Ning Lu

@xvjiarui HI, do have any good tutorial materials about Cuda programming?

模型下载？

5

您好，能否提供 vgg16_faster_rcnn_iter_5000 的下载地址？

Fix eval bugs

1

Firstly, I want to say sorry about pushing bugs. Secondly, I have double-checked the final metrics. 

Cannot reproduce the results provided in the "experiments" folder?

1

I created a training set by sampling 10000 images from the Coco datasets. I have some questions: 1. I can reproduce the "crop" experiment's training footprint with a similar scale...

Core dump when doing word embedding

2

Hi, I have an edgelist with 6M lines. When I was doing node2vec, I could learn walking successfully, but It cored when it learnt embedding process. By the way, I...

master 解码性能问题

1

https://github.com/mindspore-lab/mindocr/blob/1249bc0aea91112dd641b1767eabe458e3452173/mindocr/models/heads/rec_master_decoder.py#L148 这段代码有两个性能问题：1. 每次targets 长度会变化，会触发图编译，导致infer 的时间变长。如果steps过长会导致很慢。不利于调试。2. probs 每次会append 一个大tensor，步长过长，会导致显存占用过大，会浪费显存。 可以改成这样： ```python targets = ops.fill(ms.int32, (N, num_steps+1), self.padding_symbol) targets[:, 0] = 0 # probs = ops.zeros((N, num_steps, self.out_channels), dtype=inputs.dtype) for i in...