Python cudamat.CUDAMatrix方法代碼示例

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
def cudamat_to_cudandarray(x):
        """ take a cudamat.CUDAMatrix and make a CudaNdarray that point to its memory
        """
        if not isinstance(x, cudamat.CUDAMatrix):
            raise ValueError("We can transfer only cudamat.CUDAMatrix to CudaNdarray")
        # elif x.dtype != "float32":
        # raise ValueError("CudaNdarray support only float32")
        # We don't need this, because cudamat is always float32.
        else:
            strides = [1]
            for i in x.shape[::-1][:-1]:
                strides.append(strides[-1] * i)
            strides = tuple(strides[::-1])

            import ctypes
            ptr_long = long(ctypes.cast(x.mat.data_device, ctypes.c_void_p).value)

            # seems legit.
            z = cuda.from_gpu_pointer(ptr_long, x.shape, strides, x)
            return z 

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
def return_NumpyArray(input_array):
    '''
    If input is a numpy_array, return input
    If input is a CUDAMatrix, return numpy array
    '''
    if isinstance(input_array, np.ndarray):
        return input_array
    elif isinstance(input_array, cm.CUDAMatrix):
        if input_array.shape[1] == 1:
            return input_array.asarray().flatten()
        else:
            return input_array.asarray()
        return input_array
    else:
        raise ValueError('cannot handle input of type: %s'
                         % (type(input_array),)) 

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
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 

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
def _init_gpu():
 """ picks a board and claims it (if using cudamat aot npmat). exception if there is no board. """
 if '__gpu_inited' in globals(): return
 global _boardId
 if _useGpu=='yes':
  _boardId = ( board_id_to_use() if callable(board_id_to_use) else board_id_to_use)
  if _boardId==-1: raise GnumpyGpuUnavailableException('No gpu board is available. gnumpy will not function. Consider telling it to run on the CPU by setting environment variable GNUMPY_USE_GPU to "no".')
  _cudamat.cuda_set_device(_boardId)
  _cudamat.cublas_init()
 _cudamat.CUDAMatrix.init_random(0)
 globals()['__gpu_inited'] = None 

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
def _new_cm(sizeOrShape):
 """
 Internal.
 Returns a new CUDAMatrix object of the given size.
 This is the only proc that allocs gpu mem.
 """
 global __memoryInUse
 if type(sizeOrShape) == tuple:
  if _prodT(sizeOrShape)==0: return _new_cm(1) # cudamat workaround: cudamat can't handle size 0 arrays
  else: return _new_cm(sizeOrShape[0]*sizeOrShape[1]).reshape((sizeOrShape[1], sizeOrShape[0]))
 size = sizeOrShape
 if size==0: return _cudamat.empty((1, 1)) # cudamat workaround
 if len(_cmsForReuse[size])!=0:
  return _cm_reshape(_cmsForReuse[size].pop(), (1, size)) # re-use an abandoned cm
 _init_gpu()
 if __memoryInUse+size*4*5 > max_memory_usage: free_reuse_cache(False) # if we're somewhat close to the limit, then free what's easy to free, and hope that there are contiguous blocks available.
 if __memoryInUse+size*4 > max_memory_usage: # if we're (still) OVER the limit, then do whatever can be done to make more mem available
  free_reuse_cache(True) # gc.collect can take quite some time
  if __memoryInUse+size*4 > max_memory_usage:
   raise MemoryError('Gnumpy ran out of memory. Currently in use are %s; the maximum allowed is %s; so the present request for %s is refused. Free some memory and try again.' % (_n_bytes_str(__memoryInUse), _n_bytes_str(max_memory_usage), _n_bytes_str(size*4)))
 try:
  ret = _cudamat.empty((size, 1))
  __memoryInUse += size*4 # do this only if the malloc succeeded
  return ret
 except _cudamat.CUDAMatException, e: # this means that malloc failed
  raise MemoryError('The GPU failed to allocate the requested %d bytes of memory. This doesn\'t mean that your program is using too much memory. It does, however, mean that you should reduce the value of gnumpy.max_memory_usage (currently %s), to always have some memory unused (which is necessary to find contiguous large blocks of memory to allocate). Failing to allocate enough memory makes the GPU feel very unwell, so you are advised to restart Python now, or expect to see incoherent error messages and risk causing more serious damage.' % (size*4, str(max_memory_usage))) 

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
def seed_rand(seed=None):
 _init_gpu()
 if seed==None: seed = int(_time.time())
 _cudamat.CUDAMatrix.init_random(seed) 

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
def return_CUDAMatrix(input_array):
    '''
    If input is a numpy_array, convert to CUDAMatrix.
    If input is already CUDAMatrix, return input
    '''
    if isinstance(input_array, np.ndarray):
        if input_array.ndim == 1:
            return cm.CUDAMatrix(input_array[:, np.newaxis])
        else:
            return cm.CUDAMatrix(input_array)
    elif isinstance(input_array, cm.CUDAMatrix):
        return input_array
    else:
        raise ValueError('cannot handle input of type: %s'
                         % (type(input_array),)) 

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
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] 

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
def fromFile(self,fid):
	import cPickle as pickle
        stack = pickle.load(fid)
        self.stack = [[cm.CUDAMatrix(w),cm.CUDAMatrix(b)] 
                        for w,b in stack] 

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
def _norm(self, x, ws):
        """Calculate uj so that we can normalize it."""
        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
            y_local = y.asarray()
            del y
            del wc
            tmp_sum = np.einsum('lj,lj->j', y_local, y_local)  # TODO: Should be able to do on gpu...
        else:
            y = x.dot(ws.T)  # + noise / std Y_j^2, but it is included analytically
            tmp_sum = np.einsum('lj,lj->j', y, y)
        return np.sqrt((1 - self.eps**2) * tmp_sum / self.n_samples + self.eps**2 * np.sum(ws**2, axis=1)) 

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
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
        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 

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
def preprocess(self, x, fit=False):
        """Transform each marginal to be as close to a standard Gaussian as possible.
        'standard' (default) just subtracts the mean and scales by the std.
        'empirical' does an empirical gaussianization (but this cannot be inverted).
        'outliers' tries to squeeze in the outliers
        Any other choice will skip the transformation."""
        if self.missing_values is not None:
            x, self.n_obs = mean_impute(x, self.missing_values)  # Creates a copy
        else:
            self.n_obs = len(x)
        if self.gaussianize == 'none':
            pass
        elif self.gaussianize == 'standard':
            if fit:
                mean = np.mean(x, axis=0)
                # std = np.std(x, axis=0, ddof=0).clip(1e-10)
                std = np.sqrt(np.sum((x - mean)**2, axis=0) / self.n_obs).clip(1e-10)
                self.theta = (mean, std)
            x = ((x - self.theta[0]) / self.theta[1])
            if np.max(np.abs(x)) > 6 and self.verbose:
                print("Warning: outliers more than 6 stds away from mean. Consider using gaussianize='outliers'")
        elif self.gaussianize == 'outliers':
            if fit:
                mean = np.mean(x, axis=0)
                std = np.std(x, axis=0, ddof=0).clip(1e-10)
                self.theta = (mean, std)
            x = g((x - self.theta[0]) / self.theta[1])  # g truncates long tails
        elif self.gaussianize == 'empirical':
            print("Warning: correct inversion/transform of empirical gauss transform not implemented.")
            x = np.array([norm.ppf((rankdata(x_i) - 0.5) / len(x_i)) for x_i in x.T]).T
        if self.gpu and fit:  # Don't return GPU matrices when only transforming
            x = cm.CUDAMatrix(x)
        return x 

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
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.layerSize,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]

        if self.temporalLayer > 0:
            # dummy bias used for temporal layer
            dummy = cm.empty((1,1))
            dummy.assign(0.0)

            scale = np.sqrt(6)/np.sqrt(self.layerSize*2)
            wt = 2*scale*np.random.rand(self.layerSize,self.layerSize)-scale
            wt = cm.CUDAMatrix(wt)
            self.stack.append([wt,dummy])

            if self.train:
                dwt = cm.empty((self.layerSize,self.layerSize))
                self.grad.append([dwt,dummy])

                self.deltaTemp_M = cm.empty((self.layerSize,self.maxBatch)) 

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
def costAndGrad(self,data,labels):
        
        batchSize = data.shape[1]
        self.setViews(batchSize)
        
        # forward prop
        self.hActs[0].assign(cm.CUDAMatrix(data))

        i = 1
        for w,b in self.stack:
            cm.dot(w,self.hActs[i-1],self.hActs[i])
            self.hActs[i].add_col_vec(b)
            if i <= len(self.layerSizes):
                # hard relu
                self.hActs[i].maximum(0.0)
            i += 1

        # Subtract max activation
        self.hActs[-1].max(axis=0,target=self.rowVec)
        self.hActs[-1].add_row_mult(self.rowVec,-1.0,target=self.probs)

        # Softmax
        cm.exp(self.probs)
        self.probs.sum(axis=0,target=self.rowVec)
        cm.pow(self.rowVec,-1.0,target=self.rowVec)
        self.probs.mult_by_row(self.rowVec)

        self.probs.copy_to_host()
        cost, deltas, skip = ctc.ctc_loss(self.probs.numpy_array.astype(np.float64),
                labels,blank=0)
        self.deltasC.assign(cm.CUDAMatrix(deltas))

        if skip:
            return cost,self.grad,skip

        # back prop
        nl = len(self.layerSizes)
        i = nl 
        deltasIn,deltasOut = self.deltasC,self.deltasOut
        for w,b in reversed(self.stack):
            # compute gradient
            cm.dot(deltasIn,self.hActs[i].T,target=self.grad[i][0])
            deltasIn.sum(axis=1,target=self.grad[i][1])

            # compute next layer deltas
            if i > 0:
                self.hActs[i].sign(target=self.tmpGrad)
                cm.dot(w.T,deltasIn,target=deltasOut)
                deltasOut.mult(self.tmpGrad)

            if i == nl:
                deltasIn = self.deltasIn

            deltasIn,deltasOut = deltasOut,deltasIn
            i -= 1

        return cost,self.grad,skip 

# 需要導入模塊: import cudamat [as 別名]
# 或者: from cudamat import CUDAMatrix [as 別名]
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.layerSize,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]

        if self.temporalLayer > 0:
            # dummy bias used for temporal layer
            dummy = cm.empty((1,1))
            dummy.assign(0.0)

            scale = np.sqrt(6)/np.sqrt(self.layerSize*2)
            wtf = cm.CUDAMatrix(2*scale*np.random.rand(self.layerSize,
                    self.layerSize)-scale)
            wtb = cm.CUDAMatrix(2*scale*np.random.rand(self.layerSize,
                    self.layerSize)-scale)
            self.stack.append([wtf,dummy])
            self.stack.append([wtb,dummy])

            # forward and backward activations for temporal layer
            self.hActsFor_M = cm.empty((self.layerSize,self.maxBatch))
            self.hActsBack_M = cm.empty((self.layerSize,self.maxBatch))

            if self.train:
                dwtf = cm.empty(wtf.shape)
                self.grad.append([dwtf,dummy])
                dwtb = cm.empty(wtb.shape)
                self.grad.append([dwtb,dummy])

                self.tmpGradBack_M = cm.empty((self.layerSize,self.maxBatch))
                self.deltasFor_M = cm.empty((self.layerSize,self.maxBatch))
                self.deltasBack_M = cm.empty((self.layerSize,self.maxBatch))