def setup(shape):
"""Creates two matrices and corresponding row/column vectors"""
mat = cmt.empty(shape).fill_with_randn()
mat2 = cmt.empty(shape).fill_with_randn()
col = cmt.empty((shape[0], 1)).assign(0)
row = cmt.empty((1, shape[1])).assign(0)
return mat, mat2, col, row
def dream(self, k=10):
"""
Generate a pattern from this network.
Return
------
A new CUDAMatrix with the shape of the first layer
"""
last_layer = self.layers[-1]
v = cm.empty(last_layer.visible_bias.shape)
h = cm.empty(last_layer.hidden_bias.shape)
v_mean = cm.empty(last_layer.visible_bias.shape)
h_mean = cm.empty(last_layer.hidden_bias.shape)
h.fill_with_rand()
for _ in xrange(k):
last_layer.sample_visible(h, v_mean, v)
last_layer.sample_hidden(v, h_mean, h)
v.free_device_memory()
v_mean.free_device_memory()
h_mean.free_device_memory()
return self.reverse_transform(h)
def setup(shape):
"""Creates two matrices and corresponding row/column vectors"""
mat = cmt.empty(shape).fill_with_randn()
mat2 = cmt.empty(shape).fill_with_randn()
col = cmt.empty((shape[0], 1)).assign(0)
row = cmt.empty((1, shape[1])).assign(0)
return mat, mat2, col, row
def __init__(self):
self.m = 10
self.sizes = [3000,50,3]
#self.thetas = map(lambda x: cm.CUDAMatrix(x), neur.create_initial_thetas(self.sizes, 0.12))
self.thetas = map(lambda x: cm.empty((self.sizes[x + 1], self.sizes[x] + 1)).fill_with_rand(), range(len(self.sizes) - 1))
self.activ_layers = map(lambda x: cm.CUDAMatrix(np.zeros((self.m, x + 1))), self.sizes[0:-1])
self.activ_layers.append(cm.CUDAMatrix(np.zeros((self.m, self.sizes[-1]))))
for i in range(len(self.sizes)):
self.activ_layers[i].set_col_slice(0,1,1)
self.activ_layers_temp = map(lambda x: cm.empty((self.m, x + 1)), self.sizes[0:-1])
self.activ_layers_temp.append(cm.empty((self.m, self.sizes[-1])))
self.layer_expand_mask = map(lambda x: cm.CUDAMatrix(np.hstack([np.zeros((x, 1)), np.eye(x)])), self.sizes[0:-1])
size1 = self.sizes[0] + 1
clear = np.zeros((size1,size1))
clear[0,0] = 1
self.clear_vec = cm.CUDAMatrix(clear[0:size1,0:size1])
#print self.clear_vec.shape
self.clear_vec2 = cm.CUDAMatrix(clear[0:51,0:51])
self.z = [0] * (len(self.thetas) + 1)
self.z[1] = cm.empty((self.m, 50))
self.z[2] = cm.empty((self.m, 3))
def AllocateBatchsizeDependentMemory(self, batchsize):
self.batchsize = batchsize
if self.data:
self.data.free_device_memory()
if self.deriv:
self.deriv.free_device_memory()
dimensions = self.dimensions
numlabels = self.numlabels
if self.is_input or self.is_initialized or self.is_output:
if self.activation == deepnet_pb2.Hyperparams.REPLICATED_SOFTMAX:
self.data = cm.CUDAMatrix(np.zeros((numlabels * dimensions, batchsize)))
else:
self.data = cm.CUDAMatrix(np.zeros((dimensions, batchsize)))
self.deriv = cm.CUDAMatrix(np.zeros((numlabels * dimensions, batchsize)))
self.state = cm.CUDAMatrix(np.zeros((numlabels * dimensions, batchsize)))
self.temp = cm.CUDAMatrix(np.zeros((dimensions, batchsize)))
if self.activation == deepnet_pb2.Hyperparams.SOFTMAX or\
self.activation == deepnet_pb2.Hyperparams.REPLICATED_SOFTMAX:
self.expanded_batch = cm.CUDAMatrix(np.zeros(
(numlabels * dimensions, batchsize)))
if self.loss_function == deepnet_pb2.Layer.CROSS_ENTROPY:
if self.activation == deepnet_pb2.Hyperparams.SOFTMAX:
self.temp2 = cm.CUDAMatrix(np.zeros((dimensions, batchsize)))
self.indices = cm.CUDAMatrix(np.zeros((1, dimensions * batchsize)))
self.rowshift = cm.CUDAMatrix(
numlabels*np.arange(dimensions * batchsize).reshape(1, -1))
if self.activation == deepnet_pb2.Hyperparams.REPLICATED_SOFTMAX:
#norm_to = self.hyperparams.normalize_to
self.NN = cm.CUDAMatrix(np.ones((1, batchsize)))
self.counter = cm.empty(self.NN.shape) # Used for sampling softmaxes.
self.count_filter = cm.empty(self.NN.shape) # Used for sampling softmaxes.
if self.hyperparams.dropout:
self.mask = cm.CUDAMatrix(np.zeros(self.state.shape))
def calc_hidden_probs(data, vh, hb, batchsize):
"""
Calculate the probs in the next layer up given data, and weights
"""
if data.shape[0] % batchsize != 0:
print "WARNING! Batchsize for calc_hidden_probs is not an even divisor of example cnt."
print "Calculating probs. of hidden layer, " + str(
data.shape[0]) + " examples."
dev_data = cm.CUDAMatrix(data)
lrp_data = np.empty((data.shape[0], vh.shape[1]))
cur_data = cm.empty((batchsize, vh.shape[0]))
nex_data = cm.empty((batchsize, vh.shape[1]))
vishid = cm.CUDAMatrix(vh)
hid_bias = cm.CUDAMatrix(hb)
num_batches = data.shape[0] / batchsize
for batch in range(0, num_batches):
cur_data = dev_data.get_row_slice(batch * batchsize,
(batch + 1) * batchsize)
cm.dot(cur_data, vishid, target=nex_data)
nex_data.add_row_vec(hid_bias)
nex_data.apply_sigmoid()
nex_data.copy_to_host()
lrp_data[batch * batchsize:(batch + 1) *
batchsize, :] = nex_data.asarray().copy()
return lrp_data
def __init__(self,
input_dim,
output_dim,
spectral_radius=.9,
leak_rate=1,
input_scaling=1,
bias_scaling=0):
super(CUDAReservoirNode, self).__init__(
input_dim=input_dim,
output_dim=output_dim,
)
self.input_dim = input_dim
self.output_dim = output_dim
self.leak_rate = leak_rate
w = mdp.numx.random.normal(0, 1, (output_dim, output_dim))
w_in = mdp.numx.random.uniform(-1, 1,
(output_dim, input_dim)) * input_scaling
if output_dim < 1500:
l = mdp.numx.linalg.eigvals(w)
r = mdp.numx.amax(mdp.numx.absolute(l))
w = w * (spectral_radius / r)
self.w = cm.CUDAMatrix(w)
else:
self.w = cm.CUDAMatrix(w)
r = get_specrad(self.w)
self.w.mult(spectral_radius / r)
bias = mdp.numx.random.normal(0, 1, (output_dim, 1)) * bias_scaling
self.w_in = cm.CUDAMatrix(w_in)
self.bias = cm.CUDAMatrix(bias)
self.current_state = cm.empty((self.output_dim, 1))
self.new_state = cm.empty((self.output_dim, 1))
def __init__(self):
self.m = 10
self.sizes = [3000, 50, 3]
#self.thetas = map(lambda x: cm.CUDAMatrix(x), neur.create_initial_thetas(self.sizes, 0.12))
self.thetas = map(
lambda x: cm.empty(
(self.sizes[x + 1], self.sizes[x] + 1)).fill_with_rand(),
range(len(self.sizes) - 1))
self.activ_layers = map(
lambda x: cm.CUDAMatrix(np.zeros((self.m, x + 1))),
self.sizes[0:-1])
self.activ_layers.append(
cm.CUDAMatrix(np.zeros((self.m, self.sizes[-1]))))
for i in range(len(self.sizes)):
self.activ_layers[i].set_col_slice(0, 1, 1)
self.activ_layers_temp = map(lambda x: cm.empty((self.m, x + 1)),
self.sizes[0:-1])
self.activ_layers_temp.append(cm.empty((self.m, self.sizes[-1])))
self.layer_expand_mask = map(
lambda x: cm.CUDAMatrix(np.hstack([np.zeros(
(x, 1)), np.eye(x)])), self.sizes[0:-1])
size1 = self.sizes[0] + 1
clear = np.zeros((size1, size1))
clear[0, 0] = 1
self.clear_vec = cm.CUDAMatrix(clear[0:size1, 0:size1])
#print self.clear_vec.shape
self.clear_vec2 = cm.CUDAMatrix(clear[0:51, 0:51])
self.z = [0] * (len(self.thetas) + 1)
self.z[1] = cm.empty((self.m, 50))
self.z[2] = cm.empty((self.m, 3))
def _sample_h(self, v, x, sample=False, x_is_bias=False):
# updates self.h
#
self.h = cm.empty((v.shape[0], self.output_dim))
if x_is_bias: # Bias is precalculated
self.h.assign(x)
else:
cm.dot(x, self.bg, self.h)
self.h.add_dot(v, self.wg)
# This is a 100 times faster than calling 'add_row_vec' to add biases.
ones_cut = self._ones.get_col_slice(0, v.shape[0])
self.h.add_dot(ones_cut.T, self.bhg)
self.h.apply_sigmoid2(self.h)
if sample:
# Sample random values
sampled = cm.empty((v.shape[0], self.output_dim))
sampled.fill_with_rand()
# Sample values of hiddens
sampled.less_than(self.h, self.h)
def test_allfinite():
a = cm.empty((10, 20)).assign(1).divide(0) # NaN
b = cm.empty((10, 20)).assign(1e20).mult(1e20) # Inf
c = cm.empty((10, 20)).assign(1) # 1.0
assert (not a.allfinite()) and (not b.allfinite()) and c.allfinite(
), "CUDAMatrix.allfinite does not work"
def _sample_h(self, v, x, sample=False, x_is_bias=False):
# updates self.h
#
self.h = cm.empty((v.shape[0], self.output_dim))
if x_is_bias: # Bias is precalculated
self.h.assign(x)
else:
cm.dot(x, self.bg, self.h)
self.h.add_dot(v, self.wg)
# This is a 100 times faster than calling 'add_row_vec' to add biases.
ones_cut = self._ones.get_col_slice(0, v.shape[0])
self.h.add_dot(ones_cut.T, self.bhg)
self.h.apply_sigmoid2(self.h)
if sample:
# Sample random values
sampled = cm.empty((v.shape[0], self.output_dim))
sampled.fill_with_rand()
# Sample values of hiddens
sampled.less_than(self.h, self.h)
def AllocateBatchsizeDependentMemory(self, batchsize):
if self.data:
self.data.free_device_memory()
if self.deriv:
self.deriv.free_device_memory()
self.batchsize = batchsize
dimensions = self.dimensions
numlabels = self.numlabels
numdims = dimensions * numlabels
self.statesize = cm.CUDAMatrix(np.zeros((numdims, batchsize)))
self.batchsize_temp = cm.CUDAMatrix(np.zeros((1, batchsize)))
if self.t_op:
if self.t_op.optimizer == deepnet_pb2.Operation.PCD:
self.pos_state = cm.CUDAMatrix(np.zeros((numdims, batchsize)))
self.pos_sample = cm.CUDAMatrix(np.zeros((numdims, batchsize)))
self.neg_state = cm.CUDAMatrix(np.zeros((numdims, batchsize)))
self.neg_sample = cm.CUDAMatrix(np.zeros((numdims, batchsize)))
self.state = self.pos_state
self.sample = self.pos_sample
self.suff_stats = cm.empty((numdims, 1))
elif self.t_op.optimizer == deepnet_pb2.Operation.CD:
self.state = cm.CUDAMatrix(np.zeros((numdims, batchsize)))
self.sample = cm.CUDAMatrix(np.zeros((numdims, batchsize)))
self.suff_stats = cm.empty((numdims, 1))
else:
self.state = cm.CUDAMatrix(np.zeros((numdims, batchsize)))
self.deriv = cm.CUDAMatrix(np.zeros((numdims, batchsize)))
else:
self.state = cm.CUDAMatrix(np.zeros((numdims, batchsize)))
if self.is_input or self.is_initialized or self.is_output:
self.data = cm.CUDAMatrix(np.zeros((dimensions, batchsize)))
if self.hyperparams.dropout:
self.mask = cm.CUDAMatrix(np.zeros(self.state.shape))
def calc_cost(self, X1, X2):
P1 = cm.dot(self.Wgpu.T, X1)
P2 = cm.dot(self.Wgpu.T, X2)
Y1 = cm.dot(self.Wgpu, P1)
Y2 = cm.dot(self.Wgpu, P1)
Y1.subtract(X1)
Y2.subtract(X2)
PD = cm.empty(P1.shape)
XD = cm.empty(X1.shape)
P1.subtract(P2, target = PD)
X1.subtract(X2, target = XD)
grad = cm.dot(XD, PD.T)
grad.mult(self.lam)
grad.add(cm.dot(Y1, P1.T))
grad.add(cm.dot(Y2, P2.T))
grad.add(cm.dot(X1, cm.dot(Y1.T, self.Wgpu)))
grad.add(cm.dot(X2, cm.dot(Y2.T, self.Wgpu)))
grad.divide(X1.shape[1])
PD.mult(PD)
Y1.mult(Y1)
Y2.mult(Y2)
cost = PD.sum(axis = 0)
cost.mult(self.lam)
cost.add(Y1.sum(axis = 0))
cost.add(Y2.sum(axis = 0))
cost = cost.sum(axis = 1)
cost.divide(X1.shape[1] * 2)
return [cost, grad]
def train(self, X1, X2, max_epoch):
self.train_init()
datasz = X1.shape[1]
X1gpu = cm.CUDAMatrix(X1)
X2gpu = cm.CUDAMatrix(X2)
X1sub = cm.empty((self.data_dim, self.batchsz))
X2sub = cm.empty((self.data_dim, self.batchsz))
tic = time.time()
for epoch in range(max_epoch):
if epoch == 5:
self.moment = 0.9
for n in range(datasz / self.batchsz):
X1gpu.get_col_slice(self.batchsz * n, self.batchsz * (n+1), target = X1sub)
X2gpu.get_col_slice(self.batchsz * n, self.batchsz * (n+1), target = X2sub)
cost = self.train_step(X1sub, X2sub)
print "epoch %d cost %f" % (epoch, cost.asarray()[0,0])
toc = time.time()
print "time %s" % (toc - tic)
self.train_finalize()
def __init__(self, layer_type, input_size, output_size, learning_rate,
activation):
self.layer_type = layer_type
self.input_size = input_size
self.output_size = output_size
self.learning_rate = learning_rate
self.act = self.act_func(activation)
self.d_act = self.d_act_func(activation)
self.output = None
self.W = None
self.b = None
self.z = None
self.forward = None
self.backward = None
if self.layer_type == "fc":
self.W = np.random.rand(self.input_size,
self.output_size) * 1 - 0.5
self.b = np.random.rand(1, self.output_size) * 1 - 0.5
self.W = cm.CUDAMatrix(self.W)
self.b = cm.CUDAMatrix(self.b)
self.z = cm.empty((1, self.output_size))
self.d_act_z = cm.empty((1, self.output_size))
self.output = cm.empty((1, self.output_size))
self.delta_b = cm.empty((1, self.output_size))
self.delta_W = cm.empty((self.input_size, self.output_size))
self.output_bp = None # cm.empty((1, self.input_size))
self.forward = self.dense_fw
self.backward = self.dense_bw
def _calculate_moments_ns(self, x, ws, quick=False):
"""Calculate moments based on the weights and samples. We also calculate and save MI, TC, additivity, and
the value of the objective. Note it is assumed that <X_i^2> = 1! """
m = {} # Dictionary of moments
if self.gpu:
y = cm.empty((self.n_samples, self.m))
wc = cm.CUDAMatrix(ws)
cm.dot(x, wc.T,
target=y) # + noise, but it is included analytically
del wc
tmp_sum = np.einsum(
'lj,lj->j', y.asarray(),
y.asarray()) # TODO: Should be able to do on gpu...
else:
y = x.dot(ws.T)
tmp_sum = np.einsum('lj,lj->j', y, y)
m["uj"] = (
1 - self.eps**2) * tmp_sum / self.n_samples + self.eps**2 * np.sum(
ws**2, axis=1)
if quick and np.max(m["uj"]) >= 1.:
return False
if self.gpu:
tmp = cm.empty((self.nv, self.m))
cm.dot(x.T, y, target=tmp)
tmp_dot = tmp.asarray()
del tmp
del y
else:
tmp_dot = x.T.dot(y)
m["rho"] = (
1 - self.eps**
2) * tmp_dot.T / self.n_samples + self.eps**2 * ws # m by nv
m["ry"] = ws.dot(m["rho"].T) # normalized covariance of Y
m["Y_j^2"] = self.yscale**2 / (1. - m["uj"])
np.fill_diagonal(m["ry"], 1)
m["invrho"] = 1. / (1. - m["rho"]**2)
m["rhoinvrho"] = m["rho"] * m["invrho"]
m["Qij"] = np.dot(m['ry'], m["rhoinvrho"])
m["Qi"] = np.einsum('ki,ki->i', m["rhoinvrho"], m["Qij"])
#m["Qi-Si^2"] = np.einsum('ki,ki->i', m["rhoinvrho"], m["Qij"])
m["Si"] = np.sum(m["rho"] * m["rhoinvrho"], axis=0)
# This is the objective, a lower bound for TC
m["TC"] = np.sum(np.log(1 + m["Si"])) \
- 0.5 * np.sum(np.log(1 - m["Si"]**2 + m["Qi"])) \
+ 0.5 * np.sum(np.log(1 - m["uj"]))
if not quick:
m["MI"] = -0.5 * np.log1p(-m["rho"]**2)
m["X_i Y_j"] = m["rho"].T * np.sqrt(m["Y_j^2"])
m["X_i Z_j"] = np.linalg.solve(m["ry"], m["rho"]).T
m["X_i^2 | Y"] = (
1. - np.einsum('ij,ji->i', m["X_i Z_j"], m["rho"])).clip(1e-6)
m['I(Y_j ; X)'] = 0.5 * np.log(m["Y_j^2"]) - 0.5 * np.log(
self.yscale**2)
m['I(X_i ; Y)'] = -0.5 * np.log(m["X_i^2 | Y"])
m["TCs"] = m["MI"].sum(axis=1) - m['I(Y_j ; X)']
m["additivity"] = (m["MI"].sum(axis=0) - m['I(X_i ; Y)']).sum()
return m
def _init_params(self, batch_size):
self.next_z = cm.empty((batch_size, self.next_size))
self.next_single_z = cm.empty((1, self.next_size))
if self.level != 1:
self.my_delta = cm.empty((batch_size, self.size))
else:
self.my_delta = None
def initParams(self):
"""
Initialize parameters using 6/sqrt(fanin+fanout)
"""
sizes = [self.inputDim]+self.layerSizes+[self.outputDim]
scales = [np.sqrt(6)/np.sqrt(n+m) for n,m in zip(sizes[:-1],sizes[1:])]
self.stack = [[np.random.rand(m,n)*2*s-s,np.zeros((m,1))] \
for n,m,s in zip(sizes[:-1],sizes[1:],scales)]
self.hActs_M = [cm.empty((s,self.maxBatch)) for s in sizes]
if self.train:
# Now assuming that all layers are the same size
self.grad = [[cm.empty(w.shape),cm.empty(b.shape)] for w,b in self.stack]
self.deltasC_M = cm.empty((self.outputDim,self.maxBatch))
self.deltasOut_M = cm.empty((sizes[1],self.maxBatch))
self.deltasIn_M = cm.empty((sizes[1],self.maxBatch))
self.tmpGrad_M = cm.empty((self.layerSizes[0],self.maxBatch))
# Allocate memory once here and reuse
# Store probs
self.probs_M = cm.empty((self.outputDim,self.maxBatch))
# Store col max
self.rowVec_M = cm.empty((1,self.maxBatch))
self.stack = [[cm.CUDAMatrix(w),cm.CUDAMatrix(b)]
for w,b in self.stack]
def LoadParams(self, tied_to=None):
"""Load the parameters for this edge.
Load the parameters if present in self.proto. Otherwise initialize them
appropriately.
"""
node1 = self.node1
node2 = self.node2
proto = self.proto
self.hyperparams = proto.hyperparams
param_names = [param.name for param in proto.param]
for param in proto.param:
if param.conv or param.local:
n_locs = self.AllocateMemoryForConvolutions(
param, node1, node2)
if not param.dimensions:
if param.conv:
cv = param.conv_params
dims = [cv.num_filters, cv.size**2 * cv.num_colors]
elif param.local:
dims = [
cv.num_filters, n_locs**2 * cv.size**2 * cv.num_colors
]
else:
dims = [
node1.numlabels * node1.dimensions,
node2.numlabels * node2.dimensions
]
param.dimensions.extend(dims)
if tied_to:
if self.transpose:
self.params[param.name] = tied_to.params[param.name].T
else:
self.params[param.name] = tied_to.params[param.name]
mat = self.params[param.name]
else:
if param.mat: # and 'grad' not in param.name:
mat = util.ParameterAsNumpy(param)
else:
mat = self.InitializeParameter(param)
self.params[param.name] = cm.CUDAMatrix(mat)
if param.name == 'weight':
self.temp = cm.empty(mat.shape)
#self.temp2 = cm.empty(mat.shape)
self.gradient = cm.empty(mat.shape)
self.grad_weight = cm.empty(mat.shape)
self.gradient.assign(0)
self.grad_weight.assign(0)
if self.t_op and (self.t_op.optimizer == deepnet_pb2.Operation.PCD or \
self.t_op.optimizer == deepnet_pb2.Operation.CD):
self.suff_stats = cm.empty(
(self.node1.numlabels * self.node1.dimensions,
self.node2.numlabels * self.node2.dimensions))
def sinkhorn_lpl1_mm(a,
labels_a,
b,
M_GPU,
reg,
eta=0.1,
numItermax=10,
numInnerItermax=200,
stopInnerThr=1e-9,
verbose=False,
log=False):
p = 0.5
epsilon = 1e-3
Nfin = len(b)
indices_labels = []
classes = np.unique(labels_a)
for c in classes:
idxc, = np.where(labels_a == c)
indices_labels.append(cudamat.CUDAMatrix(idxc.reshape(1, -1)))
Mreg_GPU = cudamat.empty(M_GPU.shape)
W_GPU = cudamat.empty(M_GPU.shape).assign(0)
for cpt in range(numItermax):
Mreg_GPU.assign(M_GPU)
Mreg_GPU.add_mult(W_GPU, eta)
transp_GPU = sinkhorn(a,
b,
Mreg_GPU,
reg,
numItermax=numInnerItermax,
stopThr=stopInnerThr,
returnAsGPU=True)
# the transport has been computed. Check if classes are really
# separated
W_GPU.assign(1)
W_GPU = W_GPU.transpose()
for (i, c) in enumerate(classes):
(_, nbRow) = indices_labels[i].shape
tmpC_GPU = cudamat.empty((Nfin, nbRow)).assign(0)
transp_GPU.transpose().select_columns(indices_labels[i], tmpC_GPU)
majs_GPU = tmpC_GPU.sum(axis=1).add(epsilon)
cudamat.pow(majs_GPU, (p - 1))
majs_GPU.mult(p)
tmpC_GPU.assign(0)
tmpC_GPU.add_col_vec(majs_GPU)
W_GPU.set_selected_columns(indices_labels[i], tmpC_GPU)
W_GPU = W_GPU.transpose()
return transp_GPU.asarray()
def setVariables(self):
n, m, r = self.n, self.m, self.rank
self.H_gpu = cm.CUDAMatrix(self.H)
self.W_gpu = cm.CUDAMatrix(self.W)
self.X_gpu = cm.CUDAMatrix(self.X)
self.Wrowsum_gpu = cm.empty([r,1])
self.WH_gpu = cm.empty([n,m])
self.XWH_gpu = self.WH_gpu
self.WTXWH_gpu = cm.empty([r,m])
self.Hcolsum_gpu = cm.empty([1,r])
self.XWHHT_gpu = cm.empty([n,r])
def TranslateData(self, batch, i):
"""Applies translations to data at index i in batch."""
sizeX = self.sizeX
sizex = self.sizex
batchsize = batch[i].shape[1]
shift = (sizeX - sizex) / 2
offset_x = np.array([
random.choice(self.translate_range_x) + shift
for k in range(batchsize)
]).reshape(1, -1)
offset_y = np.array([
random.choice(self.translate_range_y) + shift
for k in range(batchsize)
]).reshape(1, -1)
num_channels = self.num_channels
d = batch[i]
if self.offset_x is None:
self.offset_x = cm.CUDAMatrix(offset_x)
else:
self.offset_x.overwrite(offset_x)
if self.offset_y is None:
self.offset_y = cm.CUDAMatrix(offset_y)
else:
self.offset_y.overwrite(offset_y)
if self.translated_d is None or self.translated_d.shape[1] != batchsize:
self.translated_d = cm.empty((sizex**2 * num_channels, batchsize))
d.generate_translations(sizeX,
sizex,
self.offset_x,
self.offset_y,
target=self.translated_d)
batch[i] = self.translated_d
def get_specrad(Ac):
"""Get spectral radius of A using the power method."""
m_size = Ac.shape[0]
x = np.random.normal(0, 1, (m_size, 1))
x = x / np.linalg.norm(x)
x = cm.CUDAMatrix(x)
y = cm.empty((m_size, 1))
diff = 200
eps = 1e-3
b = 1e10
c = 1e9
max_its = 1e6
n_its = 0
while diff > eps and n_its < max_its:
cm.dot(Ac, x, target=y)
norm = y.euclid_norm()
y.divide(norm, target=x)
a = cm.dot(y.T, x).asarray()
c = cm.dot(x.T, x).asarray()
diff = np.abs(a - b)
b = float(a)
n_its += 1
specrad = float(a / c)
print 'Spectral radius:', specrad, 'Number of iterations:', n_its
return float(a / c)
def getDataUpwards(self,loadfolder,savefolder):
# push data of visible layer upwards to form a set of new data
# because of memory issues, we have to each batch data to disc and read and combine them later
# batch mode receive data from cpu and return a matrix on cpu
datalist = os.listdir(loadfolder)
batchsize = 4096
n = 0
for dataname in datalist:
name = os.path.join(loadfolder,dataname)
mdict = scipy.io.loadmat(name)
data = mdict['data']
labels = mdict['label']
# print labels.shape
numbatch = data.shape[1]/batchsize
for batch in range(numbatch):
#print 'batch %d/%d'%(n, numbatch*len(datalist))
batchdata = data[:,batch*batchsize:(batch+1)*batchsize]
batchlabels = labels[batch*batchsize:(batch+1)*batchsize]
temp = cm.empty((self.num_hid,batchdata.shape[1]))
vis = cm.CUDAMatrix(batchdata)
self.hidActProb(vis, temp)
temp.copy_to_host()
#topdata[:,batch*batchsize:(batch+1)*batchsize] = temp.numpy_array
mdict = {}
mdict['data'] = temp.numpy_array
mdict['label'] = batchlabels
scipy.io.savemat('%s/%d.mat'%(savefolder,n),mdict)
n = n+1
def __init__(self, hidden_dim, visible_dim=None, context_dim=None,
gaussian=False, dtype=None, max_batch_size=1500):
"""
Arguments:
- hidden_dim: number of hidden variables
- visible_dim: number of observed variables
- context_dim: number of context variables
- gaussian: use gaussian visible units (default is binary)
- max_batch_size: number of datapoints to process simultaneously
The max batch size is required to be able to optimize the adding of
bias vectors by preallocating a vector of ones and taking the outer
product with the bias vector. This is faster than directly adding it
somehow.
"""
super(RBMNode, self).__init__(hidden_dim, visible_dim + context_dim, dtype)
self._input_dim = visible_dim + context_dim
self._output_dim = hidden_dim
self.context_dim = context_dim
self.visible_dim = visible_dim
self._initialized = False
self._ones = cm.empty((1, max_batch_size))
self._ones.assign(1)
self._gaussian = gaussian
def get_specrad(Ac):
"""Get spectral radius of A using the power method."""
m_size = Ac.shape[0]
x = np.random.normal(0, 1, (m_size, 1))
x = x / np.linalg.norm(x)
x = cm.CUDAMatrix(x)
y = cm.empty((m_size, 1))
diff = 200
eps = 1e-3
b = 1e10
c = 1e9
max_its = 1e6
n_its = 0
while diff > eps and n_its < max_its:
cm.dot(Ac, x, target=y)
norm = y.euclid_norm()
y.divide(norm, target=x)
a = cm.dot(y.T, x).asarray()
c = cm.dot(x.T, x).asarray()
diff = np.abs(a - b)
b = float(a)
n_its += 1
specrad = float(a / c)
print 'Spectral radius:', specrad, 'Number of iterations:', n_its
return float(a / c)
def GetErrors(self, raaX, raaY, sActivation):
# Small value to avoid log underflows
rEps = 1e-20
raaError = cudamat.empty(raaX.shape)
raaX.subtract(raaY, raaError)
raaError.mult(raaError)
raError = raaError.sum(axis=0)
rError = raError.sum(axis=1)
rSe = rError.asarray()[0,0]
#print(rSe)
# # Sum all squared errors
# rSe = numpy.sum(numpy.square(raaError))
# # Depending on the activation def type
# if(sActivation=="Logistic"):
# # Compute the average cross entropy error
# rE = -numpy.sum(numpy.multiply(raaX,numpy.log(raaY+rEps)) + numpy.multiply(1-raaX,numpy.log(1-raaY+rEps)))
# elif(sActivation=="Linear"):
# # Compute the squared error
# rE = rSe
# elif(sActivation=="Softmax"):
# # Compute the average cross entropy error
# rE = -numpy.sum(numpy.multiply(raaX,numpy.log(raaY+rEps)))
rE = 0
return(rSe, rE)
def Sparcity(M,psi):
N = cm.empty(M.shape)
N.assign(np.random.choice([0,1],M.size,p =[psi,1-psi]))
T = np.empty(M.shape)
N.mult(M,target = T)
return T
def __init__(self, hidden_dim, visible_dim=None, context_dim=None,
gaussian=False, dtype=None, max_batch_size=500):
"""
Arguments:
- hidden_dim: number of hidden variables
- visible_dim: number of observed variables
- context_dim: number of context variables
- gaussian: use gaussian visible units (default is binary)
- max_batch_size: number of datapoints to process simultaneously
The max batch size is required to be able to optimize the adding of
bias vectors by preallocating a vector of ones and taking the outer
product with the bias vector. This is faster than directly adding it
somehow.
"""
super(RBMNode, self).__init__(hidden_dim, visible_dim + context_dim, dtype)
self._input_dim = visible_dim + context_dim
self._output_dim = hidden_dim
self.context_dim = context_dim
self.visible_dim = visible_dim
self._initialized = False
self._ones = cm.empty((1, max_batch_size))
self._ones.assign(1)
self._gaussian = gaussian
# TODO: Should probably use **kwargs for these arguments and figure out
# whether to keep the leak_rate and bias in there for performance
# reasons.
self.reservoir = Oger.nodes.CUDAReservoirNode(visible_dim, context_dim,
spectral_radius=.01,
leak_rate=1,
input_scaling=.001)
def __init__(self,
hidden_dim,
visible_dim=None,
context_dim=None,
gaussian=False,
dtype=None,
max_batch_size=1500):
"""
Arguments:
- hidden_dim: number of hidden variables
- visible_dim: number of observed variables
- context_dim: number of context variables
- gaussian: use gaussian visible units (default is binary)
- max_batch_size: number of datapoints to process simultaneously
The max batch size is required to be able to optimize the adding of
bias vectors by preallocating a vector of ones and taking the outer
product with the bias vector. This is faster than directly adding it
somehow.
"""
super(RBMNode, self).__init__(hidden_dim, visible_dim + context_dim,
dtype)
self._input_dim = visible_dim + context_dim
self._output_dim = hidden_dim
self.context_dim = context_dim
self.visible_dim = visible_dim
self._initialized = False
self._ones = cm.empty((1, max_batch_size))
self._ones.assign(1)
self._gaussian = gaussian
def LoadParams(self, tied_to=None):
"""Load the parameters for this edge.
Load the parameters if present in self.proto. Otherwise initialize them
appropriately.
"""
node1 = self.node1
node2 = self.node2
proto = self.proto
self.hyperparams = proto.hyperparams
param_names = [param.name for param in proto.param]
for param in proto.param:
if param.conv or param.local:
n_locs = self.AllocateMemoryForConvolutions(param, node1, node2)
if not param.dimensions:
if param.conv:
cv = param.conv_params
dims = [cv.num_filters, cv.size**2 * cv.num_colors]
elif param.local:
dims = [cv.num_filters, n_locs**2 * cv.size**2 * cv.num_colors]
else:
dims = [node1.numlabels * node1.dimensions,
node2.numlabels * node2.dimensions]
param.dimensions.extend(dims)
if tied_to:
if self.transpose:
self.params[param.name] = tied_to.params[param.name].T
else:
self.params[param.name] = tied_to.params[param.name]
mat = self.params[param.name]
else:
if param.mat: # and 'grad' not in param.name:
mat = util.ParameterAsNumpy(param)
else:
mat = self.InitializeParameter(param)
self.params[param.name] = cm.CUDAMatrix(mat)
if param.name == 'weight':
self.temp = cm.empty(mat.shape)
#self.temp2 = cm.empty(mat.shape)
self.gradient = cm.empty(mat.shape)
self.grad_weight = cm.empty(mat.shape)
self.gradient.assign(0)
self.grad_weight.assign(0)
if self.t_op and (self.t_op.optimizer == deepnet_pb2.Operation.PCD or \
self.t_op.optimizer == deepnet_pb2.Operation.CD):
self.suff_stats = cm.empty((self.node1.numlabels * self.node1.dimensions,
self.node2.numlabels * self.node2.dimensions))
def train_init(self):
# init cudamat
cm.cublas_init()
cm.CUDAMatrix.init_random(1)
self.Wgpu = cm.CUDAMatrix(self.W)
self.speed = cm.empty(self.W.shape)
self.speed.assign(0)
def LoadParams(self):
"""Load the parameters for this edge.
Load the parameters if present in self.proto. Otherwise initialize them
appropriately.
"""
proto = self.proto
node1 = self.node1
node2 = self.node2
self.hyperparams = proto.hyperparams
param_names = [param.name for param in proto.param]
"""
for param in proto.param:
if 'grad_'+param.name not in param_names and not param.name.startswith('grad_'):
grad_p = deepnet_pb2.Parameter()
grad_p.CopyFrom(param)
grad_p.name = 'grad_' + param.name
grad_p.initialization = deepnet_pb2.Parameter.CONSTANT
grad_p.constant = 0
proto.param.extend([grad_p])
"""
for param in proto.param:
if param.conv or param.local:
n_locs = self.AllocateMemoryForConvolutions(param, node1, node2)
if not param.dimensions:
if param.conv:
cv = param.conv_params
dims = [cv.num_filters, cv.size**2 * cv.num_colors]
elif param.local:
dims = [cv.num_filters, n_locs**2 * cv.size**2 * cv.num_colors]
else:
dims = [node1.numlabels * node1.dimensions,
node2.numlabels * node2.dimensions]
param.dimensions.extend(dims)
if param.mat: # and 'grad' not in param.name:
print 'Loading saved parameters'
mat = util.ParameterAsNumpy(param)
else:
mat = self.InitializeParameter(param)
self.params[param.name] = cm.CUDAMatrix(mat)
if param.name == 'weight':
self.temp = cm.empty(mat.shape)
self.temp2 = cm.empty(mat.shape)
self.grad_weight = cm.empty(mat.shape)
self.grad_weight.assign(0)
def _init_bias(self):
assert self.use_bias
self.biases = cm.CUDAMatrix(
np.zeros((1, self.next_size))
)
self.active_biases = cm.empty(self.biases.shape)
self.biases_grad = cm.empty(self.biases.shape)
if self.use_momentum:
self.biases_update = \
cm.CUDAMatrix(np.zeros(self.biases_grad.shape))
if self.use_rmsprop:
self.biases_rmsprop_cache = \
cm.CUDAMatrix(np.zeros(self.biases_grad.shape))
self.biases_grad_square = \
cm.CUDAMatrix(np.zeros(self.biases_grad.shape))
def AllocateMemoryForConvolutions(self, param, node1, node2):
self.conv = param.conv
self.local = param.local
if self.conv:
assert not self.local
else:
assert not self.conv
self.conv_params = param.conv_params
num_colors = self.conv_params.num_colors
num_filters = self.conv_params.num_filters
size = self.conv_params.size
padding = self.conv_params.padding
stride = self.conv_params.stride
numdims, numimages = node1.state.shape
assert numdims % num_colors == 0
x = int(np.sqrt(numdims / num_colors))
assert x**2 == numdims / num_colors
n_locs = (x + 2 * padding - size) / stride + 1
input_shape = node1.state.shape[::-1]
output_shape = node2.state.shape[::-1]
self.input_t = cm.empty(input_shape)
self.input_t2 = cm.empty(input_shape)
self.output_t = cm.empty(output_shape)
self.output_t2 = cm.empty(output_shape)
if param.conv_params.max_pool:
pool_output_size = n_locs**2 * num_filters
self.unpooled_layer = cm.empty((numimages, pool_output_size))
pool_size = param.conv_params.pool_size
pool_stride = param.conv_params.pool_stride
n_pool_locs = (n_locs - pool_size) / pool_stride + 1
assert output_shape[1] == n_pool_locs**2 * num_filters
if param.conv_params.prob:
self.rnd = cm.empty(self.unpooled_layer.shape)
else:
assert output_shape[1] == n_locs**2 * num_filters
if param.conv_params.rnorm:
self.unrnormalized_layer = cm.empty(output_shape)
self.denoms = cm.empty(output_shape)
self.rnorm_temp1 = cm.empty(output_shape)
self.rnorm_temp2 = cm.empty(output_shape)
return n_locs
def __init__(self, epsilon, momentum, num_epochs, batch_size, num_batches, dim_in, dim_out, num_hid): # training parameters self.epsilon = epsilon self.momentum = momentum self.num_epochs = num_epochs self.batch_size = batch_size self.num_batches = num_batches # model parameters self.dim_in = dim_in self.dim_out = dim_out self.num_hid = num_hid # initialize weights self.w_w1 = cm.CUDAMatrix(dim_in ** -0.5 * np.random.randn(dim_in, num_hid)) self.w_b1 = cm.CUDAMatrix(np.zeros((num_hid, 1))) self.w_w2 = cm.CUDAMatrix(num_hid ** -0.5 * np.random.randn(num_hid, dim_out)) self.w_b2 = cm.CUDAMatrix(np.zeros((dim_out, 1))) # initialize weight update matrices self.wu_w1 = cm.empty(self.w_w1.shape).assign(0) self.wu_b1 = cm.empty(self.w_b1.shape).assign(0) self.wu_w2 = cm.empty(self.w_w2.shape).assign(0) self.wu_b2 = cm.empty(self.w_b2.shape).assign(0) # initialize temporary storage self.h = cm.empty((self.num_hid, self.batch_size)) self.out = cm.empty((self.dim_out, self.batch_size)) self.delta = cm.empty((self.num_hid, self.batch_size))
def AllocateMemoryForConvolutions(self, param, node1, node2):
self.conv = param.conv
self.local = param.local
if self.conv:
assert not self.local
else:
assert not self.conv
self.conv_params = param.conv_params
num_colors = self.conv_params.num_colors
num_filters = self.conv_params.num_filters
size = self.conv_params.size
padding = self.conv_params.padding
stride = self.conv_params.stride
numdims, numimages = node1.state.shape
assert numdims % num_colors == 0
x = int(np.sqrt(numdims / num_colors))
assert x**2 == numdims / num_colors
n_locs = (x + 2 * padding - size) / stride + 1
input_shape = node1.state.shape[::-1]
output_shape = node2.state.shape[::-1]
self.input_t = cm.empty(input_shape)
self.input_t2 = cm.empty(input_shape)
self.output_t = cm.empty(output_shape)
self.output_t2 = cm.empty(output_shape)
if param.conv_params.max_pool:
pool_output_size = n_locs**2 * num_filters
self.unpooled_layer = cm.empty((numimages, pool_output_size))
pool_size = param.conv_params.pool_size
pool_stride = param.conv_params.pool_stride
n_pool_locs = (n_locs - pool_size) / pool_stride + 1
assert output_shape[1] == n_pool_locs**2 * num_filters
if param.conv_params.prob:
self.rnd = cm.empty(self.unpooled_layer.shape)
else:
assert output_shape[1] == n_locs**2 * num_filters
if param.conv_params.rnorm:
self.unrnormalized_layer = cm.empty(output_shape)
self.denoms = cm.empty(output_shape)
self.rnorm_temp1 = cm.empty(output_shape)
self.rnorm_temp2 = cm.empty(output_shape)
return n_locs
def back_propagate(self,error=None): if(error is not None): self.error = cm.error #python doesn't easily allow reversed(enumerate()) - use this instead for l in reversed(self.layer): #if we're on the last layer #print(str(index)); if(l.index == len(self.layer)-1): delta_temp = cm.CUDAMatrix(np.append(self.error,np.zeros((1,self.error.shape[1])),axis=0)) else: #Possible TODO?: is there a way to get rid of this transpose? it is slow to have to do this #delta_temp = cm.empty((self.layer[l.index+1].weights.shape[1],self.layer[l.index+1].weights.shape[0])) #delta_temp = self.layer[l.index+1].weights.transpose() self.layer[l.index+1].weights.set_trans(True); delta_temp = cm.dot(self.layer[l.index+1].weights,self.layer[l.index+1].delta); self.layer[l.index+1].weights.set_trans(False); if(l.activation == 'squash'): pass #l.activation_derivative = 1.0/((1+np.abs(l.weighted_sums)**2)) elif(l.activation == 'sigmoid'): #l.activation_derivative = cm.empty(l.output.shape); l.output.apply_logistic_deriv(l.output) l.activation_derivative = l.output #elif(l.activation == 'linear_rectifier'): #1 if greater than 0, 0 otherwise. #This stores them as bools - but it doesn't matter #l.activation_derivative = np.greater(l.output,0); else: #base case is linear l.activation_derivative = cm.empty(l.output.shape); l.activation_derivitive.assign_scalar(1.0) #bottom row of activation derivative is the bias 'neuron' l.delta = cm.empty(delta_temp.shape) l.activation_derivative.mult(delta_temp,target=l.delta) #calculate weight gradient #input_t = cm.empty((l.input.shape[1],l.input.shape[0])) #input_t = l.input.transpose() l.input.set_trans(True) l.gradient.add_dot(l.delta,l.input); l.input.set_trans(False) self.epoch_size = self.epoch_size + self.input.shape[1];
def back_propagate(self, error=None):
if error is not None:
self.error = cm.error
# python doesn't easily allow reversed(enumerate()) - use this instead
for l in reversed(self.layer):
# if we're on the last layer
# print(str(index));
if l.index == len(self.layer) - 1:
delta_temp = cm.CUDAMatrix(np.append(self.error, np.zeros((1, self.error.shape[1])), axis=0))
else:
# Possible TODO?: is there a way to get rid of this transpose? it is slow to have to do this
# delta_temp = cm.empty((self.layer[l.index+1].weights.shape[1],self.layer[l.index+1].weights.shape[0]))
# delta_temp = self.layer[l.index+1].weights.transpose()
self.layer[l.index + 1].weights.set_trans(True)
delta_temp = cm.dot(self.layer[l.index + 1].weights, self.layer[l.index + 1].delta)
self.layer[l.index + 1].weights.set_trans(False)
if l.activation == "squash":
pass
# l.activation_derivative = 1.0/((1+np.abs(l.weighted_sums)**2))
elif l.activation == "sigmoid":
# l.activation_derivative = cm.empty(l.output.shape);
l.output.apply_logistic_deriv(l.output)
l.activation_derivative = l.output
# elif(l.activation == 'linear_rectifier'):
# 1 if greater than 0, 0 otherwise.
# This stores them as bools - but it doesn't matter
# l.activation_derivative = np.greater(l.output,0);
else: # base case is linear
l.activation_derivative = cm.empty(l.output.shape)
l.activation_derivitive.assign_scalar(1.0)
# bottom row of activation derivative is the bias 'neuron'
l.delta = cm.empty(delta_temp.shape)
l.activation_derivative.mult(delta_temp, target=l.delta)
# calculate weight gradient
# input_t = cm.empty((l.input.shape[1],l.input.shape[0]))
# input_t = l.input.transpose()
l.input.set_trans(True)
l.gradient.add_dot(l.delta, l.input)
l.input.set_trans(False)
self.epoch_size = self.epoch_size + self.input.shape[1]
def compute_gamma_entropy(self, G):
if not self.gpu:
Prod = G * (np.log(G) - 1)
ent = np.nan_to_num(Prod).sum()
else:
Prod = cm.empty(G.shape)
Prod = G.mult(cm.log(G.copy()).subtract(1), target=Prod)
ent = np.nan_to_num(Prod.asarray()).sum()
return ent
def _calculate_moments_syn(self, x, ws, quick=False):
"""Calculate moments based on the weights and samples. We also calculate and save MI, TC, additivity, and
the value of the objective. Note it is assumed that <X_i^2> = 1! """
m = {} # Dictionary of moments
eps = 10**-8
if self.gpu:
y = cm.empty((self.n_samples, self.m))
wc = cm.CUDAMatrix(ws)
cm.dot(x, wc.T,
target=y) # + noise, but it is included analytically
del wc
else:
y = x.dot(ws.T) # + noise, but it is included analytically
if self.gpu:
tmp_dot = cm.empty((self.nv, self.m))
cm.dot(x.T, y, target=tmp_dot)
m["X_i Y_j"] = tmp_dot.asarray(
) / self.n_samples # nv by m, <X_i Y_j>
del y
del tmp_dot
else:
m["X_i Y_j"] = x.T.dot(y) / self.n_samples
m["cy"] = ws.dot(
m["X_i Y_j"]) + self.yscale**2 * np.eye(self.m) # cov(y.T), m by m
m["Y_j^2"] = np.diag(m["cy"]).copy()
m["ry"] = m["cy"] / (np.sqrt(m["Y_j^2"]) *
np.sqrt(m["Y_j^2"][:, np.newaxis]))
m["rho"] = (m["X_i Y_j"] / np.sqrt(m["Y_j^2"])).T
m["invrho"] = 1. / (1. - m["rho"]**2)
m["rhoinvrho"] = m["rho"] * m["invrho"]
m["Qij"] = np.dot(m['ry'], m["rhoinvrho"])
m["Qi"] = np.einsum('ki,ki->i', m["rhoinvrho"], m["Qij"])
m["Si"] = np.sum(m["rho"] * m["rhoinvrho"], axis=0)
m["MI"] = -0.5 * np.log1p(-m["rho"]**2)
m["X_i Z_j"] = np.linalg.solve(m["cy"], m["X_i Y_j"].T).T
m["X_i^2 | Y"] = (
1. - np.einsum('ij,ij->i', m["X_i Z_j"], m["X_i Y_j"])).clip(1e-6)
mi_yj_x = 0.5 * np.log(m["Y_j^2"]) - 0.5 * np.log(self.yscale**2)
mi_xi_y = -0.5 * np.log(m["X_i^2 | Y"])
m["TCs"] = m["MI"].sum(axis=1) - mi_yj_x
m["additivity"] = (m["MI"].sum(axis=0) - mi_xi_y).sum()
m["TC"] = np.sum(mi_xi_y) - np.sum(mi_yj_x)
return m
def loadNN(self,filename):
d = scipy.io.loadmat(filename)
for i in range(self.H):
self.W.append(cm.CUDAMatrix(d['W%d'%i]))
self.b.append(cm.CUDAMatrix(d['b%d'%i]))
self.dW.append(cm.CUDAMatrix(np.zeros(self.W[i].shape)))
self.W_inc.append(cm.CUDAMatrix(np.zeros(self.W[i].shape)))
self.db.append(cm.CUDAMatrix(np.zeros(self.b[i].shape)))
self.b_inc.append(cm.CUDAMatrix(np.zeros(self.b[i].shape)))
self.h.append(cm.empty((self.W[i].shape[1],self.batchsize)))
def __init__(self, w, config=None):
if type(w) == np.ndarray:
self.w_ = cm.CUDAMatrix(w)
elif type(w) == tuple:
self.w_ = cm.empty(w)
else:
self.w_ = w
self.dw_ = cm.empty_like(self.w_)
self.dw_history_ = cm.empty_like(self.w_)
self.dw_history_.assign(0)
self.dw_.assign(0)
self.t_ = 0
self.rms_prop_ = config.rms_prop
self.rms_prop_factor_ = config.rms_prop_factor
if self.rms_prop_:
self.rms_prop_history_ = cm.empty_like(self.dw_)
self.rms_prop_history_.assign(1)
if config is None:
pass
elif config.init_type == config_pb2.Param.CONSTANT:
self.w_.assign(config.scale)
elif config.init_type == config_pb2.Param.GAUSSIAN:
self.w_.fill_with_randn()
self.w_.mult(config.scale)
elif config.init_type == config_pb2.Param.UNIFORM:
self.w_.fill_with_rand()
self.w_.subtract(0.5)
self.w_.mult(2 * config.scale)
elif config.init_type == config_pb2.Param.LSTM_BIAS:
init_bias = [
config.input_gate_bias, config.forget_gate_bias,
config.input_bias, config.output_gate_bias
]
self.w_.reshape((-1, 4))
for i in xrange(4):
self.w_.slice(i, (i + 1)).assign(init_bias[i])
self.w_.reshape((-1, 1))
elif config.init_type == config_pb2.Param.PRETRAINED:
f = h5py.File(config.file_name)
mat = f[config.dataset_name].value
if len(mat.shape) == 1:
mat = mat.reshape(1, -1)
assert self.w_.shape == mat.shape
self.w_.overwrite(mat)
f.close()
else:
raise Exception('Unknown parameter initialization.')
self.eps_ = config.epsilon
self.momentum_ = config.momentum
self.l2_decay_ = config.l2_decay
self.gradient_clip_ = config.gradient_clip
self.eps_decay_factor = config.eps_decay_factor
self.eps_decay_after = config.eps_decay_after
def __init__(self,
N=1000,
pz=1,
pg=0.1,
g=1.5,
alpha=1,
dt=0.1,
num_fits=1,
num_inputs=0,
state=None):
cm.cublas_init()
if state is not None:
self.from_dict(state)
else:
self.N = N
self.pg = pg
self.pz = pz
self.g = g
self.alpha = alpha
self.DT = dt
self.num_fits = num_fits
scale = 1.0 / np.sqrt(self.pg * self.N)
M_rvs = stats.norm(loc=0, scale=scale).rvs
self.M = sp.sparse.random(N, N, pg, data_rvs=M_rvs) * g
self.M = cm.CUDAMatrix(self.M.toarray())
self.P = (1.0 / self.alpha) * np.identity(N)
self.wf = cm.CUDAMatrix(np.random.uniform(-1, 1, (N, num_fits)))
#self.wo = np.expand_dims(stats.norm(loc=0,scale=(1.0/np.sqrt(N))).rvs(N),num_fits)
self.wo = cm.CUDAMatrix(np.zeros((N, num_fits)))
self.dw = np.zeros((N, num_fits))
self.woc = np.zeros((N, 1))
self.wfc = np.random.uniform(-1, 1, (N, 1))
self.x = cm.CUDAMatrix(np.expand_dims(0.5 * np.random.randn(N), 1))
self.xdt = cm.empty(self.x.shape).assign(0)
self.r = cm.tanh(self.x)
self.rdt = cm.empty(self.r.shape).assign(0)
self.z = cm.CUDAMatrix(
np.expand_dims(0.5 * np.random.randn(num_fits), 1))
self.zdt = cm.empty(self.z.shape).assign(0)
self.z_ctl = np.expand_dims(0.5 * np.random.randn(1), 1)
def alloc(self, *args):
'''
Allocate temporary GPU memory
alloc(N, K, D)
allo(key_shape_mapping)
'''
if len(args) == 3:
N, K, D = args
key_shape_mapping = {
'posteriors_NxK': (N, K), # big
'weighted_X_sum_KxD': (K, D), # medium
'vmax_Nx1': (N, 1),
'logprob_Nx1': (N, 1),
'inv_weights_Kx1': (K, 1),
'temp_NxD': (N, D), # big
'temp_KxD': (K, D), # medium
'temp_KxD_2': (K, D), # medium
'temp_Kx1': (K, 1),
'temp_Kx1_2': (K, 1),
}
elif len(args) == 1:
key_shape_mapping = args[0]
else:
ValueError(
'TempGPUMem: alloc(N, K, D) or alloc(key_shape_mapping)')
# allocate memory
for key, shape in key_shape_mapping.iteritems():
if key not in self:
logger.debug('%s: created %s at key %s',
sys._getframe().f_code.co_name,
shape,
key)
self[key] = cm.empty(shape)
elif self[key].shape != shape:
logger.debug('%s: reshaped %s from %s to %s',
sys._getframe().f_code.co_name,
key,
self[key].shape,
shape)
self[key] = cm.empty(shape)
def _sig(self, x, u):
"""Multiple the matrix u by the covariance matrix of x. We are interested in situations where
n_variables >> n_samples, so we do this without explicitly constructing the covariance matrix."""
if self.gpu:
y = cm.empty((self.n_samples, self.m))
uc = cm.CUDAMatrix(u)
cm.dot(x, uc.T, target=y)
del uc
tmp = cm.empty((self.nv, self.m))
cm.dot(x.T, y, target=tmp)
tmp_dot = tmp.asarray()
del y
del tmp
else:
y = x.dot(u.T)
tmp_dot = x.T.dot(y)
prod = (
1 - self.eps**2
) * tmp_dot.T / self.n_samples + self.eps**2 * u # nv by m, <X_i Y_j> / std Y_j
return prod
def train_step(self, X1, X2):
[cost, grad] = self.calc_cost(X1, X2)
grad.mult(self.lrate)
self.speed.mult(self.moment)
self.speed.subtract(grad)
self.Wgpu.add(self.speed)
Wmask = cm.empty(self.W.shape)
self.Wgpu.greater_than(0, target = Wmask)
self.Wgpu.mult(Wmask)
return cost
def getTestDataUpwards(self,data):
batchsize = 4096
numbatch = data.shape[1]/batchsize
topdata = np.zeros((self.num_hid,data.shape[1]))
for batch in range(numbatch):
batchdata = data[:,batch*batchsize:(batch+1)*batchsize]
temp = cm.empty((self.num_hid,batchdata.shape[1]))
vis = cm.CUDAMatrix(batchdata)
self.hidActProb(vis, temp)
temp.copy_to_host()
topdata[:,batch*batchsize:(batch+1)*batchsize] = temp.numpy_array
return topdata
def __init__(self, w, config=None):
if type(w) == np.ndarray:
self.w_ = cm.CUDAMatrix(w)
elif type(w) == tuple:
self.w_ = cm.empty(w)
else:
self.w_ = w
self.dw_ = cm.empty_like(self.w_)
self.dw_history_ = cm.empty_like(self.w_)
self.dw_history_.assign(0)
self.dw_.assign(0)
self.t_ = 0
self.rms_prop_ = config.rms_prop
self.rms_prop_factor_ = config.rms_prop_factor
if self.rms_prop_:
self.rms_prop_history_ = cm.empty_like(self.dw_)
self.rms_prop_history_.assign(1)
if config is None:
pass
elif config.init_type == config_pb2.Param.CONSTANT:
self.w_.assign(config.scale)
elif config.init_type == config_pb2.Param.GAUSSIAN:
self.w_.fill_with_randn()
self.w_.mult(config.scale)
elif config.init_type == config_pb2.Param.UNIFORM:
self.w_.fill_with_rand()
self.w_.subtract(0.5)
self.w_.mult(2 * config.scale)
elif config.init_type == config_pb2.Param.LSTM_BIAS:
init_bias = [config.input_gate_bias, config.forget_gate_bias, config.input_bias, config.output_gate_bias]
self.w_.reshape((-1, 4))
for i in xrange(4):
self.w_.slice(i, (i+1)).assign(init_bias[i])
self.w_.reshape((-1, 1))
elif config.init_type == config_pb2.Param.PRETRAINED:
f = h5py.File(config.file_name)
mat = f[config.dataset_name].value
if len(mat.shape) == 1:
mat = mat.reshape(1, -1)
assert self.w_.shape == mat.shape
self.w_.overwrite(mat)
f.close()
else:
raise Exception('Unknown parameter initialization.')
self.eps_ = config.epsilon
self.momentum_ = config.momentum
self.l2_decay_ = config.l2_decay
self.gradient_clip_ = config.gradient_clip
self.eps_decay_factor = config.eps_decay_factor
self.eps_decay_after = config.eps_decay_after
def load(self, modelnamelist):
DBN.load(self,modelnamelist)
for i in range(self.H):
# print type(self.model[i].W)
self.W.append(self.model[i].W)
# self.W[i].copy_to_host()
# print self.W[i].numpy_array
self.dW.append(cm.CUDAMatrix(np.zeros(self.model[i].W.shape)))
self.W_inc.append(cm.CUDAMatrix(np.zeros(self.model[i].W.shape)))
self.b.append(self.model[i].hb)
self.db.append(cm.CUDAMatrix(np.zeros(self.model[i].hb.shape)))
self.b_inc.append(cm.CUDAMatrix(np.zeros(self.model[i].hb.shape)))
self.h.append(cm.empty((self.model[i].num_hid,self.batchsize)))
def test_where():
m = 256
n = 128
a = np.array(np.random.randn(m, n) * 10, dtype=np.float32, order='F')
z = np.zeros_like(a)
res = np.where(a > 0, a, z)
a_d = cm.CUDAMatrix(a)
z_d = cm.CUDAMatrix(z)
res_d = cm.empty(a_d.shape)
a_d.greater_than(0, res_d)
cm.where(res_d, a_d, z_d)
assert np.abs(res - res_d.asarray()).max() < 1e-2, "Error in cudamat.where"
def test_where():
m = 256
n = 128
a = np.array(np.random.randn(m, n)*10, dtype=np.float32, order='F')
z = np.zeros_like(a)
res = np.where(a > 0, a, z);
a_d = cm.CUDAMatrix(a)
z_d = cm.CUDAMatrix(z)
res_d = cm.empty(a_d.shape)
a_d.greater_than(0, res_d)
cm.where(res_d, a_d, z_d)
assert np.abs(res-res_d.asarray()).max() < 1e-2, "Error in cudamat.where"
def __init__(self, input_dim, output_dim, spectral_radius=.9, leak_rate=1,
input_scaling=1, bias_scaling=0):
super(CUDAReservoirNode, self).__init__(input_dim=input_dim, output_dim=output_dim,)
self.input_dim = input_dim
self.output_dim = output_dim
self.leak_rate = leak_rate
w = mdp.numx.random.normal(0, 1, (output_dim, output_dim))
w_in = mdp.numx.random.uniform(-1, 1, (output_dim, input_dim)) * input_scaling
if output_dim < 1500:
l = mdp.numx.linalg.eigvals(w)
r = mdp.numx.amax(mdp.numx.absolute(l))
w = w * (spectral_radius / r)
self.w = cm.CUDAMatrix(w)
else:
self.w = cm.CUDAMatrix(w)
r = get_specrad(self.w)
self.w.mult(spectral_radius / r)
bias = mdp.numx.random.normal(0, 1, (output_dim, 1)) * bias_scaling
self.w_in = cm.CUDAMatrix(w_in)
self.bias = cm.CUDAMatrix(bias)
self.current_state = cm.empty((self.output_dim, 1))
self.new_state = cm.empty((self.output_dim, 1))
def setVariables(self):
n, m, r = self.n, self.m, self.rank
self.G_gpu = cm.CUDAMatrix(self.G)
self.W_gpu = cm.CUDAMatrix(self.W)
self.X_gpu = cm.CUDAMatrix(self.X)
self.XTX_gpu= cm.dot(self.X_gpu.T, self.X_gpu)
self.XTXpos_gpu = cm.empty((m,m))
self.XTX_gpu.greater_than(0, target=self.XTXpos_gpu)
self.XTXpos_gpu.mult(self.XTX_gpu)
self.XTXneg_gpu = cm.empty((m,m))
self.XTXpos_gpu.subtract(self.XTX_gpu, target=self.XTXneg_gpu)
self.XTXnegW_gpu = cm.empty((m,r))
self.XTXposW_gpu = cm.empty((m,r))
self.GWT_gpu = cm.empty((m,m))
self.update1_gpu = cm.empty((m,r))
self.update2_gpu = cm.empty((m,r))
self.GTG_gpu = cm.empty((r,r))
self.XTXnegG_gpu = cm.empty((m,r))
self.XTXposG_gpu = cm.empty((m,r))
def __init__(self, datafolder=None, num_hid=None, options=None):
if datafolder == None:
return
self.datafolder = datafolder
self.datalist = os.listdir(datafolder)
self.num_batchdata = len(self.datalist)
mdict = scipy.io.loadmat(os.path.join(datafolder, self.datalist[0]))
tempdata = mdict['data']
self.options = options
self.num_vis = tempdata.shape[0]
self.num_hid = num_hid
# print self.num_vis
# print self.num_hid
self.num_batches = tempdata.shape[1]/self.options.batchsize
self.batch_size = self.options.batchsize
self.doPCD = False
self.cdstep = 1
# initialize weights
self.W = cm.CUDAMatrix(0.01 * np.random.randn(self.num_vis, self.num_hid))
self.vb = cm.CUDAMatrix(np.zeros((self.num_vis,1)))# for gaussian rbm, v_bias we mean the mean of visible layer
self.hb = cm.CUDAMatrix(np.zeros((self.num_hid,1)))
# initialize weights updates
self.dW = cm.CUDAMatrix(np.zeros((self.num_vis, self.num_hid)))
self.dvb = cm.CUDAMatrix(np.zeros((self.num_vis, 1)))
self.dhb = cm.CUDAMatrix(np.zeros((self.num_hid, 1)))
self.W_inc = cm.CUDAMatrix(np.zeros((self.num_vis, self.num_hid)))
self.vb_inc = cm.CUDAMatrix(np.zeros((self.num_vis,1)))
self.hb_inc = cm.CUDAMatrix(np.zeros((self.num_hid,1)))
# initialize temporary storage
self.v = cm.empty((self.num_vis, self.batch_size))# a batch of data
self.vm = cm.empty((self.num_vis, self.batch_size))# temp storage of data-vb
self.h = cm.empty((self.num_hid, self.batch_size))
self.r = cm.empty((self.num_hid, self.batch_size))# store random number in positive phase
self.r2 = cm.empty((self.num_vis, self.batch_size))# store random number in negative phase
def Get(self, blocksize=None):
if blocksize and blocksize != self.blocksize:
self.blocksize = blocksize
if self._position >= self.current_size or self._data is None:
self._position = 0
if self.num_blocks > 1 or self._data is None:
if self.verbose:
print 'CACHE MISS in %s' % self.name
if self.gpu:
if self._data:
self._data.free_device_memory()
self._data = cm.CUDAMatrix(
self.source.Get((self.capacity)).reshape(1, -1))
self.current_size = self._data.shape[1]
else:
self._data = self.source.Get((self.capacity))
self.current_size = self._data.shape[0]
if self.randomize:
if self.permutation_link:
self.indices = self.permutation_link.indices
else:
p = np.arange(self.current_size / self.numdims)
np.random.shuffle(p)
if self.gpu:
if self.indices is not None:
self.indices.free_device_memory()
p2 = p.view()
p2.shape = 1, -1
self.indices = cm.CUDAMatrix(p2)
else:
self.indices = p
if self.gpu:
self._data.reshape((self.numdims, self.current_size / self.numdims))
shuffled_data = cm.empty(self._data.shape)
self._data.select_columns(self.indices, target=shuffled_data)
self._data.free_device_memory()
self._data = shuffled_data
self._data.reshape((1, self.current_size))
else:
view = self._data.view()
view.shape = self.current_size / self.numdims, self.numdims
self._data = view[self.indices,:].reshape(-1)
span = min(self.blocksize, self.current_size - self._position)
self._position += span
if self.gpu:
return self._data.slice(self._position - span, self._position)
else:
return self._data[self._position - span : self._position]
def LoadParams(self, proto):
self.hyperparams = proto.hyperparams
param_names = [param.name for param in proto.param]
for param in proto.param:
if not param.dimensions:
param.dimensions.extend([proto.numlabels * proto.dimensions])
if param.mat:
mat = util.ParameterAsNumpy(param).reshape(-1, 1)
else:
mat = self.InitializeParameter(param).reshape(-1, 1)
self.params[param.name] = cm.CUDAMatrix(mat)
if param.name == 'bias':
self.grad_bias = cm.empty(mat.shape)
self.grad_bias.assign(0)
self.sample_input = self.hyperparams.sample_input
