adapt to pytc 0.5

pyproject.toml

@@ -15,7 +15,7 @@ keywords = [
   "functional-programming",
   "machine-learning",
 ]
-dependencies = ["pytreeclass>=0.4.0"]
+dependencies = ["pytreeclass>=0.5.0"]
 
 classifiers = [
   "Development Status :: 5 - Production/Stable",

serket/__init__.py

@@ -18,13 +18,15 @@
     BaseKey,
     Partial,
     TreeClass,
+    autoinit,
     bcmap,
     field,
     fields,
     freeze,
     is_frozen,
     is_nondiff,
     is_tree_equal,
+    leafwise,
     tree_diagram,
     tree_flatten_with_trace,
     tree_graph,
@@ -50,6 +52,7 @@
     "is_tree_equal",
     "field",
     "fields",
+    "autoinit",
     # pprint utils
     "tree_diagram",
     "tree_graph",
@@ -74,11 +77,12 @@
     "tree_flatten_with_trace",
     "tree_repr_with_trace",
     "Partial",
+    "leafwise",
     # serket
     "nn",
     "tree_evaluation",
     "tree_state",
 )
 
 
-__version__ = "0.4.0b1"
+__version__ = "0.5.0b1"

serket/nn/activation.py

@@ -24,6 +24,7 @@
 from serket.nn.utils import IsInstance, Range, ScalarLike
 
 
+@sk.autoinit
 class AdaptiveLeakyReLU(sk.TreeClass):
     """Leaky ReLU activation function
 
@@ -39,6 +40,7 @@ def __call__(self, x: jax.Array, **k) -> jax.Array:
         return jnp.maximum(0, self.a * x) - v * jnp.maximum(0, -self.a * x)
 
 
+@sk.autoinit
 class AdaptiveReLU(sk.TreeClass):
     """ReLU activation function with learnable parameters
     Note:
@@ -51,6 +53,7 @@ def __call__(self, x: jax.Array, **k) -> jax.Array:
         return jnp.maximum(0, self.a * x)
 
 
+@sk.autoinit
 class AdaptiveSigmoid(sk.TreeClass):
     """Sigmoid activation function with learnable `a` parameter
     Note:
@@ -63,6 +66,7 @@ def __call__(self, x: jax.Array, **k) -> jax.Array:
         return 1 / (1 + jnp.exp(-self.a * x))
 
 
+@sk.autoinit
 class AdaptiveTanh(sk.TreeClass):
     """Tanh activation function with learnable parameters
     Note:
@@ -76,6 +80,7 @@ def __call__(self, x: jax.Array, **k) -> jax.Array:
         return (jnp.exp(a * x) - jnp.exp(-a * x)) / (jnp.exp(a * x) + jnp.exp(-a * x))
 
 
+@sk.autoinit
 class CeLU(sk.TreeClass):
     """Celu activation function"""
 
@@ -85,6 +90,7 @@ def __call__(self, x: jax.Array, **k) -> jax.Array:
         return jax.nn.celu(x, alpha=lax.stop_gradient(self.alpha))
 
 
+@sk.autoinit
 class ELU(sk.TreeClass):
     """Exponential linear unit"""
 
@@ -94,6 +100,7 @@ def __call__(self, x: jax.Array, **k) -> jax.Array:
         return jax.nn.elu(x, alpha=lax.stop_gradient(self.alpha))
 
 
+@sk.autoinit
 class GELU(sk.TreeClass):
     """Gaussian error linear unit"""
 
@@ -103,13 +110,15 @@ def __call__(self, x: jax.Array, **k) -> jax.Array:
         return jax.nn.gelu(x, approximate=self.approximate)
 
 
+@sk.autoinit
 class GLU(sk.TreeClass):
     """Gated linear unit"""
 
     def __call__(self, x: jax.Array, **k) -> jax.Array:
         return jax.nn.glu(x)
 
 
+@sk.autoinit
 class HardShrink(sk.TreeClass):
     """Hard shrink activation function"""
 
@@ -120,41 +129,47 @@ def __call__(self, x: jax.Array, **k) -> jax.Array:
         return jnp.where(x > alpha, x, jnp.where(x < -alpha, x, 0.0))
 
 
+@sk.autoinit
 class HardSigmoid(sk.TreeClass):
     """Hard sigmoid activation function"""
 
     def __call__(self, x: jax.Array, **k) -> jax.Array:
         return jax.nn.hard_sigmoid(x)
 
 
+@sk.autoinit
 class HardSwish(sk.TreeClass):
     """Hard swish activation function"""
 
     def __call__(self, x: jax.Array, **k) -> jax.Array:
         return jax.nn.hard_swish(x)
 
 
+@sk.autoinit
 class HardTanh(sk.TreeClass):
     """Hard tanh activation function"""
 
     def __call__(self, x: jax.Array, **k) -> jax.Array:
         return jax.nn.hard_tanh(x)
 
 
+@sk.autoinit
 class LogSigmoid(sk.TreeClass):
     """Log sigmoid activation function"""
 
     def __call__(self, x: jax.Array, **k) -> jax.Array:
         return jax.nn.log_sigmoid(x)
 
 
+@sk.autoinit
 class LogSoftmax(sk.TreeClass):
     """Log softmax activation function"""
 
     def __call__(self, x: jax.Array, **k) -> jax.Array:
         return jax.nn.log_softmax(x)
 
 
+@sk.autoinit
 class LeakyReLU(sk.TreeClass):
     """Leaky ReLU activation function"""
 
@@ -206,6 +221,7 @@ def __call__(self, x: jax.Array, **k) -> jax.Array:
         return x / (1 + jnp.abs(x))
 
 
+@sk.autoinit
 class SoftShrink(sk.TreeClass):
     """SoftShrink activation function"""
 
@@ -248,6 +264,7 @@ def __call__(self, x: jax.Array, **k) -> jax.Array:
         return x - jax.nn.tanh(x)
 
 
+@sk.autoinit
 class ThresholdedReLU(sk.TreeClass):
     """Thresholded ReLU activation function."""
 
@@ -265,6 +282,7 @@ def __call__(self, x: jax.Array, **k) -> jax.Array:
         return x * jax.nn.tanh(jax.nn.softplus(x))
 
 
+@sk.autoinit
 class PReLU(sk.TreeClass):
     """Parametric ReLU activation function"""
 
@@ -274,6 +292,7 @@ def __call__(self, x: jax.Array, **k) -> jax.Array:
         return jnp.where(x >= 0, x, x * self.a)
 
 
+@sk.autoinit
 class Snake(sk.TreeClass):
     """Snake activation function
 

serket/nn/blocks/unet.py

@@ -71,6 +71,7 @@ def __call__(self, x: jax.Array, **k) -> jax.Array:
         return self.conv(x)
 
 
+@sk.autoinit
 class UNetBlock(sk.TreeClass):
     """Vanilla UNet
 

serket/nn/containers.py

@@ -23,6 +23,7 @@
 import serket as sk
 
 
+@sk.autoinit
 class Sequential(sk.TreeClass):
     """A sequential container for layers.
 

serket/nn/contrast.py

@@ -43,6 +43,7 @@ def random_contrast_nd(
     return adjust_contrast_nd(x, contrast_factor)
 
 
+@sk.autoinit
 class AdjustContrastND(sk.TreeClass):
     """Adjusts the contrast of an NDimage by scaling the pixel values by a factor.
 
@@ -71,14 +72,12 @@ class AdjustContrast2D(AdjustContrastND):
         contrast_factor: contrast factor to adjust the image by.
     """
 
-    def __init__(self, contrast_factor=1.0):
-        super().__init__(contrast_factor=contrast_factor)
-
     @property
     def spatial_ndim(self) -> int:
         return 2
 
 
+@sk.autoinit
 class RandomContrastND(sk.TreeClass):
     """Randomly adjusts the contrast of an image by scaling the pixel
     values by a factor.
@@ -131,9 +130,6 @@ class RandomContrast2D(RandomContrastND):
         contrast_range: range of contrast factors to randomly sample from.
     """
 
-    def __init__(self, contrast_range=(0.5, 1)):
-        super().__init__(contrast_range=contrast_range)
-
     @property
     def spatial_ndim(self) -> int:
         return 2