One parameter tensor's share of DeviceBackend::adam_step_multi (HIP-6): its buffers and its own step's scalars.
More...
#include <device_backend.hpp>
|
| float * | param = nullptr |
| |
| const float * | grad = nullptr |
| |
| float * | m = nullptr |
| |
| float * | v = nullptr |
| |
| size_t | n = 0 |
| |
| float | lr = 0.0f |
| |
| float | bias_correction1 = 1.0f |
| |
| float | bias_correction2 = 1.0f |
| |
| float | decay = 1.0f |
| | AdamW: param <- decay * param before the step (1 - lr * wd); 1 for none.
|
| |
| float | l2 = 0.0f |
| | L2 weight decay: the step sees l2 * param + grad; 0 for none.
|
| |
One parameter tensor's share of DeviceBackend::adam_step_multi (HIP-6): its buffers and its own step's scalars.
◆ bias_correction1
| float pulsatrix::AdamTensorStep::bias_correction1 = 1.0f |
◆ bias_correction2
| float pulsatrix::AdamTensorStep::bias_correction2 = 1.0f |
◆ decay
| float pulsatrix::AdamTensorStep::decay = 1.0f |
AdamW: param <- decay * param before the step (1 - lr * wd); 1 for none.
◆ grad
| const float* pulsatrix::AdamTensorStep::grad = nullptr |
◆ l2
| float pulsatrix::AdamTensorStep::l2 = 0.0f |
L2 weight decay: the step sees l2 * param + grad; 0 for none.
◆ lr
| float pulsatrix::AdamTensorStep::lr = 0.0f |
| float* pulsatrix::AdamTensorStep::m = nullptr |
| size_t pulsatrix::AdamTensorStep::n = 0 |
◆ param
| float* pulsatrix::AdamTensorStep::param = nullptr |
| float* pulsatrix::AdamTensorStep::v = nullptr |
The documentation for this struct was generated from the following file: