def __init__(self, collapse='mean', maxout=False, transpose=False, **kwargs): super(TwoDToOneDLayer, self).__init__(1, **kwargs) self.set_attr('collapse', collapse) self.set_attr('transpose', transpose) Y = self.sources[0].output if transpose: Y = Y.dimshuffle(1, 0, 2, 3) #index handling def index_fn(index, size): return T.set_subtensor(index[:size], numpy.cast['int8'](1)) index_init = T.zeros((Y.shape[2],Y.shape[1]), dtype='int8') self.index, _ = theano.scan(index_fn, [index_init, T.cast(self.sources[0].output_sizes[:,1],"int32")]) self.index = self.index.dimshuffle(1, 0) n_out = self.sources[0].attrs['n_out'] if maxout: Y = Y.max(axis=3).dimshuffle(0,1,2,'x') if collapse == 'sum' or collapse == True: Y = Y.sum(axis=0) elif collapse == 'mean': Y = Y.mean(axis=0) elif collapse == 'conv': from TheanoUtil import circular_convolution Y, _ = theano.scan(lambda x_i,x_p:circular_convolution(x_i,x_p),Y,Y[0]) Y = Y[-1] elif collapse == 'flatten': self.index = T.ones((Y.shape[0] * Y.shape[1], Y.shape[2]), dtype='int8') Y = Y.reshape((Y.shape[0]*Y.shape[1],Y.shape[2],Y.shape[3])) elif str(collapse).startswith('pad_'): pad = numpy.int32(collapse.split('_')[-1]) Y = ifelse(T.lt(Y.shape[0],pad),T.concatenate([Y,T.zeros((pad-Y.shape[0],Y.shape[1],Y.shape[2],Y.shape[3]),'float32')],axis=0), ifelse(T.gt(Y.shape[0],pad),Y[:pad],Y)) Y = Y.dimshuffle(1,2,3,0).reshape((Y.shape[1],Y.shape[2],Y.shape[3]*Y.shape[0])) n_out *= pad elif collapse != False: assert False, "invalid collapse mode" if self.attrs['batch_norm']: Y = self.batch_norm(Y, n_out, force_sample=False) self.output = Y self.act = [Y, Y] self.set_attr('n_out', n_out)
def __init__(self, collapse='mean', maxout=False, transpose=False, **kwargs): super(TwoDToOneDLayer, self).__init__(1, **kwargs) self.set_attr('collapse', collapse) self.set_attr('transpose', transpose) Y = self.sources[0].output if transpose: Y = Y.dimshuffle(1, 0, 2, 3) #index handling def index_fn(index, size): return T.set_subtensor(index[:size], numpy.cast['int8'](1)) index_init = T.zeros((Y.shape[2],Y.shape[1]), dtype='int8') self.index, _ = theano.scan(index_fn, [index_init, T.cast(self.sources[0].output_sizes[:,1],"int32")]) self.index = self.index.dimshuffle(1, 0) n_out = self.sources[0].attrs['n_out'] if maxout: Y = Y.max(axis=3).dimshuffle(0,1,2,'x') if collapse == 'sum' or collapse == True: Y = Y.sum(axis=0) elif collapse == 'mean': Y = Y.mean(axis=0) elif collapse == 'conv': from TheanoUtil import circular_convolution Y, _ = theano.scan(lambda x_i,x_p:circular_convolution(x_i,x_p),Y,Y[0]) Y = Y[-1] elif collapse == 'flatten': self.index = T.ones((Y.shape[0] * Y.shape[1], Y.shape[2]), dtype='int8') Y = Y.reshape((Y.shape[0]*Y.shape[1],Y.shape[2],Y.shape[3])) elif str(collapse).startswith('pad_'): pad = numpy.int32(collapse.split('_')[-1]) Y = ifelse(T.lt(Y.shape[0],pad),T.concatenate([Y,T.zeros((pad-Y.shape[0],Y.shape[1],Y.shape[2],Y.shape[3]),'float32')],axis=0), ifelse(T.gt(Y.shape[0],pad),Y[:pad],Y)) Y = Y.dimshuffle(1,2,3,0).reshape((Y.shape[1],Y.shape[2],Y.shape[3]*Y.shape[0])) n_out *= pad elif collapse != False: assert False, "invalid collapse mode" if self.attrs['batch_norm']: Y = self.batch_norm(Y, n_out, force_sample=False) self.output = Y self.act = [Y, Y] self.set_attr('n_out', n_out)
def __init__(self, n_out, collapse_output=False, directions=4, projection='average', base=None, **kwargs): if base is None: base = [] super(TwoDLSTMLayer, self).__init__(n_out, **kwargs) assert len(self.sources) == 1 source = self.sources[0] n_in = source.attrs['n_out'] X = source.output assert X.ndim == 4 sizes = source.output_sizes self.output_sizes = sizes assert directions in [1, 2, 4], "only 1, 2 or 4 directions are supported" assert projection in ['average', 'concat'], "invalid projection" if base: self.b1 = self.add_param(base[0].b1) self.b2 = self.add_param(base[0].b2) if directions >= 1: self.b3 = self.add_param(base[0].b3) self.b4 = self.add_param(base[0].b4) self.W1, self.V_h1, self.V_v1 = self.add_param( base[0].W1), self.add_param(base[0].V_h1), self.add_param( base[0].V_v1) self.W2, self.V_h2, self.V_v2 = self.add_param( base[0].W2), self.add_param(base[0].V_h2), self.add_param( base[0].V_v2) if directions >= 1: self.W3, self.V_h3, self.V_v3 = self.add_param( base[0].W3), self.add_param(base[0].V_h3), self.add_param( base[0].V_v3) self.W4, self.V_h4, self.V_v4 = self.add_param( base[0].W4), self.add_param(base[0].V_h4), self.add_param( base[0].V_v4) #self.mass = base[0].mass #self.masks = base[0].masks #self.b1 = base[0].b1 #self.b2 = base[0].b2 #if directions >= 1: # self.b3 = base[0].b3 # self.b4 = base[0].b4 #self.W1, self.V_h1, self.V_v1 = base[0].W1, base[0].V_h1, base[0].V_v1 #self.W2, self.V_h2, self.V_v2 = base[0].W2, base[0].V_h2, base[0].V_v2 #if directions >= 1: # self.W3, self.V_h3, self.V_v3 = base[0].W3, base[0].V_h3, base[0].V_v3 # self.W4, self.V_h4, self.V_v4 = base[0].W4, base[0].V_h4, base[0].V_v4 self.mass = base[0].mass self.masks = base[0].masks else: self.b1 = self.create_and_add_bias(n_out, "1") self.b2 = self.create_and_add_bias(n_out, "2") if directions >= 1: self.b3 = self.create_and_add_bias(n_out, "3") self.b4 = self.create_and_add_bias(n_out, "4") self.W1, self.V_h1, self.V_v1 = self.create_and_add_2d_lstm_weights( n_in, n_out, "1") self.W2, self.V_h2, self.V_v2 = self.create_and_add_2d_lstm_weights( n_in, n_out, "2") if directions >= 1: self.W3, self.V_h3, self.V_v3 = self.create_and_add_2d_lstm_weights( n_in, n_out, "3") self.W4, self.V_h4, self.V_v4 = self.create_and_add_2d_lstm_weights( n_in, n_out, "4") # dropout assert len(self.masks) == 1 mask = self.masks[0] if mask is not None: X = self.mass * mask * X if str(theano.config.device).startswith('cpu'): Y = T.zeros_like(X) if projection == 'concat': Y = Y.repeat(directions, axis=-1) n_out *= directions else: if directions <= 2: Y = BidirectionalTwoDLSTMOpInstance(X, self.W1, self.W2, self.V_h1, self.V_h2, self.V_v1, self.V_v2, self.b1, self.b2, sizes) else: Y = MultiDirectionalTwoDLSTMOpInstance( X, self.W1, self.W2, self.W3, self.W4, self.V_h1, self.V_h2, self.V_h3, self.V_h4, self.V_v1, self.V_v2, self.V_v3, self.V_v4, self.b1, self.b2, self.b3, self.b4, sizes) if directions > 1: Y = T.stack(Y[:directions], axis=-1) if projection == 'average': Y = Y.mean(axis=-1) elif projection == 'concat': Y = Y.reshape((Y.shape[0], Y.shape[1], Y.shape[2], Y.shape[3] * Y.shape[4])) n_out *= directions else: Y = Y[0] Y.name = 'Y' self.set_attr('n_out', n_out) self.set_attr('collapse_output', collapse_output) self.set_attr('directions', directions) self.set_attr('projection', projection) #index handling def index_fn(index, size): return T.set_subtensor(index[:size], numpy.cast['int8'](1)) index_init = T.zeros((Y.shape[2], Y.shape[1]), dtype='int8') self.index, _ = theano.scan( index_fn, [index_init, T.cast(sizes[:, 1], "int32")]) self.index = self.index.dimshuffle(1, 0) if collapse_output == 'sum' or collapse_output == True: Y = Y.sum(axis=0) elif collapse_output == 'mean': Y = Y.mean(axis=0) elif collapse_output == 'conv': from TheanoUtil import circular_convolution Y, _ = theano.scan(lambda x_i, x_p: circular_convolution(x_i, x_p), Y, Y[0]) Y = Y[-1] elif collapse_output == 'flatten': self.index = T.ones((Y.shape[0] * Y.shape[1], Y.shape[2]), dtype='int8') Y = Y.reshape((Y.shape[0] * Y.shape[1], Y.shape[2], Y.shape[3])) elif str(collapse_output).startswith('pad_'): pad = numpy.int32(collapse_output.split('_')[-1]) Y = ifelse( T.lt(Y.shape[0], pad), T.concatenate([ Y, T.zeros( (pad - Y.shape[0], Y.shape[1], Y.shape[2], Y.shape[3]), 'float32') ], axis=0), ifelse(T.gt(Y.shape[0], pad), Y[:pad], Y)) Y = Y.dimshuffle(1, 2, 3, 0).reshape( (Y.shape[1], Y.shape[2], Y.shape[3] * Y.shape[0])) self.attrs['n_out'] *= pad elif collapse_output != False: assert False, "invalid collapse mode" if self.attrs['batch_norm']: Y = self.batch_norm( Y, self.attrs['n_out'], index=sizes if not collapse_output else self.index, force_sample=False) self.output = Y
def __init__(self, n_out, collapse_output=False, directions=4, projection='average', base=None, **kwargs): if base is None: base = [] super(TwoDLSTMLayer, self).__init__(n_out, **kwargs) assert len(self.sources) == 1 source = self.sources[0] n_in = source.attrs['n_out'] X = source.output assert X.ndim == 4 sizes = source.output_sizes if source.layer_class == "1Dto2D": #sizes has the wrong layout if coming directly from a 1Dto2D layer sizes = sizes.reshape((2, sizes.size // 2)).dimshuffle(1, 0) self.output_sizes = sizes assert directions in [1,2,4], "only 1, 2 or 4 directions are supported" assert projection in ['average', 'concat'], "invalid projection" if base: #self.b1 = self.add_param(base[0].b1) #self.b2 = self.add_param(base[0].b2) #if directions >= 1: # self.b3 = self.add_param(base[0].b3) # self.b4 = self.add_param(base[0].b4) #self.W1, self.V_h1, self.V_v1 = self.add_param(base[0].W1), self.add_param(base[0].V_h1), self.add_param(base[0].V_v1) #self.W2, self.V_h2, self.V_v2 = self.add_param(base[0].W2), self.add_param(base[0].V_h2), self.add_param(base[0].V_v2) #if directions >= 1: # self.W3, self.V_h3, self.V_v3 = self.add_param(base[0].W3), self.add_param(base[0].V_h3), self.add_param(base[0].V_v3) # self.W4, self.V_h4, self.V_v4 = self.add_param(base[0].W4), self.add_param(base[0].V_h4), self.add_param(base[0].V_v4) #self.mass = base[0].mass #self.masks = base[0].masks self.b1 = base[0].b1 self.b2 = base[0].b2 if directions >= 1: self.b3 = base[0].b3 self.b4 = base[0].b4 self.W1, self.V_h1, self.V_v1 = base[0].W1, base[0].V_h1, base[0].V_v1 self.W2, self.V_h2, self.V_v2 = base[0].W2, base[0].V_h2, base[0].V_v2 if directions >= 1: self.W3, self.V_h3, self.V_v3 = base[0].W3, base[0].V_h3, base[0].V_v3 self.W4, self.V_h4, self.V_v4 = base[0].W4, base[0].V_h4, base[0].V_v4 self.mass = base[0].mass self.masks = base[0].masks else: self.b1 = self.create_and_add_bias(n_out, "1") self.b2 = self.create_and_add_bias(n_out, "2") if directions >= 1: self.b3 = self.create_and_add_bias(n_out, "3") self.b4 = self.create_and_add_bias(n_out, "4") self.W1, self.V_h1, self.V_v1 = self.create_and_add_2d_lstm_weights(n_in, n_out, "1") self.W2, self.V_h2, self.V_v2 = self.create_and_add_2d_lstm_weights(n_in, n_out, "2") if directions >= 1: self.W3, self.V_h3, self.V_v3 = self.create_and_add_2d_lstm_weights(n_in, n_out, "3") self.W4, self.V_h4, self.V_v4 = self.create_and_add_2d_lstm_weights(n_in, n_out, "4") # dropout assert len(self.masks) == 1 mask = self.masks[0] if mask is not None: X = self.mass * mask * X if str(theano.config.device).startswith('cpu'): Y = T.zeros_like(X) if projection == 'concat': Y = Y.repeat(directions,axis=-1) n_out *= directions else: if directions <= 2: Y = BidirectionalTwoDLSTMOpInstance(X, self.W1, self.W2, self.V_h1, self.V_h2, self.V_v1, self.V_v2, self.b1, self.b2, sizes) else: Y = MultiDirectionalTwoDLSTMOpInstance(X, self.W1, self.W2, self.W3, self.W4, self.V_h1, self.V_h2, self.V_h3, self.V_h4, self.V_v1, self.V_v2, self.V_v3, self.V_v4, self.b1, self.b2, self.b3, self.b4, sizes) if directions > 1: Y = T.stack(Y[:directions],axis=-1) if projection == 'average': Y = Y.mean(axis=-1) elif projection == 'concat': Y = Y.reshape((Y.shape[0],Y.shape[1],Y.shape[2],Y.shape[3]*Y.shape[4])) n_out *= directions else: Y = Y[0] Y.name = 'Y' self.set_attr('n_out', n_out) self.set_attr('collapse_output', collapse_output) self.set_attr('directions', directions) self.set_attr('projection', projection) #index handling def index_fn(index, size): return T.set_subtensor(index[:size], numpy.cast['int8'](1)) index_init = T.zeros((Y.shape[2],Y.shape[1]), dtype='int8') self.index, _ = theano.scan(index_fn, [index_init, T.cast(sizes[:,1],"int32")]) self.index = self.index.dimshuffle(1, 0) if collapse_output == 'sum' or collapse_output == True: Y = Y.sum(axis=0) elif collapse_output == 'mean': Y = Y.mean(axis=0) elif collapse_output == 'conv': from TheanoUtil import circular_convolution Y, _ = theano.scan(lambda x_i,x_p:circular_convolution(x_i,x_p),Y,Y[0]) Y = Y[-1] elif collapse_output == 'flatten': self.index = T.ones((Y.shape[0] * Y.shape[1], Y.shape[2]), dtype='int8') Y = Y.reshape((Y.shape[0]*Y.shape[1],Y.shape[2],Y.shape[3])) elif str(collapse_output).startswith('pad_'): pad = numpy.int32(collapse_output.split('_')[-1]) Y = ifelse(T.lt(Y.shape[0],pad),T.concatenate([Y,T.zeros((pad-Y.shape[0],Y.shape[1],Y.shape[2],Y.shape[3]),'float32')],axis=0), ifelse(T.gt(Y.shape[0],pad),Y[:pad],Y)) Y = Y.dimshuffle(1,2,3,0).reshape((Y.shape[1],Y.shape[2],Y.shape[3]*Y.shape[0])) self.attrs['n_out'] *= pad elif collapse_output != False: assert False, "invalid collapse mode" if self.attrs['batch_norm']: Y = self.batch_norm(Y,self.attrs['n_out'],index=sizes if not collapse_output else self.index, force_sample=False) self.output = Y