# Add source directory to the path as Python doesn't like sub-directories
source_path = os.path.join("..", "source")
sys.path.insert(0, source_path)
from watson_studio_utils import WatsonStudioUtils
from experiment_utils import Experiment
from project_utils import ProjectUtils
# Initialize various utilities that will make our lives easier
studio_utils = WatsonStudioUtils(region="us-south")
studio_utils.configure_utilities_from_file()
project_utils = ProjectUtils(studio_utils)
# Initialize our experiment
experiment = Experiment("Fashion MNIST-dropout tests",
"Test two different dropout values", "tensorflow",
"1.5", "python", "3.5", studio_utils, project_utils)
# Add two training runs to determine which dropout is best: 0.4 or 0.9
run_1a_path = os.path.join("experiment_zips", "dropout_0.4.zip")
run_1b_path = os.path.join("experiment_zips", "dropout_0.6.zip")
# Specify different GPU types as "k80", "k80x2", "k80x4", "p100", ...
experiment.add_training_run("Run #1", "python3 experiment.py", run_1a_path,
"k80")
experiment.add_training_run("Run #2", "python3 experiment.py", run_1b_path,
"k80")
# Execute experiment
experiment.execute()
search.add_power_range("dense_neurons_1", 6, 11,
2) # 64 128 256 512 1024 2048
search_count = 5
return search.create_random_search(search_count)
# Initialize various utilities that will make our lives easier
studio_utils = WatsonStudioUtils(region="us-south")
studio_utils.configure_utilities_from_file()
project_utils = ProjectUtils(studio_utils)
# Initialize our experiment
experiment = Experiment("Fashion MNIST-Random", "Perform random grid search",
"tensorflow", "1.5", "python", "3.5", studio_utils,
project_utils)
# Create random parameters to search then create a training run for each
search = create_random_search()
experiment_zip = os.path.join("zips", "fashion_mnist_random_search.zip")
for index, run_params in enumerate(search):
run_name = "run_%d" % (index + 1)
experiment.add_training_run(run_name, run_params, "python3 experiment.py",
experiment_zip, "k80")
# Execute experiment
experiment.execute()
# Print the current status of the Experiment.
isPyTorch = False # else TensorFlow
if isPyTorch:
framework = "pytorch"
version = "0.4"
experiment_zip = "dynamic_hyperparms_pt.zip"
else:
framework = "tensorflow"
version = "1.5"
experiment_zip = "dynamic_hyperparms_tf.zip"
# Initialize our experiment
gpu_type = "k80"
experiment = Experiment( "Fashion MNIST-Custom Random-{}-{}".format(framework, gpu_type),
"Perform random grid search",
framework,
version,
"python",
"3.5",
studio_utils,
project_utils)
experiment_zip = os.path.join("experiment_zips", experiment_zip)
# Create random parameters to search then create a training run for each
search = create_random_search()
for index, run_params in enumerate(search):
# Append hyperparameters to the command
command = "python3 experiment.py"
# Specify different GPU types as "k80", "k80x2", "k80x4", "p100", ...
run_name = "run_%d" % (index + 1)
isPyTorch = True # else TensorFlow
if isPyTorch:
framework = "pytorch"
version = "0.4"
experiment_zip = "dynamic_hyperparms_pt.zip"
else:
framework = "tensorflow"
version = "1.5"
experiment_zip = "dynamic_hyperparms_tf.zip"
experiment_zip = os.path.join("experiment_zips", experiment_zip)
# Initialize our experiment
gpu_type = "k80"
experiment = Experiment(
"Fashion MNIST-RBFOpt HPO-{}-{}".format(framework,
gpu_type), "Perform RBFOpt HPO",
framework, version, "python", "3.5", studio_utils, project_utils)
# Run RBFOpt to search through the hyperparameters
rbfopt_config = get_rbfopt_config()
experiment.set_rbfopt_config(rbfopt_config)
# Specify different GPU types as "k80", "k80x2", "k80x4", "p100", ...
experiment.add_training_run("RBFOpt search", "python3 experiment.py",
experiment_zip, gpu_type)
# Execute experiment
experiment.execute()
# Print the current status of the Experiment.
#experiment.print_experiment_summary()
if GET_WEIGHTS == True:
# Get best combination of weights
inertia_weights = [0.5, 0.8, 1]
cognitive_weights = [0.5, 1, 2.05, 4]
social_weights = [0.5, 1, 2.05, 4]
for inertia_weight in inertia_weights:
for cognitive_weight in cognitive_weights:
for social_weight in social_weights:
constants.INERTIA_WEIGHT = inertia_weight
constants.COGNITIVE_WEIGHT = cognitive_weight
constants.SOCIAL_WEIGHT = social_weight
print(inertia_weight, cognitive_weight, social_weight)
# 30 experiments on a fixed population + statistics
experiment = Experiment(griewangk, constants)
max_value, min_value, mean_value = experiment.run_experiment(
30)
print('Min = ', min_value, " Max = ", max_value,
" Mean = ", mean_value)
results.add("griewangk", -1, constants, min_value,
max_value, mean_value)
else:
weight_combinations = [(0.5, 0.50, 4.00), (0.5, 1.00, 4.00)]
for (inertia_weight, cognitive_weight,
social_weight) in weight_combinations:
constants.INERTIA_WEIGHT = inertia_weight
constants.COGNITIVE_WEIGHT = cognitive_weight
constants.SOCIAL_WEIGHT = social_weight
search.add_step_range("dropout_4", 0.1, 0.9, 0.1)
search.add_power_range("dense_neurons_1", 6, 11,
2) # 64 128 256 512 1024 2048
return search
# Initialize various utilities that will make our lives easier
studio_utils = WatsonStudioUtils(region="us-south")
studio_utils.configure_utilities_from_file()
project_utils = ProjectUtils(studio_utils)
# Initialize our experiment
experiment = Experiment("Fashion MNIST-RBFOpt HPO", "Perform RBFOpt HPO",
"tensorflow", "1.5", "python", "3.5", studio_utils,
project_utils)
# Create random parameters to search then create a training run for each
run_count = 25
rbfopt_search = get_rbfopt_search()
# Note: "accuracy" is the objective variable being provided to RBFOpt. For RBFOpt to function properly
# then, we must also store the "accuracy" values to "val_dict_list.json" as shown at the end of the experiment.py file
# in the .zip experiment.
#
# To use a different objective metric, pass the object name here plus update the
# experiment.py file to pass the correct values and corresponding name to the "val_dict_list.json"
#
# Likewise if you want to use iterations instead of epochs, then you must also change that value both here as well
# as in the "val_dict_list.json".