Unexpected Keyword Argument 'jitter'

[DylanCope]

I got an error when trying to generate an attention map using visualize_saliency. On line 153 of visualization.py there is unexpected keyword argument jitter.

[eric-tramel]

I'm having the same problem, using Anaconda with a Python 3.5 virtual environment running, so I'm also making use of OpenCV3. Here is the output and offending line:

Working on filters: [3]
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-15-4d1a18352923> in <module>()
      4 
      5 
----> 6 heatmap = visualize_saliency(faceNet, 7, yclas[idx], X[idx,:,:,:], text=names[yclas[idx]])

/Users/eric/anaconda/envs/tf35/lib/python3.5/site-packages/keras_vis-0.1.2-py3.5.egg/vis/visualization.py in visualize_saliency(model, layer_idx, filter_indices, seed_img, text, overlay)
    151     ]
    152     opt = Optimizer(model.input, losses)
--> 153     grads = opt.minimize(max_iter=1, verbose=False, jitter=0, seed_img=seed_img)[1]
    154 
    155     # We are minimizing loss as opposed to maximizing output as with the paper.

TypeError: minimize() got an unexpected keyword argument 'jitter'

[eric-tramel]

Additionally, if one removes the jitter keyword from l.153 you are then presented with a new error when "smoothing the activation map"...

Working on filters: [3]
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-7-4d1a18352923> in <module>()
      4 
      5 
----> 6 heatmap = visualize_saliency(faceNet, 7, yclas[idx], X[idx,:,:,:], text=names[yclas[idx]])

/Users/eric/anaconda/envs/tf35/lib/python3.5/site-packages/keras_vis-0.1.2-py3.5.egg/vis/visualization.py in visualize_saliency(model, layer_idx, filter_indices, seed_img, text, overlay)
    164     # Smoothen activation map
    165     grads = utils.deprocess_image(grads[0])
--> 166     grads /= np.max(grads)
    167 
    168     # Convert to heatmap and zero out low probabilities for a cleaner output.

TypeError: ufunc 'true_divide' output (typecode 'd') could not be coerced to provided output parameter (typecode 'B') according to the casting rule ''same_kind''

It seems this is coming from updates in NumPy. See NumPy issue #6464 here.

[raghakot]

Fixed in latest.

[DylanCope]

@eric-tramel, I had the same issue when removing the argument, I solved it by replacing the line grads /= np.max(grads) with grads = grads / np.max(grads). Also I installed the latest version of the package and this issue persists. On top of this I get an error for heatmap[np.where(grads <= 0.2)] = 0. I fixed this by changing the shape of heatmap, which for some reason was of rank 2.

I don't know why I would be having these issues and you wouldn't, @raghakot. I would guess you've got an out of date version of numpy.

[raghakot]

Hmm, the tests pass on travis. I dont think its a numpy versioning issue. Can you post full example of what you are trying to run. Is it in theano?

[raghakot]

Feel free to reopen if you are seeing issues. There is no point in closing issues if the fix doesnt work.

[DylanCope]

It doesn't seem right to reopen the issue, as the original issue in this thread was addressed. Also I did actually get the function working in the end, my final code looked like this,

    # Smoothen activation map
    grads = np.float32(utils.deprocess_image(grads[0]))
    grads /= np.max(grads)
    grads = cv2.cvtColor(grads, cv2.COLOR_GRAY2RGB)
    blurred = np.uint8(cv2.GaussianBlur(255*grads, (3, 3), 0))
    # Convert to heatmap and zero out low probabilities for a cleaner output.
    heatmap = cv2.applyColorMap(blurred, cv2.COLORMAP_JET)

    heatmap[np.where(grads <= 0.2)] = 0

I'm not sure if its actually giving the correct output, this is an example using the Xception model with ImageNet weights. It seems reasonable, however it doesn't looked as blurred as the examples and hasn't been thresholded in the same way.

[raghakot]

Try gradCAM. it usually gives better results.