Training parameters with 1 hiddenlayer cannot be used as calculator

[Muammar El Khatib]

Let's suppose one trains an artificial neural network with a topology of 20 nodes in just 1 hidden layer. The model was set as follows:

model=NeuralNetwork(hiddenlayers=(n), fortran=True)

After optimization, if one tries to use the .amp file, the following error raises:

Traceback (most recent call last):
  File "HeCu.py", line 42, in <module>
    slab.get_potential_energy()
  File "/users/melkhati/git/ase/ase/atoms.py", line 682, in get_potential_energy
    energy = self._calc.get_potential_energy(self)
  File "/users/melkhati/git/ase/ase/calculators/calculator.py", line 422, in get_potential_energy
    energy = self.get_property('energy', atoms)
  File "/users/melkhati/git/ase/ase/calculators/calculator.py", line 461, in get_property
    self.calculate(atoms, [name], system_changes)
  File "/users/melkhati/git/amp/amp/__init__.py", line 259, in calculate
    self.descriptor.fingerprints[key])
  File "/users/melkhati/git/amp/amp/model/__init__.py", line 66, in calculate_energy
    symbol=symbol)
  File "/users/melkhati/git/amp/amp/model/neuralnetwork.py", line 363, in calculate_atomic_energy
    outputs = calculate_nodal_outputs(self.parameters, afp, symbol,)
  File "/users/melkhati/git/amp/amp/model/neuralnetwork.py", line 634, in calculate_nodal_outputs
    for hiddenlayer in hiddenlayers[1:]:
TypeError: 'int' object has no attribute '__getitem__'

The issue is related to the expected data type of hiddenlayers being a tuple, but setting one hidden layer returns an int. I solved the problem with this patch:

 diff --git a/amp/model/neuralnetwork.py b/amp/model/neuralnetwork.py
index 9e7fbc5..29b258e 100644
--- a/amp/model/neuralnetwork.py
+++ b/amp/model/neuralnetwork.py
@@ -631,6 +631,9 @@ def calculate_nodal_outputs(parameters, afp, symbol,):
         temp[0, _] = o[1][0, _]
     temp[0, np.shape(o[1])[1]] = 1.0
     ohat[1] = temp
+    if isinstance(hiddenlayers, int):
+        hiddenlayers = (hiddenlayers,)
+
     for hiddenlayer in hiddenlayers[1:]:
         layer += 1
         net[layer] = np.dot(ohat[layer - 1], weight[layer])
@@ -703,6 +706,9 @@ def calculate_dOutputs_dInputs(parameters, derafp, outputs, nsymbol,):
     dOutputs_dInputs = {}  # node values
     dOutputs_dInputs[0] = _derafp
     layer = 0  # input layer
+    if isinstance(hiddenlayers, int):
+        hiddenlayers = (hiddenlayers,)
+
     for hiddenlayer in hiddenlayers[0:]:
         layer += 1
         temp = np.dot(np.matrix(dOutputs_dInputs[layer - 1]),

Another way to avoid this (not tested) would be to set the topology of the neural network like this

model=NeuralNetwork(hiddenlayers=(n,), fortran=True)

However, it seems more natural to set hiddenlayers=(n).

Is it worth to apply that patch into main?.

[andrew_peterson]

I think that is not a bug since (20) is not a tuple but (20,) is...

[Muammar El Khatib]

You are right, this is not a bug. It may be more a usability issue because, from a user's point of view, it might be more natural to set the case of just one hidden layer with n nodes as hiddenlayer=(n) unless they know about one-element tuples. I always found that one element-tuples in Python are confusing (I will use lists in my example):

>>> a = [1]
>>> type(a)
<type 'list'>
>>> b = (1)
>>> type(b)
<type 'int'>
>>> 

Enclosing one element inside [ ] creates a list. But as Python uses ( ) for grouping, actually commas inside ( ) are those creating the tuple:

>>> c = (1,)
>>> type(c)
<type 'tuple'>

Looking at neuralnetwork.py L17 the description of the parameter is only dict. I think we should add tuple and dictionary to give more hints to users.

I have updated the severity and type of report.

[andrew_peterson]

Actually, it's the comma that makes the tuple in python.

>>> c = 1,
>>> type(c)
<type 'tuple'>

But please update the documentation so it makes sense...

[Muammar El Khatib]

Updated docstring for hiddenlayer dictionary in neuralnetwork.py. Closes #136.

→ <<cset fb84510eccd6>>

[Muammar El Khatib]

Merge branch 'master' into issue66

• master: Documentation example script correction. Updated docstring for hiddenlayer dictionary in neuralnetwork.py. Closes #136. This commit fixes the following error when using tflow module together with newer numpy versions:

→ <<cset 90e70358bf86>>

[andrew_peterson]

Merge branch 'master' into issue66

• master: Documentation example script correction. Updated docstring for hiddenlayer dictionary in neuralnetwork.py. Closes #136. This commit fixes the following error when using tflow module together with newer numpy versions:

→ <<cset 90e70358bf86>>

[Muammar El Khatib]

Merge branch 'master' into symmetryfunctions

• master: Updated docstring for hiddenlayer dictionary in neuralnetwork.py. Closes #136.

→ <<cset 7387a988f176>>

[Esben Leonhard Kolsbjerg]

I had the same problem. I would agree with Muammar that it is more user friendly to catch the mistake and correct it.