Learning Resnet Deep Neural Networks from the Ground Up

Question

I need to get some knowledge about deep neural networks.

For the very deep ResNet neural network, we can use learning transfer to train the model. But Resnet was trained in the ImageNet dataset. Therefore, their pre-prepared weights can be used to train the model with another data set. (e.g. to train a model for detecting lung cancer with CT lung images)

I believe that this approach will be inaccurate, since previously prepared weights were fully trained in other objects, but not with medical data.

Instead of transferring training, is it possible to train resnet from scratch? (but the available number of images for resnet training is about 1500). Is it possible to do this with a regular computer.

Can anyone share with me their valuable ideas

Answer 1

Is it possible to prepare scnet from scratch?

Yes, it’s possible, but the amount of time you need to get to good accuracy is highly dependent on the data. For example, training the original ResNet-50 on an NVIDIA M40 GPU took 14 days (10 ^ 18 operations with the same accuracy). The most expensive operation at CNN is convolution in the early layers.

ImageNet 14m 226x226x3 . 10000x , 10000x . , - RGB, 3 ops. . , .

, , , 1080 1080ti, ~ 100 ResNet-50 . , 2-GPU . , , .

, . , .