def fit_hyperparameters(file,
train,
train_labels,
cuda,
gpu,
save_memory=False):
"""
Creates a classifier from the given set of hyperparameters in the input
file, fits it and return it.
@param file Path of a file containing a set of hyperparemeters.
@param train Training set.
@param train_labels Labels for the training set.
@param cuda If True, enables computations on the GPU.
@param gpu GPU to use if CUDA is enabled.
@param save_memory If True, save GPU memory by propagating gradients after
each loss term, instead of doing it after computing the whole loss.
"""
classifier = scikit_wrappers.CausalCNNEncoderClassifier()
# Loads a given set of hyperparameters and fits a model with those
hf = open(os.path.join(file), 'r')
params = json.load(hf)
hf.close()
# Check the number of input channels
params['in_channels'] = numpy.shape(train)[1]
params['cuda'] = cuda
params['gpu'] = gpu
classifier.set_params(**params)
return classifier.fit(train,
train_labels,
save_memory=save_memory,
verbose=True)
def run(dataset, gpu, log_name, model):
path = '/localscratch/shared/ts/UCRArchive_2018'
save_path = '/localscratch/shared/ts/exp'
cuda = True
hyper = 'default_hyperparameters.json'
load = False
fit_classifier = False
train, train_labels, test, test_labels = load_UCR_dataset(path, dataset)
if not load and not fit_classifier:
classifier = fit_hyperparameters(hyper, train, train_labels, cuda, gpu, model)
else:
classifier = scikit_wrappers.CausalCNNEncoderClassifier(model=model)
hf = open(os.path.join(save_path, dataset + '_hyperparameters.json'), 'r')
hp_dict = json.load(hf)
hf.close()
hp_dict['cuda'] = cuda
hp_dict['gpu'] = gpu
classifier.set_params(**hp_dict)
classifier.load(os.path.join(save_path, dataset))
if not load:
if fit_classifier:
classifier.fit_classifier(classifier.encode(train), train_labels)
classifier.save(os.path.join(save_path, dataset))
with open(os.path.join(save_path, dataset + '_hyperparameters.json'), 'w') as fp:
json.dump(classifier.get_params(), fp)
print_and_log("Test accuracy: " + str(classifier.score(test, test_labels)), log_name)
def load_classifier(save_path, dataset, cuda, gpu):
"""
Loads and returns classifier from the given parameters.
@param save_path Path where the model is located.
@param dataset Name of the dataset.
@param cuda If True, enables computations on the GPU.
@param gpu GPU to use if CUDA is enabled.
"""
classifier = scikit_wrappers.CausalCNNEncoderClassifier()
hf = open(os.path.join(save_path, dataset + '_hyperparameters.json'), 'r')
hp_dict = json.load(hf)
hf.close()
hp_dict['cuda'] = cuda
hp_dict['gpu'] = gpu
classifier.set_params(**hp_dict)
classifier.load(os.path.join(save_path, dataset))
return classifier
help='activate to use CUDA')
parser.add_argument('--gpu',
type=int,
default=0,
metavar='GPU',
help='index of GPU used for computations (default: 0)')
return parser.parse_args()
if __name__ == '__main__':
args = parse_arguments()
if args.cuda and not torch.cuda.is_available():
print("CUDA is not available, proceeding without it...")
args.cuda = False
classifier = scikit_wrappers.CausalCNNEncoderClassifier()
hf = open(
os.path.join(args.save_path, args.dataset + '_hyperparameters.json'),
'r')
hp_dict = json.load(hf)
hf.close()
hp_dict['cuda'] = args.cuda
hp_dict['gpu'] = args.gpu
classifier.set_params(**hp_dict)
classifier.load(os.path.join(args.save_path, args.dataset))
print("Classification tasks...")
# List of folders / datasets in the given path
datasets = [x[0][len(args.path) + 1:] for x in os.walk(args.path)][1:]
