In this IPython notebook, we will cover more advanced aspects of the `TensorGraph` framework. In particular, we will demonstrate how to share weights between layers and show how to use `DataBag` to reduce the amount of overhead needed to train complex `TensorGraph` models.
Let's now apply a nonlinear transformation to both `left_features` and `right_features`. We can use the `Dense` layer to do so. In addition, let's make sure that we apply the same nonlinear transformation to both `left_features` and `right_features`. To this, we can use the `Layer.shared()`. We use this method by initializing a first `Dense` layer, and then calling the `Layer.shared()` method to make a copy of that layer.
Let's now combine these two transformed feature layers by addition. We will assume this network is being used to solve a regression problem, so we will introduce a `Label` that stores the true regression values. We can then define the objective function of the network via the `L2Loss` between the added output and the true label.
Let's now randomly sample an artificial dataset we can use to train this architecture. We will need to sample the `left_features`, `right_features`, and `labels` in order to be able to train the network.
%% Cell type:code id: tags:
``` python
importnumpyasnp
importnumpy.random
n_samples=100
sampled_left_features=np.random.rand(100,75)
sampled_right_features=np.random.rand(100,75)
sampled_labels=np.random.rand(75,1)
```
%% Cell type:markdown id: tags:
How can we train `TensorGraph` networks with multiple `Feature` inputs? One option is to manually construct a python generator that provides inputs. The tutorial notebook on graph convolutions does this explicitly. For simpler cases, we can use the convenience object `DataBag` which makes it easier to construct generators. A `DataBag` holds multiple datasets (added via `DataBag.add_dataset`). The method `DataBag.iterbatches()` will construct a generator that peels off batches of the desired size from each dataset and return a dictionary mapping inputs (`Feature`, `Label`, and `Weight` objects) to data for that minibatch. Let's see `DataBag` in action.
Note that we will need to wrap our sampled Numpy arrays with `NumpyDataset` objects for our call to work.
