Hicham Bellafkir comments

Results 17 comments of


                                            Hicham Bellafkir

nvmlDeviceGetHandleByPciBusId() failed with error #2

Did someone fixed that?

Validation loss vs Training loss in AudioSet training

Wouldn't it be wrong to train with Softmax and use Sigmoid for mAP? Using Softmax instead of Sigmoid gives a higher mAP value.

Validation loss vs Training loss in AudioSet training

In the case of CrossEntropyLoss, the loss is calculated with Softmax ([here](https://github.com/YuanGongND/ast/blob/e038086885c9a52c2bd8896f0058ca2f1a22d67b/src/traintest.py#L117)) as it is included in the CrossEntropyLoss operation. It is not correct to use Sigmoid for inference when...

Validation loss vs Training loss in AudioSet training

> In the case of CrossEntropyLoss, the loss is calculated with Softmax ([here](https://github.com/YuanGongND/ast/blob/e038086885c9a52c2bd8896f0058ca2f1a22d67b/src/traintest.py#L117)) as it is included in the CrossEntropyLoss operation. It is not correct to use Sigmoid for inference...

Is alignment included?

Nope

Is alignment included?

@vandana-rajan I thought alignment should include rotating the face too. but you right, alignment here means just returning the landmark points.

Passing batch of images for faster processing

@ipazc are you still woking on it, or its a dead thing ?

adjust_learning_rate can't import

did you found a solution @zyxforever ?

docker push keeps retrying and failing to local registry

I have the same issue with a custom caffe image

AlexNet backbone result (COCO)

Did u have the error in network.py that ues_hashnet is not defined? I'm running the same experiment. I will report the results after.