def __init__(self,
dataset,
train_dataset,
cls,
detname='perfect',
detector_config=None):
self.dataset = dataset
self.train_dataset = train_dataset
self.cls = cls
self.cls_ind = dataset.get_ind(cls)
self.detname = detname
self.tt = ut.TicToc()
# Check if configs exist and look up the correct config for this detname and cls
# TODO: this is inefficient because this file is re-opened for every class
loaded_detector_config = None
fname = opjoin(config.get_dets_configs_dir(train_dataset),
detname + '.txt')
if opexists(fname):
with open(fname) as f:
configs = json.load(f)
config_name = detname + '_' + cls
if config_name in configs:
loaded_detector_config = configs[config_name]
self.config = loaded_detector_config
if not self.config:
self.config = Detector.DEFAULT_CONFIG
if detector_config:
self.config.update(detector_config)
def output_det_statistics(self):
"""
Collect samples and display the statistics of times and naive and
actual_ap increases for each class, on the train dataset.
"""
det_configs = {}
all_dets,all_clses,all_samples = self.run_on_dataset(train=True,force=True)
if comm_rank==0:
cols = ['action_ind','dt','det_naive_ap','det_actual_ap','img_ind']
sample_array = np.array([[getattr(s,col) for s in all_samples] for col in cols]).T
table = Table(sample_array,cols)
# go through actions
for ind,action in enumerate(self.actions):
st = table.filter_on_column('action_ind', ind)
means = np.mean(st.arr[:,1:],0)
det_configs[self.actions[ind].name] = {}
# TODO: should depend on image size
det_configs[self.actions[ind].name]['avg_time'] = means[0]
det_configs[self.actions[ind].name]['naive_ap'] = means[1]
det_configs[self.actions[ind].name]['actual_ap'] = means[2]
if isinstance(action.obj,Detector):
img_inds = st.subset_arr('img_ind').astype(int)
cls_ind = action.obj.cls_ind
d = self.train_dataset
presence_inds = np.array([d.images[img_ind].contains_cls_ind(cls_ind) for img_ind in img_inds])
st_present = np.atleast_2d(st.arr[np.flatnonzero(presence_inds),:])
if st_present.shape[0]>0:
means = np.mean(st_present[:,2:],0)
det_configs[self.actions[ind].name]['naive_ap|present'] = means[0]
det_configs[self.actions[ind].name]['actual_ap|present'] = means[1]
else:
# this is just for testing, where there may not be enough dets
det_configs[self.actions[ind].name]['naive_ap|present'] = 0
det_configs[self.actions[ind].name]['actual_ap|present'] = 0
st_absent = np.atleast_2d(st.arr[np.flatnonzero(presence_inds==False),:])
if st_absent.shape[0]>0:
means = np.mean(st_absent[:,2:],0)
# naive_ap|absent is always 0
det_configs[self.actions[ind].name]['actual_ap|absent'] = means[1]
else:
det_configs[self.actions[ind].name]['actual_ap|absent'] = 0
detectors = '-'.join(self.detectors)
filename = opjoin(
config.get_dets_configs_dir(self.train_dataset), detectors+'.txt')
with open(filename,'w') as f:
json.dump(det_configs,f)
safebarrier(comm)
det_configs = comm.bcast(det_configs,root=0)
return det_configs
def __init__(self, dataset, train_dataset, cls, detname='perfect', detector_config=None):
self.dataset = dataset
self.train_dataset = train_dataset
self.cls = cls
self.cls_ind = dataset.get_ind(cls)
self.detname = detname
self.tt = ut.TicToc()
# Check if configs exist and look up the correct config for this detname and cls
# TODO: this is inefficient because this file is re-opened for every class
loaded_detector_config = None
fname = opjoin(config.get_dets_configs_dir(train_dataset),detname+'.txt')
if opexists(fname):
with open(fname) as f:
configs = json.load(f)
config_name = detname+'_'+cls
if config_name in configs:
loaded_detector_config = configs[config_name]
self.config = loaded_detector_config
if not self.config:
self.config = Detector.DEFAULT_CONFIG
if detector_config:
self.config.update(detector_config)
def output_det_statistics(self):
"""
Collect samples and display the statistics of times and naive and
actual_ap increases for each class, on the train dataset.
"""
det_configs = {}
all_dets, all_clses, all_samples = self.run_on_dataset(train=True,
force=True)
if comm_rank == 0:
cols = [
'action_ind', 'dt', 'det_naive_ap', 'det_actual_ap', 'img_ind'
]
sample_array = np.array([[getattr(s, col) for s in all_samples]
for col in cols]).T
table = Table(sample_array, cols)
# go through actions
for ind, action in enumerate(self.actions):
st = table.filter_on_column('action_ind', ind)
means = np.mean(st.arr[:, 1:], 0)
det_configs[self.actions[ind].name] = {}
# TODO: should depend on image size
det_configs[self.actions[ind].name]['avg_time'] = means[0]
det_configs[self.actions[ind].name]['naive_ap'] = means[1]
det_configs[self.actions[ind].name]['actual_ap'] = means[2]
if isinstance(action.obj, Detector):
img_inds = st.subset_arr('img_ind').astype(int)
cls_ind = action.obj.cls_ind
d = self.train_dataset
presence_inds = np.array([
d.images[img_ind].contains_cls_ind(cls_ind)
for img_ind in img_inds
])
st_present = np.atleast_2d(
st.arr[np.flatnonzero(presence_inds), :])
if st_present.shape[0] > 0:
means = np.mean(st_present[:, 2:], 0)
det_configs[self.actions[ind].
name]['naive_ap|present'] = means[0]
det_configs[self.actions[ind].
name]['actual_ap|present'] = means[1]
else:
# this is just for testing, where there may not be enough dets
det_configs[
self.actions[ind].name]['naive_ap|present'] = 0
det_configs[
self.actions[ind].name]['actual_ap|present'] = 0
st_absent = np.atleast_2d(
st.arr[np.flatnonzero(presence_inds == False), :])
if st_absent.shape[0] > 0:
means = np.mean(st_absent[:, 2:], 0)
# naive_ap|absent is always 0
det_configs[self.actions[ind].
name]['actual_ap|absent'] = means[1]
else:
det_configs[
self.actions[ind].name]['actual_ap|absent'] = 0
detectors = '-'.join(self.detectors)
filename = opjoin(config.get_dets_configs_dir(self.train_dataset),
detectors + '.txt')
with open(filename, 'w') as f:
json.dump(det_configs, f)
safebarrier(comm)
det_configs = comm.bcast(det_configs, root=0)
return det_configs