Oparch allows you to do some rudimentary optimization of a Sequential tensorflow model. All you need is a rough idea of a sequential model, data, and some know-how. Due to the lack of documentation, refer to oparch_tests/test_abalone_script or the functions themselves
Note: This is a personal project, and thats what it looks like.
Functions for optimization are in the oparch.optimize.py file Commonly supported ´kwargs´ for all optimization functions are:
return_metric(str) : Which metric to use to evaluate performance, default is value in oparch.configurations. LAST_LOSS usually outperforms others.
return_model"(boolean) : true or false, whether to return the model (true), or to return the tried values and results in a list of tuples (false): [(value,result), (value1,result1)...]
samples(int) : How many samples to use for the evaluation. Defaults to all if not specified or incorrect.
validation_split(float[0...1]) : how many percent of the data to use for validation (only applicable if using an evaluation based on a metric calculated from validation set). Default 0.2.
epochs(int) : How many times to go through the training data. Defaults to 5, however 1 should be used with slow computers or large datasets or models
batch_size(int) : How many data pieces should be considered before calculating gradient. Default 32. If running out of memory, consider decreasing.
verbose(int[0..3]) : What should be reported to stdout. (Not implemented)
decimals(int) : Up to how many decimal points should the evaluation metric be optimized. (Should perhaps be changed to apply to the metric, not to the result). Defaults to 5.
metrics(list(int or float)) : which metrics should be used to compile the model. Defaults to ["accuracy"]
opt_all_layer_params(model,X,y,param,**kwargs)
Loops over all layers and optimizes the given parameter with a global grid search (because often has multiple minimas)
if the layer has that parameter attribute.
Doesn't optimize the last layer.
Args:
model (Sequential): The model which layers are looped through
X (feature data): a numpy array
y (observed output): a numpy arra
param (string): a string identifier
**kwargs :
param : list with values to be tested, if not specified, uses defaults from configurations #ex. [2,4,8,16...]
Returns:
model: model with the optimal layers
example usage:
x = np.array([[0.45, 0.66, 0.77, 0.69], [0.2, 0.44, 0.55, 0.22] ....])
y = [0.5, 0.66 ....] #Could also be categorical
model = [Flatten(),Dense(1),Dense(1)]
#...build and compile model
**model** = opt_all_layer_params(model,x,y,"units")
> Optimizing 'units' for <keras.layers.core.dense.Dense object at 0x00000188BEAE3CD0> at index 1...
> units None LAST_LOSS 20.42995636533
> units 1 LAST_LOSS 23.82626426555
> ....
> Optimizing 'units' for <keras.layers.core.dense.Dense object at 0x00000188BEAE3CD0> at index 2...
> ....
opt_optimizer_parameter(model: tf.keras.models.Sequential,
X: np.ndarray,
y: np.ndarray,
param: list,
**kwargs) -> tf.keras.models.Sequential or list:
Uses a specified (default "TNC") algorithm from the scipy library, to optimize the list of parameters.
If an optimizer doesn't have a specified parameter, it is reported, removed, and then resumed with out the incorrect parameter.
All(?) optimizer parameters are between (0,1), so uses these bounds for the optimization.
It is recommended to first optimize the model structure, to get better results
Args:
model (Sequential) : model
X (array) :
y (array) :
param (list) : list of the optimizer hyperparameters to tune
**kwargs :
maxiter(int) : maximum number of iterations to do to find the optimal combination. Defaults to 50.
algo(str) : Must be one of: "TNC","L-BFGS-B" or "SLSQP", Defaults to "TNC"
Returns:
model: model with the optimal layers
example usage:
x = np.array([[0.45, 0.66, 0.77, 0.69], [0.2, 0.44, 0.55, 0.22] ....])
y = [0.5, 0.66 ....] #Could also be categorical
model = [Flatten(),Dense(1),Dense(1)]
#...build and compile model
**model** = opt_optimizer_parameter(X,y,model,["decay", "learning_rate","rho"])
> decay 0.0
> learning_rate 0.001
> rho 0.9
> LAST_LOSS 8.624521894573
>....
> Finding minima took 28 iterations.
> Best found combination:
> decay 0.01353
> learning_rate 0.00211
> rho 0.88014
> LAST_LOSS 7.3256785847764