我们从Python开源项目中，提取了以下11个代码示例，用于说明如何使用theano.tensor.split()。
def lyr_lstm( self, name_, s_x_, s_cell_, s_hid_, idim_, hdim_, axis_=-1, lyr_linear_=None, op_act_=T.tanh, op_gate_=T.nnet.sigmoid): s_inp = T.join(axis_, s_x_, s_hid_) if lyr_linear_ is None: lyr_linear_ = self.lyr_linear s_gates_lin, s_inp_lin = T.split( lyr_linear_(name_+'_rec', s_inp, idim_+hdim_, hdim_*4), [hdim_*3,hdim_], 2, axis=axis_) s_igate, s_fgate, s_ogate = T.split(op_gate_(s_gates_lin), [hdim_]*3, 3, axis=axis_) s_cell_tp1 = s_igate*op_act_(s_inp_lin) + s_fgate*s_cell_ s_hid_tp1 = op_act_(s_cell_tp1)*s_ogate return s_cell_tp1, s_hid_tp1
def lyr_gru_flat( self, name_, s_x_, s_state_, idim_, hdim_, axis_=-1, lyr_linear_=None, op_act_=T.tanh, op_gate_=T.nnet.sigmoid, params_group_='params' ): ''' GRU layer, flat version In order to use, you need to provide state variable ''' if lyr_linear_ is None: lyr_linear_ = self.lyr_linear s_igate = lyr_linear_(name_+'_igate', idim_+hdim_, idim_, params_group_=params_group_) s_inp_gated = T.join(axis_, s_x_ * op_gate_(s_igate), s_state_) s_gate_lin, s_state_tp1_lin = T.split(lyr_linear_(name_+'_gate', s_inp_gated, idim_+hdim_, hdim_*2), [hdim_,hdim_], 2, axis_) s_gate = op_gate_(s_gate_lin) return s_state_*s_gate + op_act_(s_state_tp1_lin)*(1.-s_gate)
def get_output_for(self, input, deterministic=False, **kwargs): def _phase_shift(input,r): bsize,c,a,b = input.shape[0],1,self.output_shape[2]//r,self.output_shape[3]//r X = T.reshape(input, (bsize,r,r,a,b)) X = T.transpose(X, (0, 3,4,1,2)) # bsize, a, b, r2,r1 X = T.split(x=X,splits_size=[1]*a,n_splits=a,axis=1) # a, [bsize, b, r, r] X = [T.reshape(x,(bsize,b,r,r))for x in X] X = T.concatenate(X,axis=2) # bsize, b, a*r, r X = T.split(x=X,splits_size =[1]*b,n_splits=b,axis=1) # b, [bsize, a*r, r] X = [T.reshape(x,(bsize,a*r,r))for x in X] X = T.concatenate(X,axis=2) # bsize, a*r, b*r return X.dimshuffle(0,'x',1,2) Xc = T.split(x=input,splits_size =[input.shape[1]//self.c]*self.c,n_splits=self.c,axis=1) return T.concatenate([_phase_shift(xc,self.r) for xc in Xc],axis=1) # Multiscale Dilated Convolution Block # This function (not a layer in and of itself, though you could make it one) returns a set of concatenated conv2d and dilatedconv2d layers. # Each layer uses the same basic filter W, operating at a different dilation factor (or taken as the mean of W for the 1x1 conv). # The channel-wise output of each layer is weighted by a set of coefficients, which are initialized to 1 / the total number of dilation scales, # meaning that were starting by taking an elementwise mean. These should be learnable parameters. # NOTES: - I'm considering changing the variable names to be more descriptive, and look less like ridiculous academic code. It's on the to-do list. # - I keep the bias and nonlinearity out of the default definition for this layer, as I expect it to be batchnormed and nonlinearized in the model config.
def test_local_useless_split(): x = tensor.matrix('x') splits = tensor.ivector('splits') opt = tensor.split(x, splits, n_splits=1) nonopt = tensor.split(x, splits, n_splits=3) mode = compile.get_default_mode().including("local_useless_split") f_opt = theano.function([x, splits], opt, mode=mode) f_nonopt = theano.function([x, splits], nonopt, mode=mode) f_opt(numpy.random.rand(4,4).astype(config.floatX), [4]) f_nonopt(numpy.random.rand(4,4).astype(config.floatX), [1,2,1]) graph_opt = f_opt.maker.fgraph.toposort() graph_nonopt = f_nonopt.maker.fgraph.toposort() assert isinstance(graph_opt[-1].op, DeepCopyOp) assert len(graph_nonopt)==1 assert isinstance(graph_nonopt[0].op, tensor.Split) assert check_stack_trace(f_opt, ops_to_check=[Assert]) assert check_stack_trace(f_nonopt, ops_to_check='all')
def infer_shape(self, node, in_shapes): shape_a = in_shapes[0] n = node.inputs[1] axis = node.inputs[2] if len(shape_a) == 1: return [(n,)] elif isinstance(axis, tensor.TensorConstant): out_shape = (list(shape_a[0: axis.data.item()]) + [n] + list(shape_a[axis.data + 1:])) else: l = len(shape_a) shape_a = tensor.stack(shape_a) out_shape = tensor.concatenate((shape_a[0: axis], [n], shape_a[axis + 1:])) n_splits = [1] * l out_shape = tensor.split(out_shape, n_splits, l) out_shape = [a[0] for a in out_shape] return [out_shape]
def lyr_gru( self, name_, s_x_, s_state_, idim_, hdim_, axis_=0, lyr_linear_=None, op_act_=T.tanh, op_gate_=T.nnet.sigmoid): if lyr_linear_ is None: lyr_linear_ = self.lyr_linear s_igate = lyr_linear_(name_+'_igate', idim_+hdim_, idim_) s_inp_gated = T.join(axis_, s_x_ * op_gate_(s_igate), s_state_) s_gate_lin, s_state_tp1_lin = T.split(lyr_linear_(name_+'_gate', s_inp_gated, idim_+hdim_, hdim_*2), [hdim_,hdim_], 2, axis_) s_gate = op_gate_(s_gate_lin) return s_state_*s_gate + op_act_(s_state_tp1_lin)*(1.-s_gate)
def lyr_lstm_flat( self, name_, s_x_, s_cell_, s_hid_, idim_, hdim_, axis_=-1, lyr_linear_=None, op_act_=T.tanh, op_gate_=T.nnet.sigmoid, params_group_='params' ): ''' LSTM layer, flat version In order to use, you need to provide state variable Returns: hidden_state, cell_state ''' s_inp = T.join(axis_, s_x_, s_hid_) if lyr_linear_ is None: lyr_linear_ = self.lyr_linear s_gates_lin, s_inp_lin = T.split( lyr_linear_(name_+'_rec', s_inp, idim_+hdim_, hdim_*4), [hdim_*3,hdim_], 2, axis=axis_) s_igate, s_fgate, s_ogate = T.split(op_gate_(s_gates_lin), [hdim_]*3, 3, axis=axis_) s_cell_tp1 = s_igate*op_act_(s_inp_lin) + s_fgate*s_cell_ s_hid_tp1 = op_act_(s_cell_tp1)*s_ogate return s_cell_tp1, s_hid_tp1
def assert_func_pair_optimized(self, func1, func2, data, should_copy=True, is_complex=False): """ Check that a pair of funcs is optimized properly """ x = T.cmatrix() if is_complex else T.fmatrix() o = func2(func1(x)) f = theano.function([x], o, mode=self.mode) delta = f(data) - data topo = f.maker.fgraph.toposort() if should_copy: acceptable_topo_lens = [1] else: # The 2 funcs can be split apart if they are not inverses acceptable_topo_lens = [1, 2] if should_copy: delta_condition = numpy.all(delta == 0) else: delta_condition = numpy.all(delta != 0) self.assertTrue(len(topo) in acceptable_topo_lens) self.assertTrue(delta_condition) self.assertEqual(isinstance(topo[0].op, DeepCopyOp), should_copy, "Inverse functions not removed!")
def _log_partition_symfunc(): natural_params = T.vector() size = natural_params.shape[0] // 4 np1, np2, np3, np4 = T.split(natural_params, 4 * [size], 4) log_Z = T.sum(T.gammaln(.5 * (np4 + 1))) log_Z += T.sum(- .5 * (np4 + 1) * T.log(.5 * (np1 - (np2 ** 2) / np3))) log_Z += T.sum(-.5 * T.log(np3)) func = theano.function([natural_params], log_Z) grad_func = theano.function([natural_params], T.grad(T.sum(log_Z), natural_params)) return func, grad_func
def test_local_gpu_split(): """ Test that the GpuSplit op is being applied and works """ # Construct symbolic split x = tensor.fvector() splits = tensor.lvector() ra, rb, rc = tensor.split(x, splits, n_splits=3, axis=0) # Compile function to use CPU f = theano.function([x, splits], [ra, rb, rc], mode=mode_without_gpu) # Get values for CPU version cpu_res = f([0, 1, 2, 3, 4, 5], [3, 2, 1]) l = f.maker.fgraph.toposort() # Ensure that one op is theano.tensor.Split assert any([isinstance(o.op, theano.tensor.Split) for o in l]) # GPU version f = theano.function([x, splits], [ra, rb, rc], mode=mode_with_gpu) gpu_res = f([0, 1, 2, 3, 4, 5], [3, 2, 1]) l = f.maker.fgraph.toposort() assert any([isinstance(o.op, cuda.GpuSplit) for o in l]) # Check equality assert all([(cpu == gpu).all() for cpu, gpu in zip(cpu_res, gpu_res)]) # Test the other path of the optimizer, when it is the output that # is moved to the GPU. ra = cuda.gpu_from_host(ra) f = theano.function([x, splits], [ra, rb, rc], mode=mode_with_gpu.excluding("InputToGpuOptimizer")) gpu_res = f([0, 1, 2, 3, 4, 5], [3, 2, 1]) l = f.maker.fgraph.toposort() assert any([isinstance(o.op, cuda.GpuSplit) for o in l]) # Check equality assert all([(cpu == gpu).all() for cpu, gpu in zip(cpu_res, gpu_res)]) # Test that split with only 1 output work ra = tensor.split(x, splits, n_splits=1, axis=0) f = theano.function([x, splits], [ra], mode=mode_without_gpu) cpu_res = f([0, 1, 2, 3, 4, 5], [6]) l = f.maker.fgraph.toposort() # Ensure that no op is theano.tensor.Split or GpuSplit, they get # optimized away. assert not any([isinstance(o.op, (theano.tensor.Split, cuda.GpuSplit)) for o in l]) # GPU version f = theano.function([x, splits], [ra], mode=mode_with_gpu) gpu_res = f([0, 1, 2, 3, 4, 5], [6]) l = f.maker.fgraph.toposort() assert not any([isinstance(o.op, (theano.tensor.Split, cuda.GpuSplit)) for o in l]) # Check equality assert all([(cpu == gpu).all() for cpu, gpu in zip(cpu_res, gpu_res)])
