def test_gmm_2_clusters_rand_init():
means, covars, priors, LL, posteriors = gmm(X, n_clusters=2)
assert_allclose(LL, -4341.0, atol=0.1)
assert_allclose(priors, [0.5, 0.5], atol=0.1)
assert_allclose(posteriors[0], [0.0, 1.0], atol=0.1)
assert_allclose(means, [[4, 4], [0, 0]], atol=0.2)
def test_gmm_2_clusters_rand_init():
means, covars, priors, LL, posteriors = gmm(X, n_clusters=2)
assert_allclose(LL, -4341.0, atol=0.1)
assert_allclose(priors, [0.5, 0.5], atol=0.1)
assert_allclose(posteriors[0], [0.0, 1.0], atol=0.1)
assert_allclose(means, [[4, 4], [0, 0]], atol=0.2)
def generate_gmms(X, _clusters=64):
means, covars, priors, ll, posteriors = gmm(X,
n_clusters=_clusters,
max_num_iterations=10000,
verbose=False)
model = {
'priors': priors,
'means': means,
'covars': covars,
'll': ll,
'posteriors': posteriors
}
return model
def fit(self, X, y=None):
X = X.reshape(-1, X.shape[2])
if len(X) < self.gmm_samples_number:
raise AttributeError(
'Number of samples must be greater than the number of GMM samples')
indices = np.random.choice(
X.shape[0], self.gmm_samples_number, replace=False)
X = X[indices, :]
means, covars, priors, ll, posteriors = gmm(
X,
n_clusters=self.gmm_clusters_number,
init_mode=self.init_mode,
)
means = means.transpose()
covars = covars.transpose()
self.gmm_ = (means, covars, priors)
return self
def get_gmm(kps, descs, image_size):
print("GMM started---")
# descs = StandardScaler().fit_transform(descs)
pca = PCA(n_components=64)
descs = pca.fit_transform(descs)
descs = embed_spatial_info(kps, descs, image_size)
data_mcp = gmm(descs,
n_clusters=512,
init_mode='rand',
max_num_iterations=100)
means = data_mcp[0].transpose()
covariance = data_mcp[1].transpose()
priors = data_mcp[2]
return means, covariance, priors
def test_gmm_2_clusters_custom_init_fail():
_ = gmm(X, n_clusters=2, init_mode='custom')
def fit(self,
X,
y,
gmm_init='kmeans',
svc_kernel='linear',
svc_penalty='l2',
C=1.0,
seed=None):
"""
Fit a GMM with `k` clusters using the sample images `X`. Then train a SVC on
on the Fisher vector encodings of `X`, given the class labels `y`
Parameters
----------
X : array, shape (N,H,W,C), or list of N arrays w/ shapes (H_i,W_i,C)
A set of training images from which to generate a sample of CNN
feature vectors to which a GMM will be fit. `C` should match the
`input_shape` of the CNN (must be 3 for ImageNet pretrained models).
y : array, shape (N,)
Array of class labels of images in `X`.
gmm_init : str, optional
Method to use for GMM initialization. One of {'kmeans', 'rand'}. Default = 'kmeans'.
Custom init is also possible through `cyvlfeat`, but is not implemented here.
svc_kernel : str, optional
Specifies the kernel type to be used in the support vector classifier algorithm.
It must be one of {'linear', 'poly', 'rbf', 'sigmoid', 'precomputed'} or a callable.
If none or 'linear' is given, sklearn.svm.LinearSVC (liblinear) will be used,
otherwise sklearn.svm.SVC (libsvm) will be used. (Actually, just LinearSVC for now.)
The former has more flexible penalty/loss options, and scales better to large numbers of
samples (> ~10,000). The latter obviously has more flexibility in kernel types.
svc_penalty : str, optional
One of {`l1`, `l2`}. Default='l2' is recommended.
C : float, optional
Inverse of regularization strength. Default=1.0 seems to work best in VGG+DTD tests.
seed : int, optional
Specify a random state for deterministic results. Default=None.
Returns
-------
train_score : float
Mean accuracy of trained SVC on the training set (`X`, `y`)
"""
if isinstance(X, np.ndarray):
assert X.ndim == 4, 'X must have shape (N,H,W,C) if an np.array'
feats = self.cnn.predict(X)
feats = feats.reshape(-1, feats[-1])
elif isinstance(X, list):
assert isinstance(
X[0], np.ndarray), 'X must contain numpy.ndarrays, if a list'
img_feats = [self._localfeatures(x) for x in X]
#print('(sample of) img_feats.shapes:', [i.shape for i in img_feats[0:5]])
feats = np.vstack(img_feats)
#print('all_feats.shape :', feats.shape)
else:
raise ValueError(
'GMM input X has unknown form. Should be 4D array or list of 3D arrays.'
)
# Fit the GMM
# TODO: figure out covariance_bound Buffer bug
# should be = to max(all_feats.var(axis=k_feat))*0.0001
print('Fitting GMM with %d clusters...' % self.k)
self.means, self.covars, self.priors, LL, posteriors = gmm.gmm(
feats,
n_clusters=self.k,
covariance_bound=None,
init_mode=gmm_init)
# Train the SVC
if svc_kernel == 'linear':
self.svc = LinearSVC(penalty=svc_penalty,
C=C,
class_weight='balanced',
random_state=seed)
else:
raise NotImplementedError(
'Only `linear` svc_kernel implemented right now.')
fv_X = self.encode_batch(img_feats)
self.svc.fit(fv_X, y)
return self.svc.score(fv_X, y)
def test_gmm_2_clusters_custom_init_fail():
_ = gmm(X, n_clusters=2, init_mode='custom')
def test_gmm_2_clusters_custom_init_fail():
with pytest.raises(ValueError):
_ = gmm(X, n_clusters=2, init_mode='custom')
