def tryPointMaskerCascaded(model, batch_generator, sample_names=None):
if sample_names == None:
sample_names = batch_generator.all_names
for sample_name in sample_names:
inputs, _ = batch_generator.getTrainingPairFromName(sample_name)
#X, Y = batch_generator.getBatchFromNames([sample_name])
print inputs
im = inputs[0]
width = len(im[0])
height = len(im)
print 'sample name: ' + sample_name
base_masks, residual_masks = getNormalizedCascadedMasksFromImage(
model, im)
base_coords = utils.getCoordsFromPointMasks(base_masks, width, height,
'mean')
residual_coords = utils.getCoordsFromPointMasks(
residual_masks, 28, 28, 'mean')
#max_coords = utils.getCoordsFromPointMasks(base_masks, width, height, 'max')
coords = np.add(base_coords, residual_coords) - 28 / 2.0 - 4.0
#for i in range(len(residual_masks)):
# plt.imshow(residual_masks[i])
# plt.show()
utils.visualizeCoords(im, coords)
utils.visualizeCoords(im, np.concatenate([base_coords, coords],
axis=0),
np.arange(0, len(coords)))
def tryPointMaskerDilatedOnSamples(model):
folder = 'downloads/samples/'
for fname in os.listdir(folder):
if fname.endswith(('png', 'jpg')):
im = cv2.imread(folder + fname)
im = cv2.resize(im, (224, 224))
im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
width = len(im[0])
height = len(im)
masks = getNormalizedMasksFromImage(model, im)
coords = utils.getCoordsFromPointMasks(masks, width, height,
'mean')
max_coords = utils.getCoordsFromPointMasks(masks, width, height,
'max')
#for i in range(len(masks)):
# utils.visualizeMask
# utils.visualizeCoords(, [[y_avg, x_avg]])
"""masks /= np.max(masks, axis=[1,2])
summed = np.sum(masks, axis=0)
summed = cv2.resize(summed, (height, width))
summed = np.minimum(1, summed)
"""
masks = np.moveaxis(masks, 0, -1)
#utils.visualizeCoordMasks(im, masks)
utils.visualizeCoords(im, coords)
def tryPointMaskerCascadedOnSamples(model,
folder,
output_folder,
compare_coords=None):
if not os.path.exists(output_folder):
os.mkdir(output_folder)
for fname in os.listdir(folder):
if fname.endswith(('png', 'jpg')):
im = cv2.imread(folder + '/' + fname)
im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
print folder + '/' + fname
im = cv2.resize(im, (224, 224))
width = len(im[0])
height = len(im)
t1 = time.time()
base_masks, residual_masks = getNormalizedCascadedMasksFromImage(
model, im)
print time.time() - t1
base_coords = utils.getCoordsFromPointMasks(
base_masks, width, height, 'mean')
residual_coords = utils.getCoordsFromPointMasks(
residual_masks, 28, 28, 'mean')
#max_coords = utils.getCoordsFromPointMasks(base_masks, width, height, 'max')
coords = np.add(base_coords, residual_coords) - 28 / 2.0
# scale back
full_im = cv2.imread(folder + '/' + fname)
full_im = cv2.cvtColor(full_im, cv2.COLOR_BGR2RGB)
scale_width = len(full_im[0]) / float(len(im[0]))
scale_height = len(full_im) / float(len(im))
print scale_width
coords[:, 0] *= scale_height
coords[:, 1] *= scale_width
#for i in range(len(residual_masks)):
# plt.imshow(residual_masks[i])
# plt.show()
#utils.visualizeCoords(im, base_coords)
output_path = output_folder + '/' + fname
key = fname[:-4]
if key in compare_coords:
utils.visualizeCoords(full_im,
np.concatenate(
[compare_coords[key], coords],
axis=0),
np.arange(0, len(compare_coords[key])),
output_name=output_path)
else:
utils.visualizeCoords(full_im, coords, output_name=output_path)
def trySavedFullyConnected(model, batch_generator, sample_names=None):
if sample_names == None:
sample_names = batch_generator.all_names
for sample_name in sample_names:
inputs, outputs = batch_generator.getPair(sample_name)
# alpha channel needs to be cutoff
#if im.shape[2] > 3:
# im = im[:,:,:3]
im = inputs[0]
labels = np.array(model.predict(np.array(inputs), batch_size=1))
label = labels[0]
label *= len(im)
utils.visualizeCoords(im, label)
def visualizeSamples(folder, sample_names=[], model=None, special_indices=[]):
if len(sample_names) == 0:
with open(folder + '/names.json') as fp:
sample_names = set(json.load(fp))
for sample_name in sample_names:
im = np.load(folder + '/ims/' + sample_name + '.npy')
if model == None:
coords = np.load(folder + '/coords/' + sample_name + '.npy')
else:
coords = model.predict(np.array([im]), batch_size=1)
coords = np.reshape(coords, (-1, 2))
coords *= len(im)-1
utils.visualizeCoords(im, coords, special_indices)
def tryPointMaskerVanilla(model, batch_generator, sample_names=None):
if sample_names == None:
sample_names = batch_generator.all_names
for sample_name in sample_names:
inputs, _ = batch_generator.getPair(sample_name)
X, Y = batch_generator.getBatchFromNames([sample_name])
im = inputs[0]
print_str = 'sample name: ' + sample_name
if batch_generator.isInValSet(sample_name):
print_str += ', from val set'
else:
print_str += ', from train set'
outputs = model.predict_on_batch(X)
preds = np.squeeze(outputs[1])
preds = utils.imSoftmax(preds)
preds = np.moveaxis(preds, -1, 0)
preds = [normalizeMask(pred) for pred in preds]
coords = utils.getCoordsFromPointMasks(preds, len(im[0]), len(im))
utils.visualizeCoords(im, coords)
print print_str
"""
summed = 80 * np.sum(preds, axis=0)
summed = cv2.resize(summed, (len(im), len(im[0])), interpolation=cv2.INTER_LINEAR)
helenUtils.visualizeMask(im, summed)
"""
"""for i in range(3):
pred = preds[i]
label = labels[i]
l = len(pred)
c = np.zeros((l, l, 3))
pred = np.maximum(pred - np.mean(pred), 0)
c[:,:,0] = pred / np.max(pred)
c[:,:,2] = label
plt.imshow(c)
plt.show()
#plt.show()
#plt.imshow(label)
#plt.show()
"""
"""