Generating heatmaps

[Medicmind]

Hi Leondgarse. Do you have any tips on generating heatmaps from your models eg a trained Efficientnet model from you code?

An example of a heatmap I'd be keen to generate is something like in this code
https://github.com/sidml/EfficientNet-GradCam-Visualization?

[leondgarse]

Isn't that the Make and apply gradcam heatmap example? It seems same to me... The main function is from Github keisen/tf-keras-vis and keras.io/examples Grad-CAM class activation visualization.

from keras_cv_attention_models import visualizing, test_images, efficientnet
mm = efficientnet.EfficientNetV2B0()
img = test_images.dog_cat()
superimposed_img, heatmap, preds = visualizing.make_and_apply_gradcam_heatmap(mm, img, layer_name="auto")

The 3 returned values are:

• superimposed_img is the plotted image. You can check it by plt.plot(superimposed_img).
• heatmap is the heatmap alone. Can also plot by plt.plot(heatmap)
• preds is model prediction output.

[leondgarse]

• visualizing.make_and_apply_gradcam_heatmap using tf.keras.applications.imagenet_utils.preprocess_input processing input image, and its parameter rescale_mode controls the rescaling behavior. For EfficientNetV2S, it's "tf" by default. Is your model trained using input range [0, 1]?
• The input is a single image, so don't need to expand_dims.
• Uh, right, you don't need to resize it, tf.image.resize is included in make_and_apply_gradcam_heatmap.
• This works for me:

  from keras_cv_attention_models import visualizing, efficientnet
  from PIL import Image
  
  fl='image.jpg'
  image_size=384
  mm = efficientnet.EfficientNetV2S(num_classes=2,pretrained=None)
  # mm.load_weights('checkpoints/keepefficientnet.EfficientNetV2S_384_LAMB_acrima_batchsize_16_randaug_6_mixup_0.1_cutmix_1.0_RRC_0.08_lr512_0.008_wd_0.02_latest.h5')
  input_data_type=np.float32
  image = np.array(img, dtype=input_data_type)
  superimposed_img, heatmap, preds = visualizing.make_and_apply_gradcam_heatmap(mm, image, layer_name="auto") 

[Medicmind]

Thanks. Looks like the extra dimension was the problem. The heat map looks good.

.

[leondgarse]

Great! So I can off work then. :)

[Medicmind]

Sorry, I dont know what you mean.

[leondgarse]

Never mind, just kidding. I worked late yesterday, but just for my own stuff.