flazerain

How to export the insightface mxnet model to onnx model which is same as the onnx model given to download.

6

I exported the insightface mxnet model with mxnet 1.5 and onnx 1.3(comment spatial=1) and I cannot build the tensor RT engine with tensorRT 6, the error is: [TensorRT] ERROR: (Unnamed...

关于GI(General Instance) Feature-based Distillation 实现的问题

8

我注意到您代码里的实现： idx_out = torch.ops.torchvision.nms(gibox, giscore, 0.3)[:10] return idx_out gi_idx = self.get_gi_region(soft_label, cls_score, anchors, bbox_pred, soft_targets, stride) gi_teacher = teacher_x[gi_idx] gi_student = x[gi_idx] loss_im = self.loss_im(gi_student, gi_teacher) 并没有像GI 论文里那样使用GIbox ROIAlign 进行FM。...

关于ld_loss计算的问题

2

实现中ld_loss没有除以avg_factor 如果batch_size比较大的话，ld_loss会波动很大 我训练的时候ld_loss会从刚开始1左右上升到10多，又慢慢下降。 我看您推荐的batch_size是2，如果要batch_size比较大该怎么设置呢？把ld_loss的权重 0.25 除以batchsize*2? #28 #7

怎样才能把caffe模型修改成可以tensorRT的格式？

1

我用我自己训练的mxnet模型转换为caffe模型，结构和您提供的mnet-deconv-0517.prototxt有一些不一样。 用我自己的caffe模型转换的时候会报错: Caffe Parser: Invalid axis in crop layer - only spatial cropping is supported

how can the lstm learn from request label?

5

1.how can the lstm learn from the request label given next timestep? 2. I found something that could be wrong the reward code: `a_t = tf.cast(a_t, tf.float32)` `label_t = tf.cast(label_t,...

Are there any details? Thanks

我用和https://github.com/clancylian/retinaface/blob/master/model/mnet25.prototxt 一模一样的结构进行转换(除了输入大小为640),转换之后的onnx上采样层还是没有连接到网络中.