def gmm_model(sample, k_gmm, PCA=False):
"""
Returns a tuple: (gmm,mean,pca_transform)
gmm is the ynumpy gmm model fro the sample data.
pca_tranform is None if PCA is True.
Reduces the dimensions of the sample (by 50%) if PCA is true
"""
print "Building GMM model"
# until now sample was in uint8. Convert to float32
sample = sample.astype('float32')
# compute mean and covariance matrix for the PCA
mean = sample.mean(axis = 0) #for each row
sample = sample - mean
pca_transform = None
if PCA:
cov = np.dot(sample.T, sample)
#decide to keep 1/2 of the original components, so vid_trajs_bm.shape[1]/2
#compute PCA matrix and keep only 1/2 of the dimensions.
orig_comps = sample.shape[1]
pca_dim = orig_comps/2
#eigvecs are normalized.
eigvals, eigvecs = np.linalg.eig(cov)
perm = eigvals.argsort() # sort by increasing eigenvalue
pca_transform = eigvecs[:, perm[orig_comps-pca_dim:orig_comps]] # eigenvectors for the 64 last eigenvalues
# transform sample with PCA (note that numpy imposes line-vectors,
# so we right-multiply the vectors)
sample = np.dot(sample, pca_transform)
# train GMM
gmm = ynumpy.gmm_learn(sample, k_gmm)
toReturn = (gmm,mean,pca_transform)
return toReturn
def save_model(sample, type_feature, RESULT_PATH, TYPE):
'''Save GMM models, mean, pca_transform of sample data
Parameters
----------
sample: np.ndarray type,
Sample data
type_feature: {'traj', 'hog', 'hof', 'mbh'}
One of four types of feature
RESULT_PATH: string
Path of result files
TYPE: {'video', 'sensor'}
Type of data
Returns
-------
gmm: gmm model
mean: mean value of data
pca_transform: pca transform of data
'''
# GMM part
k = 25
if type_feature == 'traj':
DIM_AFTER_PCA = 0.5 * 30
elif type_feature == 'hog':
DIM_AFTER_PCA = 0.5 * 96
elif type_feature == 'hof':
DIM_AFTER_PCA = 0.5 * 108
elif type_feature == 'mbh':
DIM_AFTER_PCA = 0.5 * 192
# compute mean and covariance matrix for the PCA
mean = sample.mean(axis = 0)
sample = sample - mean
cov = np.dot(sample.T, sample)
# compute PCA matrix and keep only 64 dimensions
eigvals, eigvecs = np.linalg.eig(cov)
# sort by increasing eigenvalue
perm = eigvals.argsort()
# eigenvectors for the 64 last eigenvalues
pca_transform = eigvecs[:, perm[-DIM_AFTER_PCA:]]
# transform sample with PCA (note that numpy imposes line-vectors,
# so we right-multiply the vectors)
sample = np.dot(sample, pca_transform)
# train GMM
gmm = ynumpy.gmm_learn(sample, k)
# save mean value to file
file_mean = open(RESULT_PATH + 'mean_' + type_feature + '_' + TYPE + '.pkl', 'wb')
pickle.dump(mean, file_mean)
file_mean.close()
# save pca transform to file
file_pca = open(RESULT_PATH + 'pca_' + type_feature + '_' + TYPE + '.pkl', 'wb')
pickle.dump(pca_transform, file_pca)
file_pca.close()
# save GMM model to file
file_gmm = open(RESULT_PATH + 'gmm_' + type_feature + '_' + TYPE + '.pkl', 'wb')
pickle.dump(gmm, file_gmm)
file_gmm.close()
return gmm, mean, pca_transform
def train_fv_gmms(tracklets_path, videonames, traintest_parts, feat_types, intermediates_path, pca_reduction=False, nt=4, verbose=False):
try:
makedirs(intermediates_path)
except OSError:
pass
for k, part in enumerate(traintest_parts):
train_inds = np.where(np.array(part) <= 0)[0] # train codebook for each possible training parition
num_samples_per_vid = int(INTERNAL_PARAMETERS['n_samples'] / float(len(train_inds)))
# process the videos
for i, feat_t in enumerate(feat_types):
D = None
# Train GMMs
output_filepath = join(intermediates_path, 'gmm' + ('_pca-' if pca_reduction else '-') + feat_t + '-' + str(k) + '.pkl')
if isfile(output_filepath):
if verbose:
print('[train_fv_gmms] %s -> OK' % output_filepath)
continue
start_time = time.time()
D = load_tracklets_sample(tracklets_path, videonames, train_inds, feat_t, num_samples_per_vid, verbose=verbose)
# (special case) trajectory features are originally positions
if feat_t == 'trj':
D = convert_positions_to_displacements(D)
if feat_t == 'mbh':
Dx = preprocessing.normalize(D[:,:D.shape[1]/2], norm='l1', axis=1)
Dy = preprocessing.normalize(D[:,D.shape[1]/2:], norm='l1', axis=1)
D = np.hstack((Dx,Dy))
else:
D = preprocessing.normalize(D, norm='l1', axis=1)
if feat_t != 'trj':
D = rootSIFT(D)
# compute PCA map and reduce dimensionality
if pca_reduction:
pca = PCA(n_components=int(INTERNAL_PARAMETERS['reduction_factor']*D.shape[1]), copy=False)
D = pca.fit_transform(D)
# train GMMs for later FV computation
D = np.ascontiguousarray(D, dtype=np.float32)
gmm = ynumpy.gmm_learn(D, INTERNAL_PARAMETERS['fv_gmm_k'], nt=nt, niter=500, redo=1, verbose=verbose)
with open(output_filepath, 'wb') as f:
cPickle.dump(dict(pca=(pca if pca_reduction else None), gmm=gmm), f)
# with open(join(intermediates_path, 'gmm-sample' + ('_pca-' if pca_reduction else '-') + feat_t + '-' + str(k) + '.pkl'), 'wb') as f:
# cPickle.dump(D,f)
elapsed_time = time.time() - start_time
if verbose:
print('[train_fv_gmms] %s -> DONE (in %.2f secs)' % (feat_t, elapsed_time))
print "vectors = "
print v
print "meta info = "
print meta
print "kmeans:"
centroids = ynumpy.kmeans(v, 3)
print "result centroids ="
print centroids[:10,:]
print "gmm:"
gmm = ynumpy.gmm_learn(v, 3)
(w, mu, sigma) = gmm
print "mu = "
print mu
print "sigma = "
print sigma
muc = numpy.vstack((mu[0, :],
mu[0, :]));
# mu[1, :],
# mu[1, :],
# compute mean and covariance matrix for the PCA mean = sample.mean(axis = 0) sample = sample - mean cov = np.dot(sample.T, sample) # compute PCA matrix and keep only 64 dimensions eigvals, eigvecs = np.linalg.eig(cov) perm = eigvals.argsort() # sort by increasing eigenvalue pca_transform = eigvecs[:, perm[64:128]] # eigenvectors for the 64 last eigenvalues # transform sample with PCA (note that numpy imposes line-vectors, # so we right-multiply the vectors) sample = np.dot(sample, pca_transform) # train GMM gmm = ynumpy.gmm_learn(sample, k) image_fvs = [] for image_desc in image_descs: # apply the PCA to the image descriptor image_desc = np.dot(image_desc - mean, pca_transform) # compute the Fisher vector, using the derivative w.r.t mu and sigma fv = ynumpy.fisher(gmm, image_desc, include = 'mu, sigma') image_fvs.append(fv) # make one matrix with all FVs image_fvs = np.vstack(image_fvs) # normalizations are done on all descriptors at once # power-normalization
print("Training set of %d local descriptors in %d dimensions" % (train_set.shape[0], train_set.shape[1]))
trainset_size = num_gmm_components * 1000
if trainset_size < train_set.shape[0]:
print("Subsampling to %d points" % trainset_size)
subset = numpy.array(
random.sample(range(train_set.shape[0]), trainset_size))
train_set = train_set[subset]
print("Training Gaussian mixture model with %d components" % num_gmm_components)
train_set = train_set.astype('float32')
gmm = ynumpy.gmm_learn(train_set, num_gmm_components)
print("Make the image index")
dataset = []
queries = []
if show:
fig = plt.figure(figsize=(10, 10))
fig.canvas.set_window_title("100 image dataset")
plot_idx = 1
for i in image_range:
filename = "%s/ukbench%05d.siftgeo" % (sift_directory, i)
print(" " + filename + "\r")
sys.stdout.flush()
def train_fv_gmms(tracklets_path, videonames, traintest_parts, feat_types, intermediates_path, pca_reduction=True, nt=4):
if not exists(intermediates_path):
makedirs(intermediates_path)
for k, part in enumerate(traintest_parts):
train_inds = np.where(part <= 0)[0] # train codebook for each possible training parition
total = len(train_inds)
num_samples_per_vid = int(INTERNAL_PARAMETERS['n_samples'] / float(total))
# process the videos
for i, feat_t in enumerate(feat_types):
output_filepath = intermediates_path + 'gmm' + ('_pca-' if pca_reduction else '-') + feat_t + '-' + str(k) + '.pkl'
if isfile(output_filepath):
print('%s -> OK' % output_filepath)
continue
start_time = time.time()
D = None # feat_t's sampled tracklets
ptr = 0
for j in range(0, total):
idx = train_inds[j]
filepath = tracklets_path + feat_t + '/' + videonames[idx] + '.pkl'
if not isfile(filepath):
sys.stderr.write('# ERROR: missing training instance'
' {}\n'.format(filepath))
sys.stderr.flush()
quit()
with open(filepath, 'rb') as f:
d = cPickle.load(f)
# init sample
if D is None:
D = np.zeros((INTERNAL_PARAMETERS['n_samples'], d.shape[1]), dtype=np.float32)
# create a random permutation for sampling some tracklets in this vids
randp = np.random.permutation(d.shape[0])
if d.shape[0] > num_samples_per_vid:
randp = randp[:num_samples_per_vid]
D[ptr:ptr+len(randp),:] = d[randp,:]
ptr += len(randp)
D = D[:ptr,:] # cut out extra reserved space
# (special case) trajectory features are originally positions
if feat_t == 'trj':
D = convert_positions_to_displacements(D)
# scale (rootSIFT)
D = rootSIFT(preprocessing.normalize(D, norm='l1', axis=1))
# compute PCA map and reduce dimensionality
if pca_reduction:
pca = PCA(n_components=int(INTERNAL_PARAMETERS['reduction_factor']*D.shape[1]), copy=False)
D = pca.fit_transform(D)
# train GMMs for later FV computation
D = np.ascontiguousarray(D, dtype=np.float32)
gmm = ynumpy.gmm_learn(D, INTERNAL_PARAMETERS['fv_gmm_k'], nt=nt, niter=100, redo=3)
with open(output_filepath, 'wb') as f:
cPickle.dump(dict(pca=(pca if pca_reduction else None), gmm=gmm), f)
elapsed_time = time.time() - start_time
print('%s -> DONE (in %.2f secs)' % (feat_t, elapsed_time))
dat = np.load('test/py/test_dat.npy')
cluster_num = len(set(label))
print cluster_num
cluster_w = np.array([(i + 2) % 5 != 0 for i in xrange(cluster_num)])
obs = np.vstack([np.vstack([dat[label == c]] * cluster_w[c]) for c in xrange(cluster_num) if cluster_w[c] != 0])
obs_w = dat.copy()
weight = cluster_w[label]
print obs_w.shape
print weight.shape
ret_original = ynumpy.gmm_learn(obs.astype(np.float32), cluster_num - 1)
ret_modified = ynumpy.gmm_learn_sw(obs_w.astype(np.float32), weight.astype(np.float32), cluster_num - 1)
print ''
print '======= result ======'
print ''
print ret_original
print ret_modified
orig_w, orig_mu, orig_sigma = ret_original
modi_w, modi_mu, modi_sigma = ret_modified
orig_i = np.argsort(orig_w)
modi_i = np.argsort(modi_w)
