implement MAPE for pytorch

Author: enigne

pinnicle/utils/data_misfit.py

@@ -13,6 +13,11 @@ def mean_squared_log_error_tf(y_true, y_pred):
     """ use tensorflow function to compute mean squared log error
     """
     return tf.keras.losses.MeanSquaredLogarithmicError()(y_true, y_pred)
+
+def mean_absolute_percentage_error_tf(y_true, y_pred):
+    """Calculates the Mean Absolute Percentage Error (MAPE) in PyTorch.
+    """
+    return tf.keras.losses.MAPE()(y_true, y_pred)
 #}}}
 # ---- jax {{{
 def surface_log_vel_misfit_jax(v_true, v_pred):
@@ -25,6 +30,13 @@ def mean_squared_log_error_jax(y_true, y_pred):
     """ use jax/numpy function to compute mean squared log error
     """
     return bkd.reduce_mean(bkd.square(jax.numpy.log(y_true+1.0) - jax.numpy.log(y_pred+1.0)))
+def mean_absolute_percentage_error_jax(y_true, y_pred):
+    """Calculates the Mean Absolute Percentage Error (MAPE) in PyTorch.
+    """
+    # Ensure y_true is not zero to avoid division by zero
+    # Add a small epsilon to the denominator to prevent NaN values
+    epsvel=2.220446049250313e-16
+    return bkd.reduce_mean(jax.numpy.abs((y_true - y_pred) / (y_true + epsvel))) * 100
 #}}}
 # ---- pytorch {{{
 def surface_log_vel_misfit_pytorch(v_true, v_pred):
@@ -37,24 +49,35 @@ def mean_squared_log_error_pytorch(y_true, y_pred):
     """ use jax/numpy function to compute mean squared log error
     """
     return bkd.reduce_mean(bkd.square(torch.log(y_true+1.0) - torch.log(y_pred+1.0)))
+
+def mean_absolute_percentage_error_pytorch(y_true, y_pred):
+    """Calculates the Mean Absolute Percentage Error (MAPE) in PyTorch.
+    """
+    # Ensure y_true is not zero to avoid division by zero
+    # Add a small epsilon to the denominator to prevent NaN values
+    epsvel=2.220446049250313e-16
+    return bkd.reduce_mean(bkd.abs((y_true - y_pred) / (y_true + epsvel))) * 100
 #}}}
 # ---------------
 def loss_dict_tf():
     return {
             "VEL_LOG": surface_log_vel_misfit_tf,
-            "MEAN_SQUARE_LOG": mean_squared_log_error_tf
+            "MEAN_SQUARE_LOG": mean_squared_log_error_tf,
+            "MAPE":mean_absolute_percentage_error_tf,
             }
 
 def loss_dict_jax():
     return {
             "VEL_LOG": surface_log_vel_misfit_jax,
-            "MEAN_SQUARE_LOG": mean_squared_log_error_jax
+            "MEAN_SQUARE_LOG": mean_squared_log_error_jax,
+            "MAPE":mean_absolute_percentage_error_jax,
             }
 
 def loss_dict_pytorch():
     return {
             "VEL_LOG": surface_log_vel_misfit_pytorch,
-            "MEAN_SQUARE_LOG": mean_squared_log_error_pytorch
+            "MEAN_SQUARE_LOG": mean_squared_log_error_pytorch,
+            "MAPE":mean_absolute_percentage_error_pytorch,
             }
 
 if backend_name ==  "tensorflow":

tests/test_utils.py

@@ -18,15 +18,17 @@ def test_save_and_load_dict(tmp_path):
 def test_data_misfit():
     with pytest.raises(Exception):
         data_misfit.get("not defined")
-    dde_loss = ["mean absolute error", "MAE", "mae", "mean squared error", "mse", "mean absolute percentage error", "MAPE", "mape", "mean l2 relative error", "softmax cross entropy", "zero"]
+    dde_loss = ["mean absolute error", "MAE", "mae", "mean squared error", "mse", "mean l2 relative error", "softmax cross entropy", "zero"]
     for l in dde_loss:
         assert data_misfit.get(l) == l
 
 def test_data_misfit_functions():
     assert data_misfit.get("VEL_LOG") != None
     assert data_misfit.get("MEAN_SQUARE_LOG") != None
+    assert data_misfit.get("MAPE") != None
     assert data_misfit.get("VEL_LOG")(backend.as_tensor([1.0]),backend.as_tensor([1.0])) == 0.0
     assert data_misfit.get("MEAN_SQUARE_LOG")(backend.as_tensor([1.0]),backend.as_tensor([1.0])) == 0.0
+    assert data_misfit.get("MAPE")(backend.as_tensor([1.0]),backend.as_tensor([1.0])) == 0.0
 
 def test_loadmat():
     filename = "flightTracks.mat"