Viredery
Viredery
你可以在tf-eager-fasterrcnn/detection/models/detectors/faster_rcnn.py中把depth=101改成50
因为Tensorflow提供一些函数,可以直接传入logits和label这两个参数去计算损失和梯度值。本质上,probs的在logits基础上多了exp计算,然后算梯度的时候再求导,相当于在prob上算log,即又变回了logits。所以TF提供了一些函数,在函数的内部实现里简化了计算量。
I am not sure what the problem is. I tried it and it ran successfully. Could you please send your program logs or screenshots?
It looks like your GPU memory is not sufficient to run this network. Maybe change backbone to ResNet50 or ResNet18. However I don't have relative pre-trained weights unfortunately.
关于第一个问题,和#4是一样的,你更新成目前最新的代码后应该没问题了,你试下。如果有问题可以再提一下 第二个权重问题。你需要先跑一下网络来进行网络的初始构造。你看下我每个jupyter文件里面加载权重之前,都有一行代码执行了一遍网络的前向传播