Add 'reapply' combinator, for iterated CNN

spacy/_ml.py

@@ -271,6 +271,28 @@ def Tok2Vec(width, embed_size, pretrained_dims=0, **kwargs):
     return tok2vec
 
 
+def reapply(layer, n_times):
+    def reapply_fwd(X, drop=0.):
+        backprops = []
+        for i in range(n_times):
+            Y, backprop = layer.begin_update(X, drop=drop)
+            X = Y
+            backprops.append(backprop)
+        def reapply_bwd(dY, sgd=None):
+            dX = None
+            for backprop in reversed(backprops):
+                dY = backprop(dY, sgd=sgd)
+                if dX is None:
+                    dX = dY
+                else:
+                    dX += dY
+            return dX
+        return Y, reapply_bwd
+    return wrap(reapply_fwd, layer)
+
+
+
+
 def asarray(ops, dtype):
     def forward(X, drop=0.):
         return ops.asarray(X, dtype=dtype), None