|
|
- "description": "Broadcasting is the process of making arrays to have compatible shapes\nfor arithmetic operations. Two shapes are compatible if for each\ndimension pair they are either equal or one of them is one. When trying\nto broadcast a Tensor to a shape, it starts with the trailing dimensions,\nand works its way forward.\n\nFor example,\n\n>>> x = tf.constant([1, 2, 3])\n>>> y = tf.broadcast_to(x, [3, 3])\n>>> print(y)\ntf.Tensor(\n [[1 2 3]\n [1 2 3]\n [1 2 3]], shape=(3, 3), dtype=int32)\n\nIn the above example, the input Tensor with the shape of `[1, 3]`\nis broadcasted to output Tensor with shape of `[3, 3]`.\n\nWhen doing broadcasted operations such as multiplying a tensor\nby a scalar, broadcasting (usually) confers some time or space\nbenefit, as the broadcasted tensor is never materialized.\n\nHowever, `broadcast_to` does not carry with it any such benefits.\nThe newly-created tensor takes the full memory of the broadcasted\nshape. (In a graph context, `broadcast_to` might be fused to\nsubsequent operation and then be optimized away, however.)",
|
|
|
+ "description": "Broadcasting is the process of making arrays to have compatible shapes\nfor arithmetic operations. Two shapes are compatible if for each\ndimension pair they are either equal or one of them is one.\n\nFor example:\n\n>>> x = tf.constant([[1, 2, 3]]) # Shape (1, 3,)\n>>> y = tf.broadcast_to(x, [2, 3])\n>>> print(y)\ntf.Tensor(\n [[1 2 3]\n [1 2 3]], shape=(2, 3), dtype=int32)\n\nIn the above example, the input Tensor with the shape of `[1, 3]`\nis broadcasted to output Tensor with shape of `[2, 3]`.\n\nWhen broadcasting, if a tensor has fewer axes than necessary its shape is\npadded on the left with ones. So this gives the same result as the previous\nexample:\n\n>>> x = tf.constant([1, 2, 3]) # Shape (3,)\n>>> y = tf.broadcast_to(x, [2, 3])\n\n\nWhen doing broadcasted operations such as multiplying a tensor\nby a scalar, broadcasting (usually) confers some time or space\nbenefit, as the broadcasted tensor is never materialized.\n\nHowever, `broadcast_to` does not carry with it any such benefits.\nThe newly-created tensor takes the full memory of the broadcasted\nshape. (In a graph context, `broadcast_to` might be fused to\nsubsequent operation and then be optimized away, however.)",
|
Lutz Roeder