def set_class_weights(self, labels, imagesMeta):
if self.wp != -1:
return
print(' Using class-specific weights!')
stats, counts = utils.getLabelStats(imagesMeta, labels)
p = counts / sum(counts)
p = np.array(p)
# print(p)
# self.wp = p
wp = 1 / (p)
self.wp = wp
def load_data(self):
cfg = basic_config(self.newDir)
if self.method == 'faster':
cfg.faster_rcnn_config()
elif self.method == 'fast':
cfg.fast_rcnn_config()
elif self.method == 'normal':
cfg.rcnn_config()
cfg = mcfg.rcnn_hoi_classes(cfg)
cfg = set_config(cfg)
cfg.get_args()
cfg.dataset = 'HICO'
cfg.update_paths()
trainMeta = utils.load_dict(cfg.data_path + 'train')
testMeta = utils.load_dict(cfg.data_path + 'test')
trainGTMeta = utils.load_dict(cfg.data_path + 'train_GT')
testGTMeta = utils.load_dict(cfg.data_path + 'test_GT')
labels = utils.load_dict(cfg.data_path + 'labels')
class_mapping = utils.load_dict(cfg.data_path + 'class_mapping')
if cfg.max_classes is not None:
# Reduce data to include only max_classes number of different classes
_, counts = utils.getLabelStats(trainGTMeta, labels)
reduced_idxs = utils.getReducedIdxs(counts, cfg.max_classes,
labels)
trainGTMeta = utils.reduceData(trainGTMeta, reduced_idxs)
testGTMeta = utils.reduceData(testGTMeta, reduced_idxs)
trainMeta = utils.reduceData(trainMeta, reduced_idxs)
testMeta = utils.reduceData(testMeta, reduced_idxs)
labels = utils.idxs2labels(reduced_idxs, labels)
cfg.nb_classes = len(labels)
cfg.set_class_weights(labels, trainGTMeta)
_, valMeta = utils.splitData(list(trainMeta.keys()), trainMeta)
self.cfg = cfg
if cfg.move:
self.move_data()
print('Data:', cfg.data_path)
print('Path:', cfg.my_results_path)
self.labels = labels
self.class_mapping = class_mapping
self.trainMeta = trainMeta
self.valMeta = valMeta
self.testMeta = testMeta
self.trainGTMeta = trainGTMeta
self.testGTMeta = testGTMeta
def __init__(self, imagesMeta, GTMeta, labels, cfg, data_type='train'):
'Initialization'
self.data_type = data_type
if data_type == 'train':
g_cfg = cfg.train_cfg
elif data_type == 'val':
g_cfg = cfg.val_cfg
else:
g_cfg = cfg.test_cfg
self.gen_type = g_cfg.type
self.batch_size = g_cfg.batch_size
self.nb_batches = g_cfg.nb_batches
self.images_per_batch = g_cfg.images_per_batch
self.nb_samples_per_image = int(self.batch_size /
self.images_per_batch)
self.shuffle = g_cfg.shuffle
self.inputs = cfg.inputs
cfg.img_out_reduction = (16, 16)
self.data_path = cfg.data_path
self.images_path = self.data_path + 'images/'
self.images_path = self.images_path + 'test/' if self.data_type == 'test' else self.images_path + 'train/'
self.nb_classes = cfg.nb_classes
self.cfg = cfg
self.dataID = list(imagesMeta.keys())
self.dataID.sort()
if self.shuffle:
r.shuffle(self.dataID)
else:
self.dataID.sort()
self.imagesMeta = imagesMeta
self.GTMeta = GTMeta
# self.gt_label, _, _ = image.getYData(self.dataID, self.imagesMeta, self.GTMeta, self.cfg)
self.nb_images = len(self.dataID)
# self.nb_samples = len(self.gt_label)
stats, _ = utils.getLabelStats(imagesMeta, labels)
self.nb_samples = stats['totalx']
if self.nb_batches is None:
self.nb_batches = self.nb_images
def load_data(self):
cfg = basic_config(self.newDir)
cfg = set_config(cfg)
cfg.get_args()
cfg.update_paths()
trainMeta = utils.load_dict(cfg.data_path + 'train')
testMeta = utils.load_dict(cfg.data_path + 'test')
trainGTMeta = utils.load_dict(cfg.data_path + 'train_GT')
testGTMeta = utils.load_dict(cfg.data_path + 'test_GT')
labels = utils.load_dict(cfg.data_path + 'labels')
if cfg.max_classes is not None:
# Reduce data to include only max_classes number of different classes
_, counts = utils.getLabelStats(trainGTMeta, labels)
trainGTMeta, reduced_idxs = utils.reduceTrainData(
trainGTMeta, counts, cfg.max_classes)
testGTMeta = utils.reduceTestData(testGTMeta, reduced_idxs)
trainMeta = utils.reduceTestData(trainMeta, reduced_idxs)
testMeta = utils.reduceTestData(testMeta, reduced_idxs)
labels = utils.idxs2labels(reduced_idxs, labels)
cfg.nb_classes = len(labels)
cfg.set_class_weights(labels, trainGTMeta)
_, valMeta = utils.splitData(list(trainMeta.keys()), trainMeta)
self.cfg = cfg
if cfg.move:
self.move_data()
print('Path:', cfg.my_results_path)
self.labels = labels
self.trainMeta = trainMeta
self.valMeta = valMeta
self.testMeta = testMeta
self.trainGTMeta = trainGTMeta
self.testGTMeta = testGTMeta
# Create batch generators
genTrain = DataGenerator(imagesMeta=data.trainMeta,
GTMeta=data.trainGTMeta,
cfg=cfg,
data_type='train')
genVal = DataGenerator(imagesMeta=data.valMeta,
GTMeta=data.trainGTMeta,
cfg=cfg,
data_type='val')
genTest = DataGenerator(imagesMeta=data.testMeta,
GTMeta=data.testGTMeta,
cfg=cfg,
data_type='test')
stats, count = utils.getLabelStats(data.trainMeta, data.labels)
if True:
# Save config
utils.saveConfig(cfg)
utils.saveSplit(cfg, list(data.trainMeta.keys()),
list(data.valMeta.keys()))
# Create model
print('Creating model...')
model = HO_RCNN(cfg)
trainer = model_trainer(model=model,
genTrain=genTrain,
genVal=genVal,
genTest=genTest,
if False:
# Save labels in file
annotations = hico.getUniqueLabels(cfg)
labels = []
for annot in annotations:
obj = annot.nname
pred = annot.vname
pred_ing = annot.vname_ing
label = {'obj': obj, 'pred': pred, 'pred_ing': pred_ing}
labels.append(label)
utils.save_dict(labels, cfg.part_data_path + 'HICO_labels')
labels = utils.load_dict(cfg.part_data_path + 'HICO_labels')
stats, counts = utils.getLabelStats(trainMeta, labels)
reduced_trainMeta, reduced_idxs = utils.reduceTrainData(
trainMeta, counts, 25)
reduced_testMeta = utils.reduceTestData(testMeta, reduced_idxs)
reduced_labels = utils.idxs2labels(reduced_idxs, labels)
reduced_stats, reduced_counts = utils.getLabelStats(
reduced_trainMeta, reduced_labels)
if False:
# Test if all images can be loaded and cropped successfully
i = 0
c = 0.0
end = len(testMeta)
for imageID, metaData in testMeta.items():
oldPath = cfg.part_data_path + 'HICO_images/test/' + metaData['imageID']
image = cv.imread(oldPath)
# Create batch generators
genTrain = DataGenerator(imagesMeta=data.trainMeta, GTMeta = data.trainGTMeta, cfg=cfg, data_type='train', labels=data.labels)
genVal = DataGenerator(imagesMeta=data.valMeta, GTMeta = data.trainGTMeta, cfg=cfg, data_type='val', labels=data.labels)
genTest = DataGenerator(imagesMeta=data.testMeta, GTMeta = data.testGTMeta, cfg=cfg, data_type='test', labels=data.labels)
y_hat = utils.load_obj(cfg.part_results_path + 'y_hat_test')
Y = genTest.getYData()
Y = utils.load_obj(cfg.part_results_path + 'y_test')
res = metrics.computeMultiLabelLoss(Y, y_hat)
print(res[1:4])
cm = metrics.computeConfusionMatrixLabels(Y, y_hat)
_, counts = utils.getLabelStats(data.trainGTMeta, data.labels)
idxs = np.concatenate([np.array([0]), np.argsort(counts, axis=0)+1], axis=0)
_, test_counts = utils.getLabelStats(data.testGTMeta, data.labels)
test_idxs = np.concatenate([np.array([0]), np.argsort(test_counts, axis=0)+ 1], axis=0)
cm_sorted = cm[idxs,:]
# cm_reduced = cm_sorted[-100:-1,0:]
plt.figure(1)
draw.plot_confusion_matrix(cm, data.labels, normalize=True)
plt.figure(2)
plt.plot([x for x in range(cm.shape[0])], cm[:,0] / np.sum(cm, axis=1))
plt.ylabel('ratio of no label instances')
plt.xlabel('classes')