def main(argv):
# Don't bother using the GPU for this
os.environ["CUDA_VISIBLE_DEVICES"] = ""
# ACM doesn't like Type 3 fonts
# https://tex.stackexchange.com/q/18687
plt.rc('pdf', fonttype=42)
plt.rc('ps', fonttype=42)
# We only want to plot this one dataset
list_of_datasets = ["wisdm_ar", "ucihar", "uwave", "ucihhar"]
dataset_names_nice = ["WISDM AR", "UCI HAR", "uWave", "UCI HHAR"]
# Get only the ones used in the SS-DA experiments
dataset_which = [
[1, 3, 4, 2, 25, 7, 21, 2, 1, 0, 11, 15, 25, 29, 30, 31, 32, 7, 8],
# [2, 7, 12, 12, 9, 14, 18, 6, 7, 17, 11, 13, 16, 18, 18, 19, 23, 24, 25],
# [2, 3, 4, 2, 1, 2, 3, 1, 4, 7, 5, 6, 7, 8],
# [1, 3, 4, 0, 1, 4, 5, 2, 3, 5, 3, 5, 6, 7, 8],
# Just do for the first two adaptation problems in SS-DA experiments
[2, 11, 7, 13],
[2, 5, 3, 5],
[1, 3, 3, 5],
]
ymaxs = [70, 35, 16, 28]
# We mostly care about WISDM AR and don't have enough space for all of them
first_ns = [None, None, None, None]
for i in range(len(list_of_datasets)):
dataset_name = list_of_datasets[i]
dataset_name_nice = dataset_names_nice[i]
which = dataset_which[i]
ymax = ymaxs[i]
first_n = first_ns[i]
# Get class balance for all users
user_source_pairs = []
for user in datasets.get_dataset_users(dataset_name):
# Note: train_on_everything=True means the training dataset consists
# of all train/valid/test data.
sources, _ = load_da(dataset_name, str(user), "",
train_on_everything=True)
# We load them one at a time
assert len(sources) == 1
source = sources[0]
user_source_pairs.append((user, source))
balance_data = compute_class_balances(dataset_name, user_source_pairs)
# Plot it
class_labels = datasets.get_dataset(dataset_name).class_labels
generate_plot(dataset_name_nice, class_labels, balance_data,
filename="class_balance_" + dataset_name + ".pdf",
which=which, ymax=ymax, first_n=first_n)
def main(argv):
counts = collections.defaultdict(dict)
for dataset_name in datasets.list_datasets():
# Note: test=False so we only look at the training samples, which is what
# we will vary in the vary-amount-of-target-data experiments
for user in datasets.get_dataset_users(dataset_name):
sources, _ = load_da(dataset_name, str(user), "", test=False)
train_count = count_training_samples(sources)
counts[dataset_name][user] = train_count
print_dictionary(counts, "dataset_target_training_sample_counts")
def main(argv):
# Don't bother using the GPU for this
os.environ["CUDA_VISIBLE_DEVICES"] = ""
for dataset_name in datasets.list_datasets():
for user in datasets.get_dataset_users(dataset_name):
# Note: test=False so we only look at the training samples, where
# train=80% of training set, test=20% of training set, i.e. the
# validation set
test = False
sources, _ = load_da(dataset_name, str(user), "", test=test)
assert len(sources) == 1
dataset = sources[0]
print_stats(dataset_name + "_" + str(user), dataset, test=test)
def main(argv):
# Don't bother using the GPU for this
os.environ["CUDA_VISIBLE_DEVICES"] = ""
for dataset_name in datasets.list_datasets():
user_source_pairs = []
for user in datasets.get_dataset_users(dataset_name):
# Note: test=False so we only look at the training samples, where
# train=80% of training set, test=20% of training set, i.e. the
# validation set
sources, _ = load_da(dataset_name, str(user), "", test=False)
# We load them one at a time
assert len(sources) == 1
source = sources[0]
user_source_pairs.append((user, source))
print_class_balances(dataset_name, user_source_pairs)
pairs.append((dataset_name, str(source_user), str(target_user)))
return pairs
if __name__ == "__main__":
# Sources-target pairs for training
pairs = []
uids = []
for name in datasets.list_datasets():
# Tune on "watch_noother" not "watch"
if name == "watch":
continue
users = datasets.get_dataset_users(name)
# Since sources-target aren't stored in filename anymore (too long), we
# would run into folder name conflicts if we didn't append a unique ID
# to each sources-target pair
uid = 0
# Make this repeatable
random.seed(42)
# Allows extra max_users for some datasets without changin uid's
#
# TODO get rid of all this confusing code once we decide what number
# to set max_users to. If we don't need to change max_users, then
# we can just increment uid's like before.
bonus_uid = 0