def computeSubSet(metadata, frame_per_seq, probTh=None, seed=42):
"""This function computes a subset of the dataset, it extracts frame_per_seq consecutive frames for each sequence.
Parameters
----------
metadata: Pandas.DataFrame
metadata of the WilFireDataset
frame_per_seq: int
frame per sequence to take
probTh: float , optional
The data set contains many more frames classified 'fire' than 'not fire', this parameter
allows to equalize the dataset. For each 'not fire' sequence, we draw a random number if
it is greater than probTh we double the number of frames used for this sequence
seed : int
you can setup a seed for repeatability
Example
-------
metadataSS = computeSubSet(metadata, 2)
wildfireSS = WildFireDataset(metadata=metadataSS, path_to_frames=path_to_frames)
"""
if not isinstance(metadata, pd.DataFrame):
try:
metadata = pd.read_csv(metadata)
except (ValueError, FileNotFoundError):
raise ValueError(f"Invalid path to CSV containing metadata. Please provide one (path={metadata})")
cryptogen = SystemRandom()
cryptogen.seed(seed)
random.seed(seed)
metadata.index = np.arange(len(metadata))
imgs = metadata['imgFile']
# Define sequences numbers
metadata.index = np.arange(len(metadata))
meta = metadata[['exploitable', 'fire', 'sequence', 'clf_confidence', 'loc_confidence',
'x', 'y', 't', 'stateStart', 'stateEnd', 'fire_id', 'fBase']]
meta = meta.drop_duplicates()
meta['seq'] = np.arange(len(meta))
metadata = pd.merge(metadata, meta, on=['exploitable', 'fire', 'sequence', 'clf_confidence',
'loc_confidence', 'x', 'y', 't', 'stateStart', 'stateEnd',
'fire_id', 'fBase'], how='inner')
# Get unique list of sequences
seq = metadata['seq']
my_set = set(seq)
uniqueSEQ = list(my_set)
random.shuffle(uniqueSEQ)
subSetImgs = []
subSetImgsEq = []
for seU in uniqueSEQ:
# For each sequence get a subSample of frame_per_seq frames
nn = [imgs[i] for i, se in enumerate(seq) if se == seU]
if(len(nn) > frame_per_seq):
nn = random.sample(nn, frame_per_seq)
nb = [float(frame.split("frame", 1)[1].split(".", 1)[0]) for frame in nn]
nb, nn = (list(t) for t in zip(*sorted(zip(nb, nn))))
subSetImgs += nn
# Equalize the dataset adding not_fire frames
if probTh is not None:
if(metadata[metadata['seq'] == seU]['fire'].values[0] == 0 and
cryptogen.random() < probTh):
nn = [imgs[i] for i, se in enumerate(seq) if se == seU]
if(len(nn) > frame_per_seq):
nn = random.sample(nn, frame_per_seq)
nb = [float(frame.split("frame", 1)[1].split(".", 1)[0]) for frame in nn]
nb, nn = (list(t) for t in zip(*sorted(zip(nb, nn))))
subSetImgsEq += nn
# Insert randomly the extra frames in the dataset
if probTh is not None:
for i in range(0, len(subSetImgsEq), 2):
idx = cryptogen.randint(0, len(subSetImgs) - 2) // 2 * 2
subSetImgs.insert(idx, subSetImgsEq[i + 1])
subSetImgs.insert(idx, subSetImgsEq[i])
# Create metadta Subset
index = [i for i, im in enumerate(metadata['imgFile'].values) if im in subSetImgs]
return metadata.iloc[index]
