def sliding_concentric_windows(self, img):
n_rows, n_cols = img.shape[:2]
y_a, x_a = 5, 5
y_b, x_b = 11, 11
new_img = np.zeros((n_rows, n_cols), dtype=np.uint)
img_a = util.pad(img, ((y_a/2, y_a/2), (x_a/2, x_a/2)), mode='constant')
img_b = util.pad(img, ((y_b/2, y_b/2), (x_b/2, x_b/2)), mode='constant')
blocks_a = util.view_as_windows(img_a, (y_a, x_a), step=1)
blocks_b = util.view_as_windows(img_b, (y_b, x_b), step=1)
for row in xrange(n_rows):
for col in xrange(n_cols):
mean_a = blocks_a[row, col].mean()
mean_b = blocks_b[row, col].mean()
r_mean = 0
if mean_a != 0:
r_mean = mean_b/float(mean_a)
if r_mean > 1.0:
new_img[row, col] = 0
else:
new_img[row, col] = 255
return new_img
def extract_patches(path, numPatchesPerImage, patchSize):
"""
:param path: path to a RGB fundus image
:param numPatchesPerImage: number of patches to extract per image
:param patchSize: patch is nxn size
:return: patches: matrix with an image patch in each row
"""
img = load(path)
img = img[:,:,1]
#contrast enhancemenet
img = equalize_adapthist(img)
windows = view_as_windows(img, (patchSize,patchSize))
j = 0
patches = np.zeros((numPatchesPerImage, patchSize*patchSize))
while(j < numPatchesPerImage):
sx = np.random.randint(0, windows.shape[0] - 1)
sy = np.random.randint(0, windows.shape[0] - 1)
x = (patchSize/2 - 1) + sx
y = (patchSize/2 - 1) + sy
r = (img.shape[0]/2) - 1
if np.sqrt((x - r) ** 2 + (y - r) **2 ) < r:
patch = windows[sx, sy, :].flatten()
patches[j,:] = patch
j += 1
else:
if j > 0:
j -= 1
return patches
def generic2dFilter(frame, shape, func, mode='reflect', out=None, padded=False
, step=1):
'''
generic2dFilter
Apply a generic function to blocks of a 2d array. Function must accept
arguments in the form (array, axis, out) where array is a 3d array of
flattened blocks from the original image, axis is the axis that sequences
the blocks and out is an optional user input array to be assigned to.
parameters
----------
mode : string
Decides how to pad the array. See numpy's pad function for a
description.
frame : 2 dimensional array
2 dimensional array
step : integer
Describes the stride used when iterating over the array.
'''
bSize= shape[0]//2, shape[1]//2
if padded:
paddedFrame = frame
reshapeSz = [(frame.shape[0]-2*bSize[0])//step
, (frame.shape[1]-2*bSize[1])//step
, -1]
else:
paddedFrame = np.pad(frame, bSize, mode=mode)
reshapeSz = [frame.shape[0]//step, frame.shape[1]//step, -1]
return func(view_as_windows(paddedFrame,shape,step=step).reshape(reshapeSz)
,axis=2, out=out)
def get_ss_ftr(self, channels, smp_loc=None):
"""
Compute self-similarity features
:param channels: shrunk channels for self-similarity features
:param smp_loc: shrunk sample locations (None for all)
:return: self-similarity features
"""
shrink = self.options["shrink"]
p_size = self.options["p_size"] / shrink
n_r, n_c, n_ch = channels.shape
ss_ftr = view_as_windows(channels, (p_size, p_size, n_ch))
if smp_loc is not None:
ss_ftr = ss_ftr.reshape((n_r - p_size + 1, n_c - p_size + 1,
p_size ** 2, n_ch))
r_pos = [r - p_size / 2 for r, _ in smp_loc]
c_pos = [c - p_size / 2 for _, c in smp_loc]
ss_ftr = ss_ftr[r_pos, c_pos]
else:
ss_ftr = ss_ftr.reshape((-1, p_size ** 2, n_ch))
n_cell = self.options["n_cell"]
half_cell_size = int(round(p_size / (2.0 * n_cell)))
grid_pos = [int(round((i + 1) * (p_size + 2 * half_cell_size - 1) / \
(n_cell + 1.0) - half_cell_size))
for i in xrange(n_cell)]
grid_pos = [r * p_size + c for r in grid_pos for c in grid_pos]
ss_ftr = ss_ftr[:, grid_pos]
ss_ftr = pdist(ss_ftr)
return ss_ftr.reshape((ss_ftr.shape[0], -1))
def im2col(im, fit_rect, fit_stride):
from skimage.util import view_as_windows
im_chan = im.shape[0]
row_num = im_chan * fit_rect[0] * fit_rect[1]
return view_as_windows(
im,
(im_chan, fit_rect[0], fit_rect[1]),
(1, fit_stride[0], fit_stride[1])).reshape((-1,row_num)).T
def dim2col(dcol, col, im, fit_rect, fit_stride):
from skimage.util import view_as_windows
im_chan = im.shape[0]
dim = np.zeros(im.shape)
v = view_as_windows(
dim,
(im_chan, fit_rect[0], fit_rect[1]),
(1, fit_stride[0], fit_stride[1]))
v += dcol.T.reshape(v.shape)
return dim * (1 / (fit_rect[0] * fit_rect[1]))
def loadFromFolder( self, folder, labelImage, patch_size = (23, 23), step_size = 4 ):
self.featureNames = []
self.features = []
for fname in glob.glob( folder + "/*.nrrd" ):
fsplit = fname.split('_');
name = "image-" + fsplit[2] + "-" + fsplit[4]
self.featureNames.append( name )
imageType = itk.Image[itk.F, 3]
reader = itk.ImageFileReader[imageType].New()
reader.SetFileName( fname )
reader.Update()
image = itk.GetArrayFromImage( reader.GetOutput() ).squeeze()
image = image - np.min(image)
image = image / np.max(image)
patches = util.view_as_windows( image, patch_size, step_size )
patch_shape = ( patches.shape[0]*patches.shape[1], patches.shape[2], patches.shape[3])
patches = np.reshape( patches, patch_shape, 'F' )
print(patches.shape)
self.features.append( patches )
print( patches.shape )
self.features = np.stack(self.features, axis=-1)
imageType = itk.Image[itk.UC, 3]
reader = itk.ImageFileReader[imageType].New()
reader.SetFileName( labelImage )
reader.Update()
labelImage = itk.GetArrayFromImage( reader.GetOutput() ).squeeze()
print("label: ")
print( labelImage.shape )
patches = util.view_as_windows( labelImage, patch_size, step_size )
patch_shape = ( patches.shape[0]*patches.shape[1], patches.shape[2], patches.shape[3])
patches = np.reshape( patches, patch_shape, 'F' )
self.labels = np.zeros( patches.shape[0] )
for i in range(0, patches.shape[0] ):
a, b = np.unique( patches[i,].flatten(), return_inverse=True )
c = np.argmax( np.bincount(b) )
self.labels[i] = a[c]
def patch_extract(image, patch_width, stride): # Ugly code, extracts r, g, b patches from the image convolutionally. patches_r = util.view_as_windows(image[:, :, 0], (patch_width, patch_width), stride) patches_g = util.view_as_windows(image[:, :, 1], (patch_width, patch_width), stride) patches_b = util.view_as_windows(image[:, :, 2], (patch_width, patch_width), stride) patches_r = numpy.reshape(patches_r, ((image.shape[0]-patch_width+1)**2, patch_width, patch_width), order="F") patches_r = numpy.reshape(patches_r, ((image.shape[0]-patch_width+1)**2, patch_width*patch_width), order="F") patches_g = numpy.reshape(patches_g, ((image.shape[0]-patch_width+1)**2, patch_width, patch_width), order="F") patches_g = numpy.reshape(patches_g, ((image.shape[0]-patch_width+1)**2, patch_width*patch_width), order="F") patches_b = numpy.reshape(patches_b, ((image.shape[0]-patch_width+1)**2, patch_width, patch_width), order="F") patches_b = numpy.reshape(patches_b, ((image.shape[0]-patch_width+1)**2, patch_width*patch_width), order="F") patches = numpy.append(patches_r, patches_g, 1); patches = numpy.append(patches, patches_b, 1); return patches
def getZMatrix(self, X):
"""
Input: (n_instances, n_channels, input_size, input_size)
Output: (n_instances, n_patchs, m)
"""
Z = view_as_windows(X, (1, X.shape[1], self.filter_size, self.filter_size))
Z = Z.reshape(np.prod(Z.shape[:4]), np.prod(Z.shape[4:]))
Q = self.rbf_feature.transform(Z).astype(np.float16)
return Q.reshape(X.shape[0], self.n_patchs, -1)
def get_training_data(noisy_image, clear_image, patch_size=3):
noisy_patches = view_as_windows(
noisy_image, (patch_size, patch_size))
training_inputs = [
noisy_patches[index].ravel()
for index in np.ndindex(
noisy_patches.shape[0], noisy_patches.shape[1])
]
patch_offset = (patch_size - 1) / 2
training_outputs = clear_image[
patch_offset:clear_image.shape[0] - patch_offset,
patch_offset:clear_image.shape[1] - patch_offset
].ravel()
return training_inputs, training_outputs
def extractLBPfeatures(img):
lbp = local_binary_pattern(img, cfg.lbp_n_points, cfg.lbp_radius, cfg.lbp_METHOD)
lbp_windows = view_as_windows(lbp, window_shape=cfg.lbp_win_shape, step=cfg.lbp_win_step)
features = []
count = 0
for windows_list in lbp_windows:
for window in windows_list:
lbp_hist, bin_edges = np.histogram(window, bins=cfg.lbp_n_bins)
lbp_hist_norm = sum(abs(lbp_hist))
lbp_hist_l1sqrtnorm = np.sqrt(lbp_hist/float(lbp_hist_norm))
features.append(lbp_hist_l1sqrtnorm)
count += 1
features_flatten = [item for sublist in features for item in sublist]
return features_flatten
def reshape_to_blocks(data, block_shape):
overlap_window = 0.5
step = (int(block_shape[0] * overlap_window),
int(block_shape[1] * overlap_window))
blocks = view_as_windows(data, block_shape, step)
axis_0 = np.expand_dims(np.arange(0, step[0] * blocks.shape[0], step[0]),
axis=-1)
axis_1 = np.expand_dims(np.arange(0, step[1] * blocks.shape[1], step[1]),
axis=0)
indices = np.stack([
np.tile(axis_0, (1, blocks.shape[1])),
np.tile(axis_1, (blocks.shape[0], 1))
],
axis=-1)
#flatten the two first dimension
blocks = blocks.reshape(-1, blocks.shape[2], blocks.shape[3], 1)
return indices, blocks #expand dims to fit keras format
def csvfft(f):
for line in f:
try:
M = 2205 #window size
f_s = 100
#initials
song = line[:-1] #song?
print('Song = ' + str(song))
rate, audio = wavfile.read(song) #read wav file
audio = np.mean(audio, axis=1) #average all the values
N = audio.shape[0] #grab mono channel
slices = util.view_as_windows(audio, window_shape=(M,), step = 2205) #get slices
win = np.hanning(M+1)[:-1]
slices = slices * win
slices = slices.T #organize into columns
spectrum = np.fft.fft(slices, axis=0) #do fft of audio
spectrum = np.abs(spectrum)
freqs = fftpack.fftfreq(len(spectrum[0]))
for i in range(len(freqs)):
freqs[i] = freqs[i] * 1000
freqs = freqs[:len(freqs)/2]
df = pd.DataFrame(index = range(len(spectrum)), columns = freqs)
print(len(freqs))
print(len(spectrum))
print(len(spectrum[0])/2)
temp = spectrum
for i in range(len(spectrum)):
#print(float(i) / len(spectrum) * 100)
temp = spectrum[i]
#print('Spectrum size = ' + str(spectrum.size/2))
#print(spectrum)
temp = temp[:(temp.size/2)]
df.loc[i] = temp
df.to_csv(line[:-4] + 'csv')
#freqs = np.fft.fftfreq(len(spectrum[300])) * f_s
#freqs = fftpack.fftfreq(len(spectrum[200]))
#for i in range(len(freqs)):
# freqs[i] = freqs[i] * 1000
#for i in range(len(spectrum[30])):
# print(freqs[i])
#plt.plot(np.abs(freqs), spectrum[250], 'r')
#plt.show()
except ValueError:
pass
def matrix_median(R, n_med=4): """ median filtering on matrix rows """ from skimage.util import view_as_windows # half median size n_h = n_med // 2 # shape m, n = R.shape # filter matrix to be filtered R_fil = view_as_windows(np.pad(R, ((0, 0), (n_h, n_h))), (1, n_med+1), step=1).reshape(m, n, n_med+1) return np.median(R_fil, axis=2)
def main():
args = parser.parse_args()
img = cv2.imread(args.file, cv2.IMREAD_GRAYSCALE)
patch_size = args.patch
patches = view_as_windows(img, (patch_size, patch_size), step=patch_size)
hmap = None
print("Total %d, %d" % patches.shape[:-2])
for i, r in enumerate(patches):
for j, template in enumerate(r):
print("Processing %d, %d" % (i, j), end="\r")
res = cv2.matchTemplate(img, template, cv2.TM_CCOEFF_NORMED)
# create a mask and threshold off only local region to the patch
cut_val = args.threshold * np.max(res)
res[res < cut_val] = 0
if hmap is not None: hmap += res
else: hmap = res
plot(img, hmap)
def predict_on_image(image,
weights,
network='unet',
n_class=1,
pad_width=16,
window_shape=(288, 288),
step=256):
"""Returns a network prediction on an image — i.e., a 2D object.
Parameters
----------
image : (M, N) array
The input data.
weights : TensorFlow weights
A file with TensorFlow weights for the desired network.
network : str, optional (default : 'unet')
The network to use in prediction, corresponding to the weights file.
n_class : int, optional (default : 1)
Number of classes to predict.
pad_width : int, optional (default : 16)
Width of the padding to add in the borders of data.
window_shape : (M, N) list, optional (default : (288, 288))
Size of the window to process.
step : int, optional (default : 16)
Size of the step used to chunk the data. Used as an overlap feature.
Returns
-------
prediction : (M, N) array
Array containing predictions on the input data.
"""
model = network_models(network, window_shape=window_shape, n_class=n_class)
if network is None:
raise (f'Network {network} not available.')
model.load_weights(weights)
image = np.pad(image, pad_width=pad_width)
image_crop = np.vstack(
util.view_as_windows(image, window_shape=window_shape, step=step))
crop_steps = image_crop.shape[0]
image_gen = tensor_generator(image_crop)
results = model.predict(image_gen, steps=crop_steps, verbose=1)
prediction = _aux_predict(results)
return prediction
def haralick_features_ssm(img, win, d):
win_sz = 2 * win + 1
window_shape = (win_sz, win_sz)
arr = np.pad(img, win, mode='reflect')
windows = util.view_as_windows(arr, window_shape)
Nd = len(d)
feats = np.zeros(shape=windows.shape[:2] + (Nd, 9), dtype=np.float64)
for m in tqdm_notebook(range(windows.shape[0]),
desc='02_rows',
leave=False):
for n in tqdm_notebook(range(windows.shape[1]),
desc='03_cols',
leave=False):
for i, di in enumerate(d):
w = windows[m, n, :, :]
feats[m, n, i, :] = calc_haralick(w, distance=di)
return feats.reshape(feats.shape[:2] + (-1, ))
def generate_real_samples(n_batch):
"""get random sampples and do the last preprocessing on them"""
while True:
# get random sample of indices from the precomputed indices
# for this we generate random indices for the index list (confusing termoonology, since we use
# indices to index the list of indices...
ixs = np.random.randint(n_samples, size=n_batch)
idcs_batch = indices_all[ixs]
# now we select the data corresponding to these indices
data_wview = view_as_windows(data, (1, 1, ndomain, ndomain))[..., 0,
0, :, :]
batch = data_wview[idcs_batch[:, 0], :, idcs_batch[:, 1],
idcs_batch[:, 2]]
# add empty channel dimension (necessary for keras, which expects a channel dimension)
batch = np.expand_dims(batch, -1)
# compute daily sum (which is the condition)
batch_cond = np.sum(batch, axis=1) # daily sum
# the data now is in mm/hour, but we want it as fractions of the daily sum for each day
for i in range(n_batch):
batch[i] = batch[i] / batch_cond[i]
# normalize daily sum
batch_cond = batch_cond / norm_scale
# add lon
batch_lon = idcs_batch[:, 2]
# normalize to [0,1]
batch_lon = (batch_lon - min_lonidx) / max_lonidx
# repeat it to ndomain x ndomain x 1 (last is channel dimension)
batch_lon = np.tile(batch_lon, (1, ndomain, ndomain, 1)).T
# now it has n_batch x ndomain x ndomain x 1
# and add it as additional varialbes
batch_cond = np.concatenate([batch_cond, batch_lon], axis=-1)
assert (batch.shape == (n_batch, nhours, ndomain, ndomain, 1))
assert (batch_cond.shape == (n_batch, ndomain, ndomain, n_channel))
assert (~np.any(np.isnan(batch)))
assert (~np.any(np.isnan(batch_cond)))
assert (np.max(batch) <= 1)
assert (np.min(batch) >= 0)
yield [batch, batch_cond]
def window_region_merge_grey(input_im, b):
block_size = (b, b)
pad_width = []
reshape_size = b * b
for i in range(len(block_size)):
if input_im.shape[i] % block_size[i] != 0:
after_width = block_size[i] - (input_im.shape[i] % block_size[i])
else:
after_width = 0
pad_width.append((0, after_width))
input_im = np.pad(input_im, pad_width=pad_width, mode='constant')
view = view_as_windows(input_im, block_size)
print(view.shape)
flatten_view = np.transpose(view, axes=(0, 1, 3, 2))
flatten_view = flatten_view.reshape(flatten_view.shape[0],
flatten_view.shape[1], reshape_size)
return np.squeeze(mode(flatten_view, axis=2)[0])
def get_sequences(melgram, _n_mels):
"""
Transform 2 or 3D mel spectrogram into 3D array of windows
:param melgram: 2 or 3D mel spectrogram (if 3D, squeezed to 2D)
:param _n_mels: number of mel buckets
:return: array of sequences. shape (n_frames, seq_len, n_mels)
"""
if len(melgram.shape) == 3:
melgram = np.squeeze(melgram)
seq_len = hyper_params.sample_frames
sequences = view_as_windows(melgram, window_shape=(seq_len, _n_mels))
sequences = np.squeeze(sequences, 1)
return sequences
def make_dataset(gesture_path, save_dir, pattern, winSize = 128,step = 2):
if not os.path.exists(save_dir):
os.makedirs(save_dir)
data = load_data(gesture_path)
# data = sp.despikeSeries(data)
x_train_signals_paths = [
os.path.join(save_dir ,signal + ".csv" )for signal in INPUT_SIGNAL_TYPES
]
x_train_files = [csv.writer(open( x_train_signals_path,"a",newline=''))for x_train_signals_path in x_train_signals_paths]
y_train_file = csv.writer(open(os.path.join(save_dir,"y.csv"),"a",newline=''))
for i ,signals in enumerate(data):
segmented_signal = util.view_as_windows(signals, (winSize,),step = step)
for signal in segmented_signal:
x_train_files[i].writerow(signal)
if i ==0:
y_train_file.writerow([pattern])
def Shrink(X, shrinkArg, max_pooling=True, padding=False):
win = shrinkArg['win']
max_pooling = shrinkArg['max_pooling']
padding = shrinkArg['padding']
if max_pooling:
X = block_reduce(X, (1, 2, 2, 1), np.average)
if padding:
new_X = []
for each in X:
each = cv2.copyMakeBorder(each, win // 2, win // 2, win // 2, win // 2, cv2.BORDER_REFLECT)
new_X.append(each)
new_X = np.array(new_X)[:, :, :, None]
else:
new_X = X
X = view_as_windows(new_X, (1, win, win, 1))
X = X.reshape(X.shape[0], X.shape[1], X.shape[2], -1)
return X
def extractLBPfeatures(img):
lbp = local_binary_pattern(img, cfg.lbp_n_points, cfg.lbp_radius,
cfg.lbp_METHOD)
lbp_windows = view_as_windows(lbp,
window_shape=cfg.lbp_win_shape,
step=cfg.lbp_win_step)
features = []
count = 0
for windows_list in lbp_windows:
for window in windows_list:
lbp_hist, bin_edges = np.histogram(window, bins=cfg.lbp_n_bins)
lbp_hist_norm = sum(abs(lbp_hist))
lbp_hist_l1sqrtnorm = np.sqrt(lbp_hist / float(lbp_hist_norm))
features.append(lbp_hist_l1sqrtnorm)
count += 1
features_flatten = [item for sublist in features for item in sublist]
return features_flatten
def generate_latent_points(n_batch):
# generate points in the latent space and a random condition
latent = np.random.normal(size=(n_batch, latent_dim))
# randomly select conditions
ixs = np.random.randint(0, n_samples, size=n_batch)
idcs_batch = indices_all[ixs]
data_wview = view_as_windows(data, (1, 1, ndomain, ndomain))[..., 0,
0, :, :]
batch = data_wview[idcs_batch[:, 0], :, idcs_batch[:, 1], idcs_batch[:, 2]]
# add empty channel dimension (necessary for keras, which expects a channel dimension)
batch = np.expand_dims(batch, -1)
batch_cond = np.sum(batch, axis=1) # daily sum
# normalize daily sum
batch_cond = batch_cond / norm_scale
assert (batch_cond.shape == (n_batch, ndomain, ndomain, 1))
assert (~np.any(np.isnan(batch_cond)))
return [latent, batch_cond]
def rolling_window(image,
clf,
scaler,
stepSize=8,
Zstep=4,
windowSize=(1, 70, 70)):
z, x, y = image.shape
rol_view = util.view_as_windows(image, (1, 70, 70), step=(4, 8, 8))
rol_view_flat = rol_view.reshape(-1, 70, 70)
a, b, c = rol_view_flat.shape
list_proba = joblib.Parallel(n_jobs=-1)(
joblib.delayed(_task_predict)(window, clf, scaler)
for window in rol_view_flat)
list_proba_array = np.asarray(list_proba)
array_proba = np.asarray(list_proba).reshape(int(a / (52 * 52)), 52, 52)
result = np.zeros((z, x - 70, y - 70))
result[::4, ::8, ::8] = array_proba
return (result)
def min_distance(patch, size, v):
print(patch, size)
x = patch.mean()
min_d = 999999999
dist = {}
for pat1 in (view_as_windows(v, size)):
for pat in pat1:
print(pat.shape)
a = abs(x - pat.mean())
a = math.pow(a, 2)
b = (size[0] * size[1])
d = a / b
dist[d] = pat
if (d < min_d):
min_d = d
dist = collections.OrderedDict(sorted(dist.items()))
dist = list(dist.values())
return min_d, (dist[0:1])
def generate_sparse_matrix(shape, op):
count = product(shape)
indices = np.arange(count).reshape(shape)
view = view_as_windows(indices, op.shape)
window_shape = view.shape[0:len(shape)]
X, Y, W = [], [], []
cnt = op.size
center = cnt // 2
opr = op.ravel()
for idx in rangeN(window_shape):
v = view[idx].ravel()
for p in range(cnt):
if opr[p] != 0:
X.append(v[center])
Y.append(v[p])
W.append(opr[p])
inside = to_csr_matrix(W, X, Y, dtype=np.float, shape=(count, count))
return inside
def sliding_window_prediction(d, step):
windows = view_as_windows(d, window_shape=(window_x, window_y), step=step)
indexes = []
for i, p in enumerate(windows):
for j, w in enumerate(p):
should_skip = white_pixel_filter(w)
if should_skip: continue
w = np.reshape(w, newshape=(1, window_x, window_y, 1))
prediction = model.predict(w)
certainty = prediction.max()
if certainty < CUTOFF: continue
predict = prediction.argmax(axis=1)
label = Labels.from_int(predict.tolist()[0])
indexes.append((i, j, label, certainty))
return indexes
def extract_patches(
image: np.ndarray,
patch_shape: Tuple[int, int] = (300, 300),
overlap: Union[Optional[int], Optional[float],
Tuple[Optional[Union[float, int]],
Optional[Union[float, int]]], ] = None,
):
if len(image.shape) > 3:
raise ValueError("Expected single image")
patch_h, patch_w = patch_shape
stride_h, stride_w = normalize_overlap(overlap, patch_shape)
window_shape = ((patch_h, patch_w, image.shape[2]) if image.ndim == 3 else
(patch_h, patch_w))
step = (stride_h, stride_w, 1) if image.ndim == 3 else (stride_h, stride_w)
patches = view_as_windows(image, window_shape, step)
# patches = patches.reshape(-1, patch_h, patch_w, image.shape[2])
return patches
def patches_true(self, iteration):
subjects_true = self.data_true(iteration)
patches_true = np.empty(
[self.batch_size * self.num_patches, self.width_patch + self.margin, self.heigth_patch + self.margin,
self.depth_patch + self.margin, 1])
i = 0
for subject in subjects_true:
patch = view_as_windows(subject, window_shape=(
(self.width_patch + self.margin), (self.heigth_patch + self.margin), (self.depth_patch + self.margin)),
step=(self.width_patch - self.margin, self.heigth_patch - self.margin,
self.depth_patch - self.margin))
for d in range(patch.shape[0]):
for v in range(patch.shape[1]):
for h in range(patch.shape[2]):
p = patch[d, v, h, :]
p = p[:, np.newaxis]
p = p.transpose((0, 2, 3, 1))
patches_true[i] = p
i = i + 1
return patches_true
def denoise_image(In, stride):
In_p = skimage.util.pad(In, 32, 'reflect')
In_p = In_p[:, :, 32:(2 * 32 + 3) - 32]
Pn_p = view_as_windows(In_p[:, :, :],
window_shape=(64, 64, 3),
step=stride)
ss = Pn_p.shape
Pn_p = Pn_p.reshape(ss[0] * ss[1], ss[3], ss[4], 3) / 255.0
Pp_p = numpy.zeros_like(Pn_p)
# r=0
for ind in range(0, ss[0] * ss[1], batch_size):
# print ind,ind+batch_size
mini_batch = Pn_p[ind:ind + batch_size]
# print "{} = {}".format("Input shape",mini_batch[0].shape)
# print mini_batch[0]
# np.save('ex1'+str(r+1)+'.npy',mini_batch)
predk = E.sess.run([y_],
feed_dict={
x_: mini_batch,
phase_train: False
})
# np.save('ez1'+str(r+1)+'.npy',predk)
# print "{} = {}".format("Shape of the batch",predk[0].shape)
# print "{} = {}".format("Max of the batch",np.max(predk[0]))
# print "{} = {}".format("Shape of the input",mini_batch.shape)
# print "{} = {}".format("Max of the input",np.max(mini_batch))
# r = r + 1
# print "{} ={}".format("Shape of the output",predk[0][0].shape)
# print predk[0][0]
# r=r+1
# print "{} = {}".format("MSE of inpit and output",((predk[0][0] - mini_batch[0])**2).mean(axis=None))
# print "{} = {}".format("PSNR of input and output",measure.compare_psnr(predk[0][0],mini_batch[0], dynamic_range = 1.0))
# if( r==2):
# sys.exit(0)
# print "{} = {}".format("MSE OF INPUT AND RECONSTRUCTED",((predk[0] - mini_batch)**2).mean(axis = None))
Pp_p[ind:ind + batch_size, :, :, :] = predk[0]
Pp_p = Pp_p * 255.0
Ip = reconstruct_image_from_patches(Pp_p, ss, In_p)
return post_process(Ip)
def form_visibility(data, rate, fc, bw, T_sti, T_stationarity):
"""
Parameter
---------
data : :py:class:`~numpy.ndarray`
(N_sample, N_channel) antenna samples. (float)
rate : int
Sample rate [Hz]
fc : float
Center frequency [Hz] around which visibility matrices are formed.
bw : float
Double-wide bandwidth [Hz] of the visibility matrix.
T_sti : float
Integration time [s]. (time-series)
T_stationarity : float
Integration time [s]. (visibility)
Returns
-------
S : :py:class:`~numpy.ndarray`
(N_slot, N_channel, N_channel) visibility matrices.
Note
----
Visibilities computed directly in the frequency domain.
For some reason visibilities are computed correctly using
`x.reshape(-1, 1).conj() @ x.reshape(1, -1)` and not the converse.
Don't know why at the moment.
"""
S_sti = (statistics.TimeSeries(data, rate).extract_visibilities(T_sti,
fc,
bw,
alpha=1.0))
N_sample, N_channel = data.shape
N_sti_per_stationary_block = int(T_stationarity / T_sti) + 1
S = (skutil.view_as_windows(
S_sti, (N_sti_per_stationary_block, N_channel, N_channel),
(N_sti_per_stationary_block, N_channel, N_channel)).squeeze(
axis=(1, 2)).sum(axis=1))
return S
def xpSample():
song = 'AlbenizI-Tango.wav'
rate, audio = wavfile.read(song)
audio = np.mean(audio, axis=1)
N = audio.shape[0]
M = 2205
slices = util.view_as_windows(audio, window_shape=(M, ), step=2205)
win = np.hanning(M + 1)[:-1]
slices = slices * win
slices = slices.T
spectrum = np.fft.fft(slices, axis=0)
spectrum = np.abs(spectrum)
#spectrum = 20 * np.log10(spectrum / spectrum.max())
f_s = 100
freqs = np.fft.fftfreq(len(spectrum[300])) * f_s
L = N / rate
print('Audio Length: {:.2f} seconds'.format(L))
print(len(audio))
#print(slices.shape)
#fig, ax = plt.subplots()
#print(len(spectrum))
#print(len(spectrum[300]))
#print(spectrum)
#spectrum = np.fft.fft(spectrum)
#spectrum = np.abs(spectrum)
#print(spectrum[i])
freqs = fftpack.fftfreq(len(spectrum[200]))
print(len(freqs))
print(len(spectrum))
print(len(spectrum[0]))
for i in range(len(freqs)):
freqs[i] = freqs[i] * 1000
for i in range(len(spectrum[30])):
print(freqs[i])
plt.plot(np.abs(freqs), spectrum[250], 'r')
#plt.plot(np.abs(spectrum[300]),freqs, 'r')
#plt.show()
#ax.plot(freqs, np.abs(spectrum[1]), 'r')
plt.show()
'''
def denoise_image(In, stride):
In_p = skimage.util.pad(In, 32, 'reflect')
Pn_p = view_as_windows(In_p, window_shape=(64, 64), step=stride)
ss = Pn_p.shape
Pn_p = Pn_p.reshape(ss[0] * ss[1], ss[2], ss[3], 1) / 255.0
Pp_p = numpy.zeros_like(Pn_p)
for ind in range(0, ss[0] * ss[1], batch_size):
#print ind,ind+batch_size
mini_batch = Pn_p[ind:ind + batch_size]
#print mini_batch.shape
predk = E.sess.run([y_],
feed_dict={
x_: mini_batch,
phase_train: False
})
#print numpy.asarray(predk).shape
Pp_p[ind:ind + batch_size, :, :, :] = predk[0]
Pp_p = Pp_p * 255.0
Ip = reconstruct_image_from_patches(Pp_p, ss, In_p)
return post_process(Ip)
def __init__(self, mat_path, patch_size=12, step_size=1, normalize=False):
dataset = sio.loadmat(mat_path)
images = np.ascontiguousarray(
dataset['IMAGES']) # shape: (512, 512, 10)
self.patches = np.squeeze(
view_as_windows(images, (patch_size, patch_size, 10),
step=step_size)) # shape: (., ., PS, PS, 10)
self.patches = self.patches.transpose((0, 1, 4, 2, 3))
self.patches = self.patches.reshape((-1, patch_size, patch_size))
if normalize:
# normalize to range [0, 1]
_min = self.patches.min()
_max = self.patches.max()
self.patches = (self.patches - _min) / (_max - _min)
if self.patches.dtype != np.float:
print('converting data type from %r to np.float' %
self.patches.dtype)
self.patches = self.patches.astype(np.float)
print('image statistics:')
print('min: %r, mean: %r, max: %r' %
(self.patches.min(), self.patches.mean(), self.patches.max()))
def image_to_crops(image, patch_size, stride, augment_fn=None):
"""
Returns a grid of patches from the given image, taking into account patch size and stride.
:param image: input image with dimensions (x, y, ch).
:param patch_size: size of crop/patch.
:param stride: size of stride in pixels. For example, 50% overlapping means stride=patch_size//2.
:param augment_fn: augmentation function applied to crops of image in float [-1, +1] format and
dimensions [n_crops, n_crops, x, y, ch].
:return: crops with dimensions [n_crops, n_crops, x, y, ch]
"""
# Get patches
crops = view_as_windows(image, (patch_size, patch_size, 3),
(stride, stride, 1)).squeeze()
# Augment
if augment_fn is not None:
crops = augment_fn(crops)
return crops
def patchify(image: np.ndarray,
patch_size: Imsize,
step: int = 1) -> np.ndarray:
"""
Split a 2D or 3D image into small patches given the patch size.
Parameters
----------
image: the image to be split. It can be 2d (m, n) or 3d (k, m, n)
patch_size: the size of a single patch
step: the step size between patches
Examples
--------
>>> image = np.array([[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12]])
>>> patches = patchify(image, (2, 2), step=1) # split image into 2*3 small 2*2 patches.
>>> assert patches.shape == (2, 3, 2, 2)
>>> reconstructed_image = unpatchify(patches, image.shape)
>>> assert (reconstructed_image == image).all()
"""
return view_as_windows(image, patch_size, step)
def bump_z(self, surface):
points = list(self.__points)
image = surface.z_image
window_shape = (3, 3)
windows = util.view_as_windows(image, window_shape)
for i, point in enumerate(points):
query_coord = (point[1] - 1, point[0] - 1)
local_window = windows[query_coord]
indices = np.where(local_window == local_window.min())
local_coords = indices[0][0], indices[1][0]
global_coords = local_coords[0] + query_coord[0], local_coords[
1] + query_coord[1]
points[i] = global_coords[1], global_coords[0]
return Path(points)
def Shrink(X, shrinkArg):
win = shrinkArg['win']
stride = shrinkArg['stride']
ith_hop = shrinkArg['ith_hop']
if ith_hop == 1:
X_new = X
else:
X_new = np.zeros([X.shape[0], X.shape[1] // 2, X.shape[2] // 2, X.shape[3]])
for i in range(0, X.shape[0]):
for j in range(0, X.shape[3]):
for m in range(0, X.shape[1], 2):
for n in range(0, X.shape[2], 2):
pixel_list = [X[i, m, n, j], X[i, m + 1, n, j],
X[i, m, n + 1, j], X[i, m + 1, n + 1, j]]
max_pixel = max(pixel_list)
X_new[i, m // 2, n // 2, j] = max_pixel
ch = X_new.shape[-1]
X_new = view_as_windows(X_new, (1, win, win, ch), (1, stride, stride, ch))
return X_new.reshape(X_new.shape[0], X_new.shape[1], X_new.shape[2], -1)
def extractLBPfeatures(img, detector):
lbp = local_binary_pattern(img, cfg.lbp_n_points, cfg.lbp_radius, cfg.lbp_METHOD)
lbp_windows = view_as_windows(lbp, window_shape=cfg.lbp_win_shape, step=cfg.lbp_win_step)
features = []
count = 0
ws = cfg.lbp_win_shape[1]
kpts = detector.detect(img)
loc = [(int(x.pt[0]),int(x.pt[1])) for x in kpts]
for poi in loc:
window = Image.fromarray(img).crop((poi[0]-ws,poi[1]-ws,poi[0]+ws,poi[1]+ws))
window = np.array(window)
# for windows_list in lbp_windows:
# for window in windows_list:
lbp_hist, bin_edges = np.histogram(window, bins=cfg.lbp_n_bins)
lbp_hist_norm = sum(abs(lbp_hist))
lbp_hist_l1sqrtnorm = np.sqrt(lbp_hist/float(lbp_hist_norm))
features.append(lbp_hist_l1sqrtnorm)
count += 1
#features_flatten = [item for sublist in features for item in sublist]
features = np.asarray(features)
return features
def parse_wave_csv(filename,
fft=False,
window_stride=0.001,
window_size=0.02,
sampling_rate=44100):
data = np.genfromtxt(filename, delimiter=',')[:30 * sampling_rate,
1:].flatten()
if fft:
fft_data = []
N = 2**int(np.ceil(np.log2(window_size * sampling_rate)))
assert N <= len(data), "Segment too short"
hann = np.hanning(N)
stride = int(round(window_stride * sampling_rate))
windows = util.view_as_windows(data, window_shape=N, step=stride)
windows = hann * windows
spectrum = np.fft.fft(windows)
return np.mean(np.abs(spectrum)[:, 1:(N - 1) // 2 + 1], axis=0)
return data
def get_dense_patches(image, patch_size = 8, sample_rate = 4):
"""Gets dense patches of pixels from an image.
This uses the view_as_windows function from scikit-image.
Parameters
----------
image: ndarray
Image to find dense patches within.
patch_size: int (default: 8)
Size of each patch. 8 x 8 patches are recommended.
sample_rate: int (default: 4)
How many pixels apart each patch should be chosen in both x and y
directions.
Returns
-------
out: ndarray
An array of which the rows make up a ravel of each dense patch.
"""
window_shape = (patch_size, patch_size)
patches = view_as_windows(image, window_shape = window_shape,
step = sample_rate)
shp = patches.shape
# Resize the array so that each row is one dense patch.
out = np.reshape(patches, (shp[0]*shp[1], shp[2]*shp[3]))
# Normalize the data
m = out.mean(axis = 1)
# Have to transpose as it won't take subtract along columns properly
out = (np.transpose(out) - m).transpose()
# Make unit length along the rows
out = normalize(out, axis = 1)
return out
def f_unsup(img):
img_ptchs = view_as_windows(img, (w, w), step=s)
orig_dim = (img_ptchs.shape[0], img_ptchs.shape[1], k_means.cluster_centers_.shape[0])
img_ptchs = numpy.reshape(img_ptchs, (img_ptchs.shape[0] * img_ptchs.shape[0], w * w))
# normalize the patches, per patch sample
img_ptchs = preprocessing.scale(img_ptchs, axis=1)
# apply whitening transform
img_ptchs = numpy.dot(img_ptchs - mu, Wh) # uses multiprocessing!
Z = k_means.transform(img_ptchs)
# sparsity constraint threshold q
uZ = numpy.percentile(Z, q, axis=1, keepdims=True)
tri_Z = numpy.maximum(0, uZ - Z)
tri_Z_image = numpy.reshape(tri_Z, orig_dim)
# pooling of averages at mesh resolution
pooled_Z = []
pool_dim1 = orig_dim[0]/mesh_r or orig_dim[0]
pool_dim2 = orig_dim[1]/mesh_r or orig_dim[1]
for i in xrange(0,orig_dim[0], pool_dim1):
for j in xrange(0,orig_dim[1], pool_dim2):
POOL_MSH = numpy.mean(tri_Z_image[i:(i + pool_dim1), j:(j + pool_dim2)], axis=(0, 1))
pooled_Z.append(POOL_MSH)
return numpy.hstack(pooled_Z)
def get_reg_ftr(self, channels, smp_loc=None):
"""
Compute regular features.
:param channels: shrunk channels for regular features
:param smp_loc: shrunk sample locations (None for all)
:return: regular features
"""
shrink = self.options["shrink"]
p_size = self.options["p_size"] / shrink
n_r, n_c, n_ch = channels.shape
reg_ftr = view_as_windows(channels, (p_size, p_size, n_ch))
reg_ftr = reg_ftr.reshape((n_r - p_size + 1, n_c - p_size + 1,
p_size ** 2 * n_ch))
if smp_loc is not None:
r_pos = [r - p_size / 2 for r, _ in smp_loc]
c_pos = [c - p_size / 2 for _, c in smp_loc]
reg_ftr = reg_ftr[r_pos, c_pos]
return reg_ftr
def process_single_file(self, filename):
'''
processes a single image file
'''
img = io.imread(filename)
try:
img = img.transpose(2,0,1)
img = transform.resize(img, (3,227,227))
except:
print "image didn't have correct shape, continuing"
return None, None
img -= self.mean_image
label = filename.split('/')[-2] # HACKY!
y = self.l2i[label]
# view is 1 x a x b x window shape.
view = util.view_as_windows(img, window_shape=self.window_shape, step=self.window_step)
# we want this to be num_windows x windows shape, so:
batches = view.reshape(np.prod(view.shape[:-3]), *view.shape[3:])
new_batches = []
for batch in batches:
new_batches.append(transform.resize(batch,(3,227,227)))
return np.array(new_batches), y
def extractRandomPatches(self, patchSize, numPatches, stride=1):
size = self.info()['entries']
data = self.next()
patches = np.zeros((numPatches*size, patchSize[0]*patchSize[1]*patchSize[2]))
j = 0
for z in range(0, size):
start = time.time()
img = np.transpose(data[1], (1,2,0))
windows = view_as_windows(img, patchSize, stride)
r = np.random.randint(0, windows.shape[0] - windows.shape[3], (numPatches,))
c = np.random.randint(0, windows.shape[1]- windows.shape[4], (numPatches,))
for i in range(0, numPatches):
patch = np.reshape(windows[r[i],c[i],0,:], windows.shape[3]*windows.shape[4]*windows.shape[5])
patches[j,:] = patch
if j % 100 == 0:
print "Extracted {0}th patch of {1} patches totally".format(j, numPatches*size)
j += 1
data = self.next()
end = time.time()
print (end - start)
return patches
def __init__(self, pixelArray, blockSizeX, blockSizeY, overlap, cellSizeX, cellSizeY):
"""Constructor from an input array. Array should be 2x2 etc.
Args:
pixelArray: The array of pixels.
blockSizeX: The number of cells on the x axis of this Block. The
size in pixels is blockSizeX*cellSizeX
blockSizeY: The number of cells on the y axis of this Block. The
size in pixels is blockSizeY*cellSizeY
overlap: Fraction of overlap between Blocks.
cellSizeX: The number of pixels on the x axis of the Cells within
this Block.
cellSizeY: The number of pixels on the y axis of the Cells within
this Block.
"""
self._pixelArray = pixelArray
# if we want no overlap then overlap = 0 but need to alter so that
# step isn't zero and is the maximum of the blockSize
if overlap == 0:
step = max(blockSizeY, blockSizeX)
else:
# Calculate step from overlap and block size
# NB This will be calculated as the minimum of either the x or y steps
step_x = np.floor(overlap * blockSizeX)
step_y = np.floor(overlap * blockSizeY)
step = min(step_x, step_y)
block_shape = (blockSizeY, blockSizeX)
self.blockArray = view_as_windows(pixelArray, block_shape, step)
# Create cell object
self.cell = Cell(self.blockArray, cellSizeX, cellSizeY)
self.cellArray = self.cell.cellArray
patches = np.zeros((total_numPatches, rfSize*rfSize*3))
whitening = True
maxIter = 50
batchSize = 1000
j = 0
values = labels_dict.values()
for each in images:
if labels_dict[each.split('.')[0]] > 0:
numPatches = 140
else:
numPatches = 40
img = load_image('../data/resized/trainOriginal/' + each)
windows = view_as_windows(img, (rfSize, rfSize, 3))
r = np.random.randint(0, windows.shape[0] - windows.shape[3], (numPatches,))
c = np.random.randint(0, windows.shape[1]- windows.shape[4], (numPatches,))
for i in range(0, numPatches):
patch = np.reshape(windows[r[i],c[i],0,:], windows.shape[3]*windows.shape[4]*windows.shape[5])
patches[j,:] = patch
if j % 100 == 0:
print "Extracted {0}th patch of {1} patches totally".format(j,total_numPatches)
j += 1
numCentroids = int(np.sqrt(total_numPatches/2))
patchesMean = np.mean(patches, axis=1, dtype=np.float32, keepdims=True)
patchesVar = np.var(patches, axis=1, dtype=np.float32, keepdims=True)
offsetMatrix = 10.0 * np.ones(patchesVar.shape)
patches = (patches - patchesMean) / np.sqrt(patchesVar + offsetMatrix)
def windows(self, size, step=1):
'''windows are overlapping sub-images'''
size = list(size)
size.append(self.channels)
return [Image(i) for i in util.view_as_windows(self, size, step)]
def doConvolution(lX,ly=None,maxPixel=None,patchSize=5,n_components=15,stride=2,pad=1,blocksize=4,showEigenImages=False):
"""
Extract patches....
"""
#@TODO incl testset....
#http://scikit-learn.org/stable/auto_examples/cluster/plot_dict_face_patches.html
#http://scikit-learn.org/stable/modules/preprocessing.html
#http://nbviewer.ipython.org/github/dmrd/hackclass-sklearn/blob/master/image_classification.ipynb
#http://scikit-learn.org/stable/auto_examples/applications/face_recognition.html
max_patches=None
whiten_pca=True
pooling=True
#data_type='float32'
print "Image width:",maxPixel
print "Image height:",maxPixel
print "Pad :",pad
print "Patch size :",patchSize
print "Stride :",stride
print "PCA components:",n_components
print "Pooling:",pooling
print "Pooling patches blocksize:",blocksize
#npatches = (maxPixel - patchSize +1+2*pad)*(maxPixel - patchSize +1+2*pad)/stride
npatchesx = (maxPixel+2*pad - patchSize)/stride +1
npatches = npatchesx * npatchesx
print "number of patches:",npatches
print "Original shape:",lX.shape, " size (MB):",float(lX.values.nbytes)/1.0E6, " dtype:",lX.values.dtype
tmpy=[]
t0 = time()
#data = np.zeros((lX.shape[0],npatches,patchSize*patchSize),dtype=np.float32)
data = np.zeros((lX.shape[0],npatches,patchSize*patchSize),dtype=np.uint8)
#Online...!?
for i,img in enumerate(lX.values):
img = np.reshape(img*255, (maxPixel, maxPixel)).astype('uint8')
#standardize patch
#img = (img -img.mean())/np.sqrt(img.var()+0.001)
#img = np.reshape(img, (maxPixel, maxPixel)).astype('float32')
if pad:
img = padImage(img,pad=pad)
patches = view_as_windows(img,(patchSize, patchSize),stride)
patches = np.reshape(patches, (patches.shape[0]*patches.shape[1],patchSize, patchSize))
patches = np.reshape(patches, (len(patches), -1))
data[i,:,:]=patches
#data.append(patches)
print("Extract patches done in %0.3fs" % (time() - t0))
data = np.reshape(data,(lX.shape[0]*npatches,patchSize*patchSize))
print "Extract patches, new shape:",data.shape, " size (MB):",float(data.nbytes)/1.0E6, " dtype:",data.dtype
#if whiten_pca:
# print "Intermediate PCA!"
# pca = RandomizedPCA(patchSize*patchSize, whiten=whiten_pca)
# data = pca.fit_transform(data)
#do a pca first
#pca = RandomizedPCA(n_components, whiten=whiten_pca)
#http://docs.scipy.org/doc/scipy/reference/generated/scipy.cluster.vq.kmeans.html
#http://ufldl.stanford.edu/wiki/index.php/Implementing_PCA/Whitening
#pca=TruncatedSVD(n_components=n_components, algorithm='randomized', n_iter=5, tol=0.0)
pca0 = MiniBatchKMeans(init='k-means++', n_clusters=n_components, n_init=5)
#pca = FastICA(n_components=n_components,whiten=whiten_pca)
pca0.fit(data)
#pca = MiniBatchSparsePCA(n_components=n_components,alpha=0.5,n_jobs=4)
#pca = MiniBatchDictionaryLearning(n_components=n_components,alpha=0.5,n_jobs=4)
#pca = SparseCoder(dictionary=pca0.cluster_centers_,transform_n_nonzero_coefs=n_components,transform_algorithm='lars',transform_alpha=None,n_jobs=4)
pca = SparseCoder(dictionary=pca0.cluster_centers_,transform_n_nonzero_coefs=n_components,transform_algorithm='omp',transform_alpha=None,n_jobs=4)
print pca
data = pca.fit_transform(data)
#transform data to 64bit!!!
#data = pca.fit(data[::4])
#data = pca.transform(data)
if isinstance(pca,RandomizedPCA):
print "PCA components shape:",pca.components_.shape
print("Explained variance",pca.explained_variance_ratio_.sum())
plt.plot(pca.explained_variance_ratio_)
plt.show()
if isinstance(pca,MiniBatchKMeans):
ck = pca.cluster_centers_
print "Cluster centers (aka dictionary D (ncomponents x nfeatures) S aka code (nsamples x ncomponents): SD=Xi):",ck.shape
if showEigenImages:
#eigenfaces = pca.components_.reshape((n_components, patchSize, patchSize))
eigenfaces = pca.cluster_centers_.reshape((n_components, patchSize, patchSize))
#eigenfaces = pca.components_
plt.figure(figsize=(4.2, 4))
for i, comp in enumerate(eigenfaces[:n_components]):
plt.subplot(n_components/3, 3, i + 1)
plt.imshow(comp, cmap=plt.cm.gray_r,
interpolation='nearest')
plt.xticks(())
plt.yticks(())
plt.suptitle('Dictionary learned from patches\n')
plt.subplots_adjust(0.08, 0.02, 0.92, 0.85, 0.08, 0.23)
plt.show()
#for img in eigenfaces:
# showImage(img,patchSize,fac=10,matrix=True)
#
#print "Eigen patches shape:",eigenfaces.shape
print "Extracted patches, after PCA shape:",data.shape, " size (MB):",float(data.nbytes)/1.0E6, " dtype:",data.dtype
##pooling
nsamples = data.shape[0]/npatches
print "number of samples",nsamples
data = np.reshape(data, (nsamples,npatches,-1))
print "New shape:",data.shape
masks = extractBlocks(npatches,blocksize=blocksize)
#print masks
arrays = np.split(data, nsamples)
tmp = []
for i,ar in enumerate(arrays):
ar = np.reshape(ar, (npatches,n_components))
if pooling:
#pool
pool = np.zeros((masks.shape[0],n_components))
for j,m in enumerate(masks):
ar_tmp = np.sum(ar[m],axis=0)#sum or mean does not matter
#ar_tmp = np.amax(ar[m],axis=0)
#ar_tmp = np.mean(ar[m],axis=0)
pool[j]=ar_tmp
pool = np.reshape(pool,(masks.shape[0]*n_components))
tmp.append(pool)
else:
ar = np.reshape(ar, (npatches*n_components))
tmp.append(ar)
#data = np.concatenate(tmp, axis=0)
data = np.asarray(tmp)
print "New shape after pooling:",data.shape, " size (MB):",float(data.nbytes)/1.0E6
lX = pd.DataFrame(data,dtype=np.float32)
print "datatype:",lX.dtypes[0]
if ly is not None:
return (lX,ly)
else:
return lX
params = suggestion.assignments
print params
w = int(params['filter_w']) # SIGOPT param (filter width in pixels)
s = int(params['slide_w']) # SIGOPT param (held fixed at 2)
q = params['sparse_p'] # SIGOPT param (percentile of active centroid distances)
n_clust = int(params['km_n_clusters']) # SIGOPT param (num of centroids to learn)
mesh_r = 2 # (what is the pooling res in pixels)
zca_eps = math.exp(params['zca_eps'])
# stack a random subset image patches, 125K
X_unlab_patches = []
random.seed(42)
print "Gathering examples..."
# Use subsample of 200K for k-means and covariance estimates
for i in random.sample(range(0, unlab_X.shape[2]), 200000):
patches = view_as_windows(unlab_X[:, :, i], (w, w), step=s)
re_shaped = numpy.reshape(patches, (patches.shape[0]*patches.shape[0], w * w))
# normalize the patches, per sample
re_shaped = preprocessing.scale(re_shaped, axis=1)
X_unlab_patches.append(re_shaped)
X_unlab_patches = numpy.vstack(X_unlab_patches)
# build whitening transform matrix
print "Fitting ZCA Whitening Transform..."
cov = LedoitWolf()
cov.fit(X_unlab_patches) # fit covariance estimate
D, U = numpy.linalg.eigh(cov.covariance_)
V = numpy.sqrt(numpy.linalg.inv(numpy.diag(D + zca_eps)))
Wh = numpy.dot(numpy.dot(U, V), U.T)
mu = numpy.mean(X_unlab_patches, axis=0)
X_unlab_patches = numpy.dot(X_unlab_patches-mu, Wh)
def generate_bounding_boxes(model, image, downscale, step, min_height, min_width, file_name, grd_truth_path,
trainingFlag, min_prob, result_file, fp_name, overlapThresh=0.5):
from skimage.util import view_as_windows
import numpy as np
from theano import tensor as T
from theano import function
arr_tensor4 = T.tensor4('arr', dtype='float32')
arr_shuffler = arr_tensor4.dimshuffle((1, 0, 2, 3))
shuffle_function = function([arr_tensor4], arr_shuffler)
boxes = []
pyramid_list, scale_list = pyramid(image, downscale, min_height, min_width)
total_window_list = []
xy_num_list = []
for i in xrange(0, len(pyramid_list)):
window_channels = []
for channels in xrange(0, len(pyramid_list[i])):
window = view_as_windows(pyramid_list[i][channels], window_shape=(min_height, min_width), step=step)
# window.shape[0]= number of columns of windows, window.shape[1]= the number of rows of windows
xy_num_list.append([window.shape[1], window.shape[0]])
# min_height & min_width info. should be discarded later for memory saving
window_reshaped = np.reshape(window, newshape=(window.shape[0] * window.shape[1], min_height, min_width))
# position of window must be calculated here : window_pos_list
window_channels.append(window_reshaped)
# window_list for one pyramid picture
window_list = np.asarray(window_channels)
# (3, n, H, W) to (n, 3, H, W)
window_list = shuffle_function(window_list)
total_window_list.extend(window_list)
total_window_list = np.asarray(total_window_list)
total_window_pos_list = cal_window_position(scale_list, xy_num_list, min_height, min_width, step)
# classification
CNN_box_list, CNN_box_pos_list, CNN_prob_list = classify_windows_with_CNN(model, total_window_list,
total_window_pos_list, accuracy=min_prob)
# NMS (overlap threshold can be modified)
sup_box_list, sup_box_pos_list, sup_box_prob_list = non_max_suppression_fast(CNN_box_list, CNN_box_pos_list,
CNN_prob_list, overlapThresh=0.6)
# temporary code
print "Suppressed box list (length: " + str(len(sup_box_pos_list)) + ")"
for i in range(len(sup_box_pos_list)):
print "box #" + str(i + 1) + " pos: " + str(sup_box_pos_list[i]) + " prob: " + str(sup_box_prob_list[i])
# save suppressed box list to "res" file format
# format : [frame, x_start, y_start, x_delta, y_delta, prob]
# annotation file name is 0 based, res frame is 1 based
# ex) I00000.txt -> frame #1
# ad-hoc frame naming : take [1:6] (5 digits) and cast as int
if trainingFlag == 0:
frame_idx = int(file_name[1:6]) + 1
for idx in xrange(0, len(sup_box_pos_list)):
x_start = sup_box_pos_list[idx][0]
x_delta = sup_box_pos_list[idx][2] - x_start
y_start = sup_box_pos_list[idx][1]
y_delta = sup_box_pos_list[idx][3] - y_start
prob = sup_box_prob_list[idx]
line = str(frame_idx)+','+str(x_start)+','+str(y_start)+\
','+str(x_delta)+','+str(y_delta)+','+str(prob)+'\n'
result_file.write(line)
ground_truth, ground_truth_with_ignore = extract_Daimler_ground_truth(file_name, grd_truth_path, trainingFlag)
# temporary code
# ground_truth=[[1,1,10,10]]
count_tp = extract_fp_examples(file_name, ground_truth, sup_box_list, sup_box_pos_list, sup_box_prob_list, fp_name=fp_name, including_ignore=0,
accThresh=0.5)
count_tp_with_ignore = extract_fp_examples(file_name, ground_truth_with_ignore, sup_box_list, sup_box_pos_list, sup_box_prob_list, fp_name=fp_name,
including_ignore=1, accThresh=0.5)
draw_rectangle(sup_box_pos_list, sup_box_prob_list, image, file_name, fp_name)
if len(sup_box_pos_list) != 0:
fppi = len(sup_box_pos_list) - count_tp_with_ignore
else:
fppi = 0
if len(ground_truth) != 0:
miss_rate_per_image = 1 - float(count_tp) / len(ground_truth)
else:
miss_rate_per_image = 0
# temporary return
return fppi, miss_rate_per_image
return matrix
window_size = 15
img_l = io.imread('HW3_images/scene_l.bmp')
img_r = io.imread('HW3_images/scene_r.bmp')
# Integer division
pad_width = window_size / 2
window_shape = (window_size, window_size)
img_l_padded = pad(img_l, pad_width=pad_width, mode='symmetric')
img_r_padded = pad(img_r, pad_width=pad_width, mode='symmetric')
img_l_view = view_as_windows(img_l_padded, window_shape=window_shape, step=1)
img_r_view = view_as_windows(img_r_padded, window_shape=window_shape, step=1)
img_height, img_width = img_l.shape
result = np.zeros(img_l.shape)
for row in xrange(img_height):
for col in xrange(img_width):
first_patch = img_l_view[row, col]
first_patch_norm = normalize_matrix(first_patch)
current_scanline = img_r_view[row, :]
scores = np.zeros(img_width)
def extract_features_img(path, centroids, rfSize, M, U, stride, normal_pooling=True):
"""
:param path: path to RGB retina image
:param centroids: learned centroids
:param rfSize: receptive field size
:param M: whitening parameter
:param P: whitening parameter
:param stride: parameter that defines the density of windows that are extracted from an image
:param normal_pooling: if true:
divide in 4 regions and pool each one
else: divide in 16 regions and pool each one
:return:feature_vector
"""
img = load(path)
try:
img = img[:,:,1]
except:
return None
#contrast enhancing
img = equalize_adapthist(img)
numFeats = img.shape[0] * img.shape[1]
numCentroids = centroids.shape[0]
#extract dense patches with predefined stride
#smaller the stride, slower the function
windows = view_as_windows(img, (rfSize, rfSize), stride)
patches = np.reshape(windows, (windows.shape[0]*windows.shape[1], rfSize*rfSize))
#data normalization
p_mean = np.mean(patches, axis=1, dtype=np.float32, keepdims=True)
p_var = np.var(patches, axis=1, dtype=np.float32, ddof=1, keepdims=True)
off_matrix = 10.0 * np.ones(p_var.shape)
patches = (patches - p_mean) / np.sqrt(p_var + off_matrix)
patches = np.dot((patches - M), U)
#calculate distance from all patches to all centroids
z = native_cdist(patches, centroids)
#mean distance from each patch to all centroids
#triangle activation function
mu = np.tile(np.array([np.mean(z, axis = 1)]).T, (1, centroids.shape[0]))
patches = np.maximum(mu - z, np.zeros(mu.shape))
rows = (img.shape[0] - rfSize + stride)/stride
columns = (img.shape[1] - rfSize + stride)/stride
patches = np.reshape(patches, (rows, columns, numCentroids))
#starting points
#central point # of the patches "image"
halfr = np.round(float(rows)/2)
halfc = np.round(float(columns)/2)
#pool quadrants
if normal_pooling:
q1 = np.array([np.sum(np.sum(patches[0:halfc, 0:halfr, :], axis = 1),axis = 0)])
q2 = np.array([np.sum(np.sum(patches[halfc:patches.shape[0], 0:halfr, :], axis = 1),axis = 0)])
q3 = np.array([np.sum(np.sum(patches[0:halfc, halfr:patches.shape[1], :], axis = 1),axis = 0)])
q4 = np.array([np.sum(np.sum(patches[halfc:patches.shape[0], halfr:patches.shape[1], :], axis = 1),axis = 0)])
feature_vector = np.vstack((q1,q2,q3,q4)).flatten()
else:
quartr = np.round(float(rows)/4)
quartc = np.round(float(columns)/2)
q1 = pool_quadrant(patches, 0, 0, quartc, quartr, halfc, halfr)
q2 = pool_quadrant(patches, halfc, 0, quartc, quartr, patches.shape[0], halfr)
q3 = pool_quadrant(patches, 0, halfr, quartc, quartr, halfc, patches.shape[1])
q4 = pool_quadrant(patches, halfc, halfr, quartc, quartr, patches.shape[0], patches.shape[1])
feature_vector = np.vstack((q1, q2, q3, q4)).flatten()
return feature_vector
def extract_features(path, keys, centroids, rfSize, ImageDim, whitening, M, P, stride):
#assert(nargin == 4 || nargin == 6)
numCentroids = centroids.shape[0]
numSamples = len(keys)
numFeats = ImageDim[0]*ImageDim[1]*ImageDim[2]
# compute features for all training images
XC = np.zeros((numSamples, numCentroids*4))
labels = np.zeros((numSamples, 1))
j = 0
for i in range(numSamples):
total_start = time.time()
print "Sample {0}".format(i)
if (np.mod(i,2000) == 0):
np.save('test_features_windowsize_{0}_iteration_{1}'.format(rfSize,i), XC)
print 'Extracting features: ' + str(i) + '/' + str(numSamples)
img = load_image(path + keys[i])
# X = np.transpose(reader.next()[1]).flatten()
X = img.flatten()
# extract overlapping sub-patches into rows of 'patches'
start = time.time()
patches = np.vstack(
(im2col(np.reshape(X[0:numFeats/3],ImageDim[0:2],'F'), (rfSize, rfSize), stride),
im2col(np.reshape(X[numFeats/3:numFeats*2/3],ImageDim[0:2],'F'), (rfSize, rfSize), stride), im2col(np.reshape(X[numFeats*2/3:numFeats],ImageDim[0:2],'F'), (rfSize, rfSize), stride))).T
end = time.time()
w = view_as_windows(img, (rfSize,rfsize, 3), stride)
w = w.reshape((w.shape[0]*w.shape[1],w.shape[3]*w.shape[4]*w.shape[5]))
print np.array_equal(patches, w)
from time import sleep
for i in xrange(w.shape[0]):
print w[i,0:20]
print patches[i,500:520]
sleep(1)
print "Extract overlapping sub-patches time:{0}".format(end - start)
# do preprocessing for each patch
# normalize for contrast
start = time.time()
patchesMean = np.mean(patches, axis=1, dtype=np.float32, keepdims=True)
patchesVar = np.var(patches, axis=1, dtype=np.float32, ddof=1, keepdims=True)
offsetMatrix = 10.0 * np.ones(patchesVar.shape)
patches = (patches - patchesMean) / np.sqrt(patchesVar + offsetMatrix)
end = time.time()
print "Preprocessing time:{0}".format(end - start)
# whiten
if (whitening):
patches = np.dot((patches - M), P)
# compute 'triangle' activation function
start = time.time()
z = native_cdist(patches, centroids)
end = time.time()
print "Triangle time:{0}".format(end - start)
start = time.time()
mu = np.tile(np.array([np.mean(z, axis = 1)]).T, (1, centroids.shape[0])) # average distance to centroids for each patch
patches = np.maximum(mu - z, np.zeros(mu.shape))
end = time.time()
print "Distance calculation time:{0}".format(end - start)
# patches is now the data matrix of activations for each patch
# reshape to numCentroids-channel image
start = time.time()
prows = (ImageDim[0]-rfSize + 1*stride)/stride
pcols = (ImageDim[1]-rfSize + 1*stride)/stride
patches = np.reshape(patches, (prows, pcols, numCentroids),'F')
end = time.time()
print "Reshaping time:{0}".format(end - start)
start = time.time()
# pool over quadrants
halfr = np.round(float(prows)/2)
halfc = np.round(float(pcols)/2)
q1 = np.array([np.sum(np.sum(patches[0:halfc, 0:halfr, :], axis = 1),axis = 0)])
q2 = np.array([np.sum(np.sum(patches[halfc:patches.shape[0], 0:halfr, :], axis = 1),axis = 0)])
q3 = np.array([np.sum(np.sum(patches[0:halfc, halfr:patches.shape[1], :], axis = 1),axis = 0)])
q4 = np.array([np.sum(np.sum(patches[halfc:patches.shape[0], halfr:patches.shape[1], :], axis = 1),axis = 0)])
end = time.time()
print "Pooling time:{0}".format(end - start)
# concatenate into feature vector
XC[j,:] = np.vstack((q1,q2,q3,q4)).flatten()
j += 1
total_end = time.time()
print "Iteration time:{0}".format(total_end - total_start)
return XC
sv = svm.SVC(verbose = True)
f=np.load('C:/users/attialex/Desktop/training_data.npz')
X=f['X']
Y=f['Y']
print 'training the forest...'
sv.fit(X,Y)
print 'done'
im = io.imread('M:/attialex/images/lfm/bernhard/template_6_2.png')
imex = im[100:300,200:350]
winds = util.view_as_windows(imex,(40,40),step=1)
h = feature.hog(winds[0][0])
print 'testing...'
td = np.zeros((winds.shape[0]*winds.shape[1],len(h)))
counter = 0
print 'creating the feature vector'
for ii in range(0,winds.shape[0]):
for jj in range(0,winds.shape[1]):
td[counter,:]=feature.hog(winds[ii][jj])
counter+=1
print 'predicting'
probs = sv.predict(td)
def im2col(img, window_size, stride=1):
arr = view_as_windows(img, window_size)
arr = arr[::stride, ::stride, :, :]
arr = arr.reshape(np.prod(arr.shape[:-2]), window_size, window_size)
return arr