Use python for loop even for tensorflow in FastMCInference because of…

… random seed is fixed issue

demo_tutorial/NN_uncertainty_analysis/README.md

@@ -12,7 +12,7 @@ Uncertainty with Dropout Variational Inference Demonstration for Regression and
 astroNN was used to do a regression task with Dropout VI in a paper 
 **Deep learning of multi-element abundances from high-resolution spectroscopic data** 
 which the code available at [https://github.com/henrysky/astroNN_spectra_paper_figures](https://github.com/henrysky/astroNN_spectra_paper_figures)
-and the paper available at [[arxiv:1808.04428](https://arxiv.org/abs/1808.04428)][[ADS](https://ui.adsabs.harvard.edu/#abs/2018arXiv180804428L/)]. 
+and the paper available at [[arxiv:1808.04428](https://ui.adsabs.harvard.edu/abs/2019MNRAS.483.3255L/abstract)][[ADS](https://ui.adsabs.harvard.edu/#abs/2018arXiv180804428L/)]. 
 We demonstrated Dropout VI can report reasonable uncertainty with high prediction 
 accuracy trained on incomplete stellar parameters and abundances data from from high-resolution stellar spectroscopic data.
 
@@ -39,7 +39,7 @@ For Dropout variational inference, related material:
 * Paper: [Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning](https://arxiv.org/abs/1506.02142)
 * Paper: [What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?](https://arxiv.org/abs/1703.04977)
 * Paper: [Bayesian Convolutional Neural Networks with Bernoulli Approximate Variational Inference](https://arxiv.org/abs/1506.02158)
-* Yarin Gal's Blog: [What My Deep Model Doesn't Know...](https://mlg.eng.cam.ac.uk/yarin/blog_3d801aa532c1ce.html)
+* Yarin Gal's Blog: [What My Deep Model Doesn't Know...](https://www.cs.ox.ac.uk/people/yarin.gal/website/blog_3d801aa532c1ce.html)
 * [Demo from Yarin Gal written in javascript](https://github.com/yaringal/HeteroscedasticDropoutUncertainty)
 
 <br>

src/astroNN/nn/layers.py

@@ -343,16 +343,22 @@ def loop_fn(i):
             return self.layer(inputs)
 
         # vectorizing operation depends on backend
-        # tensorflow vectorized_map traced operation so there is no randomness which affects e.g., dropout
-        if keras.backend.backend() == "tensorflow":
-            # outputs = backend_framework.vectorized_map(loop_fn, self.arange_n)
-            outputs = backend_framework.map_fn(loop_fn, self.arange_n)
-        elif keras.backend.backend() == "torch":
+        # TODO: tensorflow vectorized_map traced operation so there is no randomness which affects e.g., dropout
+        # if keras.backend.backend() == "tensorflow":
+        #     outputs = backend_framework.vectorized_map(loop_fn, self.arange_n)
+        if keras.backend.backend() == "torch":
             outputs = backend_framework.vmap(
                 loop_fn, randomness="different", in_dims=0
             )(self.arange_n)
         else:  # fallback to simple for loop
-            outputs = keras.ops.concatenate([self.layer(inputs) for _ in self.arange_n])
+            outputs = [self.layer(inputs) for _ in range(self.n)]
+            if isinstance(outputs[0], dict):
+                outputs = {
+                    key: keras.ops.stack([output[key] for output in outputs])
+                    for key in outputs[0].keys()
+                }
+            else:
+                outputs = keras.ops.stack(outputs)
         return outputs  # outputs can be tensor or dict of tensors
 
 

tests/test_layers.py

@@ -214,7 +214,7 @@ def test_FastMCInference(random_data):
     # ======== Simple Keras sequential Model, one input one output ======== #
     smodel = keras.models.Sequential()
     smodel.add(keras.layers.Input(shape=(7514,)))
-    smodel.add(keras.layers.Dense(32, input_shape=(7514,)))
+    smodel.add(keras.layers.Dense(32))
     smodel.add(keras.layers.Dense(10, activation="softmax"))
     smodel.compile(
         optimizer="rmsprop", loss="categorical_crossentropy", metrics=["accuracy"]