mfigurnov / sact Goto Github PK

I ran the whole validation set with batch size 50, so I got 50000 image jpgs and 50000 ponder map pngs.
I tried to find the figure in your README just to make sure everything's done well.
I only found the one below,

and I found I have couple of redundant images like:

My ImageNet raw-data seems find but I think I have a problem with tfrecord,
since I get almost 0 evaluation accuracy only in ImageNet (doing well on cifar-10).
I also posted an issue about the ImageNet script.
tensorflow/models#4051
I'm planning to reproduce part of your code to pytorch,
and if my issue is just because of the tfrecord issue, I will be fine.
Or am I missing something?
I use 2 Titan XP with python 2.7

I use 2 Titan XP with Tensorflow 1.7.
I tried to evaluate the pretrained model on ImageNet.
Even though I set the flag numexamples to 10 for quicker test,
It is running over 20 hours...
Is this normal?

Hi, First of all great paper, i am trying to run the scripts to evaluate with ImageNet data set, I am not able to get any data from the URL, 'https://raw.githubusercontent.com/tensorflow/models/master/inception/inception/data/'.
Also the link to prepare the ImageNet Dataset in TF-Slim format under 'Instructions' (https://github.com/tensorflow/models/tree/master/inception#getting-started) gives 404.

Any help would be appreciated.

regards,
anki

The code imagenet_train.py needs a pretrained resnet-101. However, Kaiming He just provide the prototxt and weights in Caffe version.

In imagenet_model you defined resnet_v2 with global_pool,
is there any reason that cifar model doesn't contain global average pooling?
Am I missing something?
or the part

I am not sure this is my computer problem or not. After opening your pdf file, there are many typos inside

Hello, I found a performance issue in the difinition of sact_image_heatmap, summary_utils.py, tf.stack will be created repeately during the program execution, resulting in reduced efficiency. So I think it should be created before the loop in the sact_image_heatmap.

The same issues exist in:

• tf.to_float(tf.logical_not(elements_finished))
• tf.expand_dims(residual_mask, 3)

Looking forward to your reply. Btw, I am very glad to create a PR to fix it if you are too busy.

Hi, recently i wanna visualize some intermediate feature maps in your programs. I use some way like
sess = tf.Session() , sess.run(images), {images are the tensor in cifar_main.py file}. It seems that it does not work, i can not convert the images tensor to numpy.
Do you have any idea to get some intermediate feature maps?

I have checked your code and find in file flopsometer.py, function conv2d you don't actually accelerating convolution but just calculate the FLOPS. However, even if FLOPS is less, you cannot say that the convolution operation must be accelerated. Because in convolution, most of the time cost is IO, for example, the im2col operation.

Recall that the residual function consists of a stack of 1 × 1, 3 × 3 and 1 × 1 convolutional layers. The first convolutional layer has to be evaluated in the positions obtained by dilating the active positions set with a 3 × 3 kernel. The second and third layers need to be evaluated just in the active positions.