Merge pull request #423 from KevinMusgrave/dev

KevinMusgrave · web-flow · commit df53e74c0998 · 2022-02-12T10:03:48.000-05:00
v1.1.1
diff --git a/conda_build/pytorch-metric-learning/meta.yaml b/conda_build/pytorch-metric-learning/meta.yaml
@@ -1,13 +1,13 @@
 {% set name = "pytorch-metric-learning" %}
-{% set version = "1.1.0" %}
+{% set version = "1.1.1" %}
 
 package:
   name: "{{ name|lower }}"
   version: "{{ version }}"
 
 source:
   url: "https://pypi.io/packages/source/{{ name[0] }}/{{ name }}/{{ name }}-{{ version }}.tar.gz"
-  sha256: d52913eee027746de928bf0e9c031f59a0915cfee4f02cdf81c198a058bd6b21
+  sha256: 6e572dc54179c762abc333fc4c6f68fcd909e800f9519ca1463235d14b9f5c44
 
 build:
   number: 0
diff --git a/src/pytorch_metric_learning/__init__.py b/src/pytorch_metric_learning/__init__.py
@@ -1 +1 @@
-__version__ = "1.1.0"
+__version__ = "1.1.1"
diff --git a/src/pytorch_metric_learning/distances/base_distance.py b/src/pytorch_metric_learning/distances/base_distance.py
@@ -27,7 +27,7 @@ def forward(self, query_emb, ref_emb=None):
         )
         mat = self.compute_mat(query_emb_normalized, ref_emb_normalized)
         if self.power != 1:
-            mat = mat ** self.power
+            mat = mat**self.power
         assert mat.size() == torch.Size((query_emb.size(0), ref_emb.size(0)))
         return mat
 
diff --git a/src/pytorch_metric_learning/losses/fast_ap_loss.py b/src/pytorch_metric_learning/losses/fast_ap_loss.py
@@ -34,7 +34,7 @@ def compute_loss(self, embeddings, labels, indices_tuple, ref_emb, ref_labels):
             return self.zero_losses()
         dist_mat = self.distance(embeddings)
 
-        histogram_max = 2 ** self.distance.power
+        histogram_max = 2**self.distance.power
         histogram_delta = histogram_max / self.num_bins
         mid_points = torch.linspace(
             0.0, histogram_max, steps=self.num_edges, device=device, dtype=dtype
diff --git a/src/pytorch_metric_learning/losses/large_margin_softmax_loss.py b/src/pytorch_metric_learning/losses/large_margin_softmax_loss.py
@@ -46,8 +46,8 @@ def get_cos_with_margin(self, cosine):
         cosine = cosine.unsqueeze(1)
         for attr in ["n_range", "margin_choose_n", "cos_powers", "alternating"]:
             setattr(self, attr, c_f.to_device(getattr(self, attr), cosine))
-        cos_powered = cosine ** self.cos_powers
-        sin_powered = (1 - cosine ** 2) ** self.n_range
+        cos_powered = cosine**self.cos_powers
+        sin_powered = (1 - cosine**2) ** self.n_range
         terms = (
             self.alternating * self.margin_choose_n * cos_powered * sin_powered
         )  # Equation 7 in the paper
diff --git a/src/pytorch_metric_learning/miners/distance_weighted_miner.py b/src/pytorch_metric_learning/miners/distance_weighted_miner.py
@@ -30,7 +30,7 @@ def mine(self, embeddings, labels, ref_emb, ref_labels):
 
         # See the first equation from Section 4 of the paper
         log_weights = (2.0 - d) * torch.log(mat) - ((d - 3) / 2) * torch.log(
-            1.0 - 0.25 * (mat ** 2.0)
+            1.0 - 0.25 * (mat**2.0)
         )
 
         inf_or_nan = torch.isinf(log_weights) | torch.isnan(log_weights)
diff --git a/src/pytorch_metric_learning/regularizers/center_invariant_regularizer.py b/src/pytorch_metric_learning/regularizers/center_invariant_regularizer.py
@@ -15,7 +15,7 @@ def compute_loss(self, weights):
         deviations_from_mean = squared_weight_norms - torch.mean(squared_weight_norms)
         return {
             "loss": {
-                "losses": (deviations_from_mean ** 2) / 4,
+                "losses": (deviations_from_mean**2) / 4,
                 "indices": c_f.torch_arange_from_size(weights),
                 "reduction_type": "element",
             }
diff --git a/src/pytorch_metric_learning/regularizers/lp_regularizer.py b/src/pytorch_metric_learning/regularizers/lp_regularizer.py
@@ -14,7 +14,7 @@ def __init__(self, p=2, power=1, **kwargs):
     def compute_loss(self, embeddings):
         reg = torch.norm(embeddings, p=self.p, dim=1)
         if self.power != 1:
-            reg = reg ** self.power
+            reg = reg**self.power
         return {
             "loss": {
                 "losses": reg,
diff --git a/src/pytorch_metric_learning/utils/accuracy_calculator.py b/src/pytorch_metric_learning/utils/accuracy_calculator.py
@@ -70,8 +70,7 @@ def r_precision(
     matches_per_row = torch.sum(same_label * relevance_mask, dim=1)
     max_possible_matches_per_row = torch.sum(relevance_mask, dim=1)
     accuracy_per_sample = (
-        c_f.to_dtype(matches_per_row, dtype=torch.float64)
-        / max_possible_matches_per_row
+        matches_per_row.type(torch.float64) / max_possible_matches_per_row
     )
     return maybe_get_avg_of_avgs(
         accuracy_per_sample, gt_labels, avg_of_avgs, return_per_class
@@ -99,9 +98,7 @@ def mean_average_precision(
     equality = is_same_label * relevance_mask
     cumulative_correct = torch.cumsum(equality, dim=1)
     k_idx = torch.arange(1, num_k + 1, device=device).repeat(num_samples, 1)
-    precision_at_ks = (
-        c_f.to_dtype(cumulative_correct * equality, dtype=torch.float64) / k_idx
-    )
+    precision_at_ks = (cumulative_correct * equality).type(torch.float64) / k_idx
     summed_precision_per_row = torch.sum(precision_at_ks * relevance_mask, dim=1)
     if at_r:
         max_possible_matches_per_row = torch.sum(relevance_mask, dim=1)
@@ -172,9 +169,7 @@ def precision_at_k(
 ):
     curr_knn_labels = knn_labels[:, :k]
     same_label = label_comparison_fn(gt_labels, curr_knn_labels)
-    accuracy_per_sample = (
-        c_f.to_dtype(torch.sum(same_label, dim=1), dtype=torch.float64) / k
-    )
+    accuracy_per_sample = torch.sum(same_label, dim=1).type(torch.float64) / k
     return maybe_get_avg_of_avgs(
         accuracy_per_sample, gt_labels, avg_of_avgs, return_per_class
     )
@@ -209,9 +204,7 @@ def get_lone_query_labels(
     unique_labels, match_counts = label_counts
     if embeddings_come_from_same_source:
         label_matches_itself = label_comparison_fn(unique_labels, unique_labels)
-        lone_condition = (
-            match_counts - c_f.to_dtype(label_matches_itself, dtype=torch.long) <= 0
-        )
+        lone_condition = match_counts - label_matches_itself.type(torch.long) <= 0
     else:
         lone_condition = match_counts == 0
     lone_query_labels = unique_labels[lone_condition]
diff --git a/src/pytorch_metric_learning/utils/common_functions.py b/src/pytorch_metric_learning/utils/common_functions.py
@@ -86,11 +86,12 @@ def get_hierarchy_label(batch_labels, hierarchy_level):
 def map_labels(label_map, labels):
     labels = to_numpy(labels)
     if labels.ndim == 2:
+        new_labels = np.zeros(labels.shape, dtype=int)
         for h in range(labels.shape[1]):
-            labels[:, h] = label_map(labels[:, h], h)
+            new_labels[:, h] = label_map(labels[:, h], h)
     else:
-        labels = label_map(labels, 0)
-    return labels
+        new_labels = label_map(labels, 0)
+    return new_labels
 
 
 def process_label(labels, hierarchy_level, label_map):
@@ -235,7 +236,9 @@ def make_label_to_rank_dict(label_set):
     Returns:
         A dictionary mapping each label to its numeric rank in the original set
     """
-    ranked = scipy.stats.rankdata(label_set) - 1
+    if len(set(label_set)) != len(label_set):
+        raise ValueError("label set must not have duplicates")
+    ranked = scipy.stats.rankdata(label_set).astype(int) - 1
     return {k: v for k, v in zip(label_set, ranked)}
 
 
diff --git a/tests/losses/test_large_margin_softmax_loss.py b/tests/losses/test_large_margin_softmax_loss.py
@@ -61,12 +61,12 @@ def test_large_margin_softmax_and_sphereface_loss(self):
                     curr_valA = (
                         c
                         * (curr_cosineA ** (margin - (2 * z)))
-                        * ((1 - curr_cosineA ** 2) ** z)
+                        * ((1 - curr_cosineA**2) ** z)
                     )
                     curr_valB = (
                         c
                         * (curr_cosineB ** (margin - (2 * z)))
-                        * ((1 - curr_cosineB ** 2) ** z)
+                        * ((1 - curr_cosineB**2) ** z)
                     )
                     if z % 2 == 1:
                         curr_valA *= -1
diff --git a/tests/regularizers/test_center_invariant_regularizer.py b/tests/regularizers/test_center_invariant_regularizer.py
@@ -49,7 +49,7 @@ def test_center_invariant_regularizer(self):
                     torch.norm(loss_func.W[:, i], p=2) ** 2
                     - average_squared_weight_norms
                 )
-                correct_reg_loss += (deviation ** 2) / 4
+                correct_reg_loss += (deviation**2) / 4
             correct_reg_loss /= num_classes
 
             correct_total_loss = correct_class_loss + (correct_reg_loss * reg_weight)
diff --git a/tests/utils/test_calculate_accuracies.py b/tests/utils/test_calculate_accuracies.py
@@ -789,3 +789,14 @@ def test_custom_knn(self):
 
         for k, v in acc1.items():
             self.assertTrue(np.isclose(v, acc2[k], rtol=1e-3))
+
+
+class TestWithinAutocast(unittest.TestCase):
+    def test_within_autocast(self):
+        if TEST_DEVICE == torch.device("cpu"):
+            return
+        AC = accuracy_calculator.AccuracyCalculator()
+        embeddings = torch.randn(1000, 32)
+        labels = torch.randint(0, 10, size=(1000,))
+        with torch.autocast(device_type="cuda", dtype=torch.float16):
+            acc = AC.get_accuracy(embeddings, embeddings, labels, labels, True)
diff --git a/tests/utils/test_common_functions.py b/tests/utils/test_common_functions.py
@@ -1,6 +1,7 @@
 import logging
 import unittest
 
+import numpy as np
 import torch
 from sklearn.preprocessing import StandardScaler
 
@@ -10,6 +11,41 @@
 from .. import WITH_COLLECT_STATS
 
 
+def process_label_helper(
+    cls,
+    dataset_labels,
+    batch_labels,
+    correct_mapped_batch_labels,
+    correct_unmapped_batch_labels,
+    hierarchy_level,
+    input_is_string=False,
+):
+    for set_min_label_to_zero in [False, True]:
+        if input_is_string and not set_min_label_to_zero:
+            continue
+        label_mapper = c_f.LabelMapper(set_min_label_to_zero, dataset_labels)
+        x = c_f.process_label(batch_labels, hierarchy_level, label_mapper.map)
+        if set_min_label_to_zero:
+            cls.assertTrue(np.array_equal(x, correct_mapped_batch_labels))
+        else:
+            cls.assertTrue(np.array_equal(x, correct_unmapped_batch_labels))
+
+
+def process_label_helper2d(
+    cls, dataset_labels, batch_labels, mapped_batch_labels, input_is_string=False
+):
+    for h in [0, 1]:
+        process_label_helper(
+            cls,
+            dataset_labels,
+            batch_labels,
+            mapped_batch_labels[h],
+            batch_labels[:, h],
+            h,
+            input_is_string=input_is_string,
+        )
+
+
 class TestCommonFunctions(unittest.TestCase):
     def test_torch_standard_scaler(self):
         torch.manual_seed(56987)
@@ -58,6 +94,72 @@ def test_logger(self):
                 self.assertTrue(c_f.LOGGER.level == level)
                 self.assertTrue(logging.getLogger().level == logging.WARNING)
 
+    def test_process_label(self):
+        # 1D number labels
+        dataset_labels = np.array([1, 1, 10, 13, 13, 13, 15])
+        batch_labels = np.array([1, 10, 1, 13, 10, 15, 15, 15])
+        mapped_batch_labels = np.array([0, 1, 0, 2, 1, 3, 3, 3])
+        process_label_helper(
+            self, dataset_labels, batch_labels, mapped_batch_labels, batch_labels, 0
+        )
+
+        # 1D string labels
+        # These work only with set_min_label_to_zero=True
+        # The labels will be sorted alphabetically. So in the following, "bear" becomes 0
+        dataset_labels = np.array(["dog", "dog", "cat", "bear", "bear", "bear", "lion"])
+        batch_labels = np.array(
+            ["dog", "cat", "dog", "bear", "cat", "lion", "lion", "lion"]
+        )
+        mapped_batch_labels = np.array([2, 1, 2, 0, 1, 3, 3, 3])
+        process_label_helper(
+            self,
+            dataset_labels,
+            batch_labels,
+            mapped_batch_labels,
+            batch_labels,
+            0,
+            input_is_string=True,
+        )
+
+        # 2D number labels
+        dataset_labels = np.array(
+            [[1, 1, 10, 13, 13, 13, 15], [10, 20, 30, 40, 50, 60, 70]]
+        ).transpose()
+        batch_labels = np.array(
+            [[1, 10, 1, 13, 10, 15, 15, 15], [30, 70, 40, 10, 70, 60, 50, 40]]
+        ).transpose()
+        mapped_batch_labels0 = np.array([0, 1, 0, 2, 1, 3, 3, 3])
+        mapped_batch_labels1 = np.array([2, 6, 3, 0, 6, 5, 4, 3])
+        process_label_helper2d(
+            self,
+            dataset_labels,
+            batch_labels,
+            [mapped_batch_labels0, mapped_batch_labels1],
+        )
+
+        # 2D string labels
+        dataset_labels = np.array(
+            [
+                ["dog", "dog", "cat", "bear", "bear", "bear", "lion"],
+                ["A.1", "A.2", "A.3", "A.4", "A.5", "A.6", "A.7"],
+            ]
+        ).transpose()
+        batch_labels = np.array(
+            [
+                ["dog", "cat", "dog", "bear", "cat", "lion", "lion", "lion"],
+                ["A.3", "A.7", "A.4", "A.1", "A.7", "A.6", "A.5", "A.4"],
+            ]
+        ).transpose()
+        mapped_batch_labels0 = np.array([2, 1, 2, 0, 1, 3, 3, 3])
+        mapped_batch_labels1 = np.array([2, 6, 3, 0, 6, 5, 4, 3])
+        process_label_helper2d(
+            self,
+            dataset_labels,
+            batch_labels,
+            [mapped_batch_labels0, mapped_batch_labels1],
+            input_is_string=True,
+        )
+
 
 if __name__ == "__main__":
     unittest.main()

Original file line number	Diff line number	Diff line change
`@@ -1 +1 @@`
`1`		`-__version__ = "1.1.0"`
	`1`	`+__version__ = "1.1.1"`
Original file line number	Diff line number	Diff line change
`@@ -27,7 +27,7 @@ def forward(self, query_emb, ref_emb=None):`
`27`	`27`	`)`
`28`	`28`	`mat = self.compute_mat(query_emb_normalized, ref_emb_normalized)`
`29`	`29`	`if self.power != 1:`
`30`		`- mat = mat ** self.power`
	`30`	`+ mat = mat**self.power`
`31`	`31`	`assert mat.size() == torch.Size((query_emb.size(0), ref_emb.size(0)))`
`32`	`32`	`return mat`
`33`	`33`
Original file line number	Diff line number	Diff line change
`@@ -30,7 +30,7 @@ def mine(self, embeddings, labels, ref_emb, ref_labels):`
`30`	`30`
`31`	`31`	`# See the first equation from Section 4 of the paper`
`32`	`32`	`log_weights = (2.0 - d) * torch.log(mat) - ((d - 3) / 2) * torch.log(`
`33`		`- 1.0 - 0.25 * (mat ** 2.0)`
	`33`	`+ 1.0 - 0.25 * (mat**2.0)`
`34`	`34`	`)`
`35`	`35`
`36`	`36`	`inf_or_nan = torch.isinf(log_weights) \| torch.isnan(log_weights)`
Original file line number	Diff line number	Diff line change
`@@ -15,7 +15,7 @@ def compute_loss(self, weights):`
`15`	`15`	`deviations_from_mean = squared_weight_norms - torch.mean(squared_weight_norms)`
`16`	`16`	`return {`
`17`	`17`	`"loss": {`
`18`		`- "losses": (deviations_from_mean ** 2) / 4,`
	`18`	`+ "losses": (deviations_from_mean**2) / 4,`
`19`	`19`	`"indices": c_f.torch_arange_from_size(weights),`
`20`	`20`	`"reduction_type": "element",`
`21`	`21`	`}`