def Main():
#matplotlib.use('Agg')
if len(sys.argv) > 1 and os.path.isfile(sys.argv[1]):
extension = os.path.splitext(sys.argv[1])[1]
fig = plt.figure()
if extension == '.ply':
#A = np.expand_dims(A,axis=0)
V, A = ply2graph(sys.argv[1],K)
V = np.expand_dims(np.array(V),axis=0)
print(V.shape)
#A = np.expand_dims(np.transpose(np.array(A),axes=(1,0,2)),axis=0)
A = np.expand_dims(np.array(A),axis=0)
print(A.shape)
#eng = matlab.engine.start_matlab()
#eng.addpath(os.path.abspath(os.path.join(os.path.dirname(__file__),'../util/pooling/graclus1.2/matlab/')))
#pooler = GraclusPoolingPyramid(len(POOL_RATIOS), GeometricAdjacencyCompander, POOL_RATIOS)#,eng)
#pooler = SpectralClusteringPoolingPyramid(len(POOL_RATIOS), MemoryAdjacencyCompander, POOL_RATIOS)
pooler = LloydPoolingPyramid(len(POOL_RATIOS), SphericalAdjacencyCompander, POOL_RATIOS)
#pooler = SBFSStrideAdjacencyPyramid(len(POOL_RATIOS), GeometricAdjacencyCompander, STRIDE, POOL_RATIOS)
#draw3DGraph(V.sum(axis=0), A.sum(axis=(0, 2)), V.shape[1], '141', fig)
#plt.show()
Plist = pooler.makeP(A.sum(axis=0),V.sum(axis=0),THETA,PHI)
#print(Plist[0].shape)
Avar = tf.placeholder(tf.float32, shape=(None, None, None, None))
Vvar = tf.placeholder(tf.float32, shape=(None, None, None))
Pvar = tf.placeholder(tf.float32, shape=(None, None, None))
Vout,Aout = make_graph_pooling_layer(Vvar,Avar,Pvar)
#vshape = V.shape
#V = tf.constant(V,dtype=tf.float32,shape=vshape)
#A = tf.constant(A,dtype=tf.float32,shape=A.shape)
#Vout,Aout,W = make_graph_embed_pooling(V, A, no_vertices=111)
#Vrev,Arev = make_graph_unpooling_layer(Vvar,Avar,Pvar)
init = tf.initialize_all_variables()
with tf.Session() as sess:
sess.run(init)
for i in range(2,2+len(POOL_RATIOS)):
subplotId = '16' + str(i)
[V,A] = sess.run([Vout,Aout],feed_dict={Vvar: V, Avar: A, Pvar: np.expand_dims(Plist[i-2],axis=0)})
print(V.shape)
if extension == '.ply':
draw3DGraph(V.sum(axis=0),A.sum(axis=(0,2)),Plist[i - 2].shape[1],subplotId,fig)
#fig2 = plt.figure(2)
#subplotId = '421'
#draw3DGraph(V.sum(axis=0), A.sum(axis=(0, 2)), A.shape[1], subplotId, fig2)
#for i in range(2,2 + len(POOL_RATIOS)):
# subplotId = '42' + str(i)
# print(V.shape)
# [V, A] = sess.run([Vrev, Arev], feed_dict={Vvar: V, Avar: A, Pvar: np.expand_dims(Plist[len(POOL_RATIOS) - (i - 2) - 1], axis=0)})
# draw3DGraph(V.sum(axis=0), A.sum(axis=(0, 2)), Plist[len(POOL_RATIOS) - (i - 2) - 1].shape[1], subplotId, fig2)
plt.show()
else:
print('Could not find path LOL!')
def processFile(filename):
V, A = ply2graph(filename, K)
V = np.expand_dims(np.array(V), axis=0)
A = np.expand_dims(np.array(A), axis=0) # BxNxLxN?
print(A.shape)
pooler = LloydPoolingPyramid(len(POOL_RATIOS), SphericalAdjacencyCompander,
POOL_RATIOS)
Plist = pooler.makeP(A.sum(axis=0), V.sum(axis=0), THETA, PHI)
for i in range(len(Plist)):
Plist[i] = np.expand_dims(np.expand_dims(Plist[i], axis=0), axis=0)
return V, A, Plist
def processFile(filename):
V, A = ply2graph(filename, K)
V = np.expand_dims(np.array(V), axis=0)
A = np.expand_dims(np.transpose(np.array(A), axes=(1, 0, 2)),
axis=0) # BxNxLxN?
pooler = LloydPoolingPyramid(len(POOL_RATIOS), SphericalAdjacencyCompander,
POOL_RATIOS)
P = np.expand_dims(np.expand_dims(np.transpose(
pooler.makeP(
np.transpose(A, axes=(0, 2, 1, 3)).sum(axis=0), V.sum(axis=0),
THETA, PHI)[0]),
axis=0),
axis=0)
return V, A, P
def processFile(filename):
V, A = ply2graph(filename, K)
V = np.expand_dims(np.array(V), axis=0)
A = np.expand_dims(np.transpose(np.array(A), axes=(1, 0, 2)),
axis=0) # BxNxLxN?
pooler = LloydPoolingPyramid(len(POOL_RATIOS), SphericalAdjacencyCompander,
POOL_RATIOS)
return V, A
def Main():
np.set_printoptions(linewidth=180)
V1, A1, P1 = processFile(FILE)
V2, A2, P2 = processFile(FILE2)
A = np.concatenate((A1, A2), axis=0)
P = np.concatenate((P1, P2), axis=0)
V = np.concatenate((V1, V2), axis=0)
print(A.shape)
indices = np.transpose(np.nonzero(A)).astype(np.int64)
values = np.ravel(A[np.nonzero(A)]).astype(np.float32)
# print(values.shape)
dense_shape = np.asarray(A.shape, dtype=np.int64)
pooler = LloydPoolingPyramid(len(POOL_RATIOS), GeometricAdjacencyCompander,
POOL_RATIOS)
Pindices = np.transpose(np.nonzero(P)).astype(np.int64)
Pvals = P[np.nonzero(P)].astype(np.float32)
Pshape = np.asarray(P.shape, dtype=np.int64)
#A = np.transpose(np.array([[[[1,1,1,1],[1,1,1,1],[1,1,1,1],[1,1,1,1]]],[[[1,1,1,1],[1,1,1,1],[1,1,1,1],[1,1,1,1]]]]),axes=(0,2,1,3))
#print(A.shape)
#indices = np.transpose(np.nonzero(A)).astype(np.int64)
#values = A[np.nonzero(A)].astype(np.float32)
#V = np.transpose(np.array([[[9,10,11,12],[5,6,7,8],[1,2,3,4]],[[1,2,3,4],[5,6,7,8],[9,10,11,12]]]),axes=(0,2,1))
#print(V.shape)
#P = np.transpose(np.array([[[[0,1,0,0.1],[1,0,1,0]]],[[[1,0,1,0],[0,1,0,0.1]]]]),axes=(0,1,3,2))
#dense_shape = np.asarray(A.shape, dtype=np.int64)
#Pindices = np.transpose(np.nonzero(P)).astype(np.int64)
#Pvals = P[np.nonzero(P)].astype(np.float32)
#Pshape = np.asarray(P.shape, dtype=np.int64)
print("PSHAPE: {0}".format(Pshape))
Avarsparse = tf.sparse_placeholder(tf.float32, shape=dense_shape)
Avar = tf.placeholder(tf.float32, shape=(None, None, None, None))
Vvar = tf.placeholder(tf.float32, shape=V.shape)
idxTensor = tf.placeholder(tf.int64, shape=(None, None))
valTensor = tf.placeholder(tf.float32, shape=(None))
PidxTensor = tf.placeholder(tf.int64, shape=(None, None))
PvalTensor = tf.placeholder(tf.float32, shape=(None))
Asparse = tf.SparseTensor(indices=idxTensor,
values=valTensor,
dense_shape=dense_shape)
Adense = tf.cast(
tf.sparse_to_dense(Asparse.indices, Asparse.dense_shape,
Asparse.values), tf.float32)
Psparse = tf.SparseTensor(indices=PidxTensor,
values=PvalTensor,
dense_shape=Pshape)
Pdense = tf.squeeze(tf.cast(
tf.sparse_to_dense(Psparse.indices, Psparse.dense_shape,
Psparse.values), tf.float32),
axis=1)
# W = make_variable('weights', [V.shape[2]*A.shape[2], 16], initializer=tf.truncated_normal_initializer(stddev=math.sqrt(1.0/(V.shape[2]*(A.shape[2]+1)*GraphCNNGlobal.GRAPHCNN_INIT_FACTOR))))
# W = tf.ones([V.shape[2]*A.shape[2], 16],dtype=tf.float32)
# W = tf.fill([V.shape[2]*A.shape[2], 16],2.0)
W = tf.random_uniform([V.shape[2] * A.shape[2], NUM_FILTERS],
dtype=tf.float32)
VmaxpoolOutSparse, Aignore = make_sparse_max_graph_pooling_layer(
Vvar, Asparse, Psparse)
dV = tf.gradients(VmaxpoolOutSparse, Vvar)
dP = tf.gradients(VmaxpoolOutSparse, PvalTensor)
init = tf.initialize_all_variables()
with tf.Session() as sess:
sess.run(init)
outputs = [VmaxpoolOutSparse, dV, dP]
[VmaxpoolOutSparseOut,dVOut,dPOut] = sess.run(outputs, \
feed_dict={idxTensor: indices, \
valTensor: values, \
PidxTensor: Pindices, \
PvalTensor: Pvals, Vvar: V})
dVMaxPoolTheory, dVMaxPoolNumeric = tf.test.compute_gradient(Vvar, \
np.array(V.shape), \
VmaxpoolOutSparse, \
np.array([V.shape[0], int(V.shape[1] * POOL_RATIOS[0]),
V.shape[2]]), extra_feed_dict={idxTensor: indices, \
valTensor: values, \
PidxTensor: Pindices, \
PvalTensor: Pvals,
Vvar: V})
dPMaxPoolTheory, dPMaxPoolNumeric = tf.test.compute_gradient(PvalTensor, \
np.array(Pvals.shape), \
VmaxpoolOutSparse, \
np.array([V.shape[0],
int(V.shape[1] * POOL_RATIOS[0]),
V.shape[2]]),
extra_feed_dict={idxTensor: indices, \
valTensor: values, \
PidxTensor: Pindices, \
PvalTensor: Pvals, Vvar: V})
dVMaxPoolError = tf.test.compute_gradient_error(Vvar, \
np.array(V.shape), \
VmaxpoolOutSparse, \
np.array([V.shape[0], int(V.shape[1] * POOL_RATIOS[0]),
V.shape[2]]),extra_feed_dict={idxTensor: indices, \
valTensor: values, \
PidxTensor: Pindices, \
PvalTensor: Pvals, Vvar: V})
dPMaxPoolError = tf.test.compute_gradient_error(PvalTensor, \
np.array(Pvals.shape), \
VmaxpoolOutSparse, \
np.array([V.shape[0],
int(V.shape[1] * POOL_RATIOS[0]),
V.shape[2]]),
extra_feed_dict={idxTensor: indices, \
valTensor: values, \
PidxTensor: Pindices, \
PvalTensor: Pvals, Vvar: V})
print(V)
print(np.transpose(P, axes=(0, 1, 3, 2)))
print('VMaxPoolResults')
print(VmaxpoolOutSparseOut)
#print(dVOut[0].shape)
print(dVOut[0])
dPOutArray = np.zeros(P.shape)
dPOutArray[np.nonzero(P)] = dPOut[0]
print(dPOutArray)
print(dVMaxPoolTheory)
print(dVMaxPoolNumeric)
print(dPMaxPoolTheory)
print(dPMaxPoolNumeric)
mse = mean_squared_error(np.ravel(dVMaxPoolTheory),
np.ravel(dVMaxPoolNumeric))
print('MSE dV: {0}'.format(mse))
print('MSE dV: {0}'.format(dVMaxPoolError))
print(np.linalg.norm(dVMaxPoolTheory))
print(np.linalg.norm(dVMaxPoolNumeric))
mse = mean_squared_error(np.ravel(dPMaxPoolTheory),
np.ravel(dPMaxPoolNumeric))
print('MSE dP: {0}'.format(mse))
print('MSE dP: {0}'.format(dPMaxPoolError))
print(np.linalg.norm(dPMaxPoolTheory))
print(np.linalg.norm(dPMaxPoolNumeric))
print(np.linalg.norm(VmaxpoolOutSparseOut))