add tweedie post_transform (#242)

add tweedie to lgbm

doc/html/hummingbird/ml/operator_converters/constants.html

@@ -27,7 +27,7 @@ <h1 class="title">Module <code>hummingbird.ml.operator_converters.constants</cod
 <details class="source">
 <summary>
 <span>Expand source code</span>
-<a href="https://github.com/microsoft/hummingbird/blob/master/hummingbird/ml/operator_converters/constants.py#L0-L47" class="git-link">Browse git</a>
+<a href="https://github.com/microsoft/hummingbird/blob/master/hummingbird/ml/operator_converters/constants.py#L0-L50" class="git-link">Browse git</a>
 </summary>
 <pre><code class="python"># -------------------------------------------------------------------------
 # Copyright (c) Microsoft Corporation. All rights reserved.
@@ -54,6 +54,9 @@ <h1 class="title">Module <code>hummingbird.ml.operator_converters.constants</cod
 SOFTMAX = &#34;SOFTMAX&#34;
 &#34;&#34;&#34;Softmax post transform&#34;&#34;&#34;
 
+TWEEDIE = &#34;TWEEDIE&#34;
+&#34;&#34;&#34;Tweedie post transform&#34;&#34;&#34;
+
 GET_PARAMETERS_FOR_TREE_TRAVERSAL = &#34;get_parameters_for_tree_trav&#34;
 &#34;&#34;&#34;Which function to use to extract the parameters for the tree traversal strategy&#34;&#34;&#34;
 
@@ -136,6 +139,10 @@ <h2 class="section-title" id="header-variables">Global variables</h2>
 <dd>
 <div class="desc"><p>The test input data for models that need to be traced</p></div>
 </dd>
+<dt id="hummingbird.ml.operator_converters.constants.TWEEDIE"><code class="name">var <span class="ident">TWEEDIE</span></code></dt>
+<dd>
+<div class="desc"><p>Tweedie post transform</p></div>
+</dd>
 </dl>
 </section>
 <section>
@@ -173,6 +180,7 @@ <h1>Index</h1>
 <li><code><a title="hummingbird.ml.operator_converters.constants.SIGMOID" href="#hummingbird.ml.operator_converters.constants.SIGMOID">SIGMOID</a></code></li>
 <li><code><a title="hummingbird.ml.operator_converters.constants.SOFTMAX" href="#hummingbird.ml.operator_converters.constants.SOFTMAX">SOFTMAX</a></code></li>
 <li><code><a title="hummingbird.ml.operator_converters.constants.TEST_INPUT" href="#hummingbird.ml.operator_converters.constants.TEST_INPUT">TEST_INPUT</a></code></li>
+<li><code><a title="hummingbird.ml.operator_converters.constants.TWEEDIE" href="#hummingbird.ml.operator_converters.constants.TWEEDIE">TWEEDIE</a></code></li>
 </ul>
 </li>
 </ul>

doc/html/hummingbird/ml/operator_converters/onnx/onnx_operator.html

@@ -327,18 +327,10 @@ <h3>Ancestors</h3>
 <h3>Methods</h3>
 <dl>
 <dt id="hummingbird.ml.operator_converters.onnx.onnx_operator.Cast.forward"><code class="name flex">
-<span>def <span class="ident">forward</span></span>(<span>self, x)</span>
+<span>def <span class="ident">forward</span></span>(<span>self, x) -> Callable[..., Any]</span>
 </code></dt>
 <dd>
-<div class="desc"><p>Defines the computation performed at every call.</p>
-<p>Should be overridden by all subclasses.</p>
-<div class="admonition note">
-<p class="admonition-title">Note</p>
-<p>Although the recipe for forward pass needs to be defined within
-this function, one should call the :class:<code>Module</code> instance afterwards
-instead of this since the former takes care of running the
-registered hooks while the latter silently ignores them.</p>
-</div></div>
+<div class="desc"></div>
 <details class="source">
 <summary>
 <span>Expand source code</span>
@@ -377,18 +369,10 @@ <h3>Ancestors</h3>
 <h3>Methods</h3>
 <dl>
 <dt id="hummingbird.ml.operator_converters.onnx.onnx_operator.Concat.forward"><code class="name flex">
-<span>def <span class="ident">forward</span></span>(<span>self, *x)</span>
+<span>def <span class="ident">forward</span></span>(<span>self, *x) -> Callable[..., Any]</span>
 </code></dt>
 <dd>
-<div class="desc"><p>Defines the computation performed at every call.</p>
-<p>Should be overridden by all subclasses.</p>
-<div class="admonition note">
-<p class="admonition-title">Note</p>
-<p>Although the recipe for forward pass needs to be defined within
-this function, one should call the :class:<code>Module</code> instance afterwards
-instead of this since the former takes care of running the
-registered hooks while the latter silently ignores them.</p>
-</div></div>
+<div class="desc"></div>
 <details class="source">
 <summary>
 <span>Expand source code</span>
@@ -429,18 +413,10 @@ <h3>Ancestors</h3>
 <h3>Methods</h3>
 <dl>
 <dt id="hummingbird.ml.operator_converters.onnx.onnx_operator.Reshape.forward"><code class="name flex">
-<span>def <span class="ident">forward</span></span>(<span>self, x)</span>
+<span>def <span class="ident">forward</span></span>(<span>self, x) -> Callable[..., Any]</span>
 </code></dt>
 <dd>
-<div class="desc"><p>Defines the computation performed at every call.</p>
-<p>Should be overridden by all subclasses.</p>
-<div class="admonition note">
-<p class="admonition-title">Note</p>
-<p>Although the recipe for forward pass needs to be defined within
-this function, one should call the :class:<code>Module</code> instance afterwards
-instead of this since the former takes care of running the
-registered hooks while the latter silently ignores them.</p>
-</div></div>
+<div class="desc"></div>
 <details class="source">
 <summary>
 <span>Expand source code</span>

doc/html/hummingbird/ml/operator_converters/sklearn/lightgbm.html

@@ -27,7 +27,7 @@ <h1 class="title">Module <code>hummingbird.ml.operator_converters.sklearn.lightg
 <details class="source">
 <summary>
 <span>Expand source code</span>
-<a href="https://github.com/microsoft/hummingbird/blob/master/hummingbird/ml/operator_converters/sklearn/lightgbm.py#L0-L100" class="git-link">Browse git</a>
+<a href="https://github.com/microsoft/hummingbird/blob/master/hummingbird/ml/operator_converters/sklearn/lightgbm.py#L0-L103" class="git-link">Browse git</a>
 </summary>
 <pre><code class="python"># -------------------------------------------------------------------------
 # Copyright (c) Microsoft Corporation. All rights reserved.
@@ -42,6 +42,7 @@ <h1 class="title">Module <code>hummingbird.ml.operator_converters.sklearn.lightg
 import numpy as np
 from onnxconverter_common.registration import register_converter
 
+from .. import constants
 from .._gbdt_commons import convert_gbdt_classifier_common, convert_gbdt_common
 from .._tree_commons import TreeParameters
 
@@ -122,6 +123,8 @@ <h1 class="title">Module <code>hummingbird.ml.operator_converters.sklearn.lightg
     # Get tree information out of the model.
     n_features = operator.raw_operator._n_features
     tree_infos = operator.raw_operator.booster_.dump_model()[&#34;tree_info&#34;]
+    if operator.raw_operator._objective == &#34;tweedie&#34;:
+        extra_config[constants.POST_TRANSFORM] = constants.TWEEDIE
 
     return convert_gbdt_common(tree_infos, _get_tree_parameters, n_features, extra_config=extra_config)
 
@@ -161,7 +164,7 @@ <h2 id="returns">Returns</h2>
 <details class="source">
 <summary>
 <span>Expand source code</span>
-<a href="https://github.com/microsoft/hummingbird/blob/master/hummingbird/ml/operator_converters/sklearn/lightgbm.py#L55-L74" class="git-link">Browse git</a>
+<a href="https://github.com/microsoft/hummingbird/blob/master/hummingbird/ml/operator_converters/sklearn/lightgbm.py#L56-L75" class="git-link">Browse git</a>
 </summary>
 <pre><code class="python">def convert_sklearn_lgbm_classifier(operator, device, extra_config):
     &#34;&#34;&#34;
@@ -207,7 +210,7 @@ <h2 id="returns">Returns</h2>
 <details class="source">
 <summary>
 <span>Expand source code</span>
-<a href="https://github.com/microsoft/hummingbird/blob/master/hummingbird/ml/operator_converters/sklearn/lightgbm.py#L77-L95" class="git-link">Browse git</a>
+<a href="https://github.com/microsoft/hummingbird/blob/master/hummingbird/ml/operator_converters/sklearn/lightgbm.py#L78-L98" class="git-link">Browse git</a>
 </summary>
 <pre><code class="python">def convert_sklearn_lgbm_regressor(operator, device, extra_config):
     &#34;&#34;&#34;
@@ -226,6 +229,8 @@ <h2 id="returns">Returns</h2>
     # Get tree information out of the model.
     n_features = operator.raw_operator._n_features
     tree_infos = operator.raw_operator.booster_.dump_model()[&#34;tree_info&#34;]
+    if operator.raw_operator._objective == &#34;tweedie&#34;:
+        extra_config[constants.POST_TRANSFORM] = constants.TWEEDIE
 
     return convert_gbdt_common(tree_infos, _get_tree_parameters, n_features, extra_config=extra_config)</code></pre>
 </details>

doc/html/hummingbird/ml/operator_converters/sklearn/skl_sv.html

@@ -304,18 +304,10 @@ <h3>Ancestors</h3>
 <h3>Methods</h3>
 <dl>
 <dt id="hummingbird.ml.operator_converters.sklearn.skl_sv.SVC.forward"><code class="name flex">
-<span>def <span class="ident">forward</span></span>(<span>self, x)</span>
+<span>def <span class="ident">forward</span></span>(<span>self, x) -> Callable[..., Any]</span>
 </code></dt>
 <dd>
-<div class="desc"><p>Defines the computation performed at every call.</p>
-<p>Should be overridden by all subclasses.</p>
-<div class="admonition note">
-<p class="admonition-title">Note</p>
-<p>Although the recipe for forward pass needs to be defined within
-this function, one should call the :class:<code>Module</code> instance afterwards
-instead of this since the former takes care of running the
-registered hooks while the latter silently ignores them.</p>
-</div></div>
+<div class="desc"></div>
 <details class="source">
 <summary>
 <span>Expand source code</span>

hummingbird/ml/operator_converters/_gbdt_commons.py

@@ -124,6 +124,9 @@ def apply_sigmoid(x):
     def apply_softmax(x):
         return torch.softmax(x, dim=1)
 
+    def apply_tweedie(x):
+        return torch.exp(x)
+
     # For models following the Sklearn API we need to build the post transform ourselves.
     if classes is not None and constants.POST_TRANSFORM not in extra_config:
         if len(classes) <= 2:
@@ -143,6 +146,11 @@ def apply_softmax(x):
                 extra_config[constants.POST_TRANSFORM] = lambda x: apply_softmax(apply_base_prediction(base_prediction)(x))
             else:
                 extra_config[constants.POST_TRANSFORM] = apply_softmax
+        elif extra_config[constants.POST_TRANSFORM] == constants.TWEEDIE:
+            if constants.BASE_PREDICTION in extra_config:
+                extra_config[constants.POST_TRANSFORM] = lambda x: apply_tweedie(apply_base_prediction(base_prediction)(x))
+            else:
+                extra_config[constants.POST_TRANSFORM] = apply_tweedie
         else:
             raise NotImplementedError("Post transform {} not implemeneted yet".format(extra_config[constants.POST_TRANSFORM]))
     elif constants.BASE_PREDICTION in extra_config:

hummingbird/ml/operator_converters/constants.py

@@ -23,6 +23,9 @@
 SOFTMAX = "SOFTMAX"
 """Softmax post transform"""
 
+TWEEDIE = "TWEEDIE"
+"""Tweedie post transform"""
+
 GET_PARAMETERS_FOR_TREE_TRAVERSAL = "get_parameters_for_tree_trav"
 """Which function to use to extract the parameters for the tree traversal strategy"""
 

hummingbird/ml/operator_converters/sklearn/lightgbm.py

@@ -11,6 +11,7 @@
 import numpy as np
 from onnxconverter_common.registration import register_converter
 
+from .. import constants
 from .._gbdt_commons import convert_gbdt_classifier_common, convert_gbdt_common
 from .._tree_commons import TreeParameters
 
@@ -91,6 +92,8 @@ def convert_sklearn_lgbm_regressor(operator, device, extra_config):
     # Get tree information out of the model.
     n_features = operator.raw_operator._n_features
     tree_infos = operator.raw_operator.booster_.dump_model()["tree_info"]
+    if operator.raw_operator._objective == "tweedie":
+        extra_config[constants.POST_TRANSFORM] = constants.TWEEDIE
 
     return convert_gbdt_common(tree_infos, _get_tree_parameters, n_features, extra_config=extra_config)
 

tests/test_lightgbm_converter.py

@@ -243,6 +243,23 @@ def test_lgbm_classifier_random_forest(self):
         self.assertIsNotNone(torch_model)
         np.testing.assert_allclose(model.predict_proba(X), torch_model.predict_proba(X), rtol=1e-06, atol=1e-06)
 
+    # Test Tweedie loss in lgbm
+    @unittest.skipIf(not lightgbm_installed(), reason="LightGBM test requires LightGBM installed")
+    def test_lgbm_tweedie(self):
+        warnings.filterwarnings("ignore")
+        model = lgb.LGBMRegressor(objective="tweedie", n_estimators=2, max_depth=5)
+
+        np.random.seed(0)
+        X = np.random.rand(100, 200)
+        X = np.array(X, dtype=np.float32)
+        y = np.random.randint(100, size=100)
+
+        model.fit(X, y)
+
+        torch_model = hummingbird.ml.convert(model, "torch")
+        self.assertIsNotNone(torch_model)
+        np.testing.assert_allclose(model.predict(X), torch_model.predict(X), rtol=1e-06, atol=1e-06)
+
     # Backend tests.
     # Test TorchScript backend regression.
     @unittest.skipIf(not lightgbm_installed(), reason="LightGBM test requires LightGBM installed")