Migrate dl code from webscaleml to spark-dl #2
Conversation
|
When defining a model, instead of directly calling the constructors of individual modules, we should use factory methods (e.g., through Scala companion object’s apply methods). This allows one to define the same model that can run using different underlying engine (e.g., MKL-DNN). |
|
Let me create a issue to log this. |
|
Please remove |
| var error = nextError | ||
| while (i > 0) { | ||
| val input = modules(i - 1).output | ||
| error = modules(i).backward(input, error) |
There was a problem hiding this comment.
Shouldn't we call each module's updateGradInput here? And then we also need to implement the updateGradInput method.
There was a problem hiding this comment.
In some special case, some module will override backward instead of updateGradInput, to do something beyond updateGradInput and accGrad(for most modules, these two operations are enough in backward).
So the container call backward.
* add github action for nano code style (#1) * Update nano-style-check.yml * Test pr (#2) * change something * make style break * Update nano-style-check.yml * Update nano-style-check.yml (#3) * Update nano-style-check.yml * Update nano-style-check.yml * Update nano-style-check.yml * Update nano-style-check.yml * Update nano-style-check.yml * Update nano-style-check.yml * fix style * change nane
2 participants
No description provided.