17 Mar 2017 |
@gitter_jonysy:matrix.org | (edited) ... it compiles (or ... => ... it _compiles_ (or ... | 17:00:06 |
@gitter_jonysy:matrix.org | (edited) ... (or translates, for ... => ... (or _translates_, for ... | 17:00:16 |
@gitter_botev:matrix.org | here we abstract the architecture to a numerical framework (opencl, cuda, arrrayfire etc) which is a level above binary code | 17:00:35 |
@gitter_jonysy:matrix.org | Right. So a sigmoid function 1.0 / (1.0 + exp(-x)) in native Rust could compile to CL like such:__kernel void sigmoid(__global float *a, __global float *b) {
uintptr_t i = get_global_id(0);
b[i] = 1.0 / (1.0 + exp(-a[i]));
} | 17:02:40 |
@gitter_botev:matrix.org | yes elementwise function would all look like that | 17:03:21 |
@gitter_jonysy:matrix.org | How does arrayfire’s JIT work? | 18:31:14 |
@gitter_jonysy:matrix.org | How does it compile a function at run time and then skip to the compiled kernel for subsequent calls? | 18:32:31 |
@gitter_jonysy:matrix.org | (edited) ... does it compile a function at run time and then skip to the compiled kernel for subsequent calls? => ... does arrayfire’s JIT engine work? | 18:32:44 |
19 Mar 2017 |
@gitter_jonysy:matrix.org | Poll: // Multilayer perceptron (MLP)
let net_cfg = {
SequentialConfig::new()
.input("data", [batch_size, 28, 28])
.force_backward(true)
.layer("reshape", Reshape::new().with_shape([batch_size, 784]))
.layer("linear-1", Linear::new().output_size(1568))
.layer("sigmoid", Sigmoid)
.layer("linear-2", Linear::new().output_size(10))
// ..
.layer("log_softmax", LogSoftmax)
};
or // Multilayer perceptron (MLP)
let net = Sequential::new([
Input("data", [batch_size, 28, 28]),
Reshape(Layer::new("reshape").with_shape([batch_size, 784])),
Linear(Layer::new("linear-1").output_size(1568)),
Sigmoid(Layer::new("sigmoid")),
Linear(Layer::new("linear-2").output_size(10)),
LogSoftmax(Layer::new("log_softmax"))
]); | 20:30:06 |
@neverfox:matrix.org | If you do the second style, why not Reshape::new("reshape")... ? | 21:23:19 |
21 Mar 2017 |
@gitter_jonysy:matrix.org | Is anyone here interested in being one of the maintainers of the leaf fork? | 19:14:54 |
22 Mar 2017 |
@gitter_jonysy:matrix.org | Why is there a dx argument in the gradient functions here | 16:38:16 |
@gitter_jonysy:matrix.org | (edited) ... functions [here](https://github.com/autumnai/collenchyma-nn/blob/master/src/frameworks/native/helper.rs#L39) => ... functions [here](https://github.com/autumnai/collenchyma-nn/blob/master/src/frameworks/native/helper.rs#L39)? | 16:38:21 |
@gitter_jonysy:matrix.org | (edited) Why is there a `dx` argument in ... => Why are there a `dx` arguments in ... | 16:40:16 |
25 Mar 2017 |
@gitter_jonysy:matrix.org | I cannot, for the life of me, figure out why the switched the parameters a and b here. Can you help me out? | 00:43:26 |
@gitter_jonysy:matrix.org | (edited) ... why the switched ... => ... why they switched ... | 00:43:37 |
@gitter_jonysy:matrix.org | (edited) ... the parameters ... => ... the order of parameters ... | 00:43:47 |
@gitter_jonysy:matrix.org | Never mind. CUDA expects a column-major memory layout. | 01:40:38 |
21 Apr 2017 |
@cathalgarvey:matrix.org | http://futhark-lang.org/ | 07:21:39 |
@cathalgarvey:matrix.org | Looks like a much funner way to write kernels :) | 07:21:49 |
29 May 2017 |
@cathalgarvey:matrix.org | A new framework appears! https://github.com/torchrs/torchrs | 08:39:26 |
@neverfox:matrix.org | Dammit that was my idea. | 14:12:19 |
@neverfox:matrix.org | Oh well, it exists now, so I think we can consider NN in Rust solved. PyTorch API is excellent and uses graph based AD. | 14:13:48 |
29 Jul 2017 |
| @aurabindo:matrix.org changed their display name from Aurabindo to jayaura. | 06:35:07 |
| @aurabindo:matrix.org changed their display name from jayaura to Aurabindo. | 06:35:19 |
15 Aug 2017 |
| @gitter_djkooks:matrix.org changed their profile picture. | 15:22:03 |
9 Nov 2017 |
| @gitter_usamec:matrix.org joined the room. | 15:11:45 |
@gitter_usamec:matrix.org | Hi. Does GIR support creation of recurrent nets? | 15:11:46 |
| @gitter_botev:matrix.org changed their profile picture. | 16:20:23 |
@gitter_botev:matrix.org | unfortunately I'm no longer developing that project due to being busy with my PhD | 16:20:24 |