keras-rs-nightly 0.0.1.dev2025040703__py3-none-any.whl → 0.0.1.dev2025040803__py3-none-any.whl

keras_rs/api/layers/__init__.py

@@ -14,6 +14,9 @@ from keras_rs.src.layers.retrieval.brute_force_retrieval import (
 from keras_rs.src.layers.retrieval.hard_negative_mining import (
     HardNegativeMining,
 )
+from keras_rs.src.layers.retrieval.remove_accidental_hits import (
+    RemoveAccidentalHits,
+)
 from keras_rs.src.layers.retrieval.sampling_probability_correction import (
     SamplingProbabilityCorrection,
 )

keras_rs/src/layers/retrieval/hard_negative_mining.py

@@ -1,13 +1,13 @@
 from typing import Any
 
 import keras
-import numpy as np
+import ml_dtypes
 from keras import ops
 
 from keras_rs.src import types
 from keras_rs.src.api_export import keras_rs_export
 
-MAX_FLOAT = np.finfo(np.float32).max / 100.0
+MAX_FLOAT = ml_dtypes.finfo("float32").max / 100.0
 
 
 def _gather_elements_along_row(

keras_rs/src/layers/retrieval/remove_accidental_hits.py

@@ -0,0 +1,84 @@
+import keras
+import ml_dtypes
+from keras import ops
+
+from keras_rs.src import types
+from keras_rs.src.api_export import keras_rs_export
+from keras_rs.src.utils import keras_utils
+
+SMALLEST_FLOAT = ml_dtypes.finfo("float32").smallest_normal / 100.0
+
+
+@keras_rs_export("keras_rs.layers.RemoveAccidentalHits")
+class RemoveAccidentalHits(keras.layers.Layer):
+    """Zeroes the logits of accidental negatives.
+
+    Zeroes the logits of negative candidates that have the same ID as the
+    positive candidate in that row.
+    """
+
+    def call(
+        self,
+        labels: types.Tensor,
+        logits: types.Tensor,
+        candidate_ids: types.Tensor,
+    ) -> types.Tensor:
+        """Zeroes selected logits.
+
+        For each row in the batch, zeroes the logits of negative candidates that
+        have the same ID as the positive candidate in that row.
+
+        Args:
+          labels: one-hot labels tensor, typically
+              `[batch_size, num_candidates]` but can have more dimensions or be
+              1D as `[num_candidates]`.
+          logits: logits tensor. Must have the same shape as `labels`.
+          candidate_ids: candidate identifiers tensor, can be `[num_candidates]`
+              or `[batch_size, num_candidates]` or have more dimensions as long
+              as they match the last dimensions of `labels`.
+
+        Returns:
+          logits: Modified logits.
+        """
+        # A more principled way is to implement
+        # `softmax_cross_entropy_with_logits` with a input mask. Here we
+        # approximate so by letting accidental hits have extremely small logits
+        # (SMALLEST_FLOAT) for ease-of-implementation.
+
+        labels_shape = ops.shape(labels)
+        labels_rank = len(labels_shape)
+        logits_shape = ops.shape(logits)
+        candidate_ids_shape = ops.shape(candidate_ids)
+        candidate_ids_rank = len(candidate_ids_shape)
+
+        if not keras_utils.check_shapes_compatible(labels_shape, logits_shape):
+            raise ValueError(
+                "`labels` and `logits` should have the same shape. Received: "
+                f"`labels.shape` = {labels_shape}, "
+                f"`logits.shape` = {logits_shape}."
+            )
+
+        if not keras_utils.check_shapes_compatible(
+            labels_shape[-candidate_ids_rank:], candidate_ids_shape
+        ):
+            raise ValueError(
+                "`candidate_ids` should have the same shape as the last "
+                "dimensions of `labels`. Received: "
+                f"`candidate_ids.shape` = {candidate_ids_shape}, "
+                f"`labels.shape` = {labels_shape}."
+            )
+
+        # Add dimensions to `candidate_ids` to have the same rank as `labels`.
+        if candidate_ids_rank < labels_rank:
+            candidate_ids = ops.expand_dims(
+                candidate_ids, list(range(labels_rank - candidate_ids_rank))
+            )
+        positive_indices = ops.expand_dims(ops.argmax(labels, axis=-1), -1)
+        positive_candidate_ids = ops.take(candidate_ids, positive_indices)
+
+        duplicate = ops.cast(
+            ops.equal(positive_candidate_ids, candidate_ids), labels.dtype
+        )
+        duplicate = ops.subtract(duplicate, labels)
+
+        return ops.add(logits, ops.multiply(duplicate, SMALLEST_FLOAT))

keras_rs/src/version.py

@@ -1,7 +1,7 @@
 from keras_rs.src.api_export import keras_rs_export
 
 # Unique source of truth for the version number.
-__version__ = "0.0.1.dev2025040703"
+__version__ = "0.0.1.dev2025040803"
 
 
 @keras_rs_export("keras_rs.version")

{keras_rs_nightly-0.0.1.dev2025040703.dist-info → keras_rs_nightly-0.0.1.dev2025040803.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: keras-rs-nightly
-Version: 0.0.1.dev2025040703
+Version: 0.0.1.dev2025040803
 Summary: Multi-backend recommender systems with Keras 3.
 Author-email: Keras RS team <keras-rs@google.com>
 License: Apache License 2.0
@@ -20,6 +20,7 @@ Classifier: Topic :: Software Development
 Requires-Python: >=3.9
 Description-Content-Type: text/markdown
 Requires-Dist: keras
+Requires-Dist: ml-dtypes
 
 # Keras Recommenders
 

{keras_rs_nightly-0.0.1.dev2025040703.dist-info → keras_rs_nightly-0.0.1.dev2025040803.dist-info}/RECORD

@@ -1,18 +1,19 @@
 keras_rs/__init__.py,sha256=X3VNKb_6VDEs5GqcbEc_l8mAsefWb5UgSu8krnQdFcM,794
 keras_rs/api/__init__.py,sha256=9Xf-uH9j_SBaTc5RU0pkxrOEgHWPwSKjf4_maySH_nU,272
-keras_rs/api/layers/__init__.py,sha256=uZgnycSMTUalc2WfFaynmV3PpvZ4XeTttkwjcjsNMvk,590
+keras_rs/api/layers/__init__.py,sha256=VvsmrU1T198D8svbjAk7bysRRfDeT71A-9wv7NyPC28,685
 keras_rs/api/losses/__init__.py,sha256=LGW7eHQh8FbQXdMV1s9zJpbloVlz_Zlo51sorWAvFwE,455
 keras_rs/src/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_rs/src/api_export.py,sha256=RsmG-DvO-cdFeAF9W6LRzms0kvtm-Yp9BAA_d-952zI,510
 keras_rs/src/types.py,sha256=UyOdgjqrqg_b58opnY8n6gTiDHKVR8z_bmEruehERBk,514
-keras_rs/src/version.py,sha256=BT4v6Amhk3aylUK7NppTKzinXu04ers7bM8r-DlIpGs,222
+keras_rs/src/version.py,sha256=a9BItWgRhEYrbx2zCLxStYnDVn5qAEy8Ovsszh2TS_o,222
 keras_rs/src/layers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_rs/src/layers/feature_interaction/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_rs/src/layers/feature_interaction/dot_interaction.py,sha256=jGHcg0EiWxth6LTxG2yWgHcyx_GXrxvA61uQqpPfnDQ,6900
 keras_rs/src/layers/feature_interaction/feature_cross.py,sha256=5OCSI0vFYzJNmgkKcuHIbVv8U2q3UvS80-qZjPimDjM,8155
 keras_rs/src/layers/retrieval/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_rs/src/layers/retrieval/brute_force_retrieval.py,sha256=mohILOt6PC6jHBztaowDbj3QBnSetuvkq55FmE39PlY,7321
-keras_rs/src/layers/retrieval/hard_negative_mining.py,sha256=8c44iEUmK_SwfHPdrVtg96ycBZzYf62ee66a49mqCbU,4115
+keras_rs/src/layers/retrieval/hard_negative_mining.py,sha256=CY8-3W52ZBIFcEfvjXJxbFltD6ulXl4-sZCRF6stIEc,4119
+keras_rs/src/layers/retrieval/remove_accidental_hits.py,sha256=fiFQLlkMBXhG8V7a8mv_hKOwlqEJeUiMBYUVQw1woTE,3270
 keras_rs/src/layers/retrieval/sampling_probability_correction.py,sha256=80vgOPfBiF-PC0dSyqS57IcIxOxi_Q_R7eSXHn1G0yI,1437
 keras_rs/src/losses/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_rs/src/losses/pairwise_hinge_loss.py,sha256=vqDGd-OnZxiqdeE6vuabE8BKDfill3D2GM0lW5JUmsg,922
@@ -23,7 +24,7 @@ keras_rs/src/losses/pairwise_soft_zero_one_loss.py,sha256=XBej5nybFXEQ-Vp6GLvNmq
 keras_rs/src/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_rs/src/utils/keras_utils.py,sha256=IjWSRieBkv7UX12qgUoI1tcOeISstCLRSTqSHpT06yE,1275
 keras_rs/src/utils/pairwise_loss_utils.py,sha256=5SAqA3z30A1awzV9l5oVbcno5Z6HXARkNcUFTPL7_jg,3380
-keras_rs_nightly-0.0.1.dev2025040703.dist-info/METADATA,sha256=W0bJDXs5KnTBjcilC9QnfkBXT-clmCaxg79UOmU83OE,3522
-keras_rs_nightly-0.0.1.dev2025040703.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
-keras_rs_nightly-0.0.1.dev2025040703.dist-info/top_level.txt,sha256=pWs8X78Z0cn6lfcIb9VYOW5UeJ-TpoaO9dByzo7_FFo,9
-keras_rs_nightly-0.0.1.dev2025040703.dist-info/RECORD,,
+keras_rs_nightly-0.0.1.dev2025040803.dist-info/METADATA,sha256=oC2tgoUja1GWeMPNM6ZFmrBJNnuPof-Eb8LioVF3kYs,3547
+keras_rs_nightly-0.0.1.dev2025040803.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
+keras_rs_nightly-0.0.1.dev2025040803.dist-info/top_level.txt,sha256=pWs8X78Z0cn6lfcIb9VYOW5UeJ-TpoaO9dByzo7_FFo,9
+keras_rs_nightly-0.0.1.dev2025040803.dist-info/RECORD,,