def run(run_id: int,
database: Database,
complete_only: bool = True,
threshold: np.float = -1,
threshold_percentile: np.float = -1,
random_seed: int = randint(1, 9999999)):
""" run clustering and returns object """
def preprocess(features: np.array):
# todo: preprocess!
preprocessed_features = features.astype(np.float)
return preprocessed_features
def fetch_threshold(df: pd.DataFrame,
percentile: np.float,
num_iter: int = 100,
num_sample: int = 1000):
seed(random_seed) # to make things reproducible
if df.shape[0] < num_sample:
num_sample = df.shape[0]
num_iter = 1
threshold = np.array([
np.percentile(
pairwise_distances(df.sample(num_sample,
random_state=randint(
0, 999999)).values,
metric='euclidean',
n_jobs=-1), percentile)
for i in range(num_iter)
]).mean()
return threshold
# set properties
properties = {
"run_id": run_id,
"random_seed": random_seed,
"method": "birch"
}
# prepare features_df
if complete_only:
selector = " AND bgc.on_contig_edge is 0"
else:
selector = ""
bgc_ids = [
row[0]
for row in database.select("bgc,run_bgc_status",
"WHERE run_bgc_status.run_id=?" +
" AND run_bgc_status.bgc_id=bgc.id" +
selector,
parameters=(run_id, ),
props=["bgc.id"],
as_tuples=True)
]
if len(bgc_ids) < 1: # check if no bgc_ids
raise Exception("Not enough input for clustering.")
hmm_ids = [
row[0] for row in database.select("hmm,run",
"WHERE hmm.db_id=run.hmm_db_id" +
" AND run.id=?",
parameters=(run_id, ),
props=["hmm.id"],
as_tuples=True)
]
features_df = pd.DataFrame(np.zeros((len(bgc_ids), len(hmm_ids)),
dtype=np.uint8),
index=bgc_ids,
columns=hmm_ids)
# fetch feature values from db
for bgc_id, hmm_id, value in database.select(
"bgc,run_bgc_status,run,hmm,bgc_features",
"WHERE run_bgc_status.bgc_id=bgc.id" + " AND run.id=?" +
" AND run_bgc_status.bgc_id=bgc.id" +
" AND run.id=run_bgc_status.run_id" +
" AND run.hmm_db_id=hmm.db_id" +
" AND bgc_features.bgc_id=bgc.id" +
" AND bgc_features.hmm_id=hmm.id" + selector,
parameters=(run_id, ),
props=[
"bgc_features.bgc_id", "bgc_features.hmm_id",
"bgc_features.value"
],
as_tuples=True):
features_df.at[bgc_id, hmm_id] = value
# initiate birch object
birch = Birch(
n_clusters=None, # no global clustering
compute_labels=False, # only calc centroids
copy=False # data already copied
)
# set threshold
if threshold >= 0:
birch.threshold = threshold
else:
if threshold_percentile < 0:
raise Exception("Threshold percentile can't be < 0.00")
# set threshold based on sampling of features
birch.threshold = fetch_threshold(features_df,
threshold_percentile)
properties["threshold"] = birch.threshold
# set flat birch
birch.branching_factor = features_df.shape[0]
# call birch
birch.fit(preprocess(features_df.values))
# save centroids
properties["centroids"] = pd.DataFrame(np.uint8(
birch.subcluster_centers_),
columns=features_df.columns)
return BirchClustering(properties)
