def forward(self, x):
radius = 1.2
adjacency_method = 'cyflann'
adjacency_kwds = {'radius': radius}
affinity_method = 'gaussian'
affinity_kwds = {'radius': radius}
laplacian_method = 'symmetricnormalized'
laplacian_kwds = {'scaling_epps': radius}
geom = Geometry(adjacency_method=adjacency_method,
adjacency_kwds=adjacency_kwds,
affinity_method=affinity_method,
affinity_kwds=affinity_kwds,
laplacian_method=laplacian_method,
laplacian_kwds=laplacian_kwds)
x = x.view(100, -1)
#print(x.shape)
geom.set_data_matrix(x)
lle = LocallyLinearEmbedding(n_components=10,
eigen_solver='dense',
geom=geom)
embed_lle = lle.fit_transform(x)
embed_lle = torch.from_numpy(embed_lle).float()
x = embed_lle.view(-1, 10)
x = self.fc(x)
return x
def forward(self, x):
#print('x in:' , x.shape)
rad1 = 0.7272
radius = rad1
adjacency_method = 'cyflann'
adjacency_kwds = {'radius': radius}
affinity_method = 'gaussian'
affinity_kwds = {'radius': radius}
laplacian_method = 'symmetricnormalized'
laplacian_kwds = {'scaling_epps': radius}
geom = Geometry(adjacency_method=adjacency_method,
adjacency_kwds=adjacency_kwds,
affinity_method=affinity_method,
affinity_kwds=affinity_kwds,
laplacian_method=laplacian_method,
laplacian_kwds=laplacian_kwds)
x = x.view(100, -1)
#print('before embedding:', x.shape)
geom.set_data_matrix(x)
spectral = SpectralEmbedding(n_components=10,
eigen_solver='amg',
geom=geom,
drop_first=False) # use 3 for spectral
embed_spectral = spectral.fit_transform(x)
embed_spectral = torch.from_numpy(embed_spectral).float()
#print('x_totorch:', embed_spectral.shape)
x = embed_spectral.view(-1, 10)
x = self.fc(x)
return x
def compute_geom_brute(self, diffusion_time, n_neighbors):
data = self.data
dim = self.dim
#set radius according to paper (i think this is dim not d)
radius = (diffusion_time * (diffusion_time * np.pi * 4)**(dim/2))**(0.5)
#set adjacency radius large enough that points beyond it have affinity close to zero
bigradius = 3 * radius
adjacency_method = 'brute'
adjacency_kwds = {'radius':bigradius}
affinity_method = 'gaussian'
affinity_kwds = {'radius':radius}
laplacian_method = 'geometric'
laplacian_kwds = {'scaling_epps':1}
geom = Geometry(adjacency_method=adjacency_method, adjacency_kwds=adjacency_kwds,
affinity_method=affinity_method, affinity_kwds=affinity_kwds,
laplacian_method=laplacian_method, laplacian_kwds=laplacian_kwds)
geom.set_data_matrix(data)
geom.adjacency_matrix = geom.compute_adjacency_matrix()
geom.laplacian_matrix = geom.compute_affinity_matrix()
geom.laplacian_matrix = self.get_laplacian(geom, radius)
return(geom)
def compute_geom(self, diffusion_time, n_neighbors):
data = self.data
dim = self.dim
#set radius according to paper (i think this is dim not d)
radius = (diffusion_time * (diffusion_time * np.pi * 4)**(dim/2))**(0.5)
#set adjacency radius large enough that points beyond it have affinity close to zero
bigradius = 3 * radius
adjacency_method = 'cyflann'
cyflann_kwds = {'index_type':'kdtrees', 'num_trees':10, 'num_checks':n_neighbors}
adjacency_kwds = {'radius':bigradius, 'cyflann_kwds':cyflann_kwds}
affinity_method = 'gaussian'
affinity_kwds = {'radius':radius}
laplacian_method = 'geometric'
laplacian_kwds = {'scaling_epps':radius}
geom = Geometry(adjacency_method=adjacency_method, adjacency_kwds=adjacency_kwds,
affinity_method=affinity_method, affinity_kwds=affinity_kwds,
laplacian_method=laplacian_method, laplacian_kwds=laplacian_kwds)
geom.set_data_matrix(data)
adjacency_matrix = geom.compute_adjacency_matrix()
laplacian_matrix = geom.compute_laplacian_matrix()
return(geom)
# Affinity an NxN (sparse) pairwise matrix insicated similarity between points
# Laplacian an NxN (sparse) pairwsie matrix containing geometric manifold information
radius = 5
adjacency_method = 'cyflann'
adjacency_kwds = {'radius': radius} # ignore distances above this radius
affinity_method = 'gaussian'
affinity_kwds = {'radius': radius} # A = exp(-||x - y||/radius^2)
laplacian_method = 'geometric'
laplacian_kwds = {
'scaling_epps': radius
} # scaling ensures convergence to Laplace-Beltrami operator
geom = Geometry(adjacency_method=adjacency_method,
adjacency_kwds=adjacency_kwds,
affinity_method=affinity_method,
affinity_kwds=affinity_kwds,
laplacian_method=laplacian_method,
laplacian_kwds=laplacian_kwds)
# You can/should also use the set_data_matrix, set_adjacency_matrix, set_affinity_matrix
# to send your data set (in whichever form it takes) this way.
geom.set_data_matrix(X)
# You can get the distance, affinity etc with e.g: Geometry.get_distance_matrix()
# you can update the keyword arguments passed inially using these functions
adjacency_matrix = geom.compute_adjacency_matrix()
# by defualt this is pass-by-reference. Use copy=True to get a copied version.
# If you don't want to pre-compute a Geometry you can pass a dictionary or geometry
# arguments to one of the embedding classes.
geom = {
random.seed(0)
workingdirectory = os.popen('git rev-parse --show-toplevel').read()[:-1]
os.chdir(workingdirectory)
radius = 50
adjacency_method = 'cyflann'
adjacency_kwds = {'radius': radius}
affinity_method = 'gaussian'
affinity_kwds = {'radius': radius}
laplacian_method = 'symmetricnormalized'
laplacian_kwds = {'scaling_epps': radius}
dataname = workingdirectory + '/untracked_data/chemistry_data/tolueneangles020619_pca50'
#dataname = '/Users/samsonkoelle/Downloads/manigrad-100818/mani-samk-gradients/untracked_data/chemistry_data/ethanolangles022119_pca50'
data = np.load(dataname + '.npy')
geom = Geometry(adjacency_method=adjacency_method,
adjacency_kwds=adjacency_kwds,
affinity_method=affinity_method,
affinity_kwds=affinity_kwds,
laplacian_method=laplacian_method,
laplacian_kwds=laplacian_kwds)
geom.set_data_matrix(data)
geom.laplacian_method = 'geometric'
geom.laplacian_kwds = {'scaling_epps': radius}
laplacian_matrix = geom.compute_laplacian_matrix()
geom.compute_laplacian_matrix()
sample = np.arange(0, data.shape[0], 1000)
distorion_vs_rad_dim2 = run_estimate_radius(data,
geom.adjacency_matrix,
sample=sample,
d=1,
rmin=1,
rmax=10,
ntry=50,
ax = fig.add_subplot(111, projection='3d')
ax.scatter(X[:, 0], X[:, 1], X[:, 2], c=color, s=1)
plt.title('orignal data')
plt.show()
plt.savefig('tutorial_data_plot.png', format='png')
plt.close()
## The Geometry Class
from megaman.geometry import Geometry
'''
Geometry is the main class that will Cache things like distance/adjacency,
affinity, and laplacian.
You instantiate the Geometry class with
'''
geom = Geometry()
'''
At class instantiation you can also pass parameters & methods for
the three main components:
Adjacency: an NxN (sparse) pairwise matrix indicating neighborhood regions
Affinity an NxN (sparse) pairwise matrix insicated similarity between points
Laplacian an NxN (sparse) pairwsie matrix containing geometric manifold information
These parameters can also be overwritten when using the individual calculation
functions so that you don't have to re-initialize every time you want to change
a parameter.
If you already have fixed parameters then you can initialize with, for example:
'''
