def theta_E(self, im, X_, Y_, w):
try:
assert (self.slip.N_X == im.shape[1])
except:
from NeuroTools.parameters import ParameterSet
from SLIP import Image
from LogGabor import LogGabor
self.slip = Image(
ParameterSet({
'N_X': im.shape[1],
'N_Y': im.shape[0]
}))
self.lg = LogGabor(self.slip)
im_ = im.sum(axis=-1)
im_ = im_ * np.exp(-.5 * ((.5 + .5 * self.slip.x - Y_)**2 +
(.5 + .5 * self.slip.y - X_)**2) / w**2)
E = np.zeros((self.N_theta, ))
for i_theta, theta in enumerate(self.thetas):
params = {
'sf_0': self.sf_0,
'B_sf': self.B_sf,
'theta': theta,
'B_theta': np.pi / self.N_theta
}
FT_lg = self.lg.loggabor(0, 0, **params)
E[i_theta] = np.sum(
np.absolute(
self.slip.FTfilter(np.rot90(im_, -1), FT_lg,
full=True))**2)
return E
def presentStimulus(win, stimulus, param, info, do_mask=True):
import time
if (param.condition == 1):
pe = ParameterSet({'N_X' : info[NS_X]/2, 'N_Y' : info[NS_Y]/2, 'figpath':'.', 'matpath':'.'})
else: pe = ParameterSet({'N_X' : info[NS_X], 'N_Y' : info[NS_Y], 'figpath':'.', 'matpath':'.'})
im = Image(pe)
if (stimulus.ndim == 3):
img1 = stim(stimulus[:, :, 0])
img2 = stim(im.translate(stimulus[:, :, 0], [param.shift, 0]))
else:
img1 = stim(stimulus)
img2 = stim(im.translate(stimulus, [param.shift, 0]))
if param.flip == -1:
img2 = 255 - img2
if do_mask:
if (param.condition == 1):
im = Image(ParameterSet({'N_X' : info[NS_X]/2, 'N_Y' : info[NS_Y]/2, 'figpath':'.', 'matpath':'.'}))
else: im = Image(ParameterSet({'N_X' : info[NS_X], 'N_Y' : info[NS_Y], 'figpath':'.', 'matpath':'.'}))
mask = im.mask[:, :, np.newaxis]
img1 = ((img1 - 127)*mask + 127).astype(int)
img2 = ((img2 - 127)*mask + 127).astype(int)
win.background.fill(RGB.Gray)
win.screen.blit(win.background, (0, 0))
winblit(img1, win, info)
winblit(img2, win, info)
def init(self):
Image.init(self)
self.n_levels = int(np.log(np.max((self.pe.N_X, self.pe.N_Y)))/np.log(self.pe.base_levels))
self.sf_0 = .5 * (1 - 1/self.n_levels) / np.logspace(0, self.n_levels-1, self.n_levels, base=self.pe.base_levels, endpoint=False)
self.theta = np.linspace(-np.pi/2, np.pi/2, self.pe.n_theta+1)[1:]
self.oc = (self.pe.N_X * self.pe.N_Y * self.pe.n_theta * self.n_levels) #(1 - self.pe.base_levels**-2)**-1)
if self.pe.use_cache is True:
self.cache = {'band':{}, 'orientation':{}}
def init(self):
Image.init(self)
self.n_levels = int(
np.log(np.max(
(self.pe.N_X, self.pe.N_Y))) / np.log(self.pe.base_levels))
self.sf_0 = 1. / np.logspace(
1, self.n_levels, self.n_levels, base=self.pe.base_levels)
self.theta = np.linspace(-np.pi / 2, np.pi / 2,
self.pe.n_theta + 1)[1:]
self.oc = (self.pe.N_X * self.pe.N_Y * self.pe.n_theta * self.n_levels
) #(1 - self.pe.base_levels**-2)**-1)
def creation_stimulus(info, screen, param, name_database='blackwhite'):
import MotionClouds as mc
from MotionClouds.display import rectif
# from libpy import lena
if (param.condition == 1):
stimulus = (np.random.rand(info[NS_X]/2, info[NS_Y]/2) > .5)
# stimulus = (np.random.rand(64, 64) > .5)
elif (param.condition == 2):
im = Image(ParameterSet({'N_X' : info[NS_X], 'N_Y' : info[NS_Y], 'figpath':'.', 'matpath':'.', 'datapath':'database/', 'do_mask':False, 'seed':None}))
stimulus, filename, croparea = im.patch(name_database)
# stimulus = lena()
stimulus = np.rot90(np.fliplr(stimulus))
stimulus = rectif(stimulus, contrast=1.)
else:
fx, fy, ft = mc.get_grids(info[NS_X], info[NS_Y], 1)
if (param.condition == 3):
t, b, B_sf, sf_0 = 0, np.pi/32, 0.1, 0.15
if (param.condition == 4):
t, b, B_sf, sf_0= 0, np.pi/8, 0.1, 0.15
if (param.condition == 5):
t, b, B_sf, sf_0= 0, np.pi/2, 0.1, 0.15
if (param.condition == 6):
t, b, B_sf, sf_0= 0, np.pi/32, 0.1, 0.03
if (param.condition == 7):
t, b, B_sf, sf_0= 0, np.pi/32, 0.1, 0.075
if (param.condition == 8):
t, b, B_sf, sf_0= 0, np.pi/32, 0.25, 0.15
if (param.condition == 9):
t, b, B_sf, sf_0= 0, np.pi/32, 0.5, 0.15
fx, fy, ft = mc.get_grids(info[NS_X], info[NS_Y], 1)
cloud = mc.random_cloud(mc.envelope_gabor(fx, fy, ft, sf_0=sf_0, B_sf=B_sf, theta=t, B_theta=b, B_V=1000.))
cloud = rectif(cloud, contrast=1.)
stimulus = cloud[:, :, 0]
return (stimulus)
def __init__(self,
height=256,
width=256,
patch_size=(12, 12),
database = 'database/',
n_components=14**2,
learning_algorithm='omp',
alpha=None,
transform_n_nonzero_coefs=20,
n_iter=5000,
eta=.01,
eta_homeo=.01,
alpha_homeo=.01,
max_patches=1000,
batch_size=100,
n_image=200,
DEBUG_DOWNSCALE=1, # set to 10 to perform a rapid experiment
verbose=0,
):
self.height = height
self.width = width
self.database = database
self.patch_size = patch_size
self.n_components = n_components
self.n_iter = int(n_iter/DEBUG_DOWNSCALE)
self.max_patches = int(max_patches/DEBUG_DOWNSCALE)
self.n_image = int(n_image/DEBUG_DOWNSCALE)
self.batch_size = batch_size
self.learning_algorithm = learning_algorithm
self.alpha=alpha
self.transform_n_nonzero_coefs = transform_n_nonzero_coefs
self.eta = eta
self.eta_homeo = eta_homeo
self.alpha_homeo = alpha_homeo
self.verbose = verbose
# Load natural images and extract patches
self.slip = Image({'N_X':height, 'N_Y':width,
'white_n_learning' : 0,
'seed': None,
'white_N' : .07,
'white_N_0' : .0, # olshausen = 0.
'white_f_0' : .4, # olshausen = 0.2
'white_alpha' : 1.4,
'white_steepness' : 4.,
'datapath': self.database,
'do_mask':True,
'N_image': n_image})
def __init__(self,
height=256,
width=256,
patch_size=(10, 10),
n_components=11**2,
learning_algorithm='omp',
transform_n_nonzero_coefs=20,
n_iter=50000,
eta=1./25,
eta_homeo=0.001,
alpha_homeo=0.02,
max_patches=10000,
batch_size=100,
n_image=200,
DEBUG_DOWNSCALE=1, # set to 10 to perform a rapid experiment<D-d>
verbose=20,
):
self.height = height
self.width = width
self.patch_size = patch_size
self.n_components = n_components
self.n_iter = int(n_iter/DEBUG_DOWNSCALE)
self.max_patches = int(max_patches/DEBUG_DOWNSCALE)
self.n_image = int(n_image/DEBUG_DOWNSCALE)
self.batch_size = batch_size
self.learning_algorithm = learning_algorithm
self.transform_n_nonzero_coefs = transform_n_nonzero_coefs
self.eta = eta
self.eta_homeo = eta_homeo
self.alpha_homeo = alpha_homeo
self.verbose = verbose
# Load natural images and extract patches
self.slip = Image(ParameterSet({'N_X':height, 'N_Y':width,
'white_n_learning' : 0,
'seed': None,
'white_N' : .07,
'white_N_0' : .0, # olshausen = 0.
'white_f_0' : .4, # olshausen = 0.2
'white_alpha' : 1.4,
'white_steepness' : 4.,
'datapath': '/Users/lolo/pool/science/PerrinetBednar15/database/',
'do_mask':True,
'N_image': n_image}))
def __init__(self, pe):
Image.__init__(self, pe)
self.init_logging(name='LogGabor')
class EdgeGrid():
def __init__(
self,
N_lame=8 * 72,
N_lame_X=None,
figsize=13,
line_width=4.,
grid_type='hex',
structure=False,
struct_angles=[-15., -65., -102.],
verb=False,
mode='both',
filename=None,
period=None,
#**kw_args
):
self.t0 = self.time(True)
self.t = self.time()
self.dt = self.t - self.t0
self.verb = verb
self.display = (mode == 'display') or (mode == 'both')
self.stream = (mode == 'stream') or (mode == 'display')
#if mode=='display': self.stream = True
self.filename = filename
self.serial = (
mode == 'serial'
) # converting a stream to the serial port to control the arduino
if self.serial: self.verb = True
# self.desired_fps = 750.
self.desired_fps = 30.
self.structure = structure
self.screenshot = True # saves a screenshot after the rendering
self.port = "5556"
# moteur:
self.serial_port, self.baud_rate = '/dev/ttyUSB0', 115200
# 1.8 deg par pas (=200 pas par tour) x 32 divisions de pas
# demultiplication : pignon1= 14 dents, pignon2 = 60 dents
self.n_pas = 200. * 32. * 60 / 14
# TODO : Vitesse Max moteur = 1 tour en 3,88
self.n_pas_max = 148 # 150
# taille installation
self.total_width = 8 # en mètres
self.lames_width = 5 # en mètres
self.lames_height = 3 # en mètres
self.background_depth = 100 # taille du 104 en profondeur
self.f = .1
self.struct_N = 6
self.struct_position = [0., 3.5]
self.struct_longueur = 3.
self.struct_angles = struct_angles
self.figsize = figsize
self.line_width = line_width
self.grid_type = grid_type
self.grid(N_lame=N_lame, N_lame_X=N_lame_X)
# self.lames[2, :] = np.pi*np.random.rand(self.N_lame)
self.N_particles_per_lame = 2**3
self.N_particles = self.struct_N * self.N_particles_per_lame
if structure: self.sample_structure()
# enregistrement / playback
self.period = period
self.load()
self.vext = '.mp4'
self.figpath = '../files/figures/elasticite/'
self.fps = 25
def load(self):
if not self.filename is None:
if os.path.isfile(self.filename):
# le fichier existe, on charge
self.z = np.load(self.filename)
self.period = self.z[:, 0].max()
def time(self, init=False):
if init: return time.time()
else: return time.time() - self.t0
def grid(self, N_lame, N_lame_X):
"""
The coordinates of the screen are centered on the (0, 0) point and axis are the classical convention:
y
^
|
+---> x
angles are from the horizontal, then in trigonometric order (anticlockwise)
"""
self.DEBUG = True
self.DEBUG = False
self.N_lame = N_lame
#if N_lame_X is None:
if self.grid_type == 'hex':
self.N_lame_X = np.int(np.sqrt(self.N_lame)) #*np.sqrt(3) / 2)
self.lames = np.zeros((4, self.N_lame))
self.lames[0, :] = np.mod(np.arange(self.N_lame), self.N_lame_X)
self.lames[0, :] += np.mod(
np.floor(np.arange(self.N_lame) / self.N_lame_X), 2) / 2
self.lames[1, :] = np.floor(np.arange(self.N_lame) / self.N_lame_X)
self.lames[1, :] *= np.sqrt(3) / 2
self.lames[0, :] /= self.N_lame_X
self.lames[1, :] /= self.N_lame_X
self.lames[0, :] += .5 / self.N_lame_X - .5
self.lames[
1, :] += 1.5 / self.N_lame_X # TODO : prove analytically
self.lames[0, :] *= self.total_width
self.lames[1, :] *= self.total_width
self.lames[1, :] -= self.total_width / 2
self.lame_length = .99 / self.N_lame_X * self.total_width * np.ones(
self.N_lame)
self.lame_width = .03 / self.N_lame_X * self.total_width * np.ones(
self.N_lame)
elif self.grid_type == 'line':
self.N_lame_X = self.N_lame
self.lames = np.zeros((4, self.N_lame))
self.lames[0, :] = np.linspace(-self.lames_width / 2,
self.lames_width / 2,
self.N_lame,
endpoint=True)
#self.lames[1, :] = self.total_width/2
self.lame_length = .12 * np.ones(self.N_lame) # en mètres
self.lame_width = .042 * np.ones(self.N_lame) # en mètres
if self.structure: self.add_structure()
self.lames_minmax = np.array([
self.lames[0, :].min(), self.lames[0, :].max(),
self.lames[1, :].min(), self.lames[1, :].max()
])
def do_structure(self):
structure_ = np.zeros((3, self.struct_N))
chain = np.zeros((2, 4))
chain[:, 0] = np.array(self.struct_position).T
for i, angle in enumerate(self.struct_angles):
chain[0, i + 1] = chain[0, i] + self.struct_longueur * np.cos(
angle * np.pi / 180.)
chain[1, i + 1] = chain[1, i] + self.struct_longueur * np.sin(
angle * np.pi / 180.)
structure_[2, 3 + i] = +angle * np.pi / 180.
structure_[2, i] = np.pi - angle * np.pi / 180
structure_[0, 3:] = .5 * (chain[0, 1:] + chain[0, :-1])
structure_[0, :3] = -.5 * (chain[0, 1:] + chain[0, :-1])
structure_[1, 3:] = .5 * (chain[1, 1:] + chain[1, :-1])
structure_[1, :3] = .5 * (chain[1, 1:] + chain[1, :-1])
return structure_
def add_structure(self):
self.N_lame += self.struct_N
self.lames = np.hstack((self.lames, np.zeros((4, self.struct_N))))
self.lames[:3, -self.struct_N:] = self.do_structure()
self.lame_length = np.hstack(
(self.lame_length,
self.struct_longueur * np.ones(self.struct_N))) # en mètres
self.lame_width = np.hstack(
(self.lame_width, .042 * np.ones(self.struct_N))) # en mètres
def sample_structure(self, N_mirror=0, alpha=.8):
struct = self.lames[:3, -self.struct_N:]
self.particles = np.ones((3, self.N_particles))
N_particles_ = self.N_particles / self.struct_N
for i, vec in enumerate(struct.T.tolist()):
x0, x1 = vec[0] - .5 * self.struct_longueur * np.cos(
vec[2]), vec[0] + .5 * self.struct_longueur * np.cos(vec[2])
y0, y1 = vec[1] - .5 * self.struct_longueur * np.sin(
vec[2]), vec[1] + .5 * self.struct_longueur * np.sin(vec[2])
self.particles[0,
int(i *
N_particles_):int((i + 1) *
N_particles_)] = np.linspace(
x0, x1, N_particles_)
self.particles[1,
int(i *
N_particles_):int((i + 1) *
N_particles_)] = np.linspace(
y0, y1, N_particles_)
# duplicate according to mirrors
for i in range(N_mirror):
particles = self.particles.copy(
) # the current structure to mirror
particles_mirror = particles.copy(
) # the new set of particles with their mirror image
for segment in self.structure_as_segments():
particles_mirror = np.hstack((particles_mirror,
mirror(particles, segment,
alpha**(i + 1))))
# print(alpha**(i+1), particles_mirror[-1, -1])
self.particles = particles_mirror
def structure_as_segments(self):
struct = self.lames[:3, -self.struct_N:]
segments = []
for i, vec in enumerate(struct.T.tolist()):
x0, x1 = vec[0] - .5 * self.struct_longueur * np.cos(
vec[2]), vec[0] + .5 * self.struct_longueur * np.cos(vec[2])
y0, y1 = vec[1] - .5 * self.struct_longueur * np.sin(
vec[2]), vec[1] + .5 * self.struct_longueur * np.sin(vec[2])
segments.append(np.array([[x0, y0], [x1, y1]]).T)
return segments
def theta_E(self, im, X_, Y_, w):
try:
assert (self.slip.N_X == im.shape[1])
except:
from NeuroTools.parameters import ParameterSet
from SLIP import Image
from LogGabor import LogGabor
self.slip = Image(
ParameterSet({
'N_X': im.shape[1],
'N_Y': im.shape[0]
}))
self.lg = LogGabor(self.slip)
im_ = im.sum(axis=-1)
im_ = im_ * np.exp(-.5 * ((.5 + .5 * self.slip.x - Y_)**2 +
(.5 + .5 * self.slip.y - X_)**2) / w**2)
E = np.zeros((self.N_theta, ))
for i_theta, theta in enumerate(self.thetas):
params = {
'sf_0': self.sf_0,
'B_sf': self.B_sf,
'theta': theta,
'B_theta': np.pi / self.N_theta
}
FT_lg = self.lg.loggabor(0, 0, **params)
E[i_theta] = np.sum(
np.absolute(
self.slip.FTfilter(np.rot90(im_, -1), FT_lg,
full=True))**2)
return E
def theta_max(self, im, X_, Y_, w):
E = self.theta_E(im, X_, Y_, w)
return self.thetas[np.argmax(E)] - np.pi / 2
def theta_sobel(self, im, N_blur):
im_ = im.copy()
sobel = np.array([[
1,
2,
1,
], [
0,
0,
0,
], [
-1,
-2,
-1,
]])
if im_.ndim == 3: im_ = im_.sum(axis=-1)
from scipy.signal import convolve2d
im_X = convolve2d(im_, sobel, 'same')
im_Y = convolve2d(im_, sobel.T, 'same')
N_X, N_Y = im_.shape
x, y = np.mgrid[0:1:1j * N_X, 0:1:1j * N_Y]
mask = np.exp(-.5 * ((x - .5)**2 + (y - .5)**2) / w**2)
im_X = convolve2d(im_X, mask, 'same')
im_Y = convolve2d(im_Y, mask, 'same')
blur = np.array([[1, 2, 1], [2, 8, 2], [1, 2, 1]])
for i in range(N_blur):
im_X = convolve2d(im_X, blur, 'same')
im_Y = convolve2d(im_Y, blur, 'same')
angle = np.arctan2(im_Y, im_X)
bord = .1
angles = np.empty(self.N_lame)
N_X, N_Y = im_.shape
for i in range(self.N_lame):
angles[i] = angle[int(
(bord + self.lames[0, i] * (1 - 2 * bord)) * N_X),
int((bord + self.lames[1, i] * (1 - 2 * bord)) *
N_Y)]
return angles - np.pi / 2
def pos_rel(self, do_torus=False):
def torus(x, w=1.):
"""
center x in the range [-w/2., w/2.]
To see what this does, try out:
>> x = np.linspace(-4,4,100)
>> pylab.plot(x, torus(x, 2.))
"""
return np.mod(x + w / 2., w) - w / 2.
dx = self.lames[0, :, np.newaxis] - self.lames[0, np.newaxis, :]
dy = self.lames[1, :, np.newaxis] - self.lames[1, np.newaxis, :]
if do_torus:
return torus(dx), torus(dy)
else:
return dx, dy
def distance(self, do_torus=False):
dx, dy = self.pos_rel(do_torus=do_torus)
return np.sqrt(dx**2 + dy**2)
def angle_relatif(self):
return self.lames[2, :, np.newaxis] - self.lames[2, np.newaxis, :]
def angle_cocir(self, do_torus=False):
dx, dy = self.pos_rel(do_torus=do_torus)
theta = self.angle_relatif()
return np.arctan2(dy, dx) - np.pi / 2 - theta
def champ(self):
if self.structure: N_lame = self.N_lame - self.struct_N
else: N_lame = self.N_lame
force = np.zeros_like(self.lames[2, :N_lame])
noise = lambda t: 0.2 * np.exp((np.cos(2 * np.pi *
(t - 0.) / 6.) - 1.) / 1.5**2)
damp = lambda t: 0.01 #* np.exp(np.cos(t / 6.) / 3.**2)
colin_t = lambda t: -.1 * np.exp((np.cos(2 * np.pi *
(t - 2.) / 6.) - 1.) / .3**2)
cocir_t = lambda t: -4. * np.exp((np.cos(2 * np.pi *
(t - 4.) / 6.) - 1.) / .5**2)
cocir_d = lambda d: np.exp(-d / .05)
colin_d = lambda d: np.exp(-d / .2)
force += colin_t(
self.t) * np.sum(np.sin(2 * (self.angle_relatif()[:N_lame])) *
colin_d(self.distance()[:N_lame]),
axis=1)
force += cocir_t(
self.t) * np.sum(np.sin(2 * (self.angle_cocir()[:N_lame])) *
cocir_d(self.distance()[:N_lame]),
axis=1)
force += noise(self.t) * np.pi * np.random.randn(N_lame)
force -= damp(self.t) * self.lames[3, :N_lame] / self.dt
return 42. * force
def update(self):
if self.structure: N_lame = self.N_lame - self.struct_N
else: N_lame = self.N_lame
self.lames[2, :N_lame] += self.lames[3, :N_lame] * self.dt / 2
self.lames[3, :N_lame] += self.champ() * self.dt
self.lames[2, :N_lame] += self.lames[3, :N_lame] * self.dt / 2
# angles are defined as non oriented between -pi/2 and pi/2
self.lames[2, :N_lame] = np.mod(self.lames[2, :N_lame] + np.pi / 2,
np.pi) - np.pi / 2
def receive(self):
if not self.filename is None:
if os.path.isfile(self.filename):
if self.structure: N_lame = self.N_lame - self.struct_N
else: N_lame = self.N_lame
self.t = self.time()
i_t = np.argmin(self.z[:, 0] < np.mod(self.t, self.period))
if self.verb:
print("playback at t=", np.mod(self.t, self.period), i_t)
self.lames[2, :N_lame] = self.z[i_t, 1:]
return
if self.stream:
if self.verb: print("Sending request")
self.socket.send(b"Hello")
if self.verb: print("Received reply ", message)
self.lames[2, :] = recv_array(self.socket)
if self.verb: print("Received reply ", Theta.shape)
else:
self.dt = self.time() - self.t
self.update()
self.t = self.time()
# if not self.filename is None:
# if not os.path.isfile(self.filename):
# # recording
# if self.verb: print("recording at t=", self.t)
# self.z = np.vstack((self.z, np.hstack((np.array(self.t), self.lames[2, :] ))))
return
def render(self,
fps=10,
W=1000,
H=618,
location=[0, 1.75, -5],
head_size=.4,
light_intensity=1.2,
reflection=1.,
look_at=[0, 1.5, 0],
fov=75,
antialiasing=0.001,
duration=5,
fname='/tmp/temp.mp4'):
def scene(t):
"""
Returns the scene at time 't' (in seconds)
"""
head_location = np.array(location) - np.array([0, 0, head_size])
import vapory
light = vapory.LightSource([15, 15, 1], 'color',
[light_intensity] * 3)
background = vapory.Box(
[0, 0, 0], [1, 1, 1],
vapory.Texture(
vapory.Pigment(
vapory.ImageMap('png',
'"../files/VISUEL_104.png"', 'once')),
vapory.Finish('ambient', 1.2)), 'scale',
[self.background_depth, self.background_depth, 0], 'translate',
[
-self.background_depth / 2, -.45 * self.background_depth,
-self.background_depth / 2
])
me = vapory.Sphere(
head_location, head_size,
vapory.Texture(vapory.Pigment('color', [1, 0, 1])))
self.t = t
self.update()
objects = [background, me, light]
for i_lame in range(self.N_lame):
#print(i_lame, self.lame_length[i_lame], self.lame_width[i_lame])
objects.append(
vapory.Box(
[
-self.lame_length[i_lame] / 2, 0,
-self.lame_width[i_lame] / 2
],
[
self.lame_length[i_lame] / 2, self.lames_height,
self.lame_width[i_lame] / 2
],
vapory.Pigment('color', [1, 1, 1]),
vapory.Finish('phong', 0.8, 'reflection', reflection),
'rotate',
(0, -self.lames[2, i_lame] * 180 / np.pi, 0), #HACK?
'translate',
(self.lames[0, i_lame], 0, self.lames[1, i_lame])))
objects.append(light)
return vapory.Scene(vapory.Camera('angle', fov, "location",
location, "look_at", look_at),
objects=objects,
included=["glass.inc"])
import moviepy.editor as mpy
if not os.path.isfile(fname):
self.dt = 1. / fps
def make_frame(t):
return scene(t).render(width=W,
height=H,
antialiasing=antialiasing)
clip = mpy.VideoClip(make_frame, duration=duration)
clip.write_videofile(fname, fps=fps)
return mpy.ipython_display(fname, fps=fps, loop=1, autoplay=1)
def plot_structure(self,
W=1000,
H=618,
fig=None,
ax=None,
border=0.0,
opts=dict(vmin=-1,
vmax=1.,
linewidths=0,
cmap=None,
alpha=.1,
s=5.),
scale='auto'): #
opts.update(cmap=plt.cm.hsv)
if fig is None:
fig = plt.figure(figsize=(self.figsize, self.figsize * H / W))
if ax is None:
ax = fig.add_axes((border, border, 1. - 2 * border,
1. - 2 * border)) #, axisbg='w')
scat = ax.scatter(self.particles[0, ::-1],
self.particles[1, ::-1],
c=self.particles[2, ::-1],
**opts)
if type(scale) is float:
ax.set_xlim([-scale, scale])
ax.set_ylim([-scale * H / W, scale * H / W])
elif not scale is 'auto':
ax.set_xlim([-self.total_width, self.total_width])
ax.set_ylim([-self.total_width * H / W, self.total_width * H / W])
else:
ax.set_xlim([
min(self.particles[0, :].min(),
self.particles[1, :].min() / H * W),
max(self.particles[0, :].max(),
self.particles[1, :].max() / H * W)
])
ax.set_ylim([
min(self.particles[1, :].min(),
self.particles[0, :].min() * H / W),
max(self.particles[1, :].max(),
self.particles[0, :].max() * H / W)
])
ax.axis('off')
return fig, ax
def animate(self,
fps=10,
W=1000,
H=618,
duration=20,
scale='auto',
fname=None):
if fname is None:
import tempfile
fname = tempfile.mktemp() + '.mp4'
import matplotlib.pyplot as plt
self.dt = 1. / fps
inches_per_pt = 1.0 / 72.27
from moviepy.video.io.bindings import mplfig_to_npimage
import moviepy.editor as mpy
def make_frame_mpl(t):
self.t = t
self.update()
fig = plt.figure(figsize=(W * inches_per_pt, H * inches_per_pt))
fig, ax = self.plot_structure(fig=fig, ax=None, scale=scale)
#ax.clear()
ax.axis('off')
#fig, ax = self.plot_structure(fig=fig, ax=ax)
return mplfig_to_npimage(fig) # RGB image of the figure
animation = mpy.VideoClip(make_frame_mpl, duration=duration)
plt.close('all')
animation.write_videofile(fname, fps=fps)
return mpy.ipython_display(fname, fps=fps, loop=1, autoplay=1, width=W)
def show_edges(self, fig=None, a=None):
self.N_theta = 12
self.thetas = np.linspace(0, np.pi, self.N_theta)
self.sf_0 = .3
self.B_sf = .3
"""
Shows the quiver plot of a set of edges, optionally associated to an image.
"""
import pylab
import matplotlib.cm as cm
if fig == None:
fig = pylab.figure(figsize=(self.figsize, self.figsize))
if a == None:
border = 0.0
a = fig.add_axes(
(border, border, 1. - 2 * border, 1. - 2 * border), axisbg='w')
else:
self.update_lines()
marge = self.lame_length * 3.
a.axis(self.lames_minmax + np.array([-marge, +marge, -marge, +marge]))
a.add_collection(self.lines)
a.axis(c='b', lw=0)
pylab.setp(a, xticks=[])
pylab.setp(a, yticks=[])
pylab.draw()
return fig, a
#def set_lines(self):
#from matplotlib.collections import LineCollection
#import matplotlib.patches as patches
# draw the segments
#segments, colors, linewidths = list(), list(), list()
#
#X, Y, Theta = self.lames[0, :], self.lames[1, :].real, self.lames[2, :]
#for x, y, theta in zip(X, Y, Theta):
#u_, v_ = np.cos(theta)*self.lame_length, np.sin(theta)*self.lame_length
#segments.append([(x - u_, y - v_), (x + u_, y + v_)])
#colors.append((0, 0, 0, 1))# black
#linewidths.append(self.line_width)
#return LineCollection(segments, linewidths=linewidths, colors=colors, linestyles='solid')
def update_lines(self):
from matplotlib.collections import LineCollection
import matplotlib.patches as patches
X, Y, Theta = self.lames[0, :], self.lames[1, :], self.lames[2, :]
segments = list()
for i, (x, y, theta) in enumerate(zip(X, Y, Theta)):
u_, v_ = np.cos(theta) * self.lame_length, np.sin(
theta) * self.lame_length
segments.append([(x - u_, y - v_), (x + u_, y + v_)])
self.lines.set_segments(segments)
def fname(self, name):
return os.path.join(self.figpath, name + self.vext)
def make_anim(self, name, make_lames, duration=3., redo=False):
if redo or not os.path.isfile(self.fname(name)):
import matplotlib.pyplot as plt
from moviepy.video.io.bindings import mplfig_to_npimage
import moviepy.editor as mpy
fig_mpl, ax = plt.subplots(1,
figsize=(self.figsize, self.figsize),
facecolor='white')
def make_frame_mpl(t):
# on ne peut changer que l'orientation des lames:
self.t = t
self.lames[2, :] = make_lames(self)
self.update_lines()
fig_mpl, ax = self.show_edges() #fig_mpl, ax)
self.t_old = t
return mplfig_to_npimage(fig_mpl) # RGB image of the figure
animation = mpy.VideoClip(make_frame_mpl, duration=duration)
animation.write_videofile(self.fname(name), fps=self.fps)
def __init__(self, pe):
Image.__init__(self, pe)
self.init_logging(name='LogGabor')
class SHL(object):
"""
Base class to define SHL experiments:
- intializing
- running learning
- visualization
- quantitative analysis
"""
def __init__(self,
height=256,
width=256,
patch_size=(10, 10),
n_components=11**2,
learning_algorithm='omp',
transform_n_nonzero_coefs=20,
n_iter=50000,
eta=1./25,
eta_homeo=0.001,
alpha_homeo=0.02,
max_patches=10000,
batch_size=100,
n_image=200,
DEBUG_DOWNSCALE=1, # set to 10 to perform a rapid experiment<D-d>
verbose=20,
):
self.height = height
self.width = width
self.patch_size = patch_size
self.n_components = n_components
self.n_iter = int(n_iter/DEBUG_DOWNSCALE)
self.max_patches = int(max_patches/DEBUG_DOWNSCALE)
self.n_image = int(n_image/DEBUG_DOWNSCALE)
self.batch_size = batch_size
self.learning_algorithm = learning_algorithm
self.transform_n_nonzero_coefs = transform_n_nonzero_coefs
self.eta = eta
self.eta_homeo = eta_homeo
self.alpha_homeo = alpha_homeo
self.verbose = verbose
# Load natural images and extract patches
self.slip = Image(ParameterSet({'N_X':height, 'N_Y':width,
'white_n_learning' : 0,
'seed': None,
'white_N' : .07,
'white_N_0' : .0, # olshausen = 0.
'white_f_0' : .4, # olshausen = 0.2
'white_alpha' : 1.4,
'white_steepness' : 4.,
'datapath': '/Users/lolo/pool/science/PerrinetBednar15/database/',
'do_mask':True,
'N_image': n_image}))
def get_data(self, name_database='serre07_distractors', seed=None):
if self.verbose:
# setup toolbar
sys.stdout.write('Extracting data...')
sys.stdout.flush()
sys.stdout.write("\b" * (toolbar_width+1)) # return to start of line, after '['
t0 = time.time()
imagelist = self.slip.make_imagelist(name_database=name_database)#, seed=seed)
for filename, croparea in imagelist:
# whitening
image, filename_, croparea_ = self.slip.patch(name_database, filename=filename, croparea=croparea, center=False)#, , seed=seed)
image = self.slip.whitening(image)
# Extract all reference patches
data_ = extract_patches_2d(image, self.patch_size, max_patches=int(self.max_patches))#, seed=seed)
data_ = data_.reshape(data_.shape[0], -1)
data_ -= np.mean(data_, axis=0)
data_ /= np.std(data_, axis=0)
try:
data = np.vstack((data, data_))
except:
data = data_.copy()
if self.verbose:
# update the bar
sys.stdout.write(filename + ", ")
sys.stdout.flush()
if self.verbose:
dt = time.time() - t0
sys.stdout.write("\n")
sys.stdout.write("Data is of shape : "+ str(data.shape))
sys.stdout.write('done in %.2fs.' % dt)
sys.stdout.flush()
return data
def learn_dico(self, name_database='serre07_distractors', **kwargs):
data = self.get_data(name_database)
# Learn the dictionary from reference patches
if self.verbose: print('Learning the dictionary...', end=' ')
t0 = time.time()
dico = SparseHebbianLearning(eta=self.eta,
n_components=self.n_components, n_iter=self.n_iter,
gain_rate=self.eta_homeo, alpha_homeo=self.alpha_homeo,
transform_n_nonzero_coefs=self.transform_n_nonzero_coefs,
batch_size=self.batch_size, verbose=self.verbose,
transform_algorithm=self.learning_algorithm,
**kwargs)
if self.verbose: print('Training on %d patches' % len(data), end='... ')
dico.fit(data)
if self.verbose:
dt = time.time() - t0
print('done in %.2fs.' % dt)
return dico
def show_dico(self, dico):
subplotpars = matplotlib.figure.SubplotParams(left=0., right=1., bottom=0., top=1., wspace=0.05, hspace=0.05,)
fig = plt.figure(figsize=(10, 10), subplotpars=subplotpars)
for i, comp in enumerate(dico.components_):
ax = fig.add_subplot(np.sqrt(self.n_components), np.sqrt(self.n_components), i + 1)
cmax = np.max(np.abs(comp))
ax.imshow(comp.reshape(self.patch_size), cmap=plt.cm.gray_r, vmin=-cmax, vmax=+cmax,
interpolation='nearest')
ax.set_xticks(())
ax.set_yticks(())
# fig.suptitle('Dictionary learned from image patches\n' +
# 'Using ' + learning_algorithm.replace('_', ' '),
# fontsize=12)
#fig.tight_layout(rect=[0, 0, .9, 1])
return fig
def code(self, data, dico, intercept, coding_algorithm='omp', **kwargs):
if self.verbose:
print('Coding data...', end=' ')
t0 = time.time()
dico.set_params(transform_algorithm=coding_algorithm, **kwargs)
code = dico.transform(data)
V = dico.components_
patches = np.dot(code, V)
if coding_algorithm == 'threshold':
patches -= patches.min()
patches /= patches.max()
patches += intercept
patches = patches.reshape(len(data), *self.patch_size)
if coding_algorithm == 'threshold':
patches -= patches.min()
patches /= patches.max()
if self.verbose:
dt = time.time() - t0
print('done in %.2fs.' % dt)
return patches
def get_data(height=256,
width=256,
n_image=200,
patch_size=(12, 12),
datapath='database/',
name_database='serre07_distractors',
max_patches=1024,
seed=None,
patch_norm=True,
verbose=0,
data_cache='/tmp/data_cache',
matname=None):
"""
Extract data:
Extract from a given database composed of image of size (height, width) a
series a random patches.
"""
if matname is None:
# Load natural images and extract patches
from SLIP import Image
slip = Image({
'N_X': height,
'N_Y': width,
'white_n_learning': 0,
'seed': seed,
'white_N': .07,
'white_N_0': .0, # olshausen = 0.
'white_f_0': .4, # olshausen = 0.2
'white_alpha': 1.4,
'white_steepness': 4.,
'datapath': datapath,
'do_mask': True,
'N_image': n_image
})
if verbose:
import sys
# setup toolbar
sys.stdout.write('Extracting data...')
sys.stdout.flush()
sys.stdout.write(
"\b" *
(toolbar_width + 1)) # return to start of line, after '['
t0 = time.time()
import os
imagelist = slip.make_imagelist(
name_database=name_database) #, seed=seed)
for filename, croparea in imagelist:
# whitening
image, filename_, croparea_ = slip.patch(
name_database,
filename=filename,
croparea=croparea,
center=False) #, seed=seed)
image = slip.whitening(image)
# Extract all reference patches and ravel them
data_ = slip.extract_patches_2d(
image, patch_size, N_patches=int(max_patches)) #, seed=seed)
data_ = data_.reshape(data_.shape[0], -1)
data_ -= np.mean(data_, axis=0)
if patch_norm:
data_ /= np.std(data_, axis=0)
# collect everything as a matrix
try:
data = np.vstack((data, data_))
except Exception:
data = data_.copy()
if verbose:
# update the bar
sys.stdout.write(filename + ", ")
sys.stdout.flush()
if verbose:
dt = time.time() - t0
sys.stdout.write("\n")
sys.stdout.write("Data is of shape : " + str(data.shape))
sys.stdout.write(' - done in %.2fs.' % dt)
sys.stdout.flush()
else:
import os
fmatname = os.path.join(data_cache, matname)
if not (os.path.isfile(fmatname + '_data.npy')):
if not (os.path.isfile(fmatname + '_data' + '_lock')):
touch(fmatname + '_data' + '_lock')
try:
if verbose:
print('No cache found {}: Extracting data...'.format(
fmatname + '_data'),
end=' ')
print(datapath)
data = get_data(height=height,
width=width,
n_image=n_image,
patch_size=patch_size,
datapath=datapath,
name_database=name_database,
max_patches=max_patches,
seed=seed,
patch_norm=patch_norm,
verbose=verbose,
matname=None)
np.save(fmatname + '_data.npy', data)
finally:
try:
os.remove(fmatname + '_data' + '_lock')
except:
print('Coud not remove ', fmatname + '_data')
else:
print('the data extraction is locked', fmatname + '_data')
return 'lock'
else:
if verbose:
print("loading the data called : {0}".format(fmatname +
'_data'))
# Une seule fois mp ici
data = np.load(fmatname + '_data.npy')
return data
FT_lg = self.loggabor(u, v, sf_0, B_sf, theta, B_theta)
fig, a1, a2 = self.im.show_FT(FT_lg * np.exp(-1j * phase))
return fig, a1, a2
def _test():
import doctest
doctest.testmod()
#####################################
#
if __name__ == '__main__':
_test()
#### Main
"""
Some examples of use for the class
"""
from pylab import imread
image = imread('database/lena512.png')[:, :, 0]
from NeuroTools.parameters import ParameterSet
pe = ParameterSet('default_param.py')
pe.N_X, pe.N_Y = image.shape
from SLIP import Image
im = Image(pe)
lg = LogGabor(im)
class SHL(object):
"""
Base class to define SHL experiments:
- intializing
- running learning
- visualization
- quantitative analysis
"""
def __init__(self,
height=256,
width=256,
patch_size=(12, 12),
database = 'database/',
n_components=14**2,
learning_algorithm='omp',
alpha=None,
transform_n_nonzero_coefs=20,
n_iter=5000,
eta=.01,
eta_homeo=.01,
alpha_homeo=.01,
max_patches=1000,
batch_size=100,
n_image=200,
DEBUG_DOWNSCALE=1, # set to 10 to perform a rapid experiment
verbose=0,
):
self.height = height
self.width = width
self.database = database
self.patch_size = patch_size
self.n_components = n_components
self.n_iter = int(n_iter/DEBUG_DOWNSCALE)
self.max_patches = int(max_patches/DEBUG_DOWNSCALE)
self.n_image = int(n_image/DEBUG_DOWNSCALE)
self.batch_size = batch_size
self.learning_algorithm = learning_algorithm
self.alpha=alpha
self.transform_n_nonzero_coefs = transform_n_nonzero_coefs
self.eta = eta
self.eta_homeo = eta_homeo
self.alpha_homeo = alpha_homeo
self.verbose = verbose
# Load natural images and extract patches
self.slip = Image({'N_X':height, 'N_Y':width,
'white_n_learning' : 0,
'seed': None,
'white_N' : .07,
'white_N_0' : .0, # olshausen = 0.
'white_f_0' : .4, # olshausen = 0.2
'white_alpha' : 1.4,
'white_steepness' : 4.,
'datapath': self.database,
'do_mask':True,
'N_image': n_image})
def get_data(self, name_database='serre07_distractors', seed=None, patch_norm=True):
if self.verbose:
# setup toolbar
sys.stdout.write('Extracting data...')
sys.stdout.flush()
sys.stdout.write("\b" * (toolbar_width+1)) # return to start of line, after '['
t0 = time.time()
imagelist = self.slip.make_imagelist(name_database=name_database)#, seed=seed)
for filename, croparea in imagelist:
# whitening
image, filename_, croparea_ = self.slip.patch(name_database, filename=filename, croparea=croparea, center=False)#, , seed=seed)
image = self.slip.whitening(image)
# Extract all reference patches and ravel them
data_ = extract_patches_2d(image, self.patch_size, max_patches=int(self.max_patches))#, seed=seed)
data_ = data_.reshape(data_.shape[0], -1)
data_ -= np.mean(data_, axis=0)
if patch_norm:
data_ /= np.std(data_, axis=0)
# collect everything as a matrix
try:
data = np.vstack((data, data_))
except:
data = data_.copy()
if self.verbose:
# update the bar
sys.stdout.write(filename + ", ")
sys.stdout.flush()
if self.verbose:
dt = time.time() - t0
sys.stdout.write("\n")
sys.stdout.write("Data is of shape : "+ str(data.shape))
sys.stdout.write('done in %.2fs.' % dt)
sys.stdout.flush()
return data
def learn_dico(self, name_database='serre07_distractors', **kwargs):
data = self.get_data(name_database)
# Learn the dictionary from reference patches
if self.verbose: print('Learning the dictionary...', end=' ')
t0 = time.time()
dico = SparseHebbianLearning(eta=self.eta,
n_components=self.n_components, n_iter=self.n_iter,
gain_rate=self.eta_homeo, alpha_homeo=self.alpha_homeo,
transform_n_nonzero_coefs=self.transform_n_nonzero_coefs,
batch_size=self.batch_size, verbose=self.verbose, n_jobs=1,
transform_algorithm=self.learning_algorithm, transform_alpha=self.alpha, **kwargs)
if self.verbose: print('Training on %d patches' % len(data), end='... ')
dico.fit(data)
if self.verbose:
dt = time.time() - t0
print('done in %.2fs.' % dt)
return dico
def show_dico(self, dico, title=None, fname=None):
subplotpars = matplotlib.figure.SubplotParams(left=0., right=1., bottom=0., top=1., wspace=0.05, hspace=0.05,)
fig = plt.figure(figsize=(10, 10), subplotpars=subplotpars)
for i, component in enumerate(dico.components_):
ax = fig.add_subplot(np.sqrt(self.n_components), np.sqrt(self.n_components), i + 1)
cmax = np.max(np.abs(component))
ax.imshow(component.reshape(self.patch_size), cmap=plt.cm.gray_r, vmin=-cmax, vmax=+cmax,
interpolation='nearest')
ax.set_xticks(())
ax.set_yticks(())
if title is not None:
fig.suptitle(title, fontsize=12, backgroundcolor = 'white', color = 'k')
#fig.tight_layout(rect=[0, 0, .9, 1])
if not fname is None: fig.savefig(fname, dpi=200)
return fig, ax
def code(self, data, dico, intercept=0., coding_algorithm='omp', **kwargs):
if self.verbose:
print('Coding data...', end=' ')
t0 = time.time()
dico.set_params(transform_algorithm=coding_algorithm, **kwargs)
code = dico.transform(data)
V = dico.components_
patches = np.dot(code, V)
if coding_algorithm == 'threshold':
patches -= patches.min()
patches /= patches.max()
patches += intercept
# patches = patches.reshape(len(data), *self.patch_size)
if coding_algorithm == 'threshold':
patches -= patches.min()
patches /= patches.max()
if self.verbose:
dt = time.time() - t0
print('done in %.2fs.' % dt)
return patches
def plot_variance(self, dico, name_database='serre07_distractors', fname=None):
data = self.get_data(name_database)
code = self.code(data, dico)
Z = np.mean(code**2)
fig = plt.figure(figsize=(12, 4))
ax = fig.add_subplot(111)
ax.bar(np.arange(self.n_components), np.mean(code**2/Z, axis=0))#, yerr=np.std(code**2/Z, axis=0))
ax.set_title('Variance of coefficients')
ax.set_ylabel('Variance')
ax.set_xlabel('#')
ax.axis('tight')
if not fname is None: fig.savefig(fname, dpi=200)
return fig, ax
def plot_variance_histogram(self, dico, name_database='serre07_distractors', fname=None):
data = self.get_data(name_database)
import pandas as pd
import seaborn as sns
code = self.code(data, dico)
fig = plt.figure(figsize=(6, 4))
ax = fig.add_subplot(111)
data = pd.DataFrame(np.mean(code**2, axis=0)/np.mean(code**2), columns=['Variance'])
with sns.axes_style("white"):
ax = sns.distplot(data['Variance'], kde_kws={'clip':(0., 5.)})
ax.set_title('distribution of the mean variance of coefficients')
ax.set_ylabel('pdf')
if not fname is None: fig.savefig(fname, dpi=200)
return fig, ax
