Add option to pass parameters to probabilistic layers (#64)

* Increase a bit the margins of the distribution * Fix missing value for mean * increase spacing around limits to 0.05 * Update to tensorflow2.16 and tf probability0.24 (#65) * Updated to tensorflow2.16 and tf probability0.24, relaxed version constraints for various libraries * update for warnings cleanup * Removed lambda functions from serializable models to remove safety warnings * Test loading probabilistic model from mlflow --------- Co-authored-by: Daniele Nerini <[email protected]> * Update packages with security alerts (#66) * Fix formatting (#67) * (WIP) Refactor a bit how we handle probabilistic layers * (WIP) Prepare change in how we handle proba layers * Small bug fix * Fix missing import * Move to another script to avoid circular import * Move to another script to avoid circular imports * Add missing import * Bug fix * Move calculation of n_params to get_proba_layer * Fix logger name * Update models.py * Modify arguments for doubly censored * Remove leftovers --------- Co-authored-by: Louis Poulain--Auzéau <[email protected]> Co-authored-by: Michele Cattaneo <[email protected]> Co-authored-by: Daniele Nerini <[email protected]>
MeteoSwiss · Nov 14, 2024 · e6bce31 · e6bce31
1 parent dc1bc21
commit e6bce31
Show file tree

Hide file tree

Showing 2 changed files with 75 additions and 24 deletions.
diff --git a/mlpp_lib/models.py b/mlpp_lib/models.py
@@ -1,5 +1,5 @@
 import logging
-from typing import Optional, Union, Any
+from typing import Optional, Union, Any, Callable
 
 import numpy as np
 import tensorflow as tf
@@ -13,7 +13,7 @@
 from tensorflow.keras import Model, initializers
 
 from mlpp_lib.physical_layers import *
-from mlpp_lib.probabilistic_layers import *
+from mlpp_lib import probabilistic_layers
 
 try:
     import tcn  # type: ignore
@@ -32,6 +32,33 @@ def call(self, inputs):
         return super().call(inputs, training=True)
 
 
+def get_probabilistic_layer(
+    output_size,
+    probabilistic_layer: Union[str, dict]
+) -> Callable:
+    """Get the probabilistic layer."""
+
+    if isinstance(probabilistic_layer, dict):
+        probabilistic_layer_name = list(probabilistic_layer.keys())[0]
+        probabilistic_layer_options = probabilistic_layer[probabilistic_layer_name]
+    else:
+        probabilistic_layer_name = probabilistic_layer
+        probabilistic_layer_options = {}
+
+    if hasattr(probabilistic_layers, probabilistic_layer_name):
+        _LOGGER.info(f"Using custom probabilistic layer: {probabilistic_layer_name}")
+        probabilistic_layer_obj = getattr(probabilistic_layers, probabilistic_layer_name)
+        n_params = getattr(probabilistic_layers, probabilistic_layer_name).params_size(output_size)
+        probabilistic_layer = (
+            probabilistic_layer_obj(output_size, name="output", **probabilistic_layer_options) if isinstance(probabilistic_layer_obj, type) 
+            else probabilistic_layer_obj(output_size, name="output")
+        )
+    else:
+        raise KeyError(f"The probabilistic layer {probabilistic_layer_name} is not available.")
+
+    return probabilistic_layer, n_params
+
+
 def _build_fcn_block(
     inputs,
     hidden_layers,
@@ -67,16 +94,15 @@ def _build_fcn_block(
 def _build_fcn_output(x, output_size, probabilistic_layer, out_bias_init):
     # probabilistic prediction
     if probabilistic_layer:
-        probabilistic_layer = globals()[probabilistic_layer]
-        n_params = probabilistic_layer.params_size(output_size)
+        probabilistic_layer, n_params = get_probabilistic_layer(output_size, probabilistic_layer)
         if isinstance(out_bias_init, np.ndarray):
             out_bias_init = np.hstack(
                 [out_bias_init, [0.0] * (n_params - out_bias_init.shape[0])]
             )
             out_bias_init = initializers.Constant(out_bias_init)
 
         x = Dense(n_params, bias_initializer=out_bias_init, name="dist_params")(x)
-        outputs = probabilistic_layer(output_size, name="output")(x)
+        outputs = probabilistic_layer(x)
 
     # deterministic prediction
     else:
@@ -247,7 +273,7 @@ def fully_connected_multibranch_network(
         )
 
     if probabilistic_layer:
-        n_params = globals()[probabilistic_layer].params_size(output_size)
+        _, n_params = get_probabilistic_layer(output_size, probabilistic_layer)
         n_branches = n_params
     else:
         n_branches = output_size
@@ -379,16 +405,15 @@ def deep_cross_network(
 
     # probabilistic prediction
     if probabilistic_layer:
-        probabilistic_layer = globals()[probabilistic_layer]
-        n_params = probabilistic_layer.params_size(output_size)
+        probabilistic_layer, n_params = get_probabilistic_layer(output_size, probabilistic_layer)
         if isinstance(out_bias_init, np.ndarray):
             out_bias_init = np.hstack(
                 [out_bias_init, [0.0] * (n_params - out_bias_init.shape[0])]
             )
             out_bias_init = initializers.Constant(out_bias_init)
 
         x = Dense(n_params, bias_initializer=out_bias_init, name="dist_params")(merge)
-        outputs = probabilistic_layer(output_size, name="output")(x)
+        outputs = probabilistic_layer(x)
 
     # deterministic prediction
     else:

diff --git a/mlpp_lib/probabilistic_layers.py b/mlpp_lib/probabilistic_layers.py
@@ -307,9 +307,18 @@ def __init__(
         # distribution function to `DistributionLambda.__init__` below as the first
         # positional argument.
         kwargs.pop("make_distribution_fn", None)
+        # get the clipping parameters and pop them
+        _clip_low = kwargs.pop("clip_low", 0.0)
+        _clip_high = kwargs.pop("clip_high", 1.0)
 
         def new_from_t(t):
-            return IndependentDoublyCensoredNormal.new(t, event_shape, validate_args)
+            return IndependentDoublyCensoredNormal.new(
+                t,
+                event_shape,
+                validate_args,
+                clip_low=_clip_low,
+                clip_high=_clip_high,
+            )
 
         super(IndependentDoublyCensoredNormal, self).__init__(
             new_from_t, convert_to_tensor_fn, **kwargs
@@ -320,7 +329,14 @@ def new_from_t(t):
         self._validate_args = validate_args
 
     @staticmethod
-    def new(params, event_shape=(), validate_args=False, name=None):
+    def new(
+        params,
+        event_shape=(),
+        validate_args=False,
+        name=None,
+        clip_low=0.0,
+        clip_high=1.0,
+    ):
         """Create the distribution instance from a `params` vector."""
         with tf.name_scope(name or "IndependentDoublyCensoredNormal"):
             params = tf.convert_to_tensor(params, name="params")
@@ -343,47 +359,55 @@ def new(params, event_shape=(), validate_args=False, name=None):
             normal_dist = tfd.Normal(loc=loc, scale=scale, validate_args=validate_args)
 
             class CustomCensored(tfd.Distribution):
-                def __init__(self, normal):
+                def __init__(self, normal, clip_low=0.0, clip_high=1.0):
                     self.normal = normal
                     super(CustomCensored, self).__init__(
                         dtype=normal.dtype,
                         reparameterization_type=tfd.FULLY_REPARAMETERIZED,
                         validate_args=validate_args,
                         allow_nan_stats=True,
                     )
+                    self.clip_low = clip_low
+                    self.clip_high = clip_high
 
                 def _sample_n(self, n, seed=None):
 
                     # Sample from normal distribution
                     samples = self.normal.sample(sample_shape=(n,), seed=seed)
 
                     # Clip values between 0 and 1
-                    chosen_samples = tf.clip_by_value(samples, 0, 1)
+                    chosen_samples = tf.clip_by_value(
+                        samples, self.clip_low, self.clip_high
+                    )
 
                     return chosen_samples
 
                 def _mean(self):
                     """
                     Original: X ~ N(mu, sigma)
-                    Censored: Y = X if 0 <= X <= 1 else 0 if X < 0 else 1
+                    Censored: Y = X if clip_low <= X <= clip_high else clip_low if X < clip_low else clip_high
                     Phi / phi: CDF / PDF of standard normal distribution
 
                     Law of total expectations:
-                        E[Y] = E[Y | X > 1] * P(X > 1) + E[Y | X < 0] * P(X < 0) + E[Y | 0 <= X <= 1] * P(0 <= X <= 1)
-                             = 1 * P(X > 1) + P(X < 0) * 0 + E[X | 0 <= X <= 1] * P(0 <= X <= 1)
-                             = 1 * P(X > 1) + E[Z ~ TruncNormal(mu, sigma, 0, 1)] * (Phi((1 - mu) / sigma) - Phi(-mu / sigma))
-                             = 1 * (1 - Phi((1 - mu) / sigma)) + mu * (Phi((1 - mu) / sigma) - Phi(-mu / sigma)) + sigma * (phi(-mu / sigma) - phi((1 - mu) / sigma))
+                        E[Y] = E[Y | X > c_h] * P(X > c_h) + E[Y | X < c_l] * P(X < c_l) + E[Y | c_l <= X <= c_h] * P(c_l <= X <= c_h)
+                             = c_h * P(X > c_h) + P(X < c_l) * c_l + E[Y | c_l <= X <= c_h] * P(c_l <= X <= c_h)
+                             = c_h * P(X > c_h) + P(X < c_l) * c_l + E[Z ~ TruncNormal(mu, sigma, c_l, c_h)] * (Phi((c_h - mu) / sigma) - Phi(c_l - mu / sigma))
+                             = c_h * (1 - Phi((c_h - mu) / sigma))
+                                + c_l * Phi((c_l - mu) / sigma)
+                                + mu * (Phi((c_h - mu) / sigma) - Phi(c_l - mu / sigma))
+                                + sigma * (phi(c_l - mu / sigma) - phi((c_h - mu) / sigma))
                     Ref for TruncatedNormal mean: https://en.wikipedia.org/wiki/Truncated_normal_distribution
                     """
                     mu, sigma = self.normal.mean(), self.normal.stddev()
-                    low_bound_standard = (0 - mu) / sigma
-                    high_bound_standard = (1 - mu) / sigma
+                    low_bound_standard = (self.clip_low - mu) / sigma
+                    high_bound_standard = (self.clip_high - mu) / sigma
 
                     cdf = lambda x: tfd.Normal(0, 1).cdf(x)
                     pdf = lambda x: tfd.Normal(0, 1).prob(x)
 
                     return (
-                        1 * (1 - cdf(high_bound_standard))
+                        self.clip_high * (1 - cdf(high_bound_standard))
+                        + self.clip_low * cdf(low_bound_standard)
                         + mu * (cdf(high_bound_standard) - cdf(low_bound_standard))
                         + sigma * (pdf(low_bound_standard) - pdf(high_bound_standard))
                     )
@@ -394,10 +418,12 @@ def _log_prob(self, value):
                     cdf = lambda x: tfd.Normal(0, 1).cdf(x)
                     pdf = lambda x: tfd.Normal(0, 1).prob(x)
 
-                    logprob_left = lambda x: tf.math.log(cdf(-mu / sigma) + 1e-3)
+                    logprob_left = lambda x: tf.math.log(
+                        cdf(self.clip_low - mu / sigma) + 1e-3
+                    )
                     logprob_middle = lambda x: self.normal.log_prob(x)
                     logprob_right = lambda x: tf.math.log(
-                        1 - cdf((1 - mu) / sigma) + 1e-3
+                        1 - cdf((self.clip_high - mu) / sigma) + 1e-3
                     )
 
                     return (
@@ -407,7 +433,7 @@ def _log_prob(self, value):
                     )
 
             return independent_lib.Independent(
-                CustomCensored(normal_dist),
+                CustomCensored(normal_dist, clip_low=clip_low, clip_high=clip_high),
                 reinterpreted_batch_ndims=tf.size(event_shape),
                 validate_args=validate_args,
             )