How can I visualize the input and output data?

[hippogriff]

How can I visualize the input and output data?

[adrelino]

I found out that labels is actually unused and just filled with 0, so can be ignore.

In my fork, I refactored the code a bit so that visualize_expressions now also returns the output of the pre-trained network in addition to the groundtruth. Can be found here:

https://github.com/adrelino/CSGNet/blob/master/visualize_expressions.py

[Hippogriff]

You can look at the test_synthetic.py file for testing the model. You can also also use visualize_expressions.py script to visualize the induced programs in the form of images.

[7675t]

Thank you, it much helps to visualize the test results. I can close this issue. Also see my PR.

Can you consider to append a simple command which just apply the network to specified single image, visualize the predicted image and CSG expression? I think it is very helpful to know the concept of your work out of the box.

Thanks,

[Hippogriff]

Thank you for the PR but it doesn't seem necessary, because visualize_expressions.py can be used by the user by slight modification.

Can you consider to append a simple command which just apply the network to specified single image, visualize the predicted image and CSG expression? I think it is very helpful to know the concept of your work out of the box.

I will add a small script.

[adrelino]

I still cannot find the script that just takes a 64x64 binary image (and the provided pre-trained model) as input and outputs a program code (that generates a similar looking image based on camfer or iou distance) for that image.

visualize_expressions.py just takes an expression and generates the ground-truth 64x64 image.

test_synthetic.py takes the pre-trained model and also the whole synthetic data-set expressions, somehow selects a validation subset and computes the mean chamfer distance metric over all program sizes. I guess that from this file, such a script that just takes an image and outputs the program could be extracted, but I am a little confused about these steps:

CSGNet/test_synthetic.py

Lines 104 to 120 in 34e9621

	data_, labels = next(test_gen_objs[k])
	one_hot_labels = prepare_input_op(labels,
	len(generator.unique_draw))
	one_hot_labels = Variable(torch.from_numpy(one_hot_labels)).cuda()
	data = Variable(torch.from_numpy(data_), volatile=True).cuda()
	labels = Variable(torch.from_numpy(labels)).cuda()
	test_output = imitate_net.test([data, one_hot_labels, max_len])
	pred_images, correct_prog, pred_prog = parser.get_final_canvas(
	test_output,
	if_just_expressions=False,
	if_pred_images=True)
	target_images = data_[-1, :, 0, :, :].astype(dtype=bool)
	targ_prog = parser.labels2exps(labels, k)
	programs_tar[jit] += targ_prog
	programs_pred[jit] += pred_prog
	distance = chamfer(target_images, pred_images)

When running inference, we should not need ground-truth labels, but then why do we need to extract one-hot-labels from the ground truth labels before feeding it to test_output = imitate_net.test([data, one_hot_labels, max_len])?
How would I proceed if instead of using the generator data_, labels = next(test_gen_objs[k]), I just input my own image (batch), let's say a logo for which I don't even have ground truth labels?

Thanks for your help!

[Hippogriff]

@adrelino Thanx for the script. The labels are unused in test mode.