Merge pull request #21 from EmanueleGhelfi/losses_and_metrics

Losses and metrics

.travis.yml

@@ -2,6 +2,8 @@ dist: xenial
 language: python
 addons:
   artifacts: true
+  apt:
+    update: true
 python:
 - '3.7'
 stages:
@@ -11,9 +13,6 @@ stages:
   - Deploy
 before_install:
   - sudo apt-get install -y --no-install-recommends bc
-addons:
-  apt:
-    update: true
 install: &requirements
 - pip install -r dev-requirements.txt
 - pip install -e .
@@ -25,10 +24,7 @@ jobs:
     - pip install codecov
     - pip --no-cache-dir install --upgrade git+https://github.com/thisch/pytest-sphinx.git pytest
     - pip --no-cache-dir install pytest-cov
-    - module=""
-    - for d in $(ls -d */); do if [ -f "$d"__init__.py ]; then module=${d::-1}; fi
-      done
-    - pytest -x -s -vvv --doctest-modules $module --cov=$module
+    - pytest -x -s -vvv --doctest-modules ashpy tests --cov=ashpy
     after_success:
     - codecov
   - stage: Black

ashpy/__init__.py

@@ -14,6 +14,7 @@
 
 """ASHPY Package."""
 
+from .modes import LogEvalMode
 from . import contexts
 from . import datasets
 from . import layers
@@ -23,7 +24,6 @@
 from . import trainers
 from . import ashtypes
 from . import keras
-from .modes import LogEvalMode
 
 __version__ = "1.0.2"
 __url__ = "https://github.com/zurutech/ashpy"

ashpy/contexts/base_context.py

@@ -78,7 +78,7 @@ def measure_metrics(self) -> None:
     def model_selection(self) -> None:
         """Use the metrics to perform model selection."""
         for metric in self._metrics:
-            metric.model_selection(self._ckpt)
+            metric.model_selection(self._ckpt, self._global_step)
 
     @property
     def log_eval_mode(self) -> LogEvalMode:

ashpy/contexts/classifier.py

@@ -21,11 +21,11 @@
 import tensorflow as tf  # pylint: disable=import-error
 
 from ashpy.contexts.base_context import BaseContext
-from ashpy.metrics import Metric
 from ashpy.modes import LogEvalMode
 
 if TYPE_CHECKING:
     from ashpy.losses.executor import Executor
+    from ashpy.metrics import Metric
 
 
 class ClassifierContext(BaseContext):

ashpy/contexts/gan.py

@@ -21,11 +21,11 @@
 import tensorflow as tf
 
 from ashpy.contexts.base_context import BaseContext
-from ashpy.metrics import Metric
 from ashpy.modes import LogEvalMode
 
 if TYPE_CHECKING:
     from ashpy.losses.executor import Executor
+    from ashpy.metrics import Metric
 
 
 class GANContext(BaseContext):

ashpy/keras/losses.py

@@ -182,3 +182,87 @@ def call(self, d_real: tf.Tensor, d_fake: tf.Tensor) -> tf.Tensor:
             self._positive_mse(tf.ones_like(d_real), d_real)
             + self._negative_mse(tf.zeros_like(d_fake), d_fake)
         )
+
+
+class DHingeLoss(tf.keras.losses.Loss):
+    r"""
+    Discriminator Hinge Loss as Keras Metric.
+    See Geometric GAN [1]_ for more details.
+
+    .. [1] https://arxiv.org/abs/1705.02894
+    """
+
+    def __init__(self) -> None:
+        """Initialize the Loss."""
+        self._hinge_loss_real = tf.keras.losses.Hinge(
+            reduction=tf.keras.losses.Reduction.NONE
+        )
+        self._hinge_loss_fake = tf.keras.losses.Hinge(
+            reduction=tf.keras.losses.Reduction.NONE
+        )
+        super().__init__()
+
+    @property
+    def reduction(self) -> tf.keras.losses.Reduction:
+        """Return the current `reduction` for this type of loss."""
+        return self._hinge_loss_fake.reduction
+
+    @reduction.setter
+    def reduction(self, value: tf.keras.losses.Reduction) -> None:
+        """
+        Set the `reduction`.
+
+        Args:
+            value (:py:class:`tf.keras.losses.Reduction`): Reduction to use for the loss.
+
+        """
+        self._hinge_loss_fake.reduction = value
+        self._hinge_loss_real.reduction = value
+
+    def call(self, d_real: tf.Tensor, d_fake: tf.Tensor) -> tf.Tensor:
+        """Compute the hinge loss"""
+        real_loss = self._hinge_loss_real(tf.ones_like(d_real), d_real)
+        fake_loss = self._hinge_loss_fake(
+            tf.math.negative(tf.ones_like(d_fake)), d_fake
+        )
+
+        loss = real_loss + fake_loss  # shape: (batch_size, 1)
+
+        return loss
+
+
+class GHingeLoss(tf.keras.losses.Loss):
+    r"""
+    Generator Hinge Loss as Keras Metric.
+    See Geometric GAN [1]_ for more details.
+
+    .. [1] https://arxiv.org/abs/1705.02894
+
+    """
+
+    def __init__(self) -> None:
+        """Initialize the Loss."""
+        super().__init__()
+        self._reduction = tf.keras.losses.Reduction.SUM_OVER_BATCH_SIZE
+
+    @property
+    def reduction(self) -> tf.keras.losses.Reduction:
+        """Return the current `reduction` for this type of loss."""
+        return self._reduction
+
+    @reduction.setter
+    def reduction(self, value: tf.keras.losses.Reduction) -> None:
+        """
+        Set the `reduction`.
+
+        Args:
+            value (:py:class:`tf.keras.losses.Reduction`): Reduction to use for the loss.
+
+        """
+        self._reduction = value
+
+    def call(self, d_real: tf.Tensor, d_fake: tf.Tensor) -> tf.Tensor:
+        """Computes the hinge loss"""
+        fake_loss = -tf.nn.relu(d_fake)
+
+        return fake_loss

ashpy/losses/__init__.py

@@ -46,18 +46,20 @@
 
     gan.GANExecutor
     gan.AdversarialLossType
-    gan.AdversarialLossG
-    gan.AdversarialLossD
+    gan.GeneratorAdversarialLoss
+    gan.DiscriminatorAdversarialLoss
     gan.GeneratorBCE
     gan.GeneratorLSGAN
     gan.GeneratorL1
+    gan.GeneratorHingeLoss
     gan.FeatureMatchingLoss
     gan.CategoricalCrossEntropy
     gan.Pix2PixLoss
     gan.Pix2PixLossSemantic
     gan.EncoderBCE
     gan.DiscriminatorMinMax
     gan.DiscriminatorLSGAN
+    gan.DiscriminatorHingeLoss
     gan.get_adversarial_loss_discriminator
     gan.get_adversarial_loss_generator
 
@@ -79,8 +81,8 @@
 from ashpy.losses.classifier import ClassifierLoss
 from ashpy.losses.executor import Executor, SumExecutor
 from ashpy.losses.gan import (
-    AdversarialLossD,
-    AdversarialLossG,
+    DiscriminatorAdversarialLoss,
+    GeneratorAdversarialLoss,
     AdversarialLossType,
     CategoricalCrossEntropy,
     DiscriminatorLSGAN,
@@ -98,22 +100,23 @@
 )
 
 __ALL__ = [
-    "AdversarialLossD",
-    "AdversarialLossD",
-    "AdversarialLossG",
+    "DiscriminatorAdversarialLoss",
+    "GeneratorAdversarialLoss",
     "AdversarialLossType",
     "CategoricalCrossEntropy",
     "ClassifierLoss",
     "ClassifierLoss",
     "DiscriminatorLSGAN",
     "DiscriminatorMinMax",
+    "DiscriminatorHingeLoss",
     "EncoderBCE",
     "Executor",
     "FeatureMatchingLoss",
     "GANExecutor",
     "GeneratorBCE",
     "GeneratorL1",
     "GeneratorLSGAN",
+    "GeneratorHingeLoss",
     "get_adversarial_loss_discriminator",
     "get_adversarial_loss_generator",
     "Pix2PixLoss",