def padding(nframes, x, y):
# dirty hacky padding
ba = (nframes - 1) / 2 # before // after
x2 = copy.deepcopy(x)
on_x2 = False
x_f = np.zeros((x.shape[0], nframes * x.shape[1]), dtype='float64')
print 'x_f shape:',x_f.shape
for i in xrange(x.shape[0]):
if y[i] == '!ENTER[2]' and y[i-1] != '!ENTER[2]': # TODO general case
on_x2 = not on_x2
if on_x2:
x2[i - ba:i,:] = 0.0
else:
x[i - ba:i,:] = 0.0
if i+ba < y.shape[0] and '!EXIT' in y[i] and not '!EXIT' in y[i+ba]: # TODO general
if on_x2:
x2[i+ba:i+2*ba+1,:] = 0.0
else:
x[i+ba:i+2*ba+1,:] = 0.0
if on_x2:
x_f[i] = np.pad(x2[max(0, i - ba):i + ba + 1].flatten(),
(max(0, (ba - i) * x.shape[1]),
max(0, ((i+ba+1) - x.shape[0]) * x.shape[1])),
'constant', constant_values=(0,0))
else:
x_f[i] = np.pad(x[max(0, i - ba):i + ba + 1].flatten(),
(max(0, (ba - i) * x.shape[1]),
max(0, ((i+ba+1) - x.shape[0]) * x.shape[1])),
'constant', constant_values=(0,0))
return x_f
def conv_backward_naive(dout, cache):
"""
A naive implementation of the backward pass for a convolutional layer.
Inputs:
- dout: Upstream derivatives.
- cache: A tuple of (x, w, b, conv_param) as in conv_forward_naive
Returns a tuple of:
- dx: Gradient with respect to x
- dw: Gradient with respect to w
- db: Gradient with respect to b
"""
dx, dw, db = None, None, None
x, w, b, conv_param = cache
stride = conv_param['stride']
pad = conv_param['pad']
N, C, H, W = x.shape
F, _, HH, WW = w.shape
Hp = 1 + (H + 2 * pad - HH) / stride
Wp = 1 + (W + 2 * pad - WW) / stride
dx = np.zeros(x.shape)
dw = np.zeros(w.shape)
db = np.zeros(b.shape)
for i in xrange(N):
# for j in xrange(F):
data = x[i]
data = np.pad(data, ((0, 0), (pad, pad), (pad, pad)), 'constant')
paded_dxi = np.pad(dx[i], ((0, 0), (pad, pad), (pad, pad)), 'constant')
filter_vert_indices = 0
filter_hori_indices = 0
for s in xrange(Hp):
filter_hori_indices = 0
for p in xrange(Wp):
data_fragment = data[:, filter_vert_indices:filter_vert_indices+HH,
filter_hori_indices:filter_hori_indices+WW]
dw += np.einsum('i, jkl->ijkl', dout[i, :, s, p], data_fragment)
# paded_dxi[:, filter_vert_indices:filter_vert_indices+HH,
# filter_hori_indices:filter_hori_indices+WW] = \
# np.einsum('ijkl,i->jkl', w, dout[i, :, s, p])
# paded_dxi[:, filter_vert_indices:filter_vert_indices+HH,
# filter_hori_indices:filter_hori_indices+WW] = \
# np.tensordot(w, dout[i, :, s, p], axes = ([0], [0]))
for f in xrange(F):
paded_dxi[:, filter_vert_indices:filter_vert_indices+HH,
filter_hori_indices:filter_hori_indices+WW] \
+= w[f] * dout[i, f, s, p]
filter_hori_indices += stride
filter_vert_indices += stride
dx[i] = paded_dxi[:, pad:-pad, pad:-pad]
db = np.einsum('ijkl->j', dout)
# print(dx)
#############################################################################
# TODO: Implement the convolutional backward pass. #
#############################################################################
#############################################################################
# END OF YOUR CODE #
#############################################################################
return dx, dw, db
def _2d_filter(mat, win2d, matsign, pad):
"""
Filtering an image using a 2D window.
Parameters
----------
mat : 2D array of floats
nrow : int
Height of the window.
ncol: int
Width of the window.
sigma: tuple of 2 floats
Sigmas of the window.
pad : int
Padding.
Returns
-------
Filtered image.
"""
matpad = np.pad(mat, ((0, 0), (pad, pad)), mode='edge')
matpad = np.pad(matpad, ((pad, pad), (0, 0)), mode='mean')
(nrow, ncol) = matpad.shape
matfilter = np.real(ifft2(fft2(matpad * matsign) * win2d) * matsign)
return matfilter[pad:nrow - pad, pad:ncol - pad]
def eval(self, data, label, lens):
predictions = []
vals = []
for i in range(data.shape[0]/self.batch_size):
D = data[range(self.batch_size*i,self.batch_size*(i+1))]
L = label[range(self.batch_size*i,self.batch_size*(i+1))]
if lens is not None:
l = lens[range(self.batch_size*i,self.batch_size*(i+1))]
feed_dict={self.dataset:D, self.labels:L, self.lengths:l}
else:
feed_dict={self.dataset:D, self.labels:L}
predictions.extend(self.sess.run(self.correct_prediction, feed_dict))
vals.extend(self.sess.run(tf.argmax(self.logits,1), feed_dict))
## DO THE EXTRA
last_chunk = self.batch_size*(i+1)
gap = self.batch_size - (data.shape[0] - last_chunk)
D = np.pad(data[last_chunk:], ((0,gap),(0,0)), mode='constant', constant_values=0)
L = np.pad(label[last_chunk:], ((0,gap),(0,0)), mode='constant', constant_values=0)
if lens is not None:
l = np.pad(lens[last_chunk:], (0,gap), mode='constant', constant_values=0)
feed_dict={self.dataset:D, self.labels:L, self.lengths:l}
else:
feed_dict={self.dataset:D, self.labels:L}
predictions.extend(self.sess.run(self.correct_prediction, feed_dict)[:self.batch_size - gap])
vals.extend(self.sess.run(tf.argmax(self.logits,1), feed_dict)[:self.batch_size - gap])
print vals
## PRINT THE PREDICTONS
return 100.0*sum(predictions)/len(predictions)
def padding(nframes, x, y):
""" Dirty hacky padding for a minimum of nframes """
b_a = (nframes - 1) / 2 # before // after
x_2 = copy.deepcopy(x)
on_x_2 = False
x_f = zeros((x.shape[0], nframes * x.shape[1]), dtype='float32')
for i in xrange(x.shape[0]):
if y[i] == '!ENTER[2]' and y[i-1] != '!ENTER[2]': # TODO general case
on_x_2 = not on_x_2
if on_x_2:
x_2[i - b_a:i, :] = 0.0
else:
x[i - b_a:i, :] = 0.0
if i+b_a < y.shape[0] and '!EXIT' in y[i] and not '!EXIT' in y[i+b_a]:
# TODO general case
if on_x_2:
x_2[i+b_a:i+2*b_a+1, :] = 0.0
else:
x[i+b_a:i+2*b_a+1, :] = 0.0
if on_x_2:
x_f[i] = pad(x_2[max(0, i - b_a):i + b_a + 1].flatten(),
(max(0, (b_a - i) * x.shape[1]),
max(0, ((i+b_a+1) - x.shape[0]) * x.shape[1])),
'constant', constant_values=(0, 0))
else:
x_f[i] = pad(x[max(0, i - b_a):i + b_a + 1].flatten(),
(max(0, (b_a - i) * x.shape[1]),
max(0, ((i+b_a+1) - x.shape[0]) * x.shape[1])),
'constant', constant_values=(0, 0))
return x_f
def test_offset():
X,Y=np.mgrid[-5:5:0.05,-5:5:0.05]
Z=np.sqrt(X**2+Y**2)+np.sin(X**2+Y**2)
Z2 = Z.copy()
for i in range(15):
dx, dy = lmi.find_offset(Z, Z2)
dx2, dy2 = lmi.find_offset(Z2, Z)
assert_array_equal([dx, dx2], [0,0])
assert_array_equal([dy, dy2], [i, -i])
Z2 = np.pad(Z2, ((0,0),(1,0)), mode='constant')[:, :-1]
Z2 = Z.copy()
for i in range(15):
dx, dy = lmi.find_offset(Z, Z2)
dx2, dy2 = lmi.find_offset(Z2, Z)
assert_array_equal([dx, dx2], [i,-i])
assert_array_equal([dy, dy2], [0,0])
Z2 = np.pad(Z2, ((1,0),(0,0)), mode='constant')[:-1, :]
Z2 = Z.copy()
for i in range(15):
dx, dy = lmi.find_offset(Z, Z2)
dx2, dy2 = lmi.find_offset(Z2, Z)
assert_array_equal([dx, dx2], [i,-i])
assert_array_equal([dy, dy2], [i, -i])
Z2 = np.pad(Z2, ((1,0),(1,0)), mode='constant')[:-1, :-1]
def SMeval(self, DWi, DU, Dlens, DWj, keep_predictions=False):
"""
Runs eval on dev/test data with the option to return predictions or performance
"""
predictions = []
for i in range(len(DWi)/self.batch_size):
batch_range = range(self.batch_size*i,self.batch_size*(i+1))
wi = DWi[batch_range]
wj = DWj[batch_range]
U = DU[batch_range]
lens = Dlens[batch_range]
feed_dict = {self.cur_world: wi, self.next_world: wj, self.inputs: U, self.lengths: lens}
if keep_predictions:
predictions.extend(self.sess.run(tf.argmax(self.logits,1), feed_dict))
else:
predictions.extend(self.sess.run(self.correct_prediction, feed_dict))
## Grab the extras
last_chunk = self.batch_size*(i+1)
gap = self.batch_size - (len(DWi) - last_chunk)
wi = np.pad(DWi[last_chunk:], ((0,gap),(0,0), (0,0), (0,0)), mode='constant', constant_values=0)
wj = np.pad(DWj[last_chunk:], ((0,gap),(0,0)), mode='constant', constant_values=0)
U = np.pad(DU[last_chunk:], ((0,gap),(0,0)), mode='constant', constant_values=0)
lens = np.pad(Dlens[last_chunk:], ((0,gap)), mode='constant', constant_values=0)
feed_dict = {self.cur_world: wi, self.next_world: wj, self.inputs: U, self.lengths: lens}
if keep_predictions:
predictions.extend(self.sess.run(tf.argmax(self.logits,1), feed_dict)[:self.batch_size - gap])
return predictions
else:
predictions.extend(self.sess.run(self.correct_prediction, feed_dict)[:self.batch_size - gap])
return 100.0*sum(predictions)/len(predictions)
def test_uninterpolated_nan_regions(boundary, normalize_kernel):
#8086
# Test NaN interpolation of contiguous NaN regions with kernels of size
# identical and greater than that of the region of NaN values.
# Test case: kernel.shape == NaN_region.shape
kernel = Gaussian2DKernel(1, 5, 5)
nan_centroid = np.full(kernel.shape, np.nan)
image = np.pad(nan_centroid, pad_width=kernel.shape[0]*2, mode='constant',
constant_values=1)
with pytest.warns(AstropyUserWarning,
match="nan_treatment='interpolate', however, NaN values detected "
"post convolution. A contiguous region of NaN values, larger "
"than the kernel size, are present in the input array. "
"Increase the kernel size to avoid this."):
result = convolve(image, kernel, boundary=boundary, nan_treatment='interpolate',
normalize_kernel=normalize_kernel)
assert(np.any(np.isnan(result)))
# Test case: kernel.shape > NaN_region.shape
nan_centroid = np.full((kernel.shape[0]-1, kernel.shape[1]-1), np.nan) # 1 smaller than kerenel
image = np.pad(nan_centroid, pad_width=kernel.shape[0]*2, mode='constant',
constant_values=1)
result = convolve(image, kernel, boundary=boundary, nan_treatment='interpolate',
normalize_kernel=normalize_kernel)
assert(~np.any(np.isnan(result))) # Note: negation
def tile_images_make_tiles(data, padsize=1, padval=0, width=None, highlights = None):
height,width = get_tiles_height_width(data.shape[0], desired_width = width)
# Old one-way padding, no highlights
#padding = ((0, width*height - data.shape[0]), (0, padsize), (0, padsize)) + ((0, 0),) * (data.ndim - 3)
#data = np.pad(data, padding, mode='constant', constant_values=(padval, padval))
# New two-way padding with highlights
#padding = ((0, width*height - data.shape[0]), (padsize, padsize), (padsize, padsize)) + ((0, 0),) * (data.ndim - 3)
#print 'tile_images: data min,max =', data.min(), data.max()
#padder = SmartPadder()
##data = np.pad(data, padding, mode=jy_pad_fn)
#data = np.pad(data, padding, mode=padder.pad_function)
#print 'padder.calls =', padder.calls
# New new way, two-way padding with highlights
if highlights is not None:
assert len(highlights) == data.shape[0]
padding = ((0, width*height - data.shape[0]), (padsize, padsize), (padsize, padsize)) + ((0, 0),) * (data.ndim - 3)
# First pad with constant vals
try:
len(padval)
except:
padval = tuple((padval,))
assert len(padval) in (1,3), 'padval should be grayscale (len 1) or color (len 3)'
if len(padval) == 1:
data = np.pad(data, padding, mode='constant', constant_values=(padval, padval))
else:
data = np.pad(data, padding, mode='constant', constant_values=(0, 0))
for cc in (0,1,2):
# Replace 0s with proper color in each channel
data[:padding[0][0], :, :, cc] = padval[cc]
if padding[0][1] > 0:
data[-padding[0][1]:, :, :, cc] = padval[cc]
data[:, :padding[1][0], :, cc] = padval[cc]
if padding[1][1] > 0:
data[:, -padding[1][1]:, :, cc] = padval[cc]
data[:, :, :padding[2][0], cc] = padval[cc]
if padding[2][1] > 0:
data[:, :, -padding[2][1]:, cc] = padval[cc]
if highlights is not None:
# Then highlight if necessary
for ii,highlight in enumerate(highlights):
if highlight is not None:
data[ii,:padding[1][0],:,:] = highlight
if padding[1][1] > 0:
data[ii,-padding[1][1]:,:,:] = highlight
data[ii,:,:padding[2][0],:] = highlight
if padding[2][1] > 0:
data[ii,:,-padding[2][1]:,:] = highlight
# tile the filters into an image
data = data.reshape((height, width) + data.shape[1:]).transpose((0, 2, 1, 3) + tuple(range(4, data.ndim + 1)))
data = data.reshape((height * data.shape[1], width * data.shape[3]) + data.shape[4:])
data = data[0:-padsize, 0:-padsize] # remove excess padding
return (height,width), data
def pad(x, nf, ma=0):
""" pad x for nf frames with margin ma. """
ba = (nf - 1) / 2 # before/after
if ma:
ret = numpy.zeros((x.shape[0] - 2 * ma, x.shape[1] * nf), dtype=theano.config.floatX)
if ba <= ma:
for j in xrange(ret.shape[0]):
ret[j] = x[j : j + 2 * ma + 1].flatten()
else:
for j in xrange(ret.shape[0]):
ret[j] = numpy.pad(
x[max(0, j - ba) : j + ba + 1].flatten(),
(max(0, (ba - j) * x.shape[1]), max(0, ((j + ba + 1) - x.shape[0]) * x.shape[1])),
"constant",
constant_values=(0, 0),
)
return ret
else:
ret = numpy.zeros((x.shape[0], x.shape[1] * nf), dtype=theano.config.floatX)
for j in xrange(x.shape[0]):
ret[j] = numpy.pad(
x[max(0, j - ba) : j + ba + 1].flatten(),
(max(0, (ba - j) * x.shape[1]), max(0, ((j + ba + 1) - x.shape[0]) * x.shape[1])),
"constant",
constant_values=(0, 0),
)
return ret
def paddingAnswers(answerSheet1, blankSheet1):
numRowsA, numColsA, numBandsA, dataTypeA = ipcv.dimensions(answerSheet1)
numRowsB, numColsB, numBandsB, dataTypeB = ipcv.dimensions(blankSheet1)
print numRowsB, numColsB
if numBandsA == 3:
answerSheet = cv2.cvtColor(answerSheet1, cv.CV_BGR2GRAY)
elif numBandsA == 1:
answerSheet = answerSheet1
if numBandsB == 3:
blankSheet = cv2.cvtColor(blankSheet1, cv.CV_BGR2GRAY)
elif numBandsB == 1:
blankSheet = blankSheet1
pad = numpy.absolute(numRowsA - numColsA)/2.0
maxCount = numpy.max(blankSheet)
if (numRowsA-numColsA) % 2 != 0:
answerSheet = numpy.pad(answerSheet, ((0,0),(pad,pad+1)), 'constant', constant_values=((maxCount, maxCount),(maxCount,maxCount)))
elif (numRowsA-numColsA) % 2 == 0:
answerSheet = numpy.pad(answerSheet, ((0,0),(pad,pad)), 'constant', constant_values=((maxCount, maxCount),(maxCount,maxCount)))
pad1 = numpy.absolute(numRowsB - numColsB)/2.0
maxCount = numpy.max(blankSheet)
if (numRowsB-numColsB) % 2 != 0:
blankSheet = numpy.pad(blankSheet, ((0,0),(pad1,pad1+1)), 'constant', constant_values=((maxCount, maxCount),(maxCount,maxCount)))
elif (numRowsA-numColsA) % 2 == 0:
blankSheet = numpy.pad(blankSheet, ((0,0),(pad1,pad1)), 'constant', constant_values=((maxCount, maxCount),(maxCount,maxCount)))
return answerSheet, blankSheet
def qea(im):
H = ss.hilbert(im,axis = 2)
H = im+1j*H
ia = np.abs(H)
ip = np.angle(H)
h1col = H[1:-1,:,:]
h0col = H[:-2,:,:]
h2col = H[2:,:,:]
ifColSign = np.sign(np.real((h0col-h2col)/(2j*h1col)))
ifCol = np.arccos((h2col+h0col)/(2*h1col))
ifCol = (np.abs(ifCol)*ifColSign)/np.pi/2
ifCol = np.pad(ifCol,((1,1),(0,0),(0,0)), mode='reflect')
h0row = H[:,:-2,:]
h1row = H[:,1:-1,:]
h2row = H[:,2:,:]
#ifxSign = np.sign(np.real((h2x-h0x)/(2j*h1x)))
ifRow = np.arccos((h2row+h0row)/(2*h1row))
ifRow = (np.abs(ifRow))/np.pi/2
ifRow = np.pad(ifRow,((0,0),(1,1),(0,0)), mode='reflect')
h0time = H[:,:,:-2]
h1time = H[:,:,1:-1]
h2time = H[:,:,2:]
#ifxSign = np.sign(np.real((h2x-h0x)/(2j*h1x)))
ifTime = np.arccos((h2time+h0time)/(2*h1time))
ifTime = (np.abs(ifTime))/np.pi/2
ifTime = np.pad(ifTime,((0,0),(0,0),(1,1)), mode='reflect')
return(ia,ip,ifRow,ifCol,ifTime)
def deepflow2( im1=None, im2=None, match=None, options=""):
"""
flow = deepflow2.deepflow2(image1, image2, match=None, options='')
Compute the flow between two images, eventually using given matches.
Images must be HxWx3 numpy arrays (convert to float32).
Match is an optional numpy array argument (None by default, ie no input match), where each row starts by x1 y1 x2 y2.
Options is an optional string argument ('' by default), to set the options. Type deepflow2() to see the list of available options.
The function returns the optical flow as a HxWx2 numpy array."""
#convert images
if None in (im1,im2):
usage_python()
return
assert im1.shape == im2.shape, "images must have the same shape"
if im1.dtype != float32:
im1 = im1.astype(float32)
if im2.dtype != float32:
im2 = im2.astype(float32)
h, w, nchannels = im1.shape
assert nchannels==3, "images must have 3 channels"
stride = 4*((w+3)//4)
im1 = pad( rollaxis(im1,2), ((0,0),(0,0),(0, stride-w)), 'constant')
im2 = pad( rollaxis(im2,2), ((0,0),(0,0),(0, stride-w)), 'constant')
# allocate flow
flowx = empty((h,stride), dtype=float32)
flowy = empty((h,stride), dtype=float32)
# compute flow
if match is not None:
assert match.shape[1]>=4
match = ascontiguousarray(match[:,:4], dtype=float32)
deepflow2_numpy( w, flowx, flowy, im1, im2, match, options)
return concatenate ( (flowx[:,:w,None], flowy[:,:w,None]), axis=2)
def _zero_pad_to_same_size(a, b):
[ay, ax], [by, bx] = a.shape, b.shape
if ax < bx or ay < by:
a = np.pad(a, ( (by-ay,0),(bx-ax,0) ), mode='constant')
elif ax > bx or ay > by:
b = np.pad(b, ( (ay-by,0),(ax-bx,0) ), mode='constant')
return a, b, [ax-bx, ay-by]
def conv_backward_naive(dout, cache):
"""
A naive implementation of the backward pass for a convolutional layer.
Inputs:
- dout: Upstream derivatives.
- cache: A tuple of (x, w, b, conv_param) as in conv_forward_naive
Returns a tuple of:
- dx: Gradient with respect to x
- dw: Gradient with respect to w
- db: Gradient with respect to b
"""
# Setting up
x, w, b, conv_param = cache
dx = np.zeros_like(x)
dw = np.zeros_like(w)
db = np.zeros_like(b)
stride = conv_param['stride']
pad = conv_param['pad']
N, C, H, W = x.shape
F, _, HH, WW = w.shape
h_out = 1 + (H + 2 * pad - HH) / stride
w_out = 1 + (W + 2 * pad - WW) / stride
#############################################################################
# TODO: Implement the convolutional backward pass. #
#############################################################################
# Padding x and dx
x_padded = np.pad(x, [(0, 0), (0, 0), (pad, pad), (pad, pad)], mode='constant')
dx_padded = np.pad(dx, [(0, 0), (0, 0), (pad, pad), (pad, pad)], mode='constant')
# Do the convolutions
# For every image, pass it through the filter and update the output
for image in range(N):
for filter in range(F):
# Then, do the convolutions in over H and W
for height in xrange(h_out):
end_point_height = height * stride
for width in xrange(w_out):
end_point_width = width * stride
# Make the convolution window
conv_window = x_padded[image, :, end_point_height:end_point_height + HH,
end_point_width:end_point_width + WW]
# And update the derivatives
db[filter] += dout[image, filter, height, width]
dw[filter] += conv_window * dout[image, filter, height, width]
# Update DX at the convolution window
dx_padded[image, :, end_point_height:end_point_height + HH, end_point_width:end_point_width + WW] += \
w[filter] * \
dout[image, filter, height, width]
# And remove the padding
dx = dx_padded[:, :, pad:pad + H, pad:pad + W]
#############################################################################
# END OF YOUR CODE #
#############################################################################
return dx, dw, db
def _DataToInputs(self, spec, labels, weighted_labels, length, filename,
truncated_length):
# This method re-implements a portion of the TensorFlow graph using numpy.
# While typically it is frowned upon to test complicated code with other
# code, there is no way around this for testing the pipeline end to end,
# which requires an actual spec computation. Furthermore, much of the
# complexity of the pipeline is due to the TensorFlow implementation,
# so comparing it against simpler numpy code still provides effective
# coverage.
truncated_length = (min(truncated_length, length)
if truncated_length else length)
# Pad or slice spec if differs from truncated_length.
if len(spec) < truncated_length:
pad_amt = truncated_length - len(spec)
spec = np.pad(spec, [(0, pad_amt), (0, 0)], 'constant')
else:
spec = spec[0:truncated_length]
# Pad or slice labels if differs from truncated_length.
if len(labels) < truncated_length:
pad_amt = truncated_length - len(labels)
labels = np.pad(labels, [(0, pad_amt), (0, 0)], 'constant')
else:
labels = labels[0:truncated_length]
inputs = [[spec, labels, truncated_length, filename]]
return inputs
def zero_padding():
samples = 77
rec_period = 4
sampling_rate = samples/rec_period
time = np.linspace(0, rec_period, samples)
sin = np.sin(time*3.75*np.pi)
win = np.hanning(len(sin))
pad_count = 23
padded_sin = np.pad(sin, (0,pad_count), 'constant')
fft = np.fft.rfft(sin)
fft_padded = np.fft.rfft(padded_sin)
bins = (np.fft.rfftfreq(len(sin))*sampling_rate)[1:]
plt.subplot(321)
plt.plot(time, sin)
plt.subplot(322)
plt.plot(bins, (np.abs(fft))[1:]*2/samples, "o")
plt.subplot(323)
plt.plot(np.linspace(0, (samples+pad_count)*rec_period/float(samples), samples+pad_count), padded_sin)
plt.subplot(324)
plt.plot((np.fft.rfftfreq(len(padded_sin))*sampling_rate)[1:], (np.abs(fft_padded))[1:]*2/samples, "o")
plt.subplot(325)
padded_sin_win = np.pad(sin*win, (0, pad_count), 'constant')
plt.plot(np.linspace(0, (samples+pad_count)*rec_period/float(samples), samples+pad_count), padded_sin_win)
plt.subplot(326)
plt.plot((np.fft.rfftfreq(len(padded_sin_win))*sampling_rate)[1:], np.abs(np.fft.rfft(padded_sin_win))[1:]*2/samples, "o")
matplotlib2tikz.save( 'myfile.tikz' )
plt.show()
def do(self,s, sigma, aArray, cArray):
# center of filter (in both directions)
c = s / 2
# extend aArray and cArray
aaArray = np.pad(aArray, ((c, c), (c, c), (c, c)),'edge')
ccArray = np.pad(cArray, ((c, c), (c, c), (c, c)), 'constant', constant_values=(False))
c = float(c)
# define gaussian function
def gaussian(cx, cy, cz, w):
return lambda x,y,z: np.exp(-(((cx-x)/w)**2 + ((cy-y)/w)**2 + ((cz-z)/w)**2)/2)
# define gaussian filter
x,y,z = np.mgrid[0:s,0:s,0:s]
filt = gaussian(c,c,c,sigma)(x,y,z)
print filt
dArray = aaArray[:,:,:];
Ashape = aaArray.shape
c = int(c)
for i in range(c, Ashape[0] - c):
for j in range(c, Ashape[1] - c):
for k in range(c, Ashape[2] - c):
if ccArray[i][j][k] == False:
continue
tempA = aaArray[(i - c):(i + c + 1), (j - c):(j + c + 1), (k - c):(k + c + 1)]
tempC = ccArray[(i - c):(i + c + 1), (j - c):(j + c + 1), (k - c):(k + c + 1)]
mask = np.where(tempC)
dArray[i][j][k] = np.sum(tempA[mask]*filt[mask])/np.sum(filt[mask])
return dArray[c:-c, c:-c, c:-c]
def _extend_data_to_include(self, url, url2):
index_1, index_2 = self.get_indexes(url, url2)
missing = index_1 < 0 or index_2 < 0
if missing:
none_index = self.get_index(None)
if none_index >= len(self.known_urls) - 2:
#extend per 500 for performance
self.known_urls = self.known_urls + [None]*500
if index_1 < 0:
self.known_urls[none_index] = url
none_index += 1
if index_2 < 0:
self.known_urls[none_index] = url2
none_index += 1
size = len(self.known_urls)
padding = size - self.click_matrix.shape[0]
self.click_matrix = \
np.matrix(np.pad(self.click_matrix,
pad_width=([0,padding],
[0,padding]),
mode='constant'))
padding = size - len(self.spend_time)
self.spend_time = np.pad(self.spend_time, pad_width=(0, padding),
mode='constant')
def __init__(self,h5_path,image_paths,max_q=None,max_mc=None):
self.h5 = h5py.File(h5_path,mode='r')
self.image_ids = self.h5['image_ids'].value
self.questions = self.h5['questions'].value
self.multiple_choice = self.h5['multiple_choice'].value
self.answers = self.h5['answers'].value
self.bounding_boxes = dict((k,v) for (k,v) in zip(self.h5['img_list'].value,
self.h5['bounding_boxes'].value))
self.N = len(self.image_ids)
if max_q:
if max_q<self.questions.shape[1]:
self.questions = self.questions[:,:max_q]
else:
self.questions = np.pad(self.questions,
((0,0),(0,max_q-self.questions.shape[-1])),
'constant',constant_values=a_w2i['</s>'])
if max_mc:
if max_mc<self.multiple_choice.shape[-1]:
self.multiple_choice = self.multiple_choice[:,:,max_mc]
else:
self.multiple_choice = np.pad(self.multiple_choice,
((0,0),(0,0),(0,max_mc-self.multiple_choice.shape[-1])),
'constant',constant_values=a_w2i['</s>'])
self.max_q = self.questions.shape[1]
self.indexes = np.arange(self.N)
self.image_paths = image_paths
def get_input_matrices(batch_data):
X_prem, X_hypo, y = process(batch_data)
batch_size = len(X_prem)
# Maximum length of premise sentence
MAX_LENGTH_PREM = max([len(entry) for entry in X_prem])
# Maximum length of hypothesis sentence
MAX_LENGTH_HYPO = max([len(entry) for entry in X_hypo])
# Mask is used in Lasagne LSTM layer
X_prem_mask = np.zeros((batch_size, MAX_LENGTH_PREM))
X_hypo_mask = np.zeros((batch_size, MAX_LENGTH_HYPO))
for i in range(batch_size):
X_prem_mask[i, :len(X_prem[i])] = 1
X_prem[i] = np.pad(X_prem[i], [(0, MAX_LENGTH_PREM - len(X_prem[i])), (0, 0)], 'constant')
for i in range(batch_size):
X_hypo_mask[i, :len(X_hypo[i])] = 1
X_hypo[i] = np.pad(X_hypo[i], [(0, MAX_LENGTH_HYPO - len(X_hypo[i])), (0, 0)], 'constant')
X_prem = np.asarray(X_prem)
X_hypo = np.asarray(X_hypo)
y = np.asarray(y)
return X_prem, X_prem_mask, X_hypo, X_hypo_mask, y
def saveSlidingWindows((im_path,filter_size,step_size,out_file_pre,idx)):
print idx;
im=scipy.misc.imread(im_path);
pad_r=getPadTuple(im.shape[0],filter_size[0],step_size);
pad_c=getPadTuple(im.shape[1],filter_size[1],step_size);
if len(im.shape)>2:
im=np.pad(im,(pad_r,pad_c,(0,0)),'edge')
else:
im=np.pad(im,(pad_r,pad_c),'edge');
start_r=0;
idx_r=0;
out_files=[];
while start_r<im.shape[0]:
start_c=0;
idx_c=0;
while start_c<im.shape[1]:
end_r=start_r+filter_size[0];
end_c=start_c+filter_size[1];
crop_curr=im[start_r:end_r,start_c:end_c];
out_file_curr=out_file_pre+'_'+str(idx_r)+'_'+str(idx_c)+'.png';
scipy.misc.imsave(out_file_curr,crop_curr);
out_files.append(out_file_curr);
start_c=start_c+step_size;
idx_c+=1;
start_r=start_r+step_size;
idx_r+=1;
return out_files;
def local_pad(x): # TODO replace with pad global function
if nf <= 1:
return x
if self._margin:
ma = self._margin
ba = (nf - 1) / 2 # before/after
if x.shape[0] - 2*ma <= 0:
print "shape[0]:", x.shape[0]
print "ma:", ma
if x.shape[1] * nf <= 0:
print "shape[1]:", x.shape[1]
print "nf:", nf
ret = numpy.zeros((x.shape[0] - 2 * ma, x.shape[1] * nf),
dtype=theano.config.floatX)
if ba <= ma:
for j in xrange(ret.shape[0]):
ret[j] = x[j:j + 2*ma + 1].flatten()
else:
for j in xrange(ret.shape[0]):
ret[j] = numpy.pad(x[max(0, j - ba):j + ba +1].flatten(),
(max(0, (ba - j) * x.shape[1]),
max(0, ((j + ba + 1) - x.shape[0]) * x.shape[1])),
'constant', constant_values=(0, 0))
return ret
else:
ret = numpy.zeros((x.shape[0], x.shape[1] * nf),
dtype=theano.config.floatX)
ba = (nf - 1) / 2 # before/after
for j in xrange(x.shape[0]):
ret[j] = numpy.pad(x[max(0, j - ba):j + ba +1].flatten(),
(max(0, (ba - j) * x.shape[1]),
max(0, ((j + ba + 1) - x.shape[0]) * x.shape[1])),
'constant', constant_values=(0, 0))
return ret
def treatArray(data):
if border_mode == 'keep':
return data
if n_dim == 3:
sh = (data.shape[0], ) + data.shape[2:] # exclude channel (z,x,y)
else:
sh = data.shape[2:] # (x,y)
if border_mode == 'crop':
excess = map(lambda x: int((x[0] - x[1]) // 2), zip(sh, ps))
if n_dim == 3:
data = data[excess[0]:excess[0] + ps[0], :, excess[1]:excess[1] + ps[1], excess[2]:excess[2] + ps[2]]
elif n_dim == 2:
data = data[:, :, excess[0]:excess[0] + ps[0], excess[1]: excess[1] + ps[1]]
else:
excess_l = map(lambda x: int(np.ceil(float(x[0] - x[1]) / 2)), zip(ps, sh))
excess_r = map(lambda x: int(np.floor(float(x[0] - x[1]) / 2)), zip(ps, sh))
if n_dim == 3:
pad_with = [(excess_l[0], excess_r[0]), (0, 0), (excess_l[1], excess_r[1]), (excess_l[2], excess_r[2])]
else:
pad_with = [(0, 0), (0, 0), (excess_l[0], excess_r[0]), (excess_l[1], excess_r[1])]
if border_mode == 'mirror':
data = np.pad(data, pad_with, mode='symmetric')
if border_mode == '0-pad':
data = np.pad(data, pad_with, mode='constant', constant_values=0)
return data
def test_find_center_vo_with_downsampling(self):
sim = read_file('sinogram.npy')
np.pad(
sim, ((1000, 1000), (0, 0), (1000, 1000)),
mode="constant", constant_values=0)
cen = find_center_vo(sim)
assert_allclose(cen, 45.28, rtol=0.015)
def calcgrad(i):
#i=cv2.imread("images.png",0)
#i=[[1,2,3,4],[5,6,7,8],[9,10,11,12],[13,14,15,16]]
#i=np.array(i)
height,width=i.shape
first=np.pad(i,((0,0),(1,0)),'constant')
second=np.pad(i,((0,1),(1,0)),'constant')
third=np.pad(i,((0,1),(0,0)),'constant')
fourth=np.pad(i,((0,1),(0,1)),'constant')
first=first[:,0:width]
second=second[1:height+1,0:width]
third=third[1:height+1,:]
fourth=fourth[1:height+1,1:width+1]
first=i-first
second=i-second
third=i-third
fourth=i-fourth
combo1=32*np.array( first >= second, dtype=int)
combo2=16*np.array( first >= third, dtype=int)
combo3=8*np.array( first >= fourth, dtype=int)
combo4=4*np.array( second >= third, dtype=int)
combo5=2*np.array( second >= fourth, dtype=int)
combo6=np.array( third >= fourth, dtype=int)
ldgp=combo1+combo2+combo3+combo4+combo5+combo6
return ldgp
def __init__(self, G_list, max_num_nodes, features='id'):
self.max_num_nodes = max_num_nodes
self.adj_all = []
self.len_all = []
self.feature_all = []
for G in G_list:
adj = nx.to_numpy_matrix(G)
# the diagonal entries are 1 since they denote node probability
self.adj_all.append(
np.asarray(adj) + np.identity(G.number_of_nodes()))
self.len_all.append(G.number_of_nodes())
if features == 'id':
self.feature_all.append(np.identity(max_num_nodes))
elif features == 'deg':
degs = np.sum(np.array(adj), 1)
degs = np.expand_dims(np.pad(degs, [0, max_num_nodes - G.number_of_nodes()], 0),
axis=1)
self.feature_all.append(degs)
elif features == 'struct':
degs = np.sum(np.array(adj), 1)
degs = np.expand_dims(np.pad(degs, [0, max_num_nodes - G.number_of_nodes()],
'constant'),
axis=1)
clusterings = np.array(list(nx.clustering(G).values()))
clusterings = np.expand_dims(np.pad(clusterings,
[0, max_num_nodes - G.number_of_nodes()],
'constant'),
axis=1)
self.feature_all.append(np.hstack([degs, clusterings]))
def _fixOddKernel(kernel):
"""Take a kernel with odd dimensions and make them even for FFT
Parameters
----------
kernel : `numpy.array`
a numpy.array
Returns
-------
out : `numpy.array`
a fixed kernel numpy.array. Returns a copy if the dimensions needed to change;
otherwise just return the input kernel.
"""
# Note this works best for the FFT if we left-pad
out = kernel
changed = False
if (out.shape[0] % 2) == 1:
out = np.pad(out, ((1, 0), (0, 0)), mode='constant')
changed = True
if (out.shape[1] % 2) == 1:
out = np.pad(out, ((0, 0), (1, 0)), mode='constant')
changed = True
if changed:
out *= (np.mean(kernel) / np.mean(out)) # need to re-scale to same mean for FFT
return out
def cross_correlation(x, y, maxlag):
"""
Cross correlation with a maximum number of lags.
`x` and `y` must be one-dimensional numpy arrays with the same length.
This computes the same result as
numpy.correlate(x, y, mode='full')[len(a)-maxlag-1:len(a)+maxlag]
The return vaue has length 2*maxlag + 1.
Author: http://stackoverflow.com/questions/30677241
Warren Weckesser
"""
from numpy.lib.stride_tricks import as_strided
def _check_arg(x, xname):
x = np.asarray(x)
if x.ndim != 1:
raise ValueError('%s must be one-dimensional.' % xname)
return x
x = _check_arg(x, 'x')
y = _check_arg(y, 'y')
py = np.pad(y.conj(), 2*maxlag, mode='constant')
T = as_strided(py[2*maxlag:], shape=(2*maxlag+1, len(y) + 2*maxlag),
strides=(-py.strides[0], py.strides[0]))
px = np.pad(x, maxlag, mode='constant')
return T.dot(px)
def __init__(self,h5_path,image_paths,max_q=None,max_mc=None):
self.h5 = h5py.File(h5_path,mode='r')
self.image_ids = self.h5['image_ids'].value
self.questions = self.h5['questions'].value
self.multiple_choice = self.h5['multiple_choices'].value
self.answers = self.h5['ground_truth'].value
self.N = len(self.image_ids)
if max_q:
if max_q<self.questions.shape[1]:
self.questions = self.questions[:,:max_q]
else:
self.questions = np.pad(self.questions,
((0,0),(0,max_q-self.questions.shape[-1])),
'constant',constant_values=w2i['</s>'])
if max_mc:
if max_mc<self.multiple_choice.shape[-1]:
self.multiple_choice = self.multiple_choice[:,:,max_mc]
else:
self.multiple_choice = np.pad(self.multiple_choice,
((0,0),(0,0),(0,max_mc-self.multiple_choice.shape[-1])),
'constant',constant_values=w2i['</s>'])
self.max_mc = self.multiple_choice.shape[1]
self.max_q = self.questions.shape[1]
self.indexes = np.arange(self.N)
self.image_paths = image_paths
if st.split('_')[0] == 'bk':
bk = get_bank_size(int(st.split('_')[1]))
st_id_list.append(bk[0])
st_list.append(bk)
else:
break
if len(st_list) == 1:
continue
real_data_id.append(st_id_list)
real_data.append(st_list)
start_id_list.append(st_id_list[0])
end_id_list.append(st_id_list[-1])
real_data_id0 = real_data_id.copy()
real_data_id = [x[:g_sequence_len] for x in real_data_id]
real_data_id1 = [np.pad(x, (0,g_sequence_len - len(x))) for x in real_data_id]
endtime = time.time(); dtime = endtime - starttime
print("\nTime for loading real world data：%.8s s" % dtime)
GENERATED_NUM = len(real_data_id1)
print('\nGENERATED_NUM,real_data_id1', GENERATED_NUM)
VOCAB_SIZE = len(x_info_ids)+1+10 # padding
print('\nVOCAB_SIZE:',VOCAB_SIZE)
print('real_vocab_size: ', len(x_info_ids))
starttime = time.time()
x_index = []
for i in range(len(x_ids)):
if x_ids[i] not in x_info_ids:
x_index.append(i)
def _pad(seq, max_len, constant_values=0):
return np.pad(seq, (0, max_len - len(seq)),
mode='constant', constant_values=constant_values)
_, model_path, img_in, img_out = sys.argv[0:4]
# For lib_maxout_theano_batch we can control batch size
# batch_size = 1024
# if len(sys.argv) > 4:
# batch_size = int(sys.argv[4])
# network = DeepNetwork(model_path, batch_size=batch_size)
network = DeepNetwork(model_path)
input_image = normalize_image_float(np.array(Image.open(img_in)))
nx, ny = input_image.shape
pad_by = network.pad_by
pad_image = np.pad(input_image, ((pad_by, pad_by), (pad_by, pad_by)), 'symmetric')
start_time = time.time()
output = network.apply_net(pad_image, perform_pad=False)
print 'Complete in {0:1.4f} seconds'.format(time.time() - start_time)
im = Image.fromarray(np.uint8(output * 255))
im.save(img_out)
print "Image saved."
import h5py
f = h5py.File(img_out.replace('.tif', '') + '.h5')
f['/probabilities'] = output
def __getitem__(self, idx):
inst = self.lineIndex[idx]
author = inst[0]
lines = inst[1]
batch = []
for line in lines:
if line >= len(self.authors[author]):
line = (line + 37) % len(self.authors[author])
img_path, gt, pad_above, pad_below = self.authors[author][line]
img = cv2.imread(img_path, 0) #read as grayscale
if img is None:
return None
if pad_above < 0:
img = img[-pad_above:, :]
pad_above = 0
if pad_below < 0:
img = img[:pad_below, :]
pad_below = 0
#if pad_above>0 or pad_below>0:
img = img = np.pad(img, ((pad_above, pad_below), (10, 10)),
'constant',
constant_values=255)
#we also pad a bit on the sides
#print('{}, {} {}'.format(img_path,pad_above,pad_below))
if img.shape[0] != self.img_height:
if img.shape[0] < self.img_height and not self.warning:
self.warning = True
print("WARNING: upsampling image to fit size")
percent = float(self.img_height) / img.shape[0]
if img.shape[1] * percent > self.max_width:
percent = self.max_width / img.shape[1]
img = cv2.resize(img, (0, 0),
fx=percent,
fy=percent,
interpolation=cv2.INTER_CUBIC)
if img.shape[0] < self.img_height:
diff = self.img_height - img.shape[0]
img = np.pad(img,
((diff // 2, diff // 2 + diff % 2), (0, 0)),
'constant',
constant_values=255)
if len(img.shape) == 2:
img = img[..., None]
if self.fg_masks_dir is not None:
fg_path = os.path.join(self.fg_masks_dir,
'{}_{}.png'.format(author, line))
fg_mask = cv2.imread(fg_path, 0)
fg_mask = fg_mask / 255
if fg_mask.shape != img[:, :, 0].shape:
print(
'Error, fg_mask ({}, {}) not the same size as image ({})'
.format(fg_path, fg_mask.shape, img[:, :, 0].shape))
th, fg_mask = cv2.threshold(
img, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
fg_mask = 255 - fg_mask
ele = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (9, 9))
fg_mask = cv2.dilate(fg_mask, ele)
fg_mask = fg_mask / 255
if self.augmentation is not None:
#img = augmentation.apply_random_color_rotation(img)
img = augmentation.apply_tensmeyer_brightness(img)
img = grid_distortion.warp_image(img)
if len(img.shape) == 2:
img = img[..., None]
img = img.astype(np.float32)
img = 1.0 - img / 128.0
if len(gt) == 0:
return None
gt_label = string_utils.str2label_single(gt, self.char_to_idx)
if self.styles:
style_i = self.npr.choice(len(self.styles[author][id]))
style = self.styles[author][id][style_i]
else:
style = None
name = img_path[img_path.rfind('/') + 1:img_path.rfind('.')]
spaced_label = None if self.spaced_by_name is None else self.spaced_by_name[
img_path]
if spaced_label is not None:
assert (spaced_label.shape[1] == 1)
toAppend = {
"image": img,
"gt": gt,
"style": style,
"gt_label": gt_label,
"spaced_label": spaced_label,
"name": name,
"center": self.center,
"author": author
}
if self.fg_masks_dir is not None:
toAppend['fg_mask'] = fg_mask
batch.append(toAppend)
#batch = [b for b in batch if b is not None]
#These all should be the same size or error
assert len(set([b['image'].shape[0] for b in batch])) == 1
assert len(set([b['image'].shape[2] for b in batch])) == 1
dim0 = batch[0]['image'].shape[0]
dim1 = max([b['image'].shape[1] for b in batch])
dim2 = batch[0]['image'].shape[2]
all_labels = []
label_lengths = []
if self.spaced_by_name is not None:
spaced_labels = []
else:
spaced_labels = None
max_spaced_len = 0
input_batch = np.full((len(batch), dim0, dim1, dim2),
PADDING_CONSTANT).astype(np.float32)
if self.fg_masks_dir is not None:
fg_masks = np.full((len(batch), dim0, dim1, 1),
0).astype(np.float32)
for i in range(len(batch)):
b_img = batch[i]['image']
toPad = (dim1 - b_img.shape[1])
if 'center' in batch[0] and batch[0]['center']:
toPad //= 2
else:
toPad = 0
input_batch[i, :, toPad:toPad + b_img.shape[1], :] = b_img
if self.fg_masks_dir is not None:
fg_masks[i, :, toPad:toPad + b_img.shape[1],
0] = batch[i]['fg_mask']
l = batch[i]['gt_label']
all_labels.append(l)
label_lengths.append(len(l))
if spaced_labels is not None:
sl = batch[i]['spaced_label']
spaced_labels.append(sl)
max_spaced_len = max(max_spaced_len, sl.shape[0])
#all_labels = np.concatenate(all_labels)
label_lengths = torch.IntTensor(label_lengths)
max_len = label_lengths.max()
all_labels = [
np.pad(l, ((0, max_len - l.shape[0]), ), 'constant')
for l in all_labels
]
all_labels = np.stack(all_labels, axis=1)
if self.spaced_by_name is not None:
spaced_labels = [
np.pad(l, ((0, max_spaced_len - l.shape[0]), (0, 0)),
'constant') for l in spaced_labels
]
ddd = spaced_labels
spaced_labels = np.concatenate(spaced_labels, axis=1)
spaced_labels = torch.from_numpy(spaced_labels)
assert (spaced_labels.size(1) == len(batch))
images = input_batch.transpose([0, 3, 1, 2])
images = torch.from_numpy(images)
labels = torch.from_numpy(all_labels.astype(np.int32))
#label_lengths = torch.from_numpy(label_lengths.astype(np.int32))
if self.fg_masks_dir is not None:
fg_masks = fg_masks.transpose([0, 3, 1, 2])
fg_masks = torch.from_numpy(fg_masks)
if batch[0]['style'] is not None:
styles = np.stack([b['style'] for b in batch], axis=0)
styles = torch.from_numpy(styles).float()
else:
styles = None
mask, top_and_bottom, center_line = makeMask(images, self.mask_post,
self.mask_random)
##DEBUG
#for i in range(5):
# mask2, top_and_bottom2 = makeMask(images,self.mask_post, self.mask_random)
# #extra_masks.append(mask2)
# mask2 = ((mask2[0,0]+1)/2).numpy().astype(np.uint8)*255
# cv2.imshow('mask{}'.format(i),mask2)
#mask = ((mask[0,0]+1)/2).numpy().astype(np.uint8)*255
#cv2.imshow('mask'.format(i),mask)
#cv2.waitKey()
toRet = {
"image": images,
"mask": mask,
"top_and_bottom": top_and_bottom,
"center_line": center_line,
"label": labels,
"style": styles,
"label_lengths": label_lengths,
"gt": [b['gt'] for b in batch],
"spaced_label": spaced_labels,
"name": [b['name'] for b in batch],
"author": [b['author'] for b in batch],
}
if self.fg_masks_dir is not None:
toRet['fg_mask'] = fg_masks
return toRet
def __init__(self, dirPath, split, config):
if 'split' in config:
split = config['split']
if split == 'test':
subdir = 'testdataset_ICDAR'
else:
subdir = 'training_WR'
self.img_height = config['img_height']
self.batch_size = config['a_batch_size']
self.max_width = config['max_width'] if 'max_width' in config else 1300
skip_pad = config['skip_pad'] if 'skip_pad' in config else False
#assert(config['batch_size']==1)
#with open(os.path.join(dirPath,'sets.json')) as f:
words_file = os.path.join(
dirPath, 'groundtruth_{}2009_pageNorm.txt'.format(split))
if not os.path.exists(words_file):
#create modified GT file with appropriate padding anotation
#the padding is so images by the same author have the same height (after normalization)
authors = defaultdict(list)
with open(
os.path.join(dirPath,
'groundtruth_{}2009.txt'.format(split))) as f:
word_list = f.readlines()
for line in word_list:
m = re.match('(lot_.+\/([^/]+)\/[^/]+.tiff) (.+)',
line.strip())
path = os.path.join(dirPath, subdir, m[1])
author = m[2]
gt = m[3]
authors[author].append((m[1], gt))
new_lines = []
for author, lines in authors.items():
above = []
below = []
n_lines = [] #to remove non-existent images
#we measure the number of pixels above and below the centerline in each image
for i, (path, gt) in enumerate(lines):
img_path = os.path.join(dirPath, subdir, path)
img = cv2.imread(img_path, 0) #read as grayscale
if img is None:
continue
img = 255 - img
img = cv2.blur(
img, (21, 21),
borderType=cv2.BORDER_CONSTANT) #borderValue=0)
#th,binarized = cv2.threshold(img,0,1,cv2.THRESH_BINARY+cv2.THRESH_OTSU)
#binarized = 1-binarized
img = img.sum(axis=1)
centerline = np.argmax(
img) #assume center is where most ink is (after blur)
assert (centerline > img.shape[0] * 0.1)
assert (centerline < img.shape[0] * 0.9)
above.append(centerline)
below.append(img.shape[0] - centerline)
n_lines.append((path, gt))
above_mean = np.mean(above)
above_std = np.std(above)
above_max = np.max(above)
below_mean = np.mean(below)
below_std = np.std(below)
below_max = np.max(below)
#we want to pad/crop to include most ascenders and descenders, but dont need all
goal_above = int(min(above_max, above_mean + above_std * 2))
goal_below = int(min(below_max, below_mean + below_std * 2))
for i, (path, gt) in enumerate(n_lines):
pad_above = goal_above - above[i]
pad_below = goal_below - below[i]
new_lines.append((path, gt, pad_above, pad_below))
with open(words_file, 'w') as f:
for path, gt, above, below in new_lines:
f.write('{} {} {} {}\n'.format(path, gt, above, below))
with open(words_file) as f:
word_list = f.readlines()
self.authors = defaultdict(list)
self.lineIndex = []
self.max_char_len = 0
for line in word_list:
m = re.match(
'(lot_.+\/([^/]+)\/[^/]+.tiff) ([^ ]+) (-?\d+) (-?\d+)',
line.strip())
path = os.path.join(dirPath, subdir, m[1])
author = m[2]
gt = m[3]
self.max_char_len = max(self.max_char_len, len(gt))
if skip_pad:
pad_above = 0
pad_below = 0
else:
pad_above = int(m[4])
pad_below = int(m[5])
self.authors[author].append((path, gt, pad_above, pad_below))
#self.lineIndex += [(author,i+authorLines) for i in range(len(lines))]
#minLines=99999
#for author,lines in self.authors.items():
#print('{} {}'.format(author,len(lines)))
#minLines = min(minLines,len(lines))
#maxCombs = int(nCr(minLines,self.batch_size)*1.2)
short = config['short'] if 'short' in config else False
for author, lines in self.authors.items():
#if split=='train':
# combs=list(itertools.combinations(list(range(len(lines))),self.batch_size))
# np.random.shuffle(combs)
# self.lineIndex += [(author,c) for c in combs[:maxCombs]]
#else:
for i in range(len(lines) // self.batch_size):
ls = []
for n in range(self.batch_size):
ls.append(self.batch_size * i + n)
inst = (author, ls)
self.lineIndex.append(inst)
if short and i >= short:
break
if short and i >= short:
continue
leftover = len(lines) % self.batch_size
fill = self.batch_size - leftover
last = []
for i in range(fill):
last.append(i % len(lines))
for i in range(leftover):
last.append(len(lines) - (1 + i))
self.lineIndex.append((author, last))
self.fg_masks_dir = config[
'fg_masks_dir'] if 'fg_masks_dir' in config else None
self.warning = False
if self.fg_masks_dir is not None:
if self.fg_masks_dir[-1] == '/':
self.fg_masks_dir = self.fg_masks_dir[:-1]
self.fg_masks_dir += '_{}'.format(self.max_width)
ensure_dir(self.fg_masks_dir)
for author, lines in self.lineIndex:
for line in lines:
img_path, gt, pad_above, pad_below = self.authors[author][
line]
fg_path = os.path.join(self.fg_masks_dir,
'{}_{}.png'.format(author, line))
if not os.path.exists(fg_path):
img = cv2.imread(img_path, 0) #read as grayscale
if img is None:
continue
if pad_above < 0:
img = img[-pad_above:, :]
pad_above = 0
if pad_below < 0:
img = img[:pad_below, :]
pad_below = 0
#if pad_above>0 or pad_below>0:
img = img = np.pad(img,
((pad_above, pad_below), (10, 10)),
'constant',
constant_values=255)
if img.shape[0] != self.img_height:
if img.shape[
0] < self.img_height and not self.warning:
self.warning = True
print("WARNING: upsampling image to fit size")
percent = float(self.img_height) / img.shape[0]
if img.shape[1] * percent > self.max_width:
percent = self.max_width / img.shape[1]
img = cv2.resize(img, (0, 0),
fx=percent,
fy=percent,
interpolation=cv2.INTER_CUBIC)
if img.shape[0] < self.img_height:
diff = self.img_height - img.shape[0]
img = np.pad(
img, ((diff // 2, diff // 2 + diff % 2),
(0, 0)),
'constant',
constant_values=255)
th, binarized = cv2.threshold(
img, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
binarized = 255 - binarized
ele = cv2.getStructuringElement(
cv2.MORPH_ELLIPSE, (9, 9))
binarized = cv2.dilate(binarized, ele)
cv2.imwrite(fg_path, binarized)
print('saved fg mask: {}'.format(fg_path))
char_set_path = config['char_file']
with open(char_set_path) as f:
char_set = json.load(f)
self.char_to_idx = char_set['char_to_idx']
self.augmentation = config[
'augmentation'] if 'augmentation' in config else None
#DEBUG
if 'overfit' in config and config['overfit']:
self.lineIndex = self.lineIndex[:10]
self.center = False #config['center_pad'] #if 'center_pad' in config else True
if 'style_loc' in config:
by_author_styles = defaultdict(list)
by_author_all_ids = defaultdict(set)
style_loc = config['style_loc']
if style_loc[-1] != '*':
style_loc += '*'
all_style_files = glob(style_loc)
assert (len(all_style_files) > 0)
for loc in all_style_files:
#print('loading '+loc)
with open(loc, 'rb') as f:
styles = pickle.load(f)
for i in range(len(styles['authors'])):
by_author_styles[styles['authors'][i]].append(
(styles['styles'][i], styles['ids'][i]))
by_author_all_ids[styles['authors'][i]].update(
styles['ids'][i])
self.styles = defaultdict(lambda: defaultdict(list))
for author in by_author_styles:
for id in by_author_all_ids[author]:
for style, ids in by_author_styles[author]:
if id not in ids:
self.styles[author][id].append(style)
for author in self.authors:
assert (author in self.styles)
else:
self.styles = None
if 'spaced_loc' in config:
with open(config['spaced_loc'], 'rb') as f:
self.spaced_by_name = pickle.load(f)
#for name,v in spaced_by_name.items():
# author, id = name.split('_')
else:
self.spaced_by_name = None
self.mask_post = config['mask_post'] if 'mask_post' in config else []
self.mask_random = config[
'mask_random'] if 'mask_random' in config else False
def __getitem__(self, index):
#---------
# Image
#---------
img_path = self.img_files[index % len(self.img_files)].rstrip()
img = np.array(Image.open(img_path))
# Handles images with less than three channels
while len(img.shape) != 3:
index += 1
img_path = self.img_files[index % len(self.img_files)].rstrip()
img = np.array(Image.open(img_path))
h, w, _ = img.shape
dim_diff = np.abs(h - w)
# Upper (left) and lower (right) padding
pad1, pad2 = dim_diff // 2, dim_diff - dim_diff // 2
# Determine padding
pad = ((pad1, pad2), (0, 0), (0, 0)) if h <= w else ((0, 0), (pad1, pad2), (0, 0))
# Add padding
input_img = np.pad(img, pad, 'constant', constant_values=128) / 255.
padded_h, padded_w, _ = input_img.shape
# Resize and normalize
input_img = resize(input_img, (*self.img_shape, 3), mode='reflect')
# Channels-first
input_img = np.transpose(input_img, (2, 0, 1))
# As pytorch tensor
input_img = torch.from_numpy(input_img).float()
#---------
# Label
#---------
label_path = self.label_files[index % len(self.img_files)].rstrip()
labels = None
if os.path.exists(label_path):
labels = np.loadtxt(label_path).reshape(-1, 5)
# Extract coordinates for unpadded + unscaled image
x1 = w * (labels[:, 1] - labels[:, 3]/2)
y1 = h * (labels[:, 2] - labels[:, 4]/2)
x2 = w * (labels[:, 1] + labels[:, 3]/2)
y2 = h * (labels[:, 2] + labels[:, 4]/2)
# Adjust for added padding
x1 += pad[1][0]
y1 += pad[0][0]
x2 += pad[1][0]
y2 += pad[0][0]
# Calculate ratios from coordinates
labels[:, 1] = ((x1 + x2) / 2) / padded_w
labels[:, 2] = ((y1 + y2) / 2) / padded_h
labels[:, 3] *= w / padded_w
labels[:, 4] *= h / padded_h
# Fill matrix
filled_labels = np.zeros((self.max_objects, 5))
if labels is not None:
filled_labels[range(len(labels))[:self.max_objects]] = labels[:self.max_objects]
filled_labels = torch.from_numpy(filled_labels)
return img_path, input_img, filled_labels
def bilateralfilter(image, texture, sigma_s, sigma_r):
r = int(np.ceil(3 * sigma_s))
# Image padding
# Symmetric padding : pads along the reflected mirror of edge of the array
# Ex: a [1,2,3,4,5]
# np.pad(a, (2,3), 'symmetric') means pad 2 elements of first axis edge and 3 elements of second axis edge
# result: [2,1,1,2,3,4,5,5,4,3]
# Pads first&second edges of each dimension with 3*sigma (r) for applying filter on borders
if image.ndim == 3:
h, w, ch = image.shape
I = np.pad(image, ((r, r), (r, r), (0, 0)),
'symmetric').astype(np.float32)
elif image.ndim == 2:
h, w = image.shape
I = np.pad(image, ((r, r), (r, r)), 'symmetric').astype(np.float32)
else:
print('Input image is not valid!')
return image
# Check texture size equals given image size then do padding
if texture.ndim == 3:
ht, wt, cht = texture.shape
# If texture shape is not equal to image shape, return
if ht != h or wt != w:
print('The guidance image is not aligned with input image!')
return image
# else pad texture
T = np.pad(texture, ((r, r), (r, r), (0, 0)),
'symmetric').astype(np.int32)
elif texture.ndim == 2:
ht, wt = texture.shape
if ht != h or wt != w:
print('The guidance image is not aligned with input image!')
return image
T = np.pad(texture, ((r, r), (r, r)), 'symmetric').astype(np.int32)
# Pre-compute
# Create np array of zeros with the same shape of the image
output = np.zeros_like(image)
# e^(- x / 2sigma^2)
scaleFactor_s = 1 / (2 * sigma_s * sigma_s)
scaleFactor_r = 1 / (2 * sigma_r * sigma_r)
# A lookup table for range kernel (COLOR)
LUT = np.exp(-np.arange(256) * np.arange(256) * scaleFactor_r)
# Generate a spatial Gaussian function (cutoff 6-sigma)
# -r for symmetric grid Ex: 0->6 becomes -3 -> 3
x, y = np.meshgrid(np.arange(2 * r + 1) - r, np.arange(2 * r + 1) - r)
# Create multi-variate gaussian distribution for spatial domain with x,y
kernel_s = np.exp(-(x * x + y * y) * scaleFactor_s)
# Main body
if I.ndim == 2 and T.ndim == 2: # I1T1 (2D Image, 2D Texture) filter
for y in range(r, r + h):
for x in range(r, r + w):
# Get gaussian values representing weights for the window
wgt = LUT[np.abs(T[y - r:y + r + 1, x - r:x + r + 1] -
T[y, x])] * kernel_s
# Calculate the intensity of the current pixel using the weighted gaussian values
# for j=-3sigma->3sigma sum(w(j) * I(j))/sum(w)
output[y - r, x - r] = np.sum(
wgt * I[y - r:y + r + 1, x - r:x + r + 1]) / np.sum(wgt)
elif I.ndim == 3 and T.ndim == 2: # I3T1 (3D Image, 2D Texture) filter
for y in range(r, r + h):
for x in range(r, r + w):
wgt = LUT[abs(T[y - r:y + r + 1, x - r:x + r + 1] -
T[y, x])] * kernel_s
wacc = np.sum(wgt)
output[y - r, x - r, 0] = np.sum(
wgt * I[y - r:y + r + 1, x - r:x + r + 1, 0]) / wacc
output[y - r, x - r, 1] = np.sum(
wgt * I[y - r:y + r + 1, x - r:x + r + 1, 1]) / wacc
output[y - r, x - r, 2] = np.sum(
wgt * I[y - r:y + r + 1, x - r:x + r + 1, 2]) / wacc
elif I.ndim == 3 and T.ndim == 3: # I3T3 (3D Image, 3D Texture) filter
for y in range(r, r + h):
for x in range(r, r + w):
# Product of 3 independent gaussians for each channel RGB
wgt = LUT[abs(T[y - r:y + r + 1, x - r:x + r + 1, 0] - T[y, x, 0])] * \
LUT[abs(T[y - r:y + r + 1, x - r:x + r + 1, 1] - T[y, x, 1])] * \
LUT[abs(T[y - r:y + r + 1, x - r:x + r + 1, 2] - T[y, x, 2])] * \
kernel_s
wacc = np.sum(wgt)
output[y - r, x - r, 0] = np.sum(
wgt * I[y - r:y + r + 1, x - r:x + r + 1, 0]) / wacc
output[y - r, x - r, 1] = np.sum(
wgt * I[y - r:y + r + 1, x - r:x + r + 1, 1]) / wacc
output[y - r, x - r, 2] = np.sum(
wgt * I[y - r:y + r + 1, x - r:x + r + 1, 2]) / wacc
elif I.ndim == 2 and T.ndim == 3: # I1T3 filter
for y in range(r, r + h):
for x in range(r, r + w):
wgt = LUT[abs(T[y - r:y + r + 1, x - r:x + r + 1, 0] - T[y, x, 0])] * \
LUT[abs(T[y - r:y + r + 1, x - r:x + r + 1, 1] - T[y, x, 1])] * \
LUT[abs(T[y - r:y + r + 1, x - r:x + r + 1, 2] - T[y, x, 2])] * \
kernel_s
output[y - r, x - r] = np.sum(
wgt * I[y - r:y + r + 1, x - r:x + r + 1]) / np.sum(wgt)
else:
print('Something wrong!')
return image
# return np.clip(output, 0, 255)
return output
#convert to numpy array
x1 = np.array(samples).astype("complex64")
# mix down with fc
fc1 = np.exp(-1.0j * 2.0 * np.pi * F_offset / Fs * np.arange(len(x1)))
x2 = x1 * fc1
f_bw = 200000
Inp1 = x2.real * np.cos(2.0 * np.pi * Fc * np.arange(len(x2)))
Qa1 = -1 * x2.imag * np.sin(2.0 * np.pi * Fc * np.arange(len(x2)))
Inp2 = np.convolve(Inp1, h)
Qa2 = np.convolve(Qa1, h)
delta_T = 155
Inp2_delay = np.pad(Inp2, (155, 0), 'constant', constant_values=(0, 0))
Qa2_delay = np.pad(Qa2, (155, 0), 'constant', constant_values=(0, 0))
Inp3 = np.pad(Inp2, (0, 155), 'constant', constant_values=(0, 0))
Qa3 = np.pad(Qa2, (0, 155), 'constant', constant_values=(0, 0))
D = Qa3 * Inp2_delay - Inp3 * Qa2_delay
y = []
d = []
k = 0
for x in D:
k += 1
if x > 0:
y.append(1)
else:
y.append(0)
if k == 30:
k = 0
def pad(img, padding, fill=0, padding_mode='constant'):
"""Pad the given PIL Image on all sides with speficified padding mode and fill value.
Args:
img (PIL Image): Image to be padded.
padding (int or tuple): Padding on each border. If a single int is provided this
is used to pad all borders. If tuple of length 2 is provided this is the padding
on left/right and top/bottom respectively. If a tuple of length 4 is provided
this is the padding for the left, top, right and bottom borders
respectively.
fill: Pixel fill value for constant fill. Default is 0. If a tuple of
length 3, it is used to fill R, G, B channels respectively.
This value is only used when the padding_mode is constant
padding_mode: Type of padding. Should be: constant, edge, reflect or symmetric. Default is constant.
constant: pads with a constant value, this value is specified with fill
edge: pads with the last value on the edge of the image
reflect: pads with reflection of image (without repeating the last value on the edge)
padding [1, 2, 3, 4] with 2 elements on both sides in reflect mode
will result in [3, 2, 1, 2, 3, 4, 3, 2]
symmetric: pads with reflection of image (repeating the last value on the edge)
padding [1, 2, 3, 4] with 2 elements on both sides in symmetric mode
will result in [2, 1, 1, 2, 3, 4, 4, 3]
square_constant: overrides padding to pad smallest edge with constant value,
specified with fill, to make a square image. 2D or 3D imgs only.
Returns:
PIL Image: Padded image.
"""
if not _is_pil_image(img):
raise TypeError('img should be PIL Image. Got {}'.format(type(img)))
if not isinstance(padding, (numbers.Number, tuple)):
raise TypeError('Got inappropriate padding arg')
if not isinstance(fill, (numbers.Number, str, tuple)):
raise TypeError('Got inappropriate fill arg')
if not isinstance(padding_mode, str):
raise TypeError('Got inappropriate padding_mode arg')
if isinstance(padding, collections.Sequence) and len(padding) not in [2, 4]:
raise ValueError("Padding must be an int or a 2, or 4 element tuple, not a " +
"{} element tuple".format(len(padding)))
assert padding_mode in ['constant', 'edge', 'reflect', 'symmetric', 'square_constant'], \
'Padding mode should be either constant, square_constant, edge, reflect or symmetric'
if padding_mode == 'constant':
return ImageOps.expand(img, border=padding, fill=fill)
elif padding_mode == 'square_constant':
if len(img.size) > 3 or len(img.size) < 2:
raise ValueError("padding_mode == 'square_constant' valid only for 2D or 3D images")
im_size = img.size if len(img.size) == 2 else img.size[:-1]
pad_amt, pad_dim = (np.max(im_size) - np.min(im_size))//2, np.argmin(im_size)
padding = np.zeros((2,), dtype=int)
padding[pad_dim] = pad_amt
return ImageOps.expand(img, border=tuple(padding), fill=fill)
else:
if isinstance(padding, int):
pad_left = pad_right = pad_top = pad_bottom = padding
if isinstance(padding, collections.Sequence) and len(padding) == 2:
pad_left = pad_right = padding[0]
pad_top = pad_bottom = padding[1]
if isinstance(padding, collections.Sequence) and len(padding) == 4:
pad_left = padding[0]
pad_top = padding[1]
pad_right = padding[2]
pad_bottom = padding[3]
img = np.asarray(img)
# RGB image
if len(img.shape) == 3:
img = np.pad(img, ((pad_top, pad_bottom), (pad_left, pad_right), (0, 0)), padding_mode)
# Grayscale image
if len(img.shape) == 2:
img = np.pad(img, ((pad_top, pad_bottom), (pad_left, pad_right)), padding_mode)
return Image.fromarray(img)
def apply(self, data):
lvl = np.array([0.1, 4, 8, 12, 30, 70, 180]) # Frequency levels in Hz
data = data.transpose()
data = np.pad(data,((0,self.splitsize -data.shape[0]),(0,0)),'constant')
nt, nc =data.shape
fs = 400
feat = []
if self.smooth:
data=resample(data,int(nt/fs*self.smooth_Hz))
nt, nc = data.shape
fs = self.smooth_Hz
lvl = np.array([0.1, 4, 8, 12, 30, 70])
if self.with_six or self.with_dy:
D = np.absolute(np.fft.rfft(data, axis=0))
D[0, :] = 0 # set the DC component to zero
for i in range(nc):
D[:,i] /= D[:,i].sum() # Normalize each channel
coorD = np.corrcoef(D.transpose())
w = Eigenvalues().apply(coorD)
tfreq = self.power_edge
ppow = 0.5
# top_freq = int(round(nt / sfreq * tfreq)) + 1
top = int(round(nt / fs * tfreq))
spedge = np.cumsum(D[:top, :], axis=0)
spedge = np.argmin(np.abs(spedge - (spedge.max(axis=0) * ppow)), axis=0)
spedge = spedge / top * tfreq
feat.append(w)
feat.append(spedge.ravel())
if self.with_six:
lseg = np.round(nt / fs * lvl).astype('int')
sixspect = np.zeros((len(lvl) - 1, nc))
for j in range(len(sixspect)):
sixspect[j, :] = 2 * np.sum(D[lseg[j]:lseg[j + 1], :], axis=0)
spentropy = -1 * np.sum(np.multiply(sixspect, np.log(sixspect)), axis=0)
feat.append(sixspect.ravel())
feat.append(spentropy.ravel())
if self.with_dy:
ldat = int(floor(nt / 2.0))
no_levels = int(floor(log(ldat, 2.0)))
dspect = np.zeros((no_levels, nc))
for j in range(no_levels - 1, -1, -1):
dspect[j, :] = 2 * np.sum(D[int(floor(ldat / 2.0)):ldat, :], axis=0)
ldat = int(floor(ldat / 2.0))
spentropyDyd = -1 * np.sum(np.multiply(dspect, np.log(dspect)), axis=0)
feat.append(dspect.ravel())
feat.append(spentropyDyd.ravel())
if self.with_mc:
mobility = np.divide(
np.std(np.diff(data, axis=0)),
np.std(data, axis=0))
complexity = np.divide(np.divide(
# std of second derivative for each channel
np.std(np.diff(np.diff(data, axis=0), axis=0), axis=0),
# std of second derivative for each channel
np.std(np.diff(data, axis=0), axis=0))
, mobility)
feat.append(mobility)
feat.append(complexity)
if self.with_time_corr:
data1 = TimeCorrelation(self.max_hz, self.scale_option).apply(data.transpose())
feat.append(data1)
if self.with_equal_freq:
data2 = FreqCorrelation(self.start, self.end, self.scale_option, self.resample_size, with_fft=True,
with_corr=True).apply(data.transpose())
feat.append(data2)
if self.onlyfd_dfa:
fd = np.zeros((2, nc))
for j in range(nc):
fd[0, j] = pyeeg.pfd(data[:, j])
fd[1, j] = pyeeg.hfd(data[:, j], 3)
DFA = np.zeros(nc)
for j in range(nc):
DFA[j] = pyeeg.dfa(data[:, j])
feat=np.concatenate((
fd.ravel(),
DFA.ravel(),
np.sqrt(DFA).ravel(),
np.square(DFA.ravel()),
np.sqrt(fd).ravel(),
np.square(fd).ravel(),
))
if self.with_square or self.with_log or self.with_sqrt:
tmp = np.concatenate(feat, axis=0)
tmp = np.absolute(tmp)
if self.with_square:
feat.append(np.square(tmp))
if self.with_log:
feat.append(np.log(tmp))
if self.with_sqrt:
feat.append(np.sqrt(tmp))
return np.concatenate(feat, axis=0)
def nms_fast(in_corners, H, W, dist_thresh):
"""
Run a faster approximate Non-Max-Suppression on numpy corners shaped:
3xN [x_i,y_i,conf_i]^T
Algo summary: Create a grid sized HxW. Assign each corner location a 1, rest
are zeros. Iterate through all the 1's and convert them either to -1 or 0.
Suppress points by setting nearby values to 0.
Grid Value Legend:
-1 : Kept.
0 : Empty or suppressed.
1 : To be processed (converted to either kept or supressed).
NOTE: The NMS first rounds points to integers, so NMS distance might not
be exactly dist_thresh. It also assumes points are within image boundaries.
Inputs
in_corners - 3xN numpy array with corners [x_i, y_i, confidence_i]^T.
H - Image height.
W - Image width.
dist_thresh - Distance to suppress, measured as an infinty norm distance.
Returns
nmsed_corners - 3xN numpy matrix with surviving corners.
nmsed_inds - N length numpy vector with surviving corner indices.
"""
grid = np.zeros((H, W)).astype(int) # Track NMS data.
inds = np.zeros((H, W)).astype(int) # Store indices of points.
# Sort by confidence and round to nearest int.
inds1 = np.argsort(-in_corners[2,:])
corners = in_corners[:,inds1]
rcorners = corners[:2,:].round().astype(int) # Rounded corners.
# Check for edge case of 0 or 1 corners.
if rcorners.shape[1] == 0:
return np.zeros((3,0)).astype(int), np.zeros(0).astype(int)
if rcorners.shape[1] == 1:
out = np.vstack((rcorners, in_corners[2])).reshape(3,1)
return out, np.zeros((1)).astype(int)
# Initialize the grid.
for i, rc in enumerate(rcorners.T):
grid[rcorners[1,i], rcorners[0,i]] = 1
inds[rcorners[1,i], rcorners[0,i]] = i
# Pad the border of the grid, so that we can NMS points near the border.
pad = dist_thresh
grid = np.pad(grid, ((pad,pad), (pad,pad)), mode='constant')
# Iterate through points, highest to lowest conf, suppress neighborhood.
count = 0
for i, rc in enumerate(rcorners.T):
# Account for top and left padding.
pt = (rc[0]+pad, rc[1]+pad)
if grid[pt[1], pt[0]] == 1: # If not yet suppressed.
grid[pt[1]-pad:pt[1]+pad+1, pt[0]-pad:pt[0]+pad+1] = 0
grid[pt[1], pt[0]] = -1
count += 1
# Get all surviving -1's and return sorted array of remaining corners.
keepy, keepx = np.where(grid==-1)
keepy, keepx = keepy - pad, keepx - pad
inds_keep = inds[keepy, keepx]
out = corners[:, inds_keep]
values = out[-1, :]
inds2 = np.argsort(-values)
out = out[:, inds2]
out_inds = inds1[inds_keep[inds2]]
return out, out_inds
def padding(_img, padd): # Add padding
_img = np.pad(_img, pad_width=[(padd, padd), (padd, padd)], mode='constant', constant_values=0)
return _img
def sde(self):
"""
Support adding kernels for sde representation
"""
import scipy.linalg as la
F = None
L = None
Qc = None
H = None
Pinf = None
P0 = None
dF = None
dQc = None
dPinf = None
dP0 = None
n = 0
nq = 0
nd = 0
# Assign models
for p in self.parts:
(Ft,Lt,Qct,Ht,Pinft,P0t,dFt,dQct,dPinft,dP0t) = p.sde()
F = la.block_diag(F,Ft) if (F is not None) else Ft
L = la.block_diag(L,Lt) if (L is not None) else Lt
Qc = la.block_diag(Qc,Qct) if (Qc is not None) else Qct
H = np.hstack((H,Ht)) if (H is not None) else Ht
Pinf = la.block_diag(Pinf,Pinft) if (Pinf is not None) else Pinft
P0 = la.block_diag(P0,P0t) if (P0 is not None) else P0t
if dF is not None:
dF = np.pad(dF,((0,dFt.shape[0]),(0,dFt.shape[1]),(0,dFt.shape[2])),
'constant', constant_values=0)
dF[-dFt.shape[0]:,-dFt.shape[1]:,-dFt.shape[2]:] = dFt
else:
dF = dFt
if dQc is not None:
dQc = np.pad(dQc,((0,dQct.shape[0]),(0,dQct.shape[1]),(0,dQct.shape[2])),
'constant', constant_values=0)
dQc[-dQct.shape[0]:,-dQct.shape[1]:,-dQct.shape[2]:] = dQct
else:
dQc = dQct
if dPinf is not None:
dPinf = np.pad(dPinf,((0,dPinft.shape[0]),(0,dPinft.shape[1]),(0,dPinft.shape[2])),
'constant', constant_values=0)
dPinf[-dPinft.shape[0]:,-dPinft.shape[1]:,-dPinft.shape[2]:] = dPinft
else:
dPinf = dPinft
if dP0 is not None:
dP0 = np.pad(dP0,((0,dP0t.shape[0]),(0,dP0t.shape[1]),(0,dP0t.shape[2])),
'constant', constant_values=0)
dP0[-dP0t.shape[0]:,-dP0t.shape[1]:,-dP0t.shape[2]:] = dP0t
else:
dP0 = dP0t
n += Ft.shape[0]
nq += Qct.shape[0]
nd += dFt.shape[2]
assert (F.shape[0] == n and F.shape[1]==n), "SDE add: Check of F Dimensions failed"
assert (L.shape[0] == n and L.shape[1]==nq), "SDE add: Check of L Dimensions failed"
assert (Qc.shape[0] == nq and Qc.shape[1]==nq), "SDE add: Check of Qc Dimensions failed"
assert (H.shape[0] == 1 and H.shape[1]==n), "SDE add: Check of H Dimensions failed"
assert (Pinf.shape[0] == n and Pinf.shape[1]==n), "SDE add: Check of Pinf Dimensions failed"
assert (P0.shape[0] == n and P0.shape[1]==n), "SDE add: Check of P0 Dimensions failed"
assert (dF.shape[0] == n and dF.shape[1]==n and dF.shape[2]==nd), "SDE add: Check of dF Dimensions failed"
assert (dQc.shape[0] == nq and dQc.shape[1]==nq and dQc.shape[2]==nd), "SDE add: Check of dQc Dimensions failed"
assert (dPinf.shape[0] == n and dPinf.shape[1]==n and dPinf.shape[2]==nd), "SDE add: Check of dPinf Dimensions failed"
assert (dP0.shape[0] == n and dP0.shape[1]==n and dP0.shape[2]==nd), "SDE add: Check of dP0 Dimensions failed"
return (F,L,Qc,H,Pinf,P0,dF,dQc,dPinf,dP0)
def dice_images(
sliceSavePath,
imageWidth,
imageHeight,
pixelPitch,
x_boundries,
y_boundries,
overlap,
progress_handle=None,
):
dicedSavePath = Path(sliceSavePath) / "diced_images"
fullSavePath = Path(sliceSavePath) / "full_sized_images"
if not Path(dicedSavePath).exists():
Path(dicedSavePath).mkdir()
if not Path(fullSavePath).exists():
Path(fullSavePath).mkdir()
images = glob.glob(str(sliceSavePath / "*.png"))
image_count = len(images)
for i, image_path in enumerate(images):
image_path = Path(image_path)
filename = image_path.stem
extention = image_path.suffix
img = np.array(Image.open(image_path))
last_x = 0
for j, x in enumerate(x_boundries):
x = int(x)
if x == 0:
continue
last_y = 0
for k, y in enumerate(y_boundries):
y = int(y)
if y == 0:
continue
sub_img = img[last_y:y, last_x:x]
# pad image
img_width = x - last_x
img_height = y - last_y
pad_x = printer.width - img_width
pad_y = printer.height - img_height
sub_img = np.pad(
sub_img,
(
(pad_y // 2, pad_y - (pad_y // 2)),
(pad_x // 2, pad_x - (pad_x // 2)),
),
)
sub_img = Image.fromarray(sub_img).convert("L")
sub_img.save(
dicedSavePath / f"{filename}_stitch_x{j-1}_y{k-1}{extention}"
)
last_y = y - overlap
last_x = x - overlap
image_path.replace(fullSavePath / f"{filename}{extention}")
if progress_handle is not None:
progress_handle("Dicing images", i + 1, image_count)
with open(sliceSavePath / "stitching_info.json", "w") as file:
info = {
"pixel_pitch": pixelPitch,
"x_boundries": x_boundries,
"y_boundries": y_boundries,
"overlap": overlap,
}
json.dump(info, file)
# test_data_path = 'F:/data/private/speech/vctk/qts/p225_001.npy'
# target_speaker_ivec = ivecs[speaker2id(target_speaker)]
receptive_field = hp.dilations[-1] * hp.size
speaker_emb = tf.placeholder(dtype=tf.float32,
shape=(hp.batch_size, len(hp.speakers)))
_input = tf.placeholder(dtype=tf.float32,
shape=(hp.batch_size, receptive_field, hp.Q))
_z_q = tf.placeholder(dtype=tf.float32,
shape=(hp.batch_size, receptive_field, hp.D))
x = decoder(_input, speaker_emb, _z_q, is_training=False)
out = tf.multinomial(tf.squeeze(x, 1), num_samples=1, output_dtype=tf.int32)
for j in range(0, len(test_data_paths)):
test_qt = np.load(test_data_paths[j][0])
test_speaker = test_data_paths[j][1]
x = np.pad(test_qt, ([0, hp.T], [0, 0]),
mode="constant",
constant_values=0)[:hp.T, :]
x = np.expand_dims(x, 0)
test_speaker = np.expand_dims(test_speaker, 0)
# target_speaker_ivec = np.expand_dims(target_speaker_ivec,0)
z_q, x, _speaker_emb = sess.run([g.z_q, g.encoder_inputs, g.speakers],
feed_dict={
g.x: x,
g.speaker_id: test_speaker
}) #shape = (B,T,K)
output = np.squeeze(x)
inputs = x[:, :receptive_field, :] # (B,r,Q)
# decode and get multinomial distribuition
for i in tqdm(range(hp.T - receptive_field - 1),
total=hp.T - receptive_field,
def zeropad_to_max_len(data, max_len=121):
return np.pad(data, [(0, 0), (0, max_len - data.shape[1]), (0, 0)],
mode="constant")
def load_umc_sheets(data_dir="/home/matthias/Data/umc_mozart", require_performance=False):
""" load unwarpped sheets """
import glob
import cv2
# initialize omr system
from omr.omr_app import OpticalMusicRecognizer
from omr.utils.data import prepare_image
from lasagne_wrapper.network import SegmentationNetwork
from omr.models import system_detector, bar_detector
net = system_detector.build_model()
system_net = SegmentationNetwork(net, print_architecture=False)
system_net.load('sheet_utils/omr_models/system_params.pkl')
net = bar_detector.build_model()
bar_net = SegmentationNetwork(net, print_architecture=False)
bar_net.load('sheet_utils/omr_models/bar_params.pkl')
piece_names = []
unwrapped_sheets = []
piece_paths = []
# get list of all pieces
piece_dirs = np.sort(glob.glob(os.path.join(data_dir, '*')))
n_pieces = len(piece_dirs)
# iterate pieces
kept_pages = 0
for i_piece, piece_dir in enumerate(piece_dirs):
piece_name = piece_dir.split('/')[-1]
# if "214_" not in piece_name:
# continue
print(col.print_colored("Processing piece %d of %d (%s)" % (i_piece + 1, n_pieces, piece_name), col.OKBLUE))
# check if there is a performance
if require_performance and len(glob.glob(os.path.join(piece_dir, "*performance*"))) == 0:
print("No performance found!")
continue
# load pages
page_paths = np.sort(glob.glob(os.path.join(piece_dir, "sheet/*.png")))
if len(page_paths) == 0:
print("No sheet available!!!")
continue
unwrapped_sheet = np.zeros((SYSTEM_HEIGHT, 0), dtype=np.uint8)
system_problem = False
for i_page, page_path in enumerate(page_paths):
kept_pages += 1
# load sheet image
I = cv2.imread(page_path, 0)
# load system coordinates
# page_id = i_page + 1
# page_systems = np.load(os.path.join(piece_dir, "coords", "systems_%02d.npy" % (i_page + 1)))
# detect systems
I_prep = prepare_image(I)
omr = OpticalMusicRecognizer(note_detector=None, system_detector=system_net, bar_detector=bar_net)
try:
page_systems = omr.detect_systems(I_prep, verbose=False)
except:
print("Problem in system detection!!!")
system_problem = True
continue
# plt.figure("System Localization")
# plt.clf()
# plt.imshow(I, cmap=plt.cm.gray)
# plt.xlim([0, I.shape[1] - 1])
# plt.ylim([I.shape[0] - 1, 0])
# for system in page_systems:
# plt.plot(system[:, 1], system[:, 0], 'mo', alpha=0.5)
# plt.show(block=True)
# unwrap sheet
for system in page_systems:
r0 = int(np.mean([system[0, 0], system[2, 0]])) - SYSTEM_HEIGHT // 2
r1 = r0 + SYSTEM_HEIGHT
c0 = int(system[0, 1])
c1 = int(system[1, 1])
# fix row slice coordinates
r0 = max(0, r0)
r1 = min(r1, I.shape[0])
r0 = max(r0, r1 - SYSTEM_HEIGHT)
staff_img = I[r0:r1, c0:c1].astype(np.uint8)
if staff_img.shape[0] < SYSTEM_HEIGHT:
to_pad = SYSTEM_HEIGHT - staff_img.shape[0]
if to_pad > (0.1 * SYSTEM_HEIGHT):
print("Problem in system padding!!!")
continue
staff_img = np.pad(staff_img, ((0, to_pad), (0, 0)), mode="edge")
unwrapped_sheet = np.hstack((unwrapped_sheet, staff_img))
# plt.figure("Unwrapped")
# plt.imshow(unwrapped_sheet)
# plt.show(block=True)
if not system_problem:
piece_names.append(piece_name)
piece_paths.append(piece_dir)
unwrapped_sheets.append(unwrapped_sheet)
print("%d pieces covering %d pages of sheet music." % (len(piece_names), kept_pages))
return piece_names, piece_paths, unwrapped_sheets
def maux(output_text, num):
print("debug -- django")
## Info & args
# parser = argparse.ArgumentParser(
# formatter_class=argparse.ArgumentDefaultsHelpFormatter
# )
# parser.add_argument("-e", "--enc_model_fpath", type=Path,
# default="D:/RemindMe/django-remindme/mysite/trained model/encoder/saved_models/pretrained.pt",
# help="Path to a saved encoder")
# parser.add_argument("-s", "--syn_model_dir", type=Path,
# default="D:/RemindMe/django-remindme/mysite/trained model/synthesizer/saved_models/logs-pretrained/",
# help="Directory containing the synthesizer model")
# parser.add_argument("-v", "--voc_model_fpath", type=Path,
# default="D:/RemindMe/django-remindme/mysite/trained model/vocoder/saved_models/pretrained/pretrained.pt",
# help="Path to a saved vocoder")
# parser.add_argument("--low_mem", action="store_true", help=\
# "If True, the memory used by the synthesizer will be freed after each use. Adds large "
# "overhead but allows to save some GPU memory for lower-end GPUs.")
# parser.add_argument("--no_sound", action="store_true", help=\
# "If True, audio won't be played.")
# args = parser.parse_args()
# print_args(args, parser)
# if not args.no_sound:
# import sounddevice as sd
## Print some environment information (for debugging purposes)
print("Running a test of your configuration...\n")
if not torch.cuda.is_available():
print(
"Your PyTorch installation is not configured to use CUDA. If you have a GPU ready "
"for deep learning, ensure that the drivers are properly installed, and that your "
"CUDA version matches your PyTorch installation. CPU-only inference is currently "
"not supported.",
file=sys.stderr)
quit(-1)
device_id = torch.cuda.current_device()
gpu_properties = torch.cuda.get_device_properties(device_id)
print(
"Found %d GPUs available. Using GPU %d (%s) of compute capability %d.%d with "
"%.1fGb total memory.\n" %
(torch.cuda.device_count(), device_id, gpu_properties.name,
gpu_properties.major, gpu_properties.minor,
gpu_properties.total_memory / 1e9))
## Load the models one by one.
print("Preparing the encoder, the synthesizer and the vocoder...")
#encoder.load_model(args.enc_model_fpath)
#synthesizer = Synthesizer(args.syn_model_dir.joinpath("taco_pretrained"), low_mem=args.low_mem)
#vocoder.load_model(args.voc_model_fpath)
encoder.load_model(
"D:/RemindMe/django-remindme/mysite/trained model/encoder/saved_models/pretrained.pt"
)
synthesizer = Synthesizer(
"D:/RemindMe/django-remindme/mysite/trained model/synthesizer/saved_models/logs-pretrained/taco_pretrained",
low_mem=False)
vocoder.load_model(
"D:/RemindMe/django-remindme/mysite/trained model/vocoder/saved_models/pretrained/pretrained.pt"
)
## Run a test
print("Testing your configuration with small inputs.")
# Forward an audio waveform of zeroes that lasts 1 second. Notice how we can get the encoder's
# sampling rate, which may differ.
# If you're unfamiliar with digital audio, know that it is encoded as an array of floats
# (or sometimes integers, but mostly floats in this projects) ranging from -1 to 1.
# The sampling rate is the number of values (samples) recorded per second, it is set to
# 16000 for the encoder. Creating an array of length <sampling_rate> will always correspond
# to an audio of 1 second.
print("\tTesting the encoder...")
encoder.embed_utterance(np.zeros(encoder.sampling_rate))
# Create a dummy embedding. You would normally use the embedding that encoder.embed_utterance
# returns, but here we're going to make one ourselves just for the sake of showing that it's
# possible.
embed = np.random.rand(speaker_embedding_size)
# Embeddings are L2-normalized (this isn't important here, but if you want to make your own
# embeddings it will be).
embed /= np.linalg.norm(embed)
# The synthesizer can handle multiple inputs with batching. Let's create another embedding to
# illustrate that
embeds = [embed, np.zeros(speaker_embedding_size)]
texts = ["test 1", "test 2"]
print(
"\tTesting the synthesizer... (loading the model will output a lot of text)"
)
mels = synthesizer.synthesize_spectrograms(texts, embeds)
# The vocoder synthesizes one waveform at a time, but it's more efficient for long ones. We
# can concatenate the mel spectrograms to a single one.
mel = np.concatenate(mels, axis=1)
# The vocoder can take a callback function to display the generation. More on that later. For
# now we'll simply hide it like this:
no_action = lambda *args: None
print("\tTesting the vocoder...")
# For the sake of making this test short, we'll pass a short target length. The target length
# is the length of the wav segments that are processed in parallel. E.g. for audio sampled
# at 16000 Hertz, a target length of 8000 means that the target audio will be cut in chunks of
# 0.5 seconds which will all be generated together. The parameters here are absurdly short, and
# that has a detrimental effect on the quality of the audio. The default parameters are
# recommended in general.
vocoder.infer_waveform(mel,
target=200,
overlap=50,
progress_callback=no_action)
print("All test passed! You can now synthesize speech.\n\n")
## Interactive speech generation
print(
"This is a GUI-less example of interface to SV2TTS. The purpose of this script is to "
"show how you can interface this project easily with your own. See the source code for "
"an explanation of what is happening.\n")
print("Interactive generation loop")
in_fpath = Path(
"D:/RemindMe/django-remindme/mysite/trained model/sam_narration2.wav")
preprocessed_wav = encoder.preprocess_wav(in_fpath)
original_wav, sampling_rate = librosa.load(in_fpath)
preprocessed_wav = encoder.preprocess_wav(original_wav, sampling_rate)
print("Loaded file succesfully")
embed = encoder.embed_utterance(preprocessed_wav)
print("Created the embedding")
embeds = [embed]
text = output_text
texts = [text]
specs = synthesizer.synthesize_spectrograms(texts, embeds)
spec = specs[0]
print("Created the mel spectrogram")
## Generating the waveform
print("Synthesizing the waveform:")
# Synthesizing the waveform is fairly straightforward. Remember that the longer the
# spectrogram, the more time-efficient the vocoder.
generated_wav = vocoder.infer_waveform(spec)
## Post-generation
# There's a bug with sounddevice that makes the audio cut one second earlier, so we
# pad it.
generated_wav = np.pad(generated_wav, (0, synthesizer.sample_rate),
mode="constant")
# Play the audio (non-blocking)
# Save it on the disk
filexpath = "D:/RemindMe/django_remindme_model/mysite/media/demo_output_%02d.wav" % num
fx = "demo_output_%02d" % num
print(generated_wav.dtype)
librosa.output.write_wav(filexpath, generated_wav.astype(np.float32),
synthesizer.sample_rate)
print("\nSaved output as %s\n\n" % filexpath)
return fx
def append_tile(array, geom, tot_array, tot_geom):
"""
Append a tile to a larger arrayset.
Args:
array: projection stack
geom: geometry descritption
tot_array: output array
tot_geom: output geometry
"""
print('Stitching a tile...')
# Assuming all projections have equal number of angles and same pixel sizes
total_shape = tot_array.shape[::2]
det_shape = array.shape[::2]
if numpy.abs(tot_geom['det_pixel'] - geom['det_pixel']) > 1e-6:
raise Exception('This array has different detector pixels! %f v.s. %f. Aborting!' % (geom['det_pixel'], tot_geom['det_pixel']))
if tot_array.shape[1] != array.shape[1]:
raise Exception('This array has different number of projections from the others. %u v.s. %u. Aborting!' % (array.shape[1], tot_array.shape[1]))
total_size = tot_geom.detector_size(total_shape)
det_size = geom.detector_size(det_shape)
# Offset from the left top corner:
y0, x0 = tot_geom.detector_centre()
y, x = geom.detector_centre()
x_offset = ((x - x0) + total_size[1] / 2 - det_size[1] / 2) / geom.pixel[1]
y_offset = ((y - y0) + total_size[0] / 2 - det_size[0] / 2) / geom.pixel[0]
# Round em up!
x_offset = int(numpy.round(x_offset))
y_offset = int(numpy.round(y_offset))
# Pad image to get the same size as the total_slice:
pad_x = tot_array.shape[2] - array.shape[2]
pad_y = tot_array.shape[0] - array.shape[0]
# Collapce both arraysets and compute residual shift
shift = _find_shift_(tot_array, array, [y_offset, x_offset])
x_offset += shift[1]
y_offset += shift[0]
# Precompute weights:
base0 = (tot_array[:, ::100, :].mean(1)) != 0
new0 = numpy.zeros_like(base0)
# Shift image:
new0[:det_shape[0], :det_shape[1]] = 1.0
new0 = interp.shift(new0, [y_offset, x_offset], order = 1)
#new0[y_offset:int(y_offset+det_shape[0]), x_offset:int(x_offset + det_shape[1])] = 1.0
base_dist = ndimage.distance_transform_bf(base0)
new_dist = ndimage.distance_transform_bf(new0)
# Trim edges to avoid interpolation errors:
base_dist -= 1
new_dist -= 1
base_dist *= base_dist > 0
new_dist *= new_dist > 0
norm = (base_dist + new_dist)
norm[norm == 0] = numpy.inf
time.sleep(0.5)
# Apply offsets:
for ii in tqdm(range(tot_array.shape[1]), unit='img'):
# Pad to match sizes:
new = numpy.pad(array[:, ii, :], ((0, pad_y), (0, pad_x)), mode = 'constant')
# Apply shift:
if (x_offset != 0) | (y_offset != 0):
# Shift image:
new = interp.shift(new, [y_offset, x_offset], order = 1)
# Add two images in a smart way:
base = tot_array[:, ii, :]
# Create distances to edge:
tot_array[:, ii, :] = ((base_dist * base) + (new_dist * new)) / norm
def _process_utterance(out_dir, index, wav_path, text):
# Load the audio to a numpy array:
wav = audio.load_wav(wav_path)
if hparams.rescaling:
wav = wav / np.abs(wav).max() * hparams.rescaling_max
# Mu-law quantize
if is_mulaw_quantize(hparams.input_type):
# [0, quantize_channels)
out = P.mulaw_quantize(wav, hparams.quantize_channels)
# Trim silences
start, end = audio.start_and_end_indices(out,
hparams.silence_threshold)
wav = wav[start:end]
out = out[start:end]
constant_values = P.mulaw_quantize(0, hparams.quantize_channels)
out_dtype = np.int16
elif is_mulaw(hparams.input_type):
# [-1, 1]
out = P.mulaw(wav, hparams.quantize_channels)
constant_values = P.mulaw(0.0, hparams.quantize_channels)
out_dtype = np.float32
else:
# [-1, 1]
out = wav
constant_values = 0.0
out_dtype = np.float32
# Compute a mel-scale spectrogram from the trimmed wav:
# (N, D)
mel_spectrogram = audio.melspectrogram(wav).astype(np.float32).T
# lws pads zeros internally before performing stft
# this is needed to adjust time resolution between audio and mel-spectrogram
l, r = audio.lws_pad_lr(wav, hparams.fft_size, audio.get_hop_size())
# zero pad for quantized signal
out = np.pad(out, (l, r), mode="constant", constant_values=constant_values)
N = mel_spectrogram.shape[0]
assert len(out) >= N * audio.get_hop_size()
# time resolution adjustment
# ensure length of raw audio is multiple of hop_size so that we can use
# transposed convolution to upsample
out = out[:N * audio.get_hop_size()]
assert len(out) % audio.get_hop_size() == 0
timesteps = len(out)
# Write the spectrograms to disk:
audio_filename = 'ljspeech-audio-%05d.npy' % index
mel_filename = 'ljspeech-mel-%05d.npy' % index
np.save(os.path.join(out_dir, audio_filename),
out.astype(out_dtype),
allow_pickle=False)
np.save(os.path.join(out_dir, mel_filename),
mel_spectrogram.astype(np.float32),
allow_pickle=False)
# Return a tuple describing this training example:
return (audio_filename, mel_filename, timesteps, text)
def resize_image(image, min_dim=None, max_dim=None, min_scale=None, mode="square"):
"""Resizes an image keeping the aspect ratio unchanged.
min_dim: if provided, resizes the image such that it's smaller
dimension == min_dim
max_dim: if provided, ensures that the image longest side doesn't
exceed this value.
min_scale: if provided, ensure that the image is scaled up by at least
this percent even if min_dim doesn't require it.
mode: Resizing mode.
none: No resizing. Return the image unchanged.
square: Resize and pad with zeros to get a square image
of size [max_dim, max_dim].
pad64: Pads width and height with zeros to make them multiples of 64.
If min_dim or min_scale are provided, it scales the image up
before padding. max_dim is ignored in this mode.
The multiple of 64 is needed to ensure smooth scaling of feature
maps up and down the 6 levels of the FPN pyramid (2**6=64).
crop: Picks random crops from the image. First, scales the image based
on min_dim and min_scale, then picks a random crop of
size min_dim x min_dim. Can be used in training only.
max_dim is not used in this mode.
Returns:
image: the resized image
window: (y1, x1, y2, x2). If max_dim is provided, padding might
be inserted in the returned image. If so, this window is the
coordinates of the image part of the full image (excluding
the padding). The x2, y2 pixels are not included.
scale: The scale factor used to resize the image
padding: Padding added to the image [(top, bottom), (left, right), (0, 0)]
"""
# Keep track of image dtype and return results in the same dtype
image_dtype = image.dtype
# Default window (y1, x1, y2, x2) and default scale == 1.
h, w = image.shape[:2]
window = (0, 0, h, w)
scale = 1
padding = [(0, 0), (0, 0), (0, 0)]
crop = None
if mode == "none":
return image, window, scale, padding, crop
# Scale?
if min_dim:
# Scale up but not down
scale = max(1, min_dim / min(h, w))
if min_scale and scale < min_scale:
scale = min_scale
# Does it exceed max dim?
if max_dim and mode == "square":
image_max = max(h, w)
if round(image_max * scale) > max_dim:
scale = max_dim / image_max
# Resize image using bilinear interpolation
if scale != 1:
image = resize(image, (round(h * scale), round(w * scale)),
preserve_range=True)
# Need padding or cropping?
if mode == "square":
# Get new height and width
h, w = image.shape[:2]
top_pad = (max_dim - h) // 2
bottom_pad = max_dim - h - top_pad
left_pad = (max_dim - w) // 2
right_pad = max_dim - w - left_pad
padding = [(top_pad, bottom_pad), (left_pad, right_pad), (0, 0)]
image = np.pad(image, padding, mode='constant', constant_values=0)
window = (top_pad, left_pad, h + top_pad, w + left_pad)
elif mode == "pad64":
h, w = image.shape[:2]
# Both sides must be divisible by 64
assert min_dim % 64 == 0, "Minimum dimension must be a multiple of 64"
# Height
if h % 64 > 0:
max_h = h - (h % 64) + 64
top_pad = (max_h - h) // 2
bottom_pad = max_h - h - top_pad
else:
top_pad = bottom_pad = 0
# Width
if w % 64 > 0:
max_w = w - (w % 64) + 64
left_pad = (max_w - w) // 2
right_pad = max_w - w - left_pad
else:
left_pad = right_pad = 0
padding = [(top_pad, bottom_pad), (left_pad, right_pad), (0, 0)]
image = np.pad(image, padding, mode='constant', constant_values=0)
window = (top_pad, left_pad, h + top_pad, w + left_pad)
elif mode == "crop":
# Pick a random crop
h, w = image.shape[:2]
y = random.randint(0, (h - min_dim))
x = random.randint(0, (w - min_dim))
crop = (y, x, min_dim, min_dim)
image = image[y:y + min_dim, x:x + min_dim]
window = (0, 0, min_dim, min_dim)
else:
raise Exception("Mode {} not supported".format(mode))
return image.astype(image_dtype), window, scale, padding, crop
def complete(self, comppars, X, parfile=None, sigma=0.0):
"""
Finds the best representation that can complete any missing
part of the ASPA code.
Input: any spectrum correctly converted into an ASPA code
"""
outDir = directory(comppars['outDir'])
maskType = comppars['maskType']
centerScale = comppars['centerScale']
nIter = int(comppars['nIter'])
outInterval = int(comppars['outInterval'])
approach = comppars['approach']
make_corner = bool(comppars['make_corner'])
beta1 = comppars['beta1']
beta2 = comppars['beta2']
lr = comppars['lr']
eps = comppars['eps']
hmcL = int(comppars['hmcL'])
hmcEps = comppars['hmcEps']
hmcBeta = comppars['hmcBeta']
hmcAnneal = comppars['hmcAnneal']
checkpoint_dir = directory(comppars['checkpointDir'])
build_directories(comppars)
if type(X) == dict:
X_to_split = copy.deepcopy(X)
elif type(X) == str:
if X[-3:] == 'dat':
X_to_split = str(X)
else:
X_to_split = np.array(X)
try:
if type(X) != np.ndarray:
true_spectrum = X
else:
true_spectrum = None
except IOError:
true_spectrum = None
with self.sess.as_default():
try:
tf.global_variables_initializer().run()
except:
tf.initialize_all_variables().run()
isLoaded = self.load(checkpoint_dir)
assert (isLoaded)
grids = Grids()
wnw_grid = grids.wnw_grid
nImgs = self.batch_size
batch_idxs = int(np.ceil(nImgs / self.batch_size))
if maskType == 'random':
fraction_masked = 0.2
mask = np.ones(self.image_shape)
mask[
np.random.random(self.image_shape[:2]) < fraction_masked] = 0.0
elif maskType == 'center':
assert (centerScale <= 0.5)
mask = np.ones(self.image_shape)
l = int(self.image_size * centerScale)
u = int(self.image_size * (1.0 - centerScale))
mask[l:u, l:u, :] = 0.0
elif maskType == 'left':
mask = np.ones(self.image_shape)
c = self.image_size // 2
mask[:, :c, :] = 0.0
elif maskType == 'full':
mask = np.ones(self.image_shape)
elif maskType == 'grid':
mask = np.zeros(self.image_shape)
mask[::4, ::4, :] = 1.0
elif maskType == 'lowres':
mask = np.zeros(self.image_shape)
elif maskType == 'parameters':
assert (centerScale <= 0.5)
mask = np.ones(self.image_shape)
mask[-3:, :, :] = 0.0
mask[:, -3:, :] = 0.0
mask[-10:, -10:, :] = 0.0
elif maskType == 'wfc3':
assert (centerScale <= 0.5)
m_size = self.image_size - 10
mask = np.ones(self.image_shape)
fake_spec = np.ones(m_size**2)
fake_spec[:334] = 0.0
fake_spec[384:] = 0.0
fake_spec = fake_spec.reshape((m_size, m_size))
mask[:m_size, :m_size, 0] = fake_spec
mask[-8:, :, :] = 0.0
mask[:, -10:, :] = 0.0
mask[-10:, -10:, :] = 0.0
else:
assert (False)
for idx in xrange(0, batch_idxs):
l = idx * self.batch_size
u = min((idx + 1) * self.batch_size, nImgs)
batchSz = u - l
if type(X) != str:
Xtrue = get_spectral_matrix(X, size=self.image_size - 10)
Xt = get_test_image(X,
sigma=sigma,
size=self.image_size,
batch_size=self.batch_size)
else:
Xtrue = get_spectral_matrix(X,
parfile=parfile,
size=self.image_size - 10)
Xt = get_test_image(X,
sigma=sigma,
size=self.image_size,
batch_size=self.batch_size,
parfile=parfile)
spec_parameters = get_parameters(Xtrue, size=self.image_size)
batch = Xt
batch_images = np.array(batch).astype(np.float32)
if batchSz < self.batch_size:
print(batchSz)
padSz = ((0, int(self.batch_size - batchSz)), (0, 0), (0, 0),
(0, 0))
batch_images = np.pad(batch_images, padSz, 'constant')
batch_images = batch_images.astype(np.float32)
zhats = np.random.uniform(-1,
1,
size=(self.batch_size, self.z_dim))
m = 0
v = 0
nImgs = 1
nRows = int(np.sqrt(nImgs))
nCols = int(np.sqrt(nImgs))
# save_images(batch_images[:nImgs, :, :, :], [nRows, nCols],
# os.path.join(config.outDir, 'before.pdf'))
plt.imsave(os.path.join(outDir, 'before.png'),
Xtrue[:, :, 0],
cmap='gist_gray',
format='png')
plt.close()
resize(os.path.join(outDir, 'before.png'))
masked_images = np.multiply(batch_images, mask)
# save_images(masked_images[:nImgs, :, :, :], [nRows, nCols],
# os.path.join(config.outDir, 'masked.pdf'))
plt.imsave(os.path.join(outDir, 'masked.png'),
masked_images[0, :, :, 0],
cmap='gist_gray',
format='png')
plt.close()
resize(os.path.join(outDir, 'masked.png'))
for img in range(batchSz):
with open(
os.path.join(outDir,
'logs/hats_{:02d}.log'.format(img)),
'a') as f:
f.write('iter loss ' + ' '.join(
['z{}'.format(zi) for zi in range(self.z_dim)]) + '\n')
for i in xrange(nIter):
fd = {
self.z: zhats,
self.mask: mask,
self.images: batch_images,
self.is_training: False
}
run = [self.complete_loss, self.grad_complete_loss, self.G]
loss, g, G_imgs = self.sess.run(run, feed_dict=fd)
for img in range(batchSz):
with open(
os.path.join(outDir,
'logs/hats_{:02d}.log'.format(img)),
'ab') as f:
f.write('{} {} '.format(i, loss[img]).encode())
np.savetxt(f, zhats[img:img + 1])
if i % outInterval == 0:
prediction_file = open(
outDir + 'predictions/prediction_{:04d}.txt'.format(i),
'w')
ranges = []
ground_truths = []
gan_avg = []
gan_p_err = []
gan_m_err = []
if type(X) == str:
"""
If the input spectrum is synthetic, you know the parameters array
and you want to compare the real value with the retrieved one, if
your spectrum does not contain a molecule, the default value is fixed
to -7.9
"""
pp = ParameterParser()
pp.read(parfile)
real_pars = pp.full_dict()
real_tp = float(real_pars['Atmosphere']['tp_iso_temp'])
real_rp = float(real_pars['Planet']['radius'])
real_mp = float(real_pars['Planet']['mass'])
atm_active_gases = np.array([
gas.upper()
for gas in real_pars['Atmosphere']['active_gases']
])
atm_active_gases_mixratios = np.array(
real_pars['Atmosphere']['active_gases_mixratios'])
real_mol = check_molecule_existence(
['CO', 'CO2', 'H2O', 'CH4'],
atm_active_gases_mixratios,
atm_active_gases,
default=-7.9)
ground_truths = np.array(real_mol +
[real_rp, real_mp, real_tp])
elif true_spectrum != None and type(X) != str:
h2o = np.log10(true_spectrum['param']['h2o_mixratio'])
ch4 = np.log10(true_spectrum['param']['ch4_mixratio'])
co2 = np.log10(true_spectrum['param']['co2_mixratio'])
co = np.log10(true_spectrum['param']['co_mixratio'])
rp = true_spectrum['param']['planet_radius'] / RJUP
mp = true_spectrum['param']['planet_mass'] / MJUP
tp = true_spectrum['param']['temperature_profile']
ground_truths = np.array(
[co, co2, h2o, ch4, rp, mp, tp])
real_mol = np.zeros(4)
else:
ground_truths = np.array([None] * 7)
real_mol = np.zeros(4)
parameters = ['CO', 'CO2', 'H2O', 'CH4', 'Rp', 'Mp', 'Tp']
labels = [
'$\log{CO}$', '$\log{CO_2}$', '$\log{H_2O}$',
'$\log{CH_4}$', '$R_p (R_j)$', '$M_p (M_j)$', '$T_p$'
]
all_hists = []
for mol in parameters:
prediction_file, gan_avg, gan_p_err, gan_m_err, ranges, all_hists = \
histogram_par(mol, G_imgs, batchSz, self.image_size,
ground_truths, all_hists,
prediction_file, gan_avg,
gan_p_err, gan_m_err, ranges)
all_hists = np.array(all_hists).T
if make_corner:
make_corner_plot(all_hists, ranges, labels,
ground_truths, comppars, i)
"""
Plot histograms
"""
hist_dict = {}
f, ax = plt.subplots(2, 4, figsize=(21, 15))
all_hists = all_hists.T
ii = 0
jj = 0
for his in range(len(all_hists)):
if his == 4:
ii = 1
jj = 4
hist_dict[labels[his]] = {}
weights = np.ones_like(all_hists[his]) / float(
len(all_hists[his]))
hist_dict[labels[his]]['histogram'] = all_hists[his]
hist_dict[labels[his]]['weights'] = weights
hist_dict[labels[his]]['bins'] = ranges[his]
ax[ii, his - jj].hist(all_hists[his],
bins=np.linspace(
min(ranges[his]),
max(ranges[his]), 20),
color='firebrick',
weights=weights)
# ax[his].set_ylim(0, 1)
ax[ii, his - jj].set_xlim(min(ranges[his]),
max(ranges[his]))
ax[ii, his - jj].axvline(gan_avg[his],
c='g',
label='ExoGAN mean')
ax[ii, his - jj].axvline(ground_truths[his],
c='b',
label='Input value')
ax[ii, his - jj].set_xlabel(labels[his] + \
' = $%1.2f_{-%1.2f}^{%1.2f}$' % (
gan_avg[his], gan_m_err[his], gan_p_err[his]))
if his == 3:
ax[ii, his - jj].legend()
# ax[his].annotate('$%1.2f_{-%1.2f}^{%1.2f}$' % (gan_avg[his], gan_p_err[his], gan_m_err[his]),
# bbox=dict(boxstyle="round4", fc="w", alpha=0.5),
# xy=(gan_avg[his], max(weights)*(0.9)),
# xycoords='data')
ax[ii, his - jj].axvline(gan_avg[his] + gan_p_err[his],
c='k',
linestyle='--')
ax[ii, his - jj].axvline(gan_avg[his] - gan_m_err[his],
c='k',
linestyle='--')
ax[-1, -1].axis('off')
plt.subplots_adjust(right=1.2)
histName = os.path.join(
outDir, 'histograms/all_par/{:04d}.pdf'.format(i))
plt.savefig(histName, bbox_inches='tight')
plt.close()
histpickle = os.path.join(
outDir, 'histograms/all_par/histogram.pickle')
with open(histpickle, 'wb') as fp:
pickle.dump(hist_dict, fp)
real_spec = Xtrue[:self.image_size, :self.image_size, :]
real_spec = real_spec[:23, :23, 0].flatten()
chi_square = []
spectra = []
f, ax = plt.subplots(sharey=True, figsize=(12, 6))
for k in range(batchSz):
spectrum = G_imgs[
k, :self.image_size, :self.image_size, :]
spectrum = spectrum[:23, :23, 0].flatten()
spectra.append(spectrum)
chi_square.append(
chisquare(spectrum[:440],
f_exp=real_spec[:440])[0])
best_ind = chi_square.index(min(chi_square))
print(i, np.mean(loss[0:batchSz]))
imgName = os.path.join(outDir,
'hats_imgs/{:04d}.png'.format(i))
# save_images(G_imgs[:nImgs, :, :, :], [nRows, nCols], imgName)
plt.imsave(imgName,
G_imgs[best_ind, :, :, 0],
cmap='gist_gray',
format='png')
plt.close()
resize(imgName)
inv_masked_hat_images = np.multiply(G_imgs, 1.0 - mask)
completed = masked_images + inv_masked_hat_images
imgName = os.path.join(outDir,
'completed/{:04d}.png'.format(i))
# save_images(completed[:nImgs, :, :, :], [nRows, nCols], imgName)
plt.imsave(imgName,
completed[best_ind, :, :, 0],
cmap='gist_gray',
format='png')
plt.close()
resize(imgName)
if spectra_int_norm:
# Compared real spectrum with the generated one
spectra_int_norm(Xtrue, self.image_size, wnw_grid,
batchSz, G_imgs, comppars, i)
if spectra_norm:
# Compare spectra with original normalisation between 0 and 1
spectra_norm(Xtrue, self.image_size, wnw_grid, batchSz,
G_imgs, comppars, i)
if spectra_real_norm:
# Compare spectra with the normalisation factor from the real spectrum
spectra_real_norm(Xtrue, self.image_size, wnw_grid,
batchSz, G_imgs, comppars, i)
if approach == 'adam':
# Optimize single completion with Adam
m_prev = np.copy(m)
v_prev = np.copy(v)
m = beta1 * m_prev + (1 - beta1) * g[0]
v = beta2 * v_prev + (1 - beta2) * np.multiply(g[0], g[0])
m_hat = m / (1 - beta1**(i + 1))
v_hat = v / (1 - beta2**(i + 1))
zhats += -np.true_divide(lr * m_hat,
(np.sqrt(v_hat) + eps))
zhats = np.clip(zhats, -1, 1)
elif approach == 'hmc':
# Sample example completions with HMC (not in paper)
zhats_old = np.copy(zhats)
loss_old = np.copy(loss)
v = np.random.randn(self.batch_size, self.z_dim)
v_old = np.copy(v)
for steps in range(hmcL):
v -= hmcEps / 2 * hmcBeta * g[0]
zhats += hmcEps * v
np.copyto(zhats, np.clip(zhats, -1, 1))
loss, g, _, _ = self.sess.run(run, feed_dict=fd)
v -= hmcEps / 2 * hmcBeta * g[0]
for img in range(batchSz):
logprob_old = hmcBeta * loss_old[img] + np.sum(
v_old[img]**2) / 2
logprob = hmcBeta * loss[img] + np.sum(v[img]**2) / 2
accept = np.exp(logprob_old - logprob)
if accept < 1 and np.random.uniform() > accept:
np.copyto(zhats[img], zhats_old[img])
hmcBeta *= hmcAnneal
else:
assert (False)
for each in original_features:
sm = []
for beach in cover_features:
sm.append(euclidean(each, beach))
similarity_matrix.append(sm)
return np.array(similarity_matrix)
model = load_model(sys.argv[1])
original_song = sys.argv[2]
cover_song = sys.argv[3]
original_signal = librosa.load(original_song)[0]
cover_signal = librosa.load(cover_song)[0]
original_features = extract_features(original_signal)
cover_features = extract_features(cover_signal)
oti_cover_features = oti_func(original_features, cover_features)
mat = sim_matrix(original_features, oti_cover_features)
if mat.shape[0] < 180:
mat = np.pad(mat, ((0,180 - mat.shape[0]),(0,0)), mode = 'constant', constant_values=0)
if mat.shape[1] < 180:
mat = np.pad(mat, ((0,0),(0,180 - mat.shape[1])), mode = 'constant', constant_values=0)
ans = model.predict(mat.reshape(1,180,180,1))
if ans[0][0] < ans[0][1]: print("The song is a cover pair with probability of : {}".format(ans[0][1]))
else: print("The song is not a cover pair with probability of : {}".format(ans[0][0]))
def coarseness(image, voxelspacing = None, mask = slice(None)):
r"""
Takes a simple or multi-spectral image and returns the coarseness of the texture.
Step1 At each pixel, compute six averages for the windows of size 2**k x 2**k,
k=0,1,...,5, around the pixel.
Step2 At each pixel, compute absolute differences E between the pairs of non
overlapping averages in every directions.
step3 At each pixel, find the value of k that maximises the difference Ek in either
direction and set the best size Sbest=2**k
step4 Compute the coarseness feature Fcrs by averaging Sbest over the entire image.
Parameters
----------
image : array_like or list/tuple of array_like
A single image or a list/tuple of images (for multi-spectral case).
voxelspacing : sequence of floats
The side-length of each voxel.
mask : array_like
A binary mask for the image or a slice object
Returns
-------
coarseness : float
The size of coarseness of the given texture. It is basically the size of
repeating elements in the image.
See Also
--------
"""
# Step1: At each pixel (x,y), compute six averages for the windows
# of size 2**k x 2**k, k=0,1,...,5, around the pixel.
image = numpy.asarray(image, dtype=numpy.float32)
# set default mask or apply given mask
if not type(mask) is slice:
if not type(mask[0] is slice):
mask = numpy.array(mask, copy=False, dtype = numpy.bool)
image = image[mask]
# set default voxel spacing if not suppliec
if None == voxelspacing:
voxelspacing = tuple([1.] * image.ndim)
if len(voxelspacing) != image.ndim:
print "Voxel spacing and image dimensions do not fit."
return None
# set padding for image border control
padSize = tuple((numpy.rint((2**5.0) * voxelspacing[jj]),0) for jj in xrange(image.ndim))
Apad = numpy.pad(image,pad_width=padSize, mode='reflect')
# Allocate memory
E = numpy.empty((6,image.ndim)+image.shape)
# prepare some slicer
rawSlicer = [slice(None)] * image.ndim
slicerForImageInPad = [slice(padSize[d][0],None)for d in xrange(image.ndim)]
for k in xrange(6):
size_vs = tuple(numpy.rint((2**k) * voxelspacing[jj]) for jj in xrange(image.ndim))
A = uniform_filter(Apad, size = size_vs, mode = 'mirror')
# Step2: At each pixel, compute absolute differences E(x,y) between
# the pairs of non overlapping averages in the horizontal and vertical directions.
for d in xrange(image.ndim):
borders = numpy.rint((2**k) * voxelspacing[d])
slicerPad_k_d = slicerForImageInPad[:]
slicerPad_k_d[d]= slice((padSize[d][0]-borders if borders < padSize[d][0] else 0),None)
A_k_d = A[slicerPad_k_d]
AslicerL = rawSlicer[:]
AslicerL[d] = slice(0, -borders)
AslicerR = rawSlicer[:]
AslicerR[d] = slice(borders, None)
E[k,d,...] = numpy.abs(A_k_d[AslicerL] - A_k_d[AslicerR])
# step3: At each pixel, find the value of k that maximises the difference Ek(x,y)
# in either direction and set the best size Sbest(x,y)=2**k
k_max = E.max(1).argmax(0)
dim = E.argmax(1)
dim_vox_space = numpy.asarray([voxelspacing[dim[k_max.flat[i]].flat[i]] for i in xrange(k_max.size)]).reshape(k_max.shape)
S = (2**k_max) * dim_vox_space
# step4: Compute the coarseness feature Fcrs by averaging Sbest(x,y) over the entire image.
return S.mean()
def pad_nd_img(image,
new_shape=None,
mode="edge",
kwargs=None,
return_slicer=False,
shape_must_be_divisible_by=None):
"""
one padder to pad them all. Documentation? Well okay. A little bit
"""
if kwargs is None:
kwargs = {}
if new_shape is not None:
old_shape = np.array(image.shape[-len(new_shape):])
else:
assert shape_must_be_divisible_by is not None
assert isinstance(shape_must_be_divisible_by,
(list, tuple, np.ndarray))
new_shape = image.shape[-len(shape_must_be_divisible_by):]
old_shape = new_shape
num_axes_nopad = len(image.shape) - len(new_shape)
new_shape = [
max(new_shape[i], old_shape[i]) for i in range(len(new_shape))
]
if not isinstance(new_shape, np.ndarray):
new_shape = np.array(new_shape)
if shape_must_be_divisible_by is not None:
if not isinstance(shape_must_be_divisible_by,
(list, tuple, np.ndarray)):
shape_must_be_divisible_by = [shape_must_be_divisible_by
] * len(new_shape)
else:
assert len(shape_must_be_divisible_by) == len(new_shape)
for i in range(len(new_shape)):
if new_shape[i] % shape_must_be_divisible_by[i] == 0:
new_shape[i] -= shape_must_be_divisible_by[i]
new_shape = np.array([
new_shape[i] + shape_must_be_divisible_by[i] -
new_shape[i] % shape_must_be_divisible_by[i]
for i in range(len(new_shape))
])
difference = new_shape - old_shape
pad_below = difference // 2
pad_above = difference // 2 + difference % 2
pad_list = [[0, 0]] * num_axes_nopad + list(
[list(i) for i in zip(pad_below, pad_above)])
res = np.pad(image, pad_list, mode, **kwargs)
if not return_slicer:
return res
else:
pad_list = np.array(pad_list)
pad_list[:, 1] = np.array(res.shape) - pad_list[:, 1]
slicer = list(slice(*i) for i in pad_list)
return res, slicer
def _get_compiled_theano_functions(N_QUAD_PTS):
# Planet masses: m1,m2
m1, m2 = T.dscalars(2)
mstar = 1
mu1 = m1 * mstar / (mstar + m1)
mu2 = m2 * mstar / (mstar + m2)
eta1 = mstar + m1
eta2 = mstar + m2
beta1 = mu1 * T.sqrt(eta1 / mstar) / (mu1 + mu2)
beta2 = mu2 * T.sqrt(eta2 / mstar) / (mu1 + mu2)
j, k = T.lscalars('jk')
s = (j - k) / k
# Angle variable for averaging over
psi = T.dvector('psi')
# Quadrature weights
quad_weights = T.dvector('w')
# Dynamical variables:
Ndof = 3
Nconst = 1
dyvars = T.vector()
s1, s2, phi, I1, I2, Phi, dRtilde = [
dyvars[i] for i in range(2 * Ndof + Nconst)
]
a20 = T.constant(1.)
a10 = ((j - k) / j)**(2 / 3) * (eta1 / eta2)**(1 / 3) * a20
L10 = beta1 * T.sqrt(a10)
L20 = beta2 * T.sqrt(a20)
Psi = s * L20 + (1 + s) * L10
Rtilde = dRtilde - L10 - L20
####
# angles
####
rtilde = T.constant(0.)
Omega = -1 * rtilde
l1 = phi + k * (1 + s) * psi + Omega
l2 = phi + k * s * psi + Omega
gamma1 = s1 - phi - Omega
gamma2 = s2 - phi - Omega
q1 = 0.5 * np.pi - Omega
q2 = -0.5 * np.pi - Omega
pomega1 = -1 * gamma1
pomega2 = -1 * gamma2
Omega1 = -1 * q1
Omega2 = -1 * q2
omega1 = pomega1 - Omega1
omega2 = pomega2 - Omega2
###
# actions
###
Gamma1 = I1
Gamma2 = I2
L1 = Psi / k - s * (I1 + I2) - s * Phi
L2 = -1 * Psi / k + (1 + s) * (I1 + I2) + (1 + s) * Phi
Cz = -1 * Rtilde
R = L1 + L2 - Gamma1 - Gamma2 - Cz
G1 = L1 - Gamma1
G2 = L2 - Gamma2
r2_by_r1 = (L2 - L1 - Gamma2 + Gamma1) / (L1 + L2 - Gamma1 - Gamma2 - R)
rho1 = 0.5 * R * (1 + r2_by_r1)
rho2 = 0.5 * R * (1 - r2_by_r1)
a1 = (L1 / beta1)**2
e1 = T.sqrt(1 - (1 - (Gamma1 / L1))**2)
a2 = (L2 / beta2)**2
e2 = T.sqrt(1 - (1 - (Gamma2 / L2))**2)
cos_inc1 = 1 - rho1 / G1
cos_inc2 = 1 - rho2 / G2
inc1 = T.arccos(cos_inc1)
inc2 = T.arccos(cos_inc2)
Hkep = -0.5 * T.sqrt(eta1) * beta1 / a1 - 0.5 * T.sqrt(eta2) * beta2 / a2
ko = KeplerOp()
M1 = l1 - pomega1
M2 = l2 - pomega2
sinf1, cosf1 = ko(M1, e1 + T.zeros_like(M1))
sinf2, cosf2 = ko(M2, e2 + T.zeros_like(M2))
#
n1 = T.sqrt(eta1 / mstar) * a1**(-3 / 2)
n2 = T.sqrt(eta2 / mstar) * a2**(-3 / 2)
Hint_dir, Hint_ind, r1, r2, v1, v2 = calc_Hint_components_sinf_cosf(
a1, a2, e1, e2, inc1, inc2, omega1, omega2, Omega1, Omega2, n1, n2,
sinf1, cosf1, sinf2, cosf2)
eps = m1 * m2 / (mu1 + mu2) / T.sqrt(mstar)
Hpert = (Hint_dir + Hint_ind / mstar)
Hpert_av = Hpert.dot(quad_weights)
Htot = Hkep + eps * Hpert_av
#####################################################
# Set parameters for compiling functions with Theano
#####################################################
# Get numerical quadrature nodes and weights
nodes, weights = np.polynomial.legendre.leggauss(N_QUAD_PTS)
# Rescale for integration interval from [-1,1] to [-pi,pi]
nodes = nodes * np.pi
weights = weights * 0.5
# 'givens' will fix some parameters of Theano functions compiled below
givens = [(psi, nodes), (quad_weights, weights)]
# 'ins' will set the inputs of Theano functions compiled below
# Note: 'extra_ins' will be passed as values of object attributes
# of the 'ResonanceEquations' class 'defined below
extra_ins = [m1, m2, j, k]
ins = [dyvars] + extra_ins
orbels = [a1, e1, inc1, k * s1, a2, e2, inc2, k * s2, phi, Omega]
orbels_dict = dict(
zip([
'a1', 'e1', 'inc1', 'theta1', 'a2', 'e2', 'inc2', 'theta2', 'phi'
], orbels))
actions = [L1, L2, Gamma1, Gamma2, rho1, rho2]
actions_dict = dict(
zip(['L1', 'L2', 'Gamma1', 'Gamma2', 'Q1', 'Q2'], actions))
# Conservative flow
gradHtot = T.grad(Htot, wrt=dyvars)
hessHtot = theano.gradient.hessian(Htot, wrt=dyvars)
Jtens = T.as_tensor(
np.pad(_get_Omega_matrix(Ndof), (0, Nconst), 'constant'))
H_flow_vec = Jtens.dot(gradHtot)
H_flow_jac = Jtens.dot(hessHtot)
##########################
# Compile Theano functions
##########################
orbels_fn = theano.function(inputs=ins,
outputs=orbels_dict,
givens=givens,
on_unused_input='ignore')
actions_fn = theano.function(inputs=ins,
outputs=actions_dict,
givens=givens,
on_unused_input='ignore')
Rtilde_fn = theano.function(inputs=ins,
outputs=Rtilde,
givens=givens,
on_unused_input='ignore')
Htot_fn = theano.function(inputs=ins,
outputs=Htot,
givens=givens,
on_unused_input='ignore')
Hpert_fn = theano.function(inputs=ins,
outputs=Hpert_av,
givens=givens,
on_unused_input='ignore')
Hpert_components_fn = theano.function(
inputs=ins,
outputs=[Hint_dir.dot(quad_weights),
Hint_ind.dot(quad_weights)],
givens=givens,
on_unused_input='ignore')
H_flow_vec_fn = theano.function(inputs=ins,
outputs=H_flow_vec,
givens=givens,
on_unused_input='ignore')
H_flow_jac_fn = theano.function(inputs=ins,
outputs=H_flow_jac,
givens=givens,
on_unused_input='ignore')
return dict({
'orbital_elements': orbels_fn,
'actions': actions_fn,
'Rtilde': Rtilde_fn,
'Hamiltonian': Htot_fn,
'Hpert': Hpert_fn,
'Hpert_components': Hpert_components_fn,
'Hamiltonian_flow': H_flow_vec_fn,
'Hamiltonian_flow_jacobian': H_flow_jac_fn
})
def compute_pyramid(net, psize, step, interval, image, mean_pixel=None):
"""Compute a pyramid of CNN-derived features for the given image. Similar
to ``impyra.m``, except we haven't bothered upscaling, since Chen & Yuille
don't upscale anyway.
:param net: a ``caffe.Net`` instance corresponding to the fully
convolutional "deploy" model.
:param psize: parameter from Chen & Yuille. It's actually ``step * tsize``,
where ``tsize`` is a kind of "natural" template size computed from the
dimensions of skeletons in the training set. Unlike Chen & Yuille, we
use **ROW MAJOR ORDER** for ``psize``!
:param step: yet another parameter from Chen & Yuille. I actually have no
idea what this corresponds to, intuitively.
:param interval: how many pyramid levels does it take to halve the data
size?
:param image: ``h * w * c`` ``ndarray`` representing a single input image.
:param mean_pixel: optional mean pixel argument.
:returns: list of dictionaries with ``output_size``,
``{width,height}_pad``, ``scale`` and ``features`` keys. Each entry in
the list corresponds to a level of the feature pyramid (largest scale
first). The ``features`` key is an "image" representing the fully
convolutional netowrk output, where the number of channels in the image
is equal to the number of softmax outputs in the CNN."""
assert image.ndim == 3 and image.shape[2] == 3
if mean_pixel is None:
mean_pixel = 128 * np.ones((3,))
else:
# Flip the mean pixel to BGR
mean_pixel = mean_pixel[::-1]
height_pad, width_pad = np.maximum(np.ceil((psize - 1) / 2.0), 0)\
.astype('int')
scale = 2 ** (1.0 / interval)
image_size = np.array(image.shape[:2])
max_scale = int(1 + np.floor(np.log(np.min(image_size)) / np.log(scale)))
# This will have keys 'output_size', 'scale', 'height_pad', 'width_pad',
# 'features'
rv = [{} for _ in xrange(max_scale)]
# A natural size, I guess
max_batch_size = interval
for batch_level in xrange(0, max_scale, max_batch_size):
batch_size = np.min(max_batch_size, max_scale - batch_level)
base_dims = image_size / scale ** (batch_level)
scaled = cf.io.resize(image, base_dims.astype('int'))
# This next array will be passed to Caffe
caffe_input = np.zeros((
batch_size,
3,
scaled.shape[1] + 2 * height_pad,
scaled.shape[0] + 2 * width_pad,
))
for sublevel in xrange(batch_level, batch_level + batch_size):
# Pad and add to Caffe input
pad_dims = (2 * (height_pad,), 2 * (width_pad,), 2 * (0,))
padded = np.pad(scaled, pad_dims, mode='edge') - mean_pixel
max_row, max_col = padded.shape[:2]
caffe_input[sublevel - batch_level, :, :max_row, :max_col] = \
padded.transpose((2, 0, 1))
# Store metadata
info = rv[sublevel]
info['output_size'] = np.floor(
(padded.shape[:2] - psize) / float(step)
).astype('int') + 1
info['scale'] = step * scale ** (sublevel - 1)
info['width_pad'] = width_pad / float(step)
info['height_pad'] = height_pad / float(step)
# Resize for the next step
base_dims /= scale
scaled = cf.io.resize(image, base_dims.astype('int'))
# To do a fully convolutional forward pass, we just reshape the data
# layer and let the rest follow
net.blobs['data'].reshape(*caffe_input.shape)
net.blobs['data'].data[...] = caffe_input
# TODO: What does result contain? Apparently it's a dictionary mapping
# blob names to ndarrays for those blobs. In this case, I guess we'll
# have a batch_size * softmax_outputs * something * something_else
# ndarray, where something and something_else will be decided by
# some annoying arithmetic on strides, pads and steps. Ugh, gross.
result = net.forward()['prob']
for sublevel in xrange(batch_level, batch_level + batch_size):
info = rv[sublevel]
max_row, max_col = info['output_size']
info['features'] = result[
sublevel - batch_level, :, :max_row, :max_col
].transpose((1, 2, 0))
return rv
def MakeNdarray(tensor):
"""Create a numpy ndarray from a tensor.
Create a numpy ndarray with the same shape and data as the tensor.
For example:
```python
# Tensor a has shape (2,3)
a = tf.constant([[1,2,3],[4,5,6]])
proto_tensor = tf.make_tensor_proto(a) # convert `tensor a` to a proto tensor
tf.make_ndarray(proto_tensor) # output: array([[1, 2, 3],
# [4, 5, 6]], dtype=int32)
# output has shape (2,3)
```
Args:
tensor: A TensorProto.
Returns:
A numpy array with the tensor contents.
Raises:
TypeError: if tensor has unsupported type.
"""
shape = [d.size for d in tensor.tensor_shape.dim]
num_elements = np.prod(shape, dtype=np.int64)
tensor_dtype = dtypes.as_dtype(tensor.dtype)
dtype = tensor_dtype.as_numpy_dtype
if tensor.tensor_content:
return (np.frombuffer(tensor.tensor_content,
dtype=dtype).copy().reshape(shape))
if tensor_dtype == dtypes.string:
# np.pad throws on these arrays of type np.object.
values = list(tensor.string_val)
padding = num_elements - len(values)
if padding > 0:
last = values[-1] if values else ""
values.extend([last] * padding)
return np.array(values, dtype=dtype).reshape(shape)
if tensor_dtype == dtypes.float16 or tensor_dtype == dtypes.bfloat16:
# the half_val field of the TensorProto stores the binary representation
# of the fp16: we need to reinterpret this as a proper float16
values = np.fromiter(tensor.half_val, dtype=np.uint16)
values.dtype = tensor_dtype.as_numpy_dtype
elif tensor_dtype == dtypes.float32:
values = np.fromiter(tensor.float_val, dtype=dtype)
elif tensor_dtype == dtypes.float64:
values = np.fromiter(tensor.double_val, dtype=dtype)
elif tensor_dtype in [
dtypes.int32, dtypes.uint8, dtypes.uint16, dtypes.int16, dtypes.int8,
dtypes.qint32, dtypes.quint8, dtypes.qint8, dtypes.qint16, dtypes.quint16
]:
values = np.fromiter(tensor.int_val, dtype=dtype)
elif tensor_dtype == dtypes.int64:
values = np.fromiter(tensor.int64_val, dtype=dtype)
elif tensor_dtype == dtypes.complex64:
it = iter(tensor.scomplex_val)
values = np.array([complex(x[0], x[1]) for x in zip(it, it)], dtype=dtype)
elif tensor_dtype == dtypes.complex128:
it = iter(tensor.dcomplex_val)
values = np.array([complex(x[0], x[1]) for x in zip(it, it)], dtype=dtype)
elif tensor_dtype == dtypes.bool:
values = np.fromiter(tensor.bool_val, dtype=dtype)
else:
raise TypeError("Unsupported tensor type: %s" % tensor.dtype)
if values.size == 0:
return np.zeros(shape, dtype)
if values.size != num_elements:
values = np.pad(values, (0, num_elements - values.size), "edge")
return values.reshape(shape)
# Read image
image = cv2.imread('test_image.jpg')
# Resize image, if need
#(row_num_im, col_num_im, chan) = image.shape
#image = cv2.resize(image,(int(col_num_im/2), int(row_num_im/2)))
A = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY).astype('float64')
# Matrix for output
(row_num, col_num) = A.shape
B = np.zeros((row_num, col_num))
# padding for convinience
A_padding = np.pad(A, pad_width=1, mode='constant', constant_values=0)
# Sobel operator
filt_size = 3
Sx = np.array([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]])
Sy = np.array([[1, 2, 1], [0, 0, 0], [-1, -2, -1]])
for i in range(row_num):
for j in range(col_num):
# because we do padding, index changed
idx_i_A = i + 1
idx_j_A = j + 1
A_block = A_padding[idx_i_A - 1:idx_i_A + 2, idx_j_A - 1:idx_j_A + 2]
B[i, j] = np.sum(np.multiply(A_block, Sx)) + np.sum(
np.multiply(A_block, Sy))
B_min = np.amin(B)
def pad_x_single(x_single, max_length):
assert x_single.shape[0] <= max_length
pad_width_head = (0, max_length - x_single.shape[0])
pad_width_tail = ((0, 0),) * (x_single.ndim - 1)
return np.pad(x_single, (pad_width_head,) + pad_width_tail,
mode='constant', constant_values=pad_value)
