_dataset_test_folder"
 RUN_DATASET_TESTS = os.environ.get("RUN_DATASET_TESTS", False)
+RUN_DOMAINNET_DATASET_TESTS = os.environ.get("RUN_DOMAINNET_DATASET_TESTS", False)
 
 TEST_DTYPES = [getattr(torch, x) for x in dtypes_from_environ]
 TEST_DEVICE = torch.device(device_from_environ)

tests/datasets/test_domainnet.py

@@ -4,18 +4,18 @@
 
 from pytorch_adapt.datasets import DomainNet, DomainNet126, DomainNet126Full
 
-from .. import DATASET_FOLDER, RUN_DATASET_TESTS
+from .. import DATASET_FOLDER, RUN_DOMAINNET_DATASET_TESTS
 from .utils import (
     check_full,
     check_train_test_disjoint,
     check_train_test_matches_full,
     loop_through_dataset,
-    skip_reason,
+    skip_reason_domainnet,
 )
 
 
 class TestDomainNet(unittest.TestCase):
-    @unittest.skipIf(not RUN_DATASET_TESTS, skip_reason)
+    @unittest.skipIf(not RUN_DOMAINNET_DATASET_TESTS, skip_reason_domainnet)
     def test_domainnet(self):
         transform = torch_transforms.Compose(
             [
@@ -39,7 +39,7 @@ def test_domainnet(self):
             self.assertTrue(len(dataset) == length)
             loop_through_dataset(dataset)
 
-    @unittest.skipIf(not RUN_DATASET_TESTS, skip_reason)
+    @unittest.skipIf(not RUN_DOMAINNET_DATASET_TESTS, skip_reason_domainnet)
     def test_domainnet126(self):
         check_train_test_matches_full(
             self,
@@ -50,7 +50,7 @@ def test_domainnet126(self):
             DATASET_FOLDER,
         )
 
-    @unittest.skipIf(not RUN_DATASET_TESTS, skip_reason)
+    @unittest.skipIf(not RUN_DOMAINNET_DATASET_TESTS, skip_reason_domainnet)
     def test_domainnet126_full(self):
         check_full(
             self,

tests/datasets/test_getters.py

@@ -18,13 +18,37 @@
 
 
 class TestGetters(unittest.TestCase):
-    def helper(self, datasets, src_class, target_class, sizes):
+    def helper(
+        self,
+        datasets,
+        src_class,
+        target_class,
+        sizes,
+        target_with_labels=False,
+        supervised=False,
+    ):
         for k in ["src_train", "src_val"]:
             self.assertTrue(isinstance(datasets[k].dataset.datasets[0], src_class))
+            self.assertTrue(isinstance(datasets[k], SourceDataset))
             self.assertTrue(len(datasets[k]) == sizes[k])
-        for k in ["target_train", "target_val"]:
+
+        target_splits = ["target_train", "target_val"]
+        if target_with_labels:
+            target_splits += ["target_train_with_labels", "target_val_with_labels"]
+        target_sizes = {k for k in sizes.keys() if k.startswith("target")}
+
+        self.assertTrue(set(target_splits) == target_sizes)
+        for k in target_splits:
             self.assertTrue(isinstance(datasets[k].dataset.datasets[0], target_class))
+            self.assertTrue(isinstance(datasets[k], TargetDataset))
             self.assertTrue(len(datasets[k]) == sizes[k])
+            if supervised:
+                # target_train and target_val will be supervised
+                # and if target_with_labels is true, the with_labels will also be supervised
+                self.assertTrue(datasets[k].supervised)
+            else:
+                # otherwise, only the ones with labels are supervised
+                self.assertTrue(datasets[k].supervised == k.endswith("with_labels"))
 
     @unittest.skipIf(not RUN_DATASET_TESTS, skip_reason)
     def test_empty_array(self):
@@ -41,21 +65,30 @@ def test_empty_array(self):
         self.assertTrue(isinstance(datasets["train"], TargetDataset))
 
     @unittest.skipIf(not RUN_DATASET_TESTS, skip_reason)
-    def test_get_mnist_mnistm(self):
-        datasets = get_mnist_mnistm(
-            ["mnist"], ["mnistm"], folder=DATASET_FOLDER, download=True
-        )
-        self.helper(
-            datasets,
-            MNIST,
-            MNISTM,
-            {
-                "src_train": 60000,
-                "src_val": 10000,
-                "target_train": 59001,
-                "target_val": 9001,
-            },
-        )
+    def test_get_mnist_mnistm_return_targets(self):
+        for target_with_labels in [False, True]:
+            for supervised in [False, True]:
+                datasets = get_mnist_mnistm(
+                    ["mnist"],
+                    ["mnistm"],
+                    folder=DATASET_FOLDER,
+                    download=True,
+                    return_target_with_labels=target_with_labels,
+                    supervised=supervised,
+                )
+                len_dict = {
+                    "src_train": 60000,
+                    "src_val": 10000,
+                    "target_train": 59001,
+                    "target_val": 9001,
+                }
+                if target_with_labels:
+                    len_dict["target_train_with_labels"] = len_dict["target_train"]
+                    len_dict["target_val_with_labels"] = len_dict["target_val"]
+
+                self.helper(
+                    datasets, MNIST, MNISTM, len_dict, target_with_labels, supervised
+                )
 
     @unittest.skipIf(not RUN_DATASET_TESTS, skip_reason)
     def test_officehome(self):

tests/datasets/utils.py

@@ -6,6 +6,7 @@
 from torchvision import transforms as torch_transforms
 
 skip_reason = "RUN_DATASET_TESTS is False"
+skip_reason_domainnet = "RUN_DOMAINNET_DATASET_TESTS is False"
 
 
 def simple_transform():