上QQ阅读APP看书，第一时间看更新

Visualizing training

Since we've written log data from both the models in Chapter 2, Using Deep Learning to Solve Regression Problems, we can use TensorBoard to compare the two models graphically. Open up TensorBoard and head to the SCALARS tab. You should see something like this. You may need to click loss and val_loss to expand the graphs:

Tensorboard displaying the loss and val_loss plots for the model

If you look at the bottom-left corner of the screen, you should notice that each directory we created has a run associated with it. Both are currently selected. This means that on our graphs, we will see output for both models.

TensorBoard can accommodate many, many runs, and you can filter them via a regular expression (for example ^dnn* would show all runs that start with dnn). This means that if you're searching for the best model through many experiments or runs (such as hyperparameter optimization), you can quickly navigate them if you explicitly and consistently name your runs and include meaningful hyperparameter and architecture information in the name, so do that!

The default X scale on these graphs is epochs. The Y value is the loss function we chose, which was MAE. You can click on the graphs to explore them and drag to zoom.

Seeing the graphs like this, we can really see the relative bias and variance of each network. While there is a good separation between the models in train loss, the deep neural network only gets marginally better on the validation set, suggesting that we've headed into overfitting territory.