Jiyoung Giuliana Yun comments

Results 130 comments of


                                            Jiyoung Giuliana Yun

[onert][training] Changing classification layer of pre-trained model and fine-tuning

Thank you for your interest and suggestion. TF-lite preprocessing provides image processing library to make it easier to use input and output of TensorFlow lite model in android. I think...

[onert][training] Changing classification layer of pre-trained model and fine-tuning

If we introduce new API for inserting ops to graph, I would proceed in the following direction: 1. Insert op after output of the graph. - This is because I...

[onert][training] Changing classification layer of pre-trained model and fine-tuning

It seems that the function of inserting new ops to the graph is not required in the inference feature. Therefore, related work needs to be done with focus on training....

[onert][training] Changing classification layer of pre-trained model and fine-tuning

@mbencer I'm sorry for the late reply. I was thinking about how valuable this job is on the On-Device Training scope, but I didn't find the clear reason so I...

[onert][training] Changing classification layer of pre-trained model and fine-tuning

@mbencer > So do you think here that it's enough to operate here only on in-memory representation and conversion to device format (like 'tvn')? Btw. Is it possible to create...

[onert][training] Changing classification layer of pre-trained model and fine-tuning

@mbencer I didn't think to use nnfw api **on host machine**. I thought model developers would develop models on host machine using Python API. This is because nnfw api is...

[onert][training] Changing classification layer of pre-trained model and fine-tuning

@mbencer I didn't notice that you were working on the fine tuning feature. Your draft code is similar to what I was trying to implement. I will look at the...

[onert][training] Changing classification layer of pre-trained model and fine-tuning

@mbencer When I tested your PR, it works well. I tried to compare TensorFlow and ONERT training using the same model file containing the same weight. And the result is...

[onert][training] Changing classification layer of pre-trained model and fine-tuning

I made [a draft PR](https://github.com/Samsung/ONE/pull/12984) to introduce the `trainable` attribute in circle+ training info. And I tested this code with your draft PR and it works fine. I fixed your...

[onert][training] Changing classification layer of pre-trained model and fine-tuning

@mbencer > 1. Am I right that trainable indexes set via `nnfw_train_disable_node_update` and `nnfw_train_enable_node_update` should overrides settings from circle plus? Right. The default training info is from the circle+ file....