FrancescoSaverioZuppichini
commented
2 years ago Hey @Boraly
Very cool project, I don't know actually the GradCam has always worked for me, I was working on an integration with torchcam[https://github.com/frgfm/torch-cam] but it's not there yet. I would recommend to try with the GradCam there and seeing if there is a problem.
When you create a cam using it you can all pass a reference to the layer, that should make using glasses super easy
from torchcam.methods import GradCAM
cam_extractor = GradCAM(model, [model.encoder])
Let me know, feel free to create a jupyter-notebook sharing your code/issues
Hey,
I trained a model to classify polyps. Now I tried to visualize using GradCAM. I copied the code from the tutorial, but I think the heatmap is inverted. I will add a example Image.
My Code:
``directory = '/content/polyp_paris_export/images/' fnames = list(fname for fname in listdir(directory) if fname.endswith('.png')) cfg = AutoTransform.from_name('efficientnet_b0') norm = cfg.transforms[-1] postprocessing = Normalize(-norm.mean / norm.std, (1.0 / norm.std)) model.eval() for i in fnames: img = directory + i img_1 = Image.open(img)
batch_1 = cfg(img_1).unsqueeze(0)
plt.imshow(img_1)
sa_1 = model.interpret(batch_1, using=SaliencyMap()).show()
grad_1 = model.interpret(batch_1, using=GradCam(), postprocessing=postprocessing).show() plt.show() sleep(1) clear_output(wait=True) input("Press Enter to continue...")``
You can cleary see, that the SaliencyMap recognize the polyps, but the GradCAM just focused on the area around them.
I tried to just invert the Image by using 1-img. But the results doesn't look clear.
How can I fix that?