enas

enas copied to clipboard

Sorry for my rude, I'm a programmer with Pytorch and Keras , I like this concept about AutoML so I want to learn more about your code, but Tensorflow 2.0...

heartInsert

[` if self.data_format == "NHCW":`](https://github.com/melodyguan/enas/blob/master/src/cifar10/micro_child.py#L253) I believe this needs to be channel last format and hence NHWC?

ambrishrawat

Hi, I've fixed the following bugs when I was reproducing your results and modifying the code: 1. Change average pooling kernel size from 1 to 2 - It's more reasonable...

a00012025

Hi, could anyone tell me how to understand the attention mechanism in the code: [https://github.com/melodyguan/enas/blob/master/src/cifar10/general_controller.py#L121-L124](https://github.com/melodyguan/enas/blob/master/src/cifar10/general_controller.py#L121-L124)? I did not find that in the paper. Thank you.

JiahaoYao

Possible Bug in in micro_child.py

2

Hi, It appears to me that while building the input to the next pooling layer, we are throwing away layers[0] element with this line. https://github.com/melodyguan/enas/blob/d1a90ac915301198f2a30ce136e9040e6c4235ff/src/cifar10/micro_child.py#L277 I think it should be...

yashkant

Not really an issue, but for the sake of clarity, would you mind giving a slightly deep description of the arguments inputted to the main.py of cifar-10? ` python src/cifar10/main.py...

JadBatmobile

Graphs of accuracy and loss during final training process

1

Compared to the normal training process, train acc continues to increase and losses should continue to decline. However, the ENAS process shows that there are periodically periods in which the...

dhlee-jubilee

How to transfer to the larger data, like 300*300?

10

marsggbo

Question about a detail of micro child

2

I notice in cifar10/micro_child.py in line 522  I wonder why num_possible_inputs equals curr_cell +1 since in enas_conv num_possible_inputs = curr_cell +2. as I understand curr_cell can choose from previous...

Stopless-K

How many number filters for macro child to reproduce the results in the paper

1

You mention that the architecture in the paper has 12 layers with 2 fixed pooling layer (_factorized_reduction as written in the code), but didn't mention about how many out_filters is...

nott0