leondgarse
leondgarse
I think a temporary easiest way is using `tflite`, just works if model is not too complicated. There are times we cannot save a `h5` format: ```py # A test...
Same with @discort . Previously I have `libnccl-dev=2.11.4-1+cuda11.5` installed, but my CUDA version is `11.2.152-1`. Re-install `libnccl-dev=2.8.3-1+cuda11.2` worked for me. ```py sudo apt install libnccl2=2.8.3-1+cuda11.2 libnccl-dev=2.8.3-1+cuda11.2 pip uninstall -y horovod...
Ya, `SCRFD` is trained on `640x480` resolution, so onnx also expecting `det_shape=640`. It's said `SCRFD` only works better at `640`: [SCRFD vs TinaFace vs Retinaface #1639](https://github.com/deepinsight/insightface/issues/1639).
I only tried the vanilla mixup augment method very earlier, and the results not good. I think it's not compatible with margin loss functions like `ArcFace`: - Mixup will expect...
I think an imlementation is possible, will try and check if will be better.
VPL mode is added. It can be enabled by `tt = train.Train(..., use_vpl=True)`. I think it should be same with the official implementation. Here is my test results using basic...
Ya, I have compared them several times. It seems the main parts are: - [insightface/prepare_queue_lambda](https://github.com/deepinsight/insightface/blob/master/recognition/vpl/vpl.py#L100) ```py def prepare_queue_lambda(self, label, iters): self.queue_lambda[:] = 0.0 if iters>self.cfg['start_iters']: allowed_delta = self.cfg['allowed_delta'] if self.vpl_mode==0:...
Here is the result using `start_iters=8000 / 4, allowed_delta=200 / 4` for `batch_size=512`:  **Results** | VPL | lfw | cfp_fp | agedb_30 | IJBB 1e-4 | IJBC 1e-4 |...
It's 2 parameters now, `vpl_start_iters` and `vpl_allowed_delta`, `use_vpl` is abandoned. VPL mode is enabled by setting `vpl_start_iters > 0` now, like `tt = train.Train(..., vpl_start_iters=8000)`. Default `vpl_start_iters=-1, vpl_allowed_delta=200`.
I'm using my `IJB_evals.py`. Just ran a bunch of `1:N` tests, `VPL start 8000, delta 200` performs not bad in this test: - **VPL False** ```py >>>> Gallery 1 top1:...