def __init__(self,cf,input_sheet,x=0.0,y=0.0,template=BoundingBox(radius=0.1),
mask=patterngenerator.Constant(),
min_matrix_radius=1):
"""
From an existing copy of ConnectionField (CF) that acts as a
template, create a new CF that shares weights with the
template CF. Copies all the properties of CF to stay
identical except the weights variable that actually contains
the data.
The only difference from a normal CF is that the weights of
the CF are implemented as a numpy view into the single master
copy of the weights stored in the CF template.
"""
# CEBALERT: There's no call to super's __init__; see JAHACKALERT
# below.
template = copy(template)
if not isinstance(template,Slice):
template = Slice(template,input_sheet,force_odd=True,
min_matrix_radius=min_matrix_radius)
# Note: if passed in, mask is shared between CFs (but not if created here)
if not hasattr(mask,'view'):
mask = _create_mask(patterngenerator.Constant(),
template.compute_bounds(input_sheet),
input_sheet,True,0.5)
self._has_norm_total=False
self.mask=mask
weights_slice = self._create_input_sheet_slice(input_sheet,x,y,template,min_matrix_radius=min_matrix_radius)
self.weights = weights_slice.submatrix(cf.weights)
def __init__(self,cf,input_sheet,x=0.0,y=0.0,template=BoundingBox(radius=0.1),
mask=patterngenerator.Constant(),
min_matrix_radius=1):
"""
From an existing copy of ConnectionField (CF) that acts as a
template, create a new CF that shares weights with the
template CF. Copies all the properties of CF to stay
identical except the weights variable that actually contains
the data.
The only difference from a normal CF is that the weights of
the CF are implemented as a numpy view into the single master
copy of the weights stored in the CF template.
"""
# CEBALERT: There's no call to super's __init__; see JAHACKALERT
# below.
template = copy(template)
if not isinstance(template,Slice):
template = Slice(template,input_sheet,force_odd=True,
min_matrix_radius=min_matrix_radius)
# Note: if passed in, mask is shared between CFs (but not if created here)
if not hasattr(mask,'view'):
mask = _create_mask(patterngenerator.Constant(),
template.compute_bounds(input_sheet),
input_sheet,True,0.5)
self._has_norm_total=False
self.mask=mask
weights_slice = self._create_input_sheet_slice(input_sheet,x,y,template,min_matrix_radius=min_matrix_radius)
self.weights = weights_slice.submatrix(cf.weights)
def test_connection_field_like(self):
# test a ConnectionField-like example
sheet = Sheet(nominal_density=10,nominal_bounds=BoundingBox(radius=0.5))
cf_bounds = BoundingBox(points=((0.3,0.3),(0.6,0.6)))
slice_ = Slice(cf_bounds,sheet)
slice_.crop_to_sheet(sheet)
# check it's been cropped to fit onto sheet...
self.assertEqual(slice_.tolist(),[0,2,8,10])
# now check that it gives the correct bounds...
cropped_bounds = slice_.compute_bounds(sheet)
true_cropped_bounds = BoundingBox(points=((0.3,0.3),(0.5,0.5)))
for a,b in zip(cropped_bounds.lbrt(),true_cropped_bounds.lbrt()):
self.assertAlmostEqual(a,b)
def test_connection_field_like(self):
# test a ConnectionField-like example
sheet = Sheet(nominal_density=10,nominal_bounds=BoundingBox(radius=0.5))
cf_bounds = BoundingBox(points=((0.3,0.3),(0.6,0.6)))
slice_ = Slice(cf_bounds,sheet)
slice_.crop_to_sheet(sheet)
# check it's been cropped to fit onto sheet...
self.assertEqual(slice_.tolist(),[0,2,8,10])
# now check that it gives the correct bounds...
cropped_bounds = slice_.compute_bounds(sheet)
true_cropped_bounds = BoundingBox(points=((0.3,0.3),(0.5,0.5)))
for a,b in zip(cropped_bounds.lbrt(),true_cropped_bounds.lbrt()):
self.assertAlmostEqual(a,b)
class SparseCFProjection(CFProjection):
"""
A projection composed of SparseConnectionFields from a Sheet into
a ProjectionSheet.
SparseCFProjection computes its activity using a response_fn which
can either be an optimized function implemented as part of the
sparse matrix class or an unoptimized function, which requests the
weights in dense format. The initial contents of the
SparseConnectionFields mapping from the input Sheet into the
target ProjectionSheet are controlled by the weights_generator,
cf_shape, and weights_output_fn parameters, while the location of
the ConnectionField is controlled by the coord_mapper parameter.
Any subclass has to implement the interface activate(self) that
computes the response from the input and stores it in the activity
array.
"""
cf_type = param.Parameter(default=SparseConnectionField,doc="""
Type of ConnectionField to use when creating individual CFs.""")
learning_fn = param.Callable(default=CFPLF_Hebbian_Sparse,doc="""
Function for computing changes to the weights based on one activation step.""")
response_fn = param.Callable(default=CFPRF_DotProduct_Sparse,doc="""
Function for computing the Projection response to an input pattern.""")
weights_output_fns = param.HookList(default=[CFPOF_DivisiveNormalizeL1_Sparse],doc="""
Functions applied to each CF after learning.""")
initialized = param.Boolean(default=False)
def __init__(self,initialize_cfs=True,**params):
"""
Initialize the Projection with a set of cf_type objects
(typically SparseConnectionFields), each located at the
location in the source sheet corresponding to the unit in the
target sheet. The cf_type objects are stored in the 'cfs'
array.
The nominal_bounds_template specified may be altered: the
bounds must be fitted to the Sheet's matrix, and the weights
matrix must have odd dimensions. These altered bounds are
passed to the individual connection fields.
A mask for the weights matrix is constructed. The shape is
specified by cf_shape; the size defaults to the size
of the nominal_bounds_template.
"""
super(CFProjection,self).__init__(**params)
self.weights_generator.set_dynamic_time_fn(None,sublistattr='generators')
# get the actual bounds_template by adjusting a copy of the
# nominal_bounds_template to ensure an odd slice, and to be
# cropped to sheet if necessary
self._slice_template = Slice(copy(self.nominal_bounds_template),
self.src,force_odd=True,
min_matrix_radius=self.min_matrix_radius)
self.bounds_template = self._slice_template.compute_bounds(self.src)
self.mask_template = _create_mask(self.cf_shape,self.bounds_template,
self.src,self.autosize_mask,
self.mask_threshold)
self.n_units = self._calc_n_units()
self.activity = np.array(self.dest.activity)
self.norm_total = np.array(self.dest.activity,dtype=np.float64)
self.has_norm_total = False
if initialize_cfs:
self._create_cfs()
self.input_buffer = None
def __getstate__(self):
"""
Method to support pickling of sparse weights object.
"""
state_dict = self.__dict__.copy()
state_dict['triplets'] = state_dict['weights'].getTriplets()
state_dict['weight_shape'] = (self.src.activity.shape,self.dest.activity.shape)
del state_dict['weights']
return state_dict
def __setstate__(self,state_dict):
"""
Method to support unpickling of sparse weights object.
"""
self.__dict__.update(state_dict)
self.weights = sparse.csarray_float(self.weight_shape[0],self.weight_shape[1])
rowInds, colInds, values = self.triplets
self.weights.setTriplets(rowInds,colInds,values)
del self.triplets
del self.weight_shape
def _create_cfs(self):
"""
Creates the CF objects, initializing the weights one by one
and adding them to the sparse weights object in chunks.
"""
vectorized_create_cf = simple_vectorize(self._create_cf)
self.cfs = vectorized_create_cf(*self._generate_coords())
self.flatcfs = list(self.cfs.flat)
self.weights = sparse.csarray_float(self.src.activity.shape,self.dest.activity.shape)
cf_x,cf_y = self.dest.activity.shape
src_x,src_y = self.src.activity.shape
y_array = np.zeros((src_x*src_y*cf_y),dtype=np.int32)
x_array = np.zeros((src_x*src_y*cf_y),dtype=np.int32)
val_array = np.zeros((src_x*src_y*cf_y),dtype=np.float32)
# Iterate over the CFs
for x in range(cf_x):
temp_sparse = sparse.csarray_float(self.src.activity.shape,self.dest.activity.shape)
idx = 0
for y in range(cf_y):
x1,x2,y1,y2 = self.cfs[x][y].input_sheet_slice.tolist()
if self.same_cf_shape_for_all_cfs:
mask_template = self.mask_template
else:
mask_template = _create_mask(self.cf_shape,self.bounds_template,
self.src,self.autosize_mask,
self.mask_threshold)
weights = self.cfs[x][y]._init_weights(mask_template)
cn_x,cn_y = weights.shape
y_val = x * cf_y + y
for cnx in range(cn_x):
val_array[idx:idx+cn_y] = weights[cnx,:]
x_val = (x1+cnx) * src_y + y1
x_array[idx:idx+cn_y] = range(x_val,x_val+cn_y)
y_array[idx:idx+cn_y] = y_val
idx += cn_y
nnz_idx = val_array.nonzero()
temp_sparse.setTriplets(x_array[nnz_idx],y_array[nnz_idx],val_array[nnz_idx])
self.weights += temp_sparse
x_array *= 0; y_array *= 0; val_array *= 0.0
del temp_sparse
self.weights.compress()
self.apply_learn_output_fns()
print self.name , "loaded"
def _create_cf(self,x,y):
"""
Create a ConnectionField at x,y in the src sheet.
"""
try:
CF = self.cf_type(template=self._slice_template,projection=self,input_sheet=self.src,x=x,y=y,
weights_generator=self.weights_generator,
min_matrix_radius=self.min_matrix_radius)
except NullCFError:
if self.allow_null_cfs:
CF = None
else:
raise
return CF
def activate(self,input_activity):
"""Activate using the specified response_fn and output_fn."""
if self.input_fns:
input_activity = input_activity.copy()
for iaf in self.input_fns:
iaf(input_activity)
self.input_buffer = input_activity
self.activity *=0.0
self.response_fn(self)
for of in self.output_fns:
of(self.activity)
def learn(self):
"""
For a SparseCFProjection, learn consists of calling the learning_fn.
"""
# Learning is performed if the input_buffer has already been set,
# i.e. there is an input to the Projection.
if self.input_buffer != None:
self.learning_fn(self)
def apply_learn_output_fns(self,active_units_mask=True):
"""
Apply the weights_output_fns to each unit.
"""
for of in self.weights_output_fns: of(self)
def n_bytes(self):
"""
Estimates the size on the basis of the number non-zeros in the
sparse matrix, asssuming indices and values are stored using
32-bit integers and floats respectively.
"""
return self.n_conns() * (3 * 4)
def n_conns(self):
"""
Returns number of nonzero weights.
"""
return self.weights.getnnz()
class SparseCFProjection(CFProjection):
"""
A projection composed of SparseConnectionFields from a Sheet into
a ProjectionSheet.
SparseCFProjection computes its activity using a response_fn which
can either be an optimized function implemented as part of the
sparse matrix class or an unoptimized function, which requests the
weights in dense format. The initial contents of the
SparseConnectionFields mapping from the input Sheet into the
target ProjectionSheet are controlled by the weights_generator,
cf_shape, and weights_output_fn parameters, while the location of
the ConnectionField is controlled by the coord_mapper parameter.
Any subclass has to implement the interface activate(self) that
computes the response from the input and stores it in the activity
array.
"""
cf_type = param.Parameter(default=SparseConnectionField,doc="""
Type of ConnectionField to use when creating individual CFs.""")
learning_fn = param.Callable(default=CFPLF_Hebbian_Sparse,doc="""
Function for computing changes to the weights based on one activation step.""")
response_fn = param.Callable(default=CFPRF_DotProduct_Sparse,doc="""
Function for computing the Projection response to an input pattern.""")
weights_output_fns = param.HookList(default=[CFPOF_DivisiveNormalizeL1_Sparse],doc="""
Functions applied to each CF after learning.""")
initialized = param.Boolean(default=False)
def __init__(self,initialize_cfs=True,**params):
"""
Initialize the Projection with a set of cf_type objects
(typically SparseConnectionFields), each located at the
location in the source sheet corresponding to the unit in the
target sheet. The cf_type objects are stored in the 'cfs'
array.
The nominal_bounds_template specified may be altered: the
bounds must be fitted to the Sheet's matrix, and the weights
matrix must have odd dimensions. These altered bounds are
passed to the individual connection fields.
A mask for the weights matrix is constructed. The shape is
specified by cf_shape; the size defaults to the size
of the nominal_bounds_template.
"""
super(CFProjection,self).__init__(**params)
self.weights_generator.set_dynamic_time_fn(None,sublistattr='generators')
# get the actual bounds_template by adjusting a copy of the
# nominal_bounds_template to ensure an odd slice, and to be
# cropped to sheet if necessary
self._slice_template = Slice(copy(self.nominal_bounds_template),
self.src,force_odd=True,
min_matrix_radius=self.min_matrix_radius)
self.bounds_template = self._slice_template.compute_bounds(self.src)
self.mask_template = _create_mask(self.cf_shape,self.bounds_template,
self.src,self.autosize_mask,
self.mask_threshold)
self.n_units = self._calc_n_units()
self.activity = np.array(self.dest.activity)
self.norm_total = np.array(self.dest.activity,dtype=np.float64)
self.has_norm_total = False
if initialize_cfs:
self._create_cfs()
if self.apply_output_fns_init:
self.apply_learn_output_fns()
self.input_buffer = None
def __getstate__(self):
"""
Method to support pickling of sparse weights object.
"""
state_dict = self.__dict__.copy()
state_dict['triplets'] = state_dict['weights'].getTriplets()
state_dict['weight_shape'] = (self.src.activity.shape,self.dest.activity.shape)
del state_dict['weights']
return state_dict
def __setstate__(self,state_dict):
"""
Method to support unpickling of sparse weights object.
"""
self.__dict__.update(state_dict)
self.weights = sparse.csarray_float(self.weight_shape[0],self.weight_shape[1])
rowInds, colInds, values = self.triplets
self.weights.setTriplets(rowInds,colInds,values)
del self.triplets
del self.weight_shape
def _create_cfs(self):
"""
Creates the CF objects, initializing the weights one by one
and adding them to the sparse weights object in chunks.
"""
vectorized_create_cf = simple_vectorize(self._create_cf)
self.cfs = vectorized_create_cf(*self._generate_coords())
self.flatcfs = list(self.cfs.flat)
self.weights = sparse.csarray_float(self.src.activity.shape,self.dest.activity.shape)
cf_x,cf_y = self.dest.activity.shape
src_x,src_y = self.src.activity.shape
y_array = np.zeros((src_x*src_y*cf_y),dtype=np.int32)
x_array = np.zeros((src_x*src_y*cf_y),dtype=np.int32)
val_array = np.zeros((src_x*src_y*cf_y),dtype=np.float32)
# Iterate over the CFs
for x in range(cf_x):
temp_sparse = sparse.csarray_float(self.src.activity.shape,self.dest.activity.shape)
idx = 0
for y in range(cf_y):
cf = self.cfs[x][y]
label = cf.label + ('-%d' % self.seed if self.seed is not None else '')
name = "%s_CF (%.5f, %.5f)" % ('' if label is None else label, cf.x,cf.y)
x1,x2,y1,y2 = cf.input_sheet_slice.tolist()
if self.same_cf_shape_for_all_cfs:
mask_template = self.mask_template
else:
mask_template = _create_mask(self.cf_shape,self.bounds_template,
self.src,self.autosize_mask,
self.mask_threshold, name=name)
weights = self.cfs[x][y]._init_weights(mask_template)
cn_x,cn_y = weights.shape
y_val = x * cf_y + y
for cnx in range(cn_x):
val_array[idx:idx+cn_y] = weights[cnx,:]
x_val = (x1+cnx) * src_y + y1
x_array[idx:idx+cn_y] = range(x_val,x_val+cn_y)
y_array[idx:idx+cn_y] = y_val
idx += cn_y
nnz_idx = val_array.nonzero()
temp_sparse.setTriplets(x_array[nnz_idx],y_array[nnz_idx],val_array[nnz_idx])
self.weights += temp_sparse
x_array *= 0; y_array *= 0; val_array *= 0.0
del temp_sparse
self.weights.compress()
self.debug("Sparse projection %r loaded" % self.name)
def _create_cf(self,x,y):
"""
Create a ConnectionField at x,y in the src sheet.
"""
label = self.hash_format.format(name=self.name,
src=self.src.name,
dest=self.dest.name)
try:
CF = self.cf_type(template=self._slice_template,
projection=self,input_sheet=self.src,x=x,y=y,
weights_generator=self.weights_generator,
min_matrix_radius=self.min_matrix_radius,
label=label)
except NullCFError:
if self.allow_null_cfs:
CF = None
else:
raise
return CF
def get_sheet_mask(self):
return np.ones(self.activity.shape, dtype=self.activity.dtype)
def get_active_units_mask(self):
return np.ones(self.activity.shape, dtype=self.activity.dtype)
def activate(self,input_activity):
"""Activate using the specified response_fn and output_fn."""
if self.input_fns:
input_activity = input_activity.copy()
for iaf in self.input_fns:
iaf(input_activity)
self.input_buffer = input_activity
self.activity *=0.0
self.response_fn(self)
for of in self.output_fns:
of(self.activity)
def learn(self):
"""
For a SparseCFProjection, learn consists of calling the learning_fn.
"""
# Learning is performed if the input_buffer has already been set,
# i.e. there is an input to the Projection.
if self.input_buffer is not None:
self.learning_fn(self)
def apply_learn_output_fns(self,active_units_mask=True):
"""
Apply the weights_output_fns to each unit.
"""
for of in self.weights_output_fns: of(self)
def n_bytes(self):
"""
Estimates the size on the basis of the number non-zeros in the
sparse matrix, asssuming indices and values are stored using
32-bit integers and floats respectively.
"""
return self.n_conns() * (3 * 4)
def n_conns(self):
"""
Returns number of nonzero weights.
"""
return self.weights.getnnz()
