def evaluate_feats(db,
N,
feat_pools=feat_pools,
d_type='d1',
depths=[None, 300, 200, 100, 50, 30, 10, 5, 3, 1]):
result = open(
os.path.join(result_dir,
'feature_fusion-{}-{}feats.csv'.format(d_type, N)), 'w')
for i in range(N):
result.write("feat{},".format(i))
result.write("depth,distance,MMAP")
combinations = itertools.combinations(feat_pools, N)
for combination in combinations:
fusion = FeatureFusion(features=list(combination))
for d in depths:
APs = evaluate_class(db, f_instance=fusion, d_type=d_type, depth=d)
cls_MAPs = []
for cls, cls_APs in APs.items():
MAP = np.mean(cls_APs)
cls_MAPs.append(MAP)
r = "{},{},{},{}".format(",".join(combination), d, d_type,
np.mean(cls_MAPs))
print(r)
result.write('\n' + r)
print()
result.close()
return a
np.random.seed(0)
IMG = sigmoid(np.random.randn(2, 2, 3)) * 255
IMG = IMG.astype(int)
hist = color.histogram(IMG, type='global', n_bin=4)
assert np.equal(np.where(hist > 0)[0], np.array([37, 43, 58, 61])).all(), "global histogram implement failed"
hist = color.histogram(IMG, type='region', n_bin=4, n_slice=2)
assert np.equal(np.where(hist > 0)[0], np.array([58, 125, 165, 235])).all(), "region histogram implement failed"
# examinate distance
np.random.seed(1)
IMG = sigmoid(np.random.randn(4, 4, 3)) * 255
IMG = IMG.astype(int)
hist = color.histogram(IMG, type='region', n_bin=4, n_slice=2)
IMG2 = sigmoid(np.random.randn(4, 4, 3)) * 255
IMG2 = IMG2.astype(int)
hist2 = color.histogram(IMG2, type='region', n_bin=4, n_slice=2)
assert distance(hist, hist2, d_type='d1') == 2, "d1 implement failed"
assert distance(hist, hist2, d_type='d2-norm') == 2, "d2 implement failed"
# evaluate database
APs = evaluate_class(db, f_class=Color, d_type=d_type, depth=depth)
cls_MAPs = []
for cls, cls_APs in APs.items():
MAP = np.mean(cls_APs)
print("Class {}, MAP {}".format(cls, MAP))
cls_MAPs.append(MAP)
print("MMAP", np.mean(cls_MAPs))
inputs = torch.autograd.Variable(
torch.from_numpy(img).float())
d_hist = vgg_model(inputs)[pick_layer]
d_hist = np.sum(d_hist.data.cpu().numpy(), axis=0)
d_hist /= np.sum(d_hist) # normalize
samples.append({
'img': d_img,
'cls': d_cls,
'hist': d_hist
})
except BaseException:
pass
cPickle.dump(
samples, open(os.path.join(cache_dir, sample_cache), "wb",
True))
return samples
if __name__ == "__main__":
# evaluate database
DB = Database()
APs = evaluate_class(DB, f_class=VGGNetFeat, d_type=d_type, depth=depth)
cls_MAPs = []
for cls, cls_APs in APs.items():
MAP = np.mean(cls_APs)
print("Class {}, MAP {}".format(cls, MAP))
cls_MAPs.append(MAP)
print("MMAP", np.mean(cls_MAPs))
except:
if verbose:
print(
"Counting histogram..., config=%s, distance=%s, depth=%s" %
(sample_cache, d_type, depth))
samples = []
data = db.get_data()
for d in data.itertuples():
d_img, d_cls = getattr(d, "img"), getattr(d, "cls")
d_hist = self.histogram(d_img, type=h_type, n_slice=n_slice)
samples.append({'img': d_img, 'cls': d_cls, 'hist': d_hist})
cPickle.dump(
samples, open(os.path.join(cache_dir, sample_cache), "wb",
True))
return samples
if __name__ == "__main__":
db = Database()
# evaluate database
APs = evaluate_class(db, f_class=Daisy, d_type=d_type, depth=depth)
cls_MAPs = []
for cls, cls_APs in APs.items():
MAP = np.mean(cls_APs)
print("Class {}, MAP {}".format(cls, MAP))
cls_MAPs.append(MAP)
print("MMAP", np.mean(cls_MAPs))
else:
inputs = torch.autograd.Variable(
torch.from_numpy(img).float())
d_hist = res_model(inputs)[pick_layer]
d_hist = d_hist.data.cpu().numpy().flatten()
d_hist /= np.sum(d_hist) # normalize
samples.append({
'img': d_img,
'cls': d_cls,
'hist': d_hist
})
except:
pass
cPickle.dump(
samples, open(os.path.join(cache_dir, sample_cache), "wb",
True))
return samples
if __name__ == "__main__":
# evaluate database
db = Database()
APs = evaluate_class(db, f_class=ResNetFeat, d_type=d_type, depth=depth)
cls_MAPs = []
for cls, cls_APs in APs.items():
MAP = np.mean(cls_APs)
print("Class {}, MAP {}".format(cls, MAP))
cls_MAPs.append(MAP)
print("MMAP", np.mean(cls_MAPs))
"Counting histogram..., config=%s, distance=%s, depth=%s" %
(sample_cache, d_type, depth))
samples = []
data = db.get_data()
for d in data.itertuples():
d_img, d_cls = getattr(d, "img"), getattr(d, "cls")
d_hist = self.histogram(d_img, type=h_type, n_slice=n_slice)
samples.append({'img': d_img, 'cls': d_cls, 'hist': d_hist})
cPickle.dump(
samples, open(os.path.join(cache_dir, sample_cache), "wb",
True))
return samples
if __name__ == "__main__":
db = Database()
# check shape
assert edge_kernels.shape == (5, 2, 2)
# evaluate database
APs = evaluate_class(db, f_class=Edge, d_type=d_type, depth=depth)
cls_MAPs = []
for cls, cls_APs in APs.items():
MAP = np.mean(cls_APs)
print("Class {}, MAP {}".format(cls, MAP))
cls_MAPs.append(MAP)
print("MMAP", np.mean(cls_MAPs))
db = Database()
# evaluate features double-wise
evaluate_feats(db, N=2, d_type='d1')
# evaluate features triple-wise
evaluate_feats(db, N=3, d_type='d1')
# evaluate features quadra-wise
evaluate_feats(db, N=4, d_type='d1')
# evaluate features penta-wise
evaluate_feats(db, N=5, d_type='d1')
# evaluate features hexa-wise
evaluate_feats(db, N=6, d_type='d1')
# evaluate features hepta-wise
evaluate_feats(db, N=7, d_type='d1')
# evaluate database
fusion = FeatureFusion(features=['color', 'daisy'])
APs = evaluate_class(db, f_instance=fusion, d_type=d_type, depth=depth)
cls_MAPs = []
for cls, cls_APs in APs.items():
MAP = np.mean(cls_APs)
print("Class {}, MAP {}".format(cls, MAP))
cls_MAPs.append(MAP)
print("MMAP", np.mean(cls_MAPs))
asy_result.append(
pool.apply_async(self.lbp_map,
args=(item.img, radius, n_points)))
for result, item in tqdm(zip(asy_result, data.itertuples()),
desc="get results"):
samples.append({
'img': item.img,
'cls': item.cls,
'hist': result.get()
})
# save and return
pickle.dump(samples,
open(os.path.join(cache_dir, sample_cache), "wb",
True)) # 保存在cache_dir/sample_cache下
return samples
if __name__ == "__main__":
db = Database()
# evaluate database
APs, match = evaluate_class(db, f_class=LBP, d_type=d_type, depth=depth)
cls_MAPs = []
for cls, cls_APs in APs.items():
MAP = np.mean(cls_APs)
print("Class {}, MAP {}".format(cls, MAP))
cls_MAPs.append(MAP)
print("MMAP", np.mean(cls_MAPs))