def plot_saliency():
with CustomObjectScope({
'Dilation2D': Dilation2D,
'Erosion2D': Erosion2D,
'attach_attention_module': attach_attention_module
}):
model = load_model(
'/home/CVL1/Shaobo/PneSeg/model2/RSNA_seg_Shaobo.h5')
model.summary()
gb = get_batch(valid_filenames, 250)
abatch = next(gb)
imgs = abatch[0]['image_in']
msks = abatch[1]['segmentation']
labels = abatch[1]['classification']
plt.figure(figsize=(40, 20))
n = 3
for i in range(n):
img = imgs[i]
msk = msks[i]
class_label = labels[i]
img1 = np.reshape(img, [height, width])
msk = np.reshape(msk, [height, width])
comp = msk > 0.5 / 1.
comp1 = measure.label(comp)
predImg = np.reshape(imgs[i], [1, height, width, 1])
seg_pre, class_pre = model.predict(predImg)
pred = np.reshape(seg_pre, [height, width])
pred = (pred > 0.5) / 1.
pred1 = measure.label(pred)
class_idx = np.where(class_label == 1.)[0]
ax1 = plt.subplot(3, n, i + 1)
ax1.imshow(img1)
ax1.set_title('Seg_Result ' + str(class_idx))
#ax1.get_xaxis().set_visible(False)
#ax1.get_yaxis().set_visible(False)
for region in measure.regionprops(comp1):
# retrieve x, y, height and width
y11, x11, y12, x12 = region.bbox
height1 = y12 - y11
width1 = x12 - x11
ax1.add_patch(
patches.Rectangle((x11, y11),
width1,
height1,
linewidth=1,
edgecolor='b',
facecolor='none'))
for region in measure.regionprops(pred1):
# retrieve x, y, height and width
y21, x21, y22, x22 = region.bbox
height2 = y22 - y21
width2 = x22 - x21
ax1.add_patch(
patches.Rectangle((x21, y21),
width2,
height2,
linewidth=1,
edgecolor='r',
facecolor='none'))
#class_idx = np.where(class_label == 1.)[0]
print(class_label)
print(class_idx[0])
layer_idx = utils.find_layer_idx(model, 'classification')
print(layer_idx)
model.layers[layer_idx].activation = activations.linear
model = utils.apply_modifications(model,
custom_objects={
'Dilation2D': Dilation2D,
'Erosion2D': Erosion2D
})
img2 = resize(imgs[i], (height, width))
grads = visualize_saliency(model,
layer_idx,
filter_indices=class_idx[0],
seed_input=img2)
#grads = visualize_cam(model1, layer_idx, filter_indices= class_idx[0], seed_input = img2, backprop_modifier='guided')
ax2 = plt.subplot(3, n, i + 1 + n)
ax2.imshow(grads, cmap='jet')
ax2.set_title('Saliency Map')
#ax4.get_xaxis().set_visible(False)
# ax4.get_yaxis().set_visible(False)
for region in measure.regionprops(comp1):
y11, x11, y12, x12 = region.bbox
height1 = y12 - y11
width1 = x12 - x11
ax2.add_patch(
patches.Rectangle((x11, y11),
width1,
height1,
linewidth=1,
edgecolor='g',
facecolor='none'))
for region in measure.regionprops(pred1):
# retrieve x, y, height and width
y21, x21, y22, x22 = region.bbox
height2 = y22 - y21
width2 = x22 - x21
ax2.add_patch(
patches.Rectangle((x21, y21),
width2,
height2,
linewidth=1,
edgecolor='r',
facecolor='none'))
model.layers[layer_idx].activation = keras.activations.linear
model = utils.apply_modifications(model,
custom_objects={
'Dilation2D': Dilation2D,
'Erosion2D': Erosion2D
})
#print(layer_idx)
#model.layers[layer_idx].activation = activations.linear
penultimate_layer = utils.find_layer_idx(model, 'down4')
print(penultimate_layer)
#model = utils.apply_modifications(model)
img2 = resize(imgs[i], (height, width))
grads = visualize_cam(model,
layer_idx,
filter_indices=class_idx[0],
seed_input=img2,
penultimate_layer_idx=penultimate_layer)
ax3 = plt.subplot(3, n, i + 1 + 2 * n)
ax3.imshow(grads, cmap='jet', alpha=0.8)
#ax4.set_title('vanilla')
#ax4.get_xaxis().set_visible(False)
#ax4.get_yaxis().set_visible(False)
for region in measure.regionprops(comp1):
y11, x11, y12, x12 = region.bbox
height1 = y12 - y11
width1 = x12 - x11
ax3.add_patch(
patches.Rectangle((x11, y11),
width1,
height1,
linewidth=1,
edgecolor='g',
facecolor='none'))
for region in measure.regionprops(pred1):
# retrieve x, y, height and width
y21, x21, y22, x22 = region.bbox
height2 = y22 - y21
width2 = x22 - x21
ax3.add_patch(
patches.Rectangle((x21, y21),
width2,
height2,
linewidth=1,
edgecolor='r',
facecolor='none'))
plt.show()
plt.savefig('/home/CVL1/Shaobo/PneSeg/model2/test.jpg')
def plot_saliency():
with CustomObjectScope(
{'attach_attention_module': attach_attention_module}):
model = load_model('/home/CVL1/Shaobo/mri/h5File/mri.h5')
model.summary()
gb = get_batch(valid_filenames, 5)
abatch = next(gb)
imgs = abatch[0]['image_in']
msks = abatch[1]['segmentation']
labels = abatch[1]['classification']
plt.figure(figsize=(40, 20))
n = 3
for i in range(n):
img = imgs[i]
msk = msks[i]
class_label = labels[i]
img1 = np.reshape(img, [height, width])
msk = np.reshape(msk, [height, width])
comp = msk > 0.65 / 1.
comp1 = measure.label(comp)
predImg = np.reshape(imgs[i], [1, height, width, 1])
seg_pre, class_pre = model.predict(predImg)
pred = np.reshape(seg_pre, [height, width])
pred = (pred > 0.65) / 1.
pred1 = measure.label(pred)
class_idx = np.where(class_label == 1.)[0]
ax1 = plt.subplot(4, n, i + 1)
ax1.imshow(img1)
ax1.set_title('Seg_Result ' + str(class_idx))
#ax1.get_xaxis().set_visible(False)
#ax1.get_yaxis().set_visible(False)
print(class_label)
print(class_idx[0])
layer_idx = utils.find_layer_idx(model, 'classification')
print(layer_idx)
model.layers[layer_idx].activation = activations.linear
model = utils.apply_modifications(model,
custom_objects={
'attach_attention_module':
attach_attention_module
})
img2 = resize(imgs[i], (height, width))
grads = visualize_saliency(model,
layer_idx,
filter_indices=class_idx[0],
seed_input=img2)
#grads = visualize_cam(model1, layer_idx, filter_indices= class_idx[0], seed_input = img2, backprop_modifier='guided')
ax2 = plt.subplot(4, n, i + 1 + n)
ax2.imshow(grads, cmap='jet')
ax2.set_title('Saliency Map')
#ax4.get_xaxis().set_visible(False)
# ax4.get_yaxis().set_visible(False)
model.layers[layer_idx].activation = keras.activations.linear
model = utils.apply_modifications(model,
custom_objects={
'attach_attention_module':
attach_attention_module
})
#print(layer_idx)
#model.layers[layer_idx].activation = activations.linear
penultimate_layer = utils.find_layer_idx(model, 'down3')
print(penultimate_layer)
#model = utils.apply_modifications(model)
img2 = resize(imgs[i], (height, width))
grads = visualize_cam(model,
layer_idx,
filter_indices=class_idx[0],
seed_input=img2,
penultimate_layer_idx=penultimate_layer)
ax3 = plt.subplot(4, n, i + 1 + 2 * n)
ax3.imshow(grads, cmap='jet', alpha=0.8)
ax4 = plt.subplot(4, n, i + 1 + 3 * n)
plt.imshow(msk, cmap='gray')
ax4.get_xaxis().set_visible(False)
ax4.get_yaxis().set_visible(False)
ax4.set_title("Mask")
plt.show()
plt.savefig('/home/CVL1/Shaobo/mri/outputs/test1.jpg')