using_explore_trace.txt

An outline for interactively inspecting what the model is doing:

0. train and produce model checkpoints
    run_40k_gatedpushpop_normalize_no_op.sh
1. run the trained model at an epoch of choice and log activations
    trace_40k_gatedpushpop_normalize_no_op.sh
2. interactively select and visualize model activations for train/test examples
    explore_trace.sh

Some additional tips for using explore_trace.sh:

After explore_trace.sh is run, you will see an interactive prompt, we try
to emulate a folder structure for grouping the sets of examples 
in our data. 

For example, you can type 'ls' to see the current top level folders, and 'cd' to
select one. At the top level there is train, validation, and test. The next level
is divided into bins based on whether an equality is predicted correctly by our model.
For instance, to see correctly predicted examples, you can do 'cd (True,)', 
and do 'cd (False,)' for incorrect predictions. 'cd ..' will get you to the previous
level.

A bin is the lowest level folder, once you are in a particular bin, all examples 
are essentially in a flat list. The initial printed statistics include the count
for the bin and a percentage N_bin/N_set, (e.g. correct_train/all_train).
You can select a random batch with the sel command,
e.g. 'sel mode=random n=10' to select 10 random examples from the bin. 
Finally, typing the 'plot' command will create visualizations of the equation tree
and save it to the 'visualizations' directory. 

In the visualization documents, the 'mem' rows of a node represents the stack, 
the 'act' rows represent push/pop  action (or if the particular version of the 
model has no-op action, then it will be in the 3rd row). 
'rep' row of a node represents the output of the node.

You can change the default parameters (--load-epoch and --trace-path in 
trace_40k_gatedpushpop_normalize_no_op.sh) and (--trace-files in 
trace_40k_gatedpushpop_normalize_no_op.sh) to explore the model activations for a 
different epoch.