def __init__(self, num_subfunctions_ratio=0.05):
self.name = 'GLM'
print('Initializing parameters...')
#self.params = glm.setParameters(m=5000, dh=10, ds=50) # small
self.params = glm.setParameters(m=5000, dh=10, ds=49) # small
#self.params = glm.setParameters(m=20000, dh=100, ds=500) # large
#self.params = glm.setParameters(m=20000, dh=25, ds=256, n=5) # huge
#self.params = glm.setParameters(m=1e5, dh=10, ds=256, n=50) # Jascha huge
#self.params = glm.setParameters(m=1e5, n=100, ds=256, dh=10) # shared
## initialize parameters
print('Generating model...')
self.theta_true = glm.generateModel(self.params)
self.theta_init = glm.generateModel(self.params)
for key in self.theta_init.keys():
self.theta_init[key] /= 1e3
print('Simulating model to generate data...')
self.data = glm.generateData(self.theta_true, self.params)
batch_size = self.data['x'].shape[0]
trueMinimum, trueGrad = self.f_df(self.theta_true, (0, batch_size))
print('Minimum for true parameters %g'%(trueMinimum))
print('Norm of the gradient at the minimum:')
for key in trueGrad.keys():
print('grad[' + key + ']: %g'%(np.linalg.norm(trueGrad[key])))
# print out some network information
glm.visualizeNetwork(self.theta_true, self.params, self.data)
# break the data up into minibatches
self.N = int(np.ceil(np.sqrt(batch_size)*num_subfunctions_ratio))
self.subfunction_references = []
samples_per_subfunction = int(np.floor(batch_size/self.N))
for mb in range(self.N):
start_idx = mb*samples_per_subfunction
end_idx = (mb+1)*samples_per_subfunction
self.subfunction_references.append((start_idx, end_idx,))
self.full_objective_references = self.subfunction_references
def __init__(self, num_subfunctions_ratio=0.05, baseDir='/home/nirum/data/retina/glm-feb-19/'):
#def __init__(self, num_subfunctions_ratio=0.05, baseDir='/home/nirum/data/retina/glm-feb-19/small'):
self.name = 'GLM'
print('Initializing parameters...')
## FOR CUSTOM SIZES
#self.params = glm.setParameters(m=5000, dh=10, ds=50) # small
#self.params = glm.setParameters(m=5000, dh=10, ds=49) # small
#self.params = glm.setParameters(m=20000, dh=100, ds=500) # large
#self.params = glm.setParameters(m=20000, dh=25, ds=256, n=5) # huge
#self.params = glm.setParameters(m=1e5, dh=10, ds=256, n=50) # Jascha huge
#self.params = glm.setParameters(m=1e5, n=100, ds=256, dh=10) # shared
## FROM EXTERNAL DATA FILES
# load sizes of external data
shapes = np.load(join(baseDir, 'metadata.npz'))
# set up GLM parameters
self.params = glm.setParameters(dh=40, ds=shapes['stimSlicedShape'][1], n=shapes['rateShape'][1])
#self.params = glm.setParameters(dh=40, ds=shapes['stimSlicedShape'][1], n=2)
## initialize parameters
print('Generating model...')
#self.theta_true = glm.generateModel(self.params)
self.theta_init = glm.generateModel(self.params)
for key in self.theta_init.keys():
self.theta_init[key] /= 1e3
#print('Simulating model to generate data...')
#self.data = glm.generateData(self.theta_true, self.params)
# load external data files as memmapped arrays
print('Loading external data...')
self.data = glm.loadExternalData('stimulus_sliced.dat', 'rates.dat', shapes, baseDir=baseDir)
# all the data
batch_size = self.data['x'].shape[0]
#batch_size = 10000
print('ready to go!')
#trueMinimum, trueGrad = self.f_df(self.theta_true, (0, batch_size))
#print('Minimum for true parameters %g'%(trueMinimum))
#print('Norm of the gradient at the minimum:')
#for key in trueGrad.keys():
#print('grad[' + key + ']: %g'%(np.linalg.norm(trueGrad[key])))
# print out some network information
#glm.visualizeNetwork(self.theta_true, self.params, self.data)
# break the data up into minibatches
self.N = int(np.ceil(np.sqrt(batch_size)*num_subfunctions_ratio))
self.subfunction_references = []
samples_per_subfunction = int(np.floor(batch_size/self.N))
for mb in range(self.N):
print(mb, self.N)
start_idx = mb*samples_per_subfunction
end_idx = (mb+1)*samples_per_subfunction
self.subfunction_references.append((start_idx, end_idx,))
self.full_objective_references = self.subfunction_references
print('initialized ...')
def __init__(self,
num_subfunctions_ratio=0.05,
baseDir='/home/nirum/data/retina/glm-feb-19/'):
#def __init__(self, num_subfunctions_ratio=0.05, baseDir='/home/nirum/data/retina/glm-feb-19/small'):
self.name = 'GLM'
print('Initializing parameters...')
## FOR CUSTOM SIZES
#self.params = glm.setParameters(m=5000, dh=10, ds=50) # small
#self.params = glm.setParameters(m=5000, dh=10, ds=49) # small
#self.params = glm.setParameters(m=20000, dh=100, ds=500) # large
#self.params = glm.setParameters(m=20000, dh=25, ds=256, n=5) # huge
#self.params = glm.setParameters(m=1e5, dh=10, ds=256, n=50) # Jascha huge
#self.params = glm.setParameters(m=1e5, n=100, ds=256, dh=10) # shared
## FROM EXTERNAL DATA FILES
# load sizes of external data
shapes = np.load(join(baseDir, 'metadata.npz'))
# set up GLM parameters
self.params = glm.setParameters(dh=40,
ds=shapes['stimSlicedShape'][1],
n=shapes['rateShape'][1])
#self.params = glm.setParameters(dh=40, ds=shapes['stimSlicedShape'][1], n=2)
## initialize parameters
print('Generating model...')
#self.theta_true = glm.generateModel(self.params)
self.theta_init = glm.generateModel(self.params)
for key in self.theta_init.keys():
self.theta_init[key] /= 1e3
#print('Simulating model to generate data...')
#self.data = glm.generateData(self.theta_true, self.params)
# load external data files as memmapped arrays
print('Loading external data...')
self.data = glm.loadExternalData('stimulus_sliced.dat',
'rates.dat',
shapes,
baseDir=baseDir)
# all the data
batch_size = self.data['x'].shape[0]
#batch_size = 10000
print('ready to go!')
#trueMinimum, trueGrad = self.f_df(self.theta_true, (0, batch_size))
#print('Minimum for true parameters %g'%(trueMinimum))
#print('Norm of the gradient at the minimum:')
#for key in trueGrad.keys():
#print('grad[' + key + ']: %g'%(np.linalg.norm(trueGrad[key])))
# print out some network information
#glm.visualizeNetwork(self.theta_true, self.params, self.data)
# break the data up into minibatches
self.N = int(np.ceil(np.sqrt(batch_size) * num_subfunctions_ratio))
self.subfunction_references = []
samples_per_subfunction = int(np.floor(batch_size / self.N))
for mb in range(self.N):
print(mb, self.N)
start_idx = mb * samples_per_subfunction
end_idx = (mb + 1) * samples_per_subfunction
self.subfunction_references.append((
start_idx,
end_idx,
))
self.full_objective_references = self.subfunction_references
print('initialized ...')