Defines all state transitions, populates q_values where actions are valid Legal Transitions: conv -> conv, pool

Returns new image size given previous image size and filter parameters

Creates a single LMDB file. Path is the full path of filename to store

Update a single Q-Value for start_state given the state we transitioned to and the reward.

Takes a parsed model string and returns a recursive list of states.

Return a list of possible FC sizes given the current state

Reset the state for a new random walk

Have Q-Learning agent sample current policy to generate a network

Deletes uneeded log files and model saves: args: replay - with standard replay dic columns. Deletes only model saves that aren't

If root_path is None, we download the data set from internet. Either save path or root path must not be None and not both. Returns

load single batch of cifar

Epsilon defaults to the minimum

X and y must be the same length vectors

start_state: Should be the actual start_state, not a bucketed state action: valid action returns: next state, not bucketed

Convert the list of strings to a string we can train from according to the grammar

Returns the string asociated with state.

Updates self.state according to an epsilon-greedy strategy

How to define reward from accuracy

Add drop out every 2 layers and after each fully connected layer Sets dropout rate to be between 0 and 0.5 at a linear rate

Returns True if the experiment is complete

Splits Xtr in validation and train sets. Also saves a full train lmdb. full train and split train are both first shuffled before being saved

Order replay by iteration, then remove duplicate nets keeping the first

Update all Q-Values for a sequence.