Add possibility to use custom padding (address issue #458) (#489)

* add possibility to use custom padding

* Add padding_dict to actual function

* Simplify arguments and fix docs

* typo

torchani/data/__init__.py

@@ -35,7 +35,11 @@ Available transformations are listed below:
 - `shuffle`
 - `cache` cache the result of previous transformations.
 - `collate` pad the dataset, convert it to tensor, and stack them
-    together to get a batch.
+    together to get a batch. `collate` uses a default padding dictionary
+    ``{'species': -1, 'coordinates': 0.0, 'forces': 0.0, 'energies': 0.0}`` for
+    padding, but a custom padding dictionary can be passed as an optional
+    parameter, which overrides this default padding.
+
 - `pin_memory` copy the tensor to pinned memory so that later transfer
     to cuda could be faster.
 
@@ -94,8 +98,8 @@ if PKBAR_INSTALLED:
 
 verbose = True
 
-
 PROPERTIES = ('energies',)
+
 PADDING = {
     'species': -1,
     'coordinates': 0.0,
@@ -104,8 +108,11 @@ PADDING = {
 }
 
 
-def collate_fn(samples):
-    return utils.stack_with_padding(samples, PADDING)
+def collate_fn(samples, padding=None):
+    if padding is None:
+        padding = PADDING
+
+    return utils.stack_with_padding(samples, padding)
 
 
 class IterableAdapter:
@@ -241,8 +248,8 @@ class Transformations:
         return ret
 
     @staticmethod
-    def collate(reenterable_iterable, batch_size):
-        def reenterable_iterable_factory():
+    def collate(reenterable_iterable, batch_size, padding=None):
+        def reenterable_iterable_factory(padding=None):
             batch = []
             i = 0
             for d in reenterable_iterable:
@@ -250,10 +257,13 @@ class Transformations:
                 i += 1
                 if i == batch_size:
                     i = 0
-                    yield collate_fn(batch)
+                    yield collate_fn(batch, padding)
                     batch = []
             if len(batch) > 0:
-                yield collate_fn(batch)
+                yield collate_fn(batch, padding)
+
+        reenterable_iterable_factory = functools.partial(reenterable_iterable_factory,
+                                                         padding)
         try:
             length = (len(reenterable_iterable) + batch_size - 1) // batch_size
             return IterableAdapterWithLength(reenterable_iterable_factory, length)