Relax tensorflow version limit

Author: kbattocchi

econml/iv/nnet/_deepiv.py

@@ -89,9 +89,13 @@ def mog_loss_model(n_components, d_t):
     # LL = C - log(sum(pi_i/sig^d * exp(-d2/(2*sig^2))))
     # Use logsumexp for numeric stability:
     # LL = C - log(sum(exp(-d2/(2*sig^2) + log(pi_i/sig^d))))
-    # TODO: does the numeric stability actually make any difference?
     def make_logloss(d2, sig, pi):
-        return -K.logsumexp(-d2 / (2 * K.square(sig)) + K.log(pi / K.pow(sig, d_t)), axis=-1)
+        # logsumexp doesn't exist in keras 2.4; simulate it
+        values = - d2 / (2 * K.square(sig)) + K.log(pi / K.pow(sig, d_t))
+        # logsumexp(a,b,c) = log(exp(a)+exp(b)+exp(c)) = log((exp(a-k)+exp(b-k)+exp(c-k))*exp(k))
+        # = log((exp(a-k)+exp(b-k)+exp(c-k))) + k
+        mx = K.max(values, axis=-1)
+        return -K.log(K.sum(K.exp(values - L.Reshape((-1, 1))(mx)), axis=-1)) - mx
 
     ll = L.Lambda(lambda dsp: make_logloss(*dsp), output_shape=(1,))([d2, sig, pi])
 

econml/tests/test_deepiv.py

@@ -34,6 +34,26 @@ def test_stop_grad(self):
         model.compile('nadam')
         model.fit([np.array([[1]]), np.array([[2]]), np.array([[0]])], [])
 
+    def test_mog_loss(self):
+        inputs = [keras.layers.Input(shape=s) for s in [(3,), (3, 2), (3,), (2,)]]
+        ll_model = keras.engine.Model(inputs, mog_loss_model(3, 2)(inputs))
+
+        for n in range(10):
+            ps = -np.log(np.random.uniform(size=(3,)))
+            pi = ps / np.sum(ps)
+            mu = np.random.normal(size=(3, 2))
+            sig = np.exp(np.random.normal(size=3,))
+            t = np.random.normal(size=(2,))
+
+            pred = ll_model.predict([pi.reshape(1, 3), mu.reshape(1, 3, 2), sig.reshape(1, 3), t.reshape(1, 2)])
+
+            # LL = C - log(sum(pi_i/sig^d * exp(-d2/(2*sig^2))))
+            d = mu - t.reshape(-1, 2)
+            d2 = np.sum(d * d, axis=-1)
+            ll = -np.log(np.sum(pi / (sig * sig) * np.exp(-d2 / (2 * sig * sig)), axis=0))
+
+            assert np.allclose(ll, pred[0])
+
     @pytest.mark.slow
     def test_deepiv_shape(self):
         fit_opts = {"epochs": 2}

setup.cfg

@@ -43,9 +43,9 @@ install_requires =
     numpy
     scipy > 1.4.0
     scikit-learn >= 0.24
-    keras < 2.4
+    keras
     sparse
-    tensorflow > 1.10, < 2.3
+    tensorflow > 1.10
     joblib >= 0.13.0
     numba != 0.42.1
     statsmodels >= 0.9