Python API for Myelin #302
Conversation
| // Get single-dimensional index. | ||
| int idx = PyInt_AsLong(index); | ||
| if (idx == -1 && PyErr_Occurred()) return nullptr; | ||
| if (idx < 0 || idx >= format->dim(0)) { |
There was a problem hiding this comment.
Suggestion: you can support the Python convention here:
if (idx < 0) idx += format->dim(0);
python/myelin/gradient.py
Outdated
| def neg_grad(op, g): | ||
| x = op.inputs[0] | ||
| y = op.outputs[0] | ||
| g.add(x, g.expr.mul(g.d(y), g.expr.neg(g.v(x)))) |
There was a problem hiding this comment.
Shouldn't this just be:
g.add(x, g.expr.neg(g.d(y))) ?
There was a problem hiding this comment.
Yes. This bug has already been fixed in the C++ version.
python/myelin/gradient.py
Outdated
| def reciprocal_grad(op, g): | ||
| x = op.inputs[0] | ||
| y = op.outputs[0] | ||
| g.add(x, g.expr.mul(g.d(y), g.expr.neg(g.expr.rcp(g.expr.square(g.v(x)))))) |
There was a problem hiding this comment.
g.expr.neg(g.expr.square(g.v(y))) might be more efficient than g.expr.neg(g.expr.rcp(g.expr.square(g.v(x))))))
There was a problem hiding this comment.
Oops. This is my autograd experiment in python which was included in this PR by mistake.
However, the optimization is still good since you save a division, so I have changed this in the C++ version. Thanks for spotting this.
doc/guide/myelin.md
Outdated
|
|
||
|  | ||
|
|
||
| The graph only shows the input and output variables (gren and blue), and the |
doc/guide/myelin.md
Outdated
|
|
||
| The weights in W and b can be initialized from NumPy arrays or any other | ||
| objects that support the | ||
| [Python bufffer protocol](https://docs.python.org/2/c-api/buffer.html): |
python/myelin/gradient.py
Outdated
| def reciprocal_grad(op, g): | ||
| x = op.inputs[0] | ||
| y = op.outputs[0] | ||
| g.add(x, g.expr.mul(g.d(y), g.expr.neg(g.expr.rcp(g.expr.square(g.v(x)))))) |
There was a problem hiding this comment.
Oops. This is my autograd experiment in python which was included in this PR by mistake.
However, the optimization is still good since you save a division, so I have changed this in the C++ version. Thanks for spotting this.
python/myelin/gradient.py
Outdated
| def neg_grad(op, g): | ||
| x = op.inputs[0] | ||
| y = op.outputs[0] | ||
| g.add(x, g.expr.mul(g.d(y), g.expr.neg(g.v(x)))) |
There was a problem hiding this comment.
Yes. This bug has already been fixed in the C++ version.
doc/guide/myelin.md
Outdated
|
|
||
|  | ||
|
|
||
| The graph only shows the input and output variables (gren and blue), and the |
doc/guide/myelin.md
Outdated
|
|
||
| The weights in W and b can be initialized from NumPy arrays or any other | ||
| objects that support the | ||
| [Python bufffer protocol](https://docs.python.org/2/c-api/buffer.html): |
You can now run Myelin computations from Python using pysling. I have updated flow.py to support flow version 5 format. You can now compile these flows to a network to run computations. I have updated the documentation with examples of running Myelin in Python.
The Myelin tensors support the Python buffer interface, so these can be shared with other Python modules with buffer support, like numpy.