Softmax + cross entropy for multi-category classification. More...
#include <loss.h>
Public Member Functions | |
| Tensor | Forward (int flag, const Tensor &prediction, const Tensor &target) override |
| Compute the loss values for each sample/instance given the prediction and the target. More... | |
| Tensor | Backward () override |
| Compute the gradients of the loss values w.r.t. More... | |
 Public Member Functions inherited from singa::Loss | |
| void | Setup (const string &conf) |
| virtual void | ToDevice (std::shared_ptr< Device > device) |
| virtual void | Setup (const LossConf &conf) |
| Set meta fields from user configurations. | |
| float | Evaluate (int flag, const Tensor &prediction, const Tensor &target) |
| Average loss values for all samples in the mini-batch It calls Forward() internally. More... | |
Detailed Description
Softmax + cross entropy for multi-category classification.
Member Function Documentation
◆ Backward()
|
overridevirtual |
Compute the gradients of the loss values w.r.t.
the prediction, which is: p[i] - t[i]/ t[j]
Implements singa::Loss.
◆ Forward()
|
overridevirtual |
Compute the loss values for each sample/instance given the prediction and the target.
If the target consists one integer per instance, i.e. the label index (dentoed as idx_truth), the loss is -log(p[idx_truth]), p[] is the probability for each category, computed from Softmax(prediction). If the target consists one array per instance (e.g., for multiple labels), the loss is - (t[i] * log(p[i]) / t[j], t[i] is the weight of the i-th label (e.g., 1: the instance has this label, 0: the instance does not have this label).
Users can call Average(const Tensor&) to get the average loss value over all samples in the batch.
Implements singa::Loss.
The documentation for this class was generated from the following file:
- /home/moaz/incubator-singa/include/singa/model/loss.h
