我是TensorFlow的新手。我想构建一个具有以下属性的双射器：它接受一个n维概率分布p(x1, x2, …, xn)，并且只转换两个特定的维度i和j，使得xi’ = xi, xj’ = xj*exp(s(xi)) + t(xj)，其中s和t是通过神经网络实现的两个函数。它输出p(x1, x2, …, xi’, .., xj’, .., xn)。我有一个基本的代码如下所示：
def net(x, out_size, block_w_id, block_d_id, layer_id): x = tf.contrib.layers.fully_connected(x, 256, reuse=tf.AUTO_REUSE, scope='x1_block_w_{}_block_d_{}_layer_{}'.format(block_w_id, \ block_d_id,\ layer_id)) x = tf.contrib.layers.fully_connected(x, 256, reuse=tf.AUTO_REUSE, scope='x2_block_w_{}_block_d_{}_layer_{}'.format(block_w_id,\ block_d_id,\ layer_id)) y = tf.contrib.layers.fully_connected(x, out_size, reuse=tf.AUTO_REUSE, scope='y_block_w_{}_block_d_{}_layer_{}'.format(block_w_id,\ block_d_id,\ layer_id))# return layers.stack(x, layers.fully_connected(reuse=tf.AUTO_REUSE), [512, 512, out_size]) return yclass NVPCoupling(tfb.Bijector): """NVP仿射耦合层，用于2D单元。 """ def __init__(self, input_idx1, input_idx2, block_w_id = 0, block_d_id = 0, layer_id = 0, validate_args = False\ , name="NVPCoupling"): """ NVPCoupling仅操作两个输入，索引为idx1和idx2。 """ super(NVPCoupling, self).__init__(\ event_ndims = 1, validate_args = validate_args, name = name) self.idx1 = input_idx1 self.idx2 = input_idx2 self.block_w_id = block_w_id self.block_d_id = block_d_id self.layer_id = layer_id # 创建变量 tmp = tf.placeholder(dtype=DTYPE, shape = [1, 1]) self.s(tmp) self.t(tmp) def s(self, xd): with tf.variable_scope('s_block_w_id_{}_block_d_id_{}_layer_{}'.format(self.block_w_id,\ self.block_d_id,\ self.layer_id),\ reuse = tf.AUTO_REUSE): return net(xd, 1, self.block_w_id, self.block_d_id, self.layer_id) def t(self, xd): with tf.variable_scope('t_block_w_id_{}_block_d_id_{}_layer_{}'.format(self.block_w_id,\ self.block_d_id,\ self.layer_id),\ reuse = tf.AUTO_REUSE): return net(xd, 1, self.block_w_id, self.block_d_id, self.layer_id) def _forward(self, x): x_left, x_right = x[:, self.idx1:(self.idx1 + 1)], x[:, self.idx2:(self.idx2 + 1)] y_right = x_right * tf.exp(self.s(x_left)) + self.t(x_left) output_tensor = tf.concat([ x[:,0:self.idx1], x_left, x[:, self.idx1+1:self.idx2]\ , y_right, x[:, (self.idx2+1):]], axis = 1) return output_tensor def _inverse(self, y): y_left, y_right = y[:, self.idx1:(self.idx1 + 1)], y[:, self.idx2:(self.idx2 + 1)] x_right = (y_right - self.t(y_left)) * tf.exp(-self.s(y_left)) output_tensor = tf.concat([ y[:, 0:self.idx1], y_left, y[:, self.idx1+1 : self.idx2]\ , x_right, y[:, (self.idx2+1):]], axis = 1) return output_tensor def _forward_log_det_jacobian(self, x): event_dims = self._event_dims_tensor(x) x_left = x[:, self.idx1:(self.idx1+1)] return tf.reduce_sum(self.s(x_left), axis=event_dims)但它并没有按我预期的那样工作。当我使用这个类时，出现了一个错误：
base_dist = tfd.MultivariateNormalDiag(loc=tf.zeros([2], DTYPE))num_bijectors = 4bijectors = []bijectors.append(NVPCoupling(input_idx1=0, input_idx2=1, \ block_w_id=0, block_d_id=0, layer_id=0))bijectors.append(NVPCoupling(input_idx1=1, input_idx2=0, \ block_w_id=0, block_d_id=0, layer_id=1))bijectors.append(NVPCoupling(input_idx1=0, input_idx2=1, \ block_w_id=0, block_d_id=0, layer_id=2))bijectors.append(NVPCoupling(input_idx1=0, input_idx2=1, \ block_w_id=0, block_d_id=0, layer_id=3))flow_bijector = tfb.Chain(list(reversed(bijectors)))dist = tfd.TransformedDistribution( distribution=base_dist, bijector=flow_bijector)dist.sample(1000)出现错误如下：
---------------------------------------------------------------------------ValueError Traceback (most recent call last)<ipython-input-16-04da05d30f8d> in <module>()----> 1 dist.sample(1000)/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/tensorflow/python/ops/distributions/distribution.pyc in sample(self, sample_shape, seed, name) 708 samples: a `Tensor` with prepended dimensions `sample_shape`. 709 """--> 710 return self._call_sample_n(sample_shape, seed, name) 711 712 def _log_prob(self, value):/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/tensorflow/python/ops/distributions/transformed_distribution.pyc in _call_sample_n(self, sample_shape, seed, name, **kwargs) 412 # returned result. 413 y = self.bijector.forward(x, **kwargs)--> 414 y = self._set_sample_static_shape(y, sample_shape) 415 416 return y/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/tensorflow/python/ops/distributions/distribution.pyc in _set_sample_static_shape(self, x, sample_shape) 1220 shape = tensor_shape.TensorShape( 1221 [None]*(ndims - event_ndims)).concatenate(self.event_shape)-> 1222 x.set_shape(x.get_shape().merge_with(shape)) 1223 1224 # Infer batch shape./Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/tensorflow/python/framework/tensor_shape.pyc in merge_with(self, other) 671 return TensorShape(new_dims) 672 except ValueError:--> 673 raise ValueError("Shapes %s and %s are not compatible" % (self, other)) 674 675 def concatenate(self, other):ValueError: Shapes (1000, 4) and (?, 2) are not compatible真心希望有专家能帮助我理解我哪里做错了以及如何修复这个问题。非常感谢！H.
回答：
我认为问题出在这里（为了清晰起见稍作重新格式化）：
output_tensor = tf.concat([ x[:,0:self.idx1], x_left, x[:, self.idx1+1:self.idx2], y_right, x[:, (self.idx2+1):]], axis = 1)这假设idx2 > idx1，但在你给idx1=1和idx2=0的情况下这不是真的。这会导致你拼接的元素比你预期的要多，使得第二维度变成4而不是2。
我在_forward中打印了形状如下：
print("self.idx1: %s" % self.idx1)print("self.idx2: %s" % self.idx2)print("x[:,0:self.idx1]: %s" % x[:,0:self.idx1].shape)print("x_left: %s" % x_left.shape)print("x[:, self.idx1+1:self.idx2]: %s" % x[:, self.idx1+1:self.idx2].shape)print("x_right.shape: %s" % x_right.shape)print("y_right: %s" % y_right.shape)print("x[:, (self.idx2+1):]: %s" % x[:, (self.idx2+1):].shape)print("output_tensor.shape: %s" % output_tensor.shape)得到的输出是：
self.idx1: 0self.idx2: 1x[:,0:self.idx1]: (1000, 0)x_left: (1000, 1)x[:, self.idx1+1:self.idx2]: (1000, 0)x_right.shape: (1000, 1)y_right: (1000, 1)x[:, (self.idx2+1):]: (1000, 0)output_tensor.shape: (1000, 2)self.idx1: 1self.idx2: 0x[:,0:self.idx1]: (1000, 1)x_left: (1000, 1)x[:, self.idx1+1:self.idx2]: (1000, 0)x_right.shape: (1000, 1)y_right: (1000, 1)x[:, (self.idx2+1):]: (1000, 1)output_tensor.shape: (1000, 4)self.idx1: 0self.idx2: 1x[:,0:self.idx1]: (1000, 0)x_left: (1000, 1)x[:, self.idx1+1:self.idx2]: (1000, 0)x_right.shape: (1000, 1)y_right: (1000, 1)x[:, (self.idx2+1):]: (1000, 2)output_tensor.shape: (1000, 4)self.idx1: 0self.idx2: 1x[:,0:self.idx1]: (1000, 0)x_left: (1000, 1)x[:, self.idx1+1:self.idx2]: (1000, 0)x_right.shape: (1000, 1)y_right: (1000, 1)x[:, (self.idx2+1):]: (1000, 2)output_tensor.shape: (1000, 4)我认为你需要在idx1 > idx2的情况下更仔细地考虑如何重新组装拼接的部分。
希望这能让你重新回到正轨！
