Update implementation_notes.md (#112)

Update implementation_notes.md


Co-authored-by: lazaratan <lazar.atanackovic@recursionpharma.com>

docs/implementation_notes.md

@@ -1,6 +1,6 @@
 # Implementation notes
 
-This repo is centered around training GFlowNets that produce graphs. While we intend to specialize towards building molecules, we've tried to keep the implementation moderately agnostic to that fact, which makes it able to support other graph-generation environments.
+This repo is centered around training GFlowNets that produce graphs, although sequences are also supported. While we intend to specialize towards building molecules, we've tried to keep the implementation moderately agnostic to that fact, which makes it able to support other graph-generation environments.
 
 ## Environment, Context, Task, Trainers
 
@@ -33,4 +33,4 @@ The code contains a specific categorical distribution type for graph actions, `G
 
 Consider for example the `AddNode` and `SetEdgeAttr` actions, one applies to nodes and one to edges. An efficient way to produce logits for these actions would be to take the node/edge embeddings and project them (e.g. via an MLP) to a `(n_nodes, n_node_actions)` and `(n_edges, n_edge_actions)` tensor respectively. We thus obtain a list of tensors representing the logits of different actions, but logits are mixed between graphs in the minibatch, so one cannot simply apply a `softmax` operator on the tensor. 
 
-The `GraphActionCategorical` class handles this and can be used to compute various other things, such as entropy, log probabilities, and so on; it can also be used to sample from the distribution.
+The `GraphActionCategorical` class handles this and can be used to compute various other things, such as entropy, log probabilities, and so on; it can also be used to sample from the distribution.