def preprocess(filename, num_resamplings=25):
# read data
#filename = "../data/MarieTherese_jul31_and_Aug07_all.pkl"
pkl_file = open(filename, 'rb')
data1 = cPickle.load(pkl_file)
num_strokes = len(data1)
# get the unique stroke labels, map to class labels (ints) for later using dictionary
stroke_dict = dict()
value_index = 0
for i in range(0, num_strokes):
current_key = data1[i][0]
if current_key not in stroke_dict:
stroke_dict[current_key] = value_index
value_index = value_index + 1
# save the dictionary to file, for later use
dict_filename = "../data/stroke_label_mapping.pkl"
dict_file = open(dict_filename, 'wb')
pickle.dump(stroke_dict, dict_file)
# - smooth data
# for each stroke, get the vector of data, smooth/interpolate it over time, store sampling from smoothed signal in vector
# - sample at regular intervals (1/30 of total time, etc.) -> input vector X
num_params = len(data1[0][1][0]) #accelx, accely, etc.
#num_params = 16 #accelx, accely, etc.
# re-sample the interpolated spline this many times (25 or so seems ok, since most letters have this many points)
#num_resamplings = 25
# the square of the number of resamplings, for shaping the X_2 into a vector
num_resamplings_2 = num_resamplings * num_resamplings
# build an output array large enough to hold the vectors for each stroke and the (unicode -> int) stroke value (1 elts)
# output_array = np.zeros((num_strokes, (num_resamplings_2 + num_resamplings) * num_params + 1))
output_array = np.zeros(
(num_strokes, (2 * num_resamplings) * num_params + 1))
print output_array.size
print filename
print num_params
print num_resamplings_2
print
for i in range(0, num_strokes):
# how far?
# if (i % 100 == 0):
# print float(i)/num_strokes
X_matrix = np.zeros(
(num_params, num_resamplings * 2)
) # the array to store in (using original data and 2 derivs, 2 integrals)
# the array to store reshaped resampled vector in
X_2_vector_scaled = np.zeros((num_params, num_resamplings_2))
# the array to store the above 2 concatenated
# concatenated_X_X_2 = np.zeros((num_params, num_resamplings_2 + num_resamplings))
concatenated_X_X_2 = np.zeros(
(num_params, num_resamplings * 2)
) # the array to store in (using original data and 2 derivs, 2 integrals)
# for each parameter (accelX, accelY, ...)
# map the unicode character to int
curr_stroke_val = stroke_dict[data1[i][0]]
#print(len(curr_stroke))
#print(curr_stroke[0])
#print(curr_stroke[1])
curr_data = data1[i][1]
# fix if too short for interpolation - pad current data with 3 zeros
if (len(curr_data) <= 3):
curr_data = np.concatenate([curr_data, np.zeros((3, num_params))])
time = np.arange(0, len(curr_data),
1) # the sample 'times' (0 to number of samples)
time_new = np.arange(0, len(curr_data),
float(len(curr_data)) /
num_resamplings) # the resampled time points
for j in range(0, num_params): # iterate through parameters
signal = curr_data[:,
j] # one signal (accelx, etc.) to interpolate
# interpolate the signal using a spline or so, so that arbitrary points can be used
# (~30 seems reasonable based on data, for example)
#tck = interpolate.splrep(time, signal, s=0) # the interpolation represenation
tck = UnivariateSpline(time, signal, s=0)
# sample the interpolation num_resamplings times to get values
# resampled_data = interpolate.splev(time_new, tck, der=0) # the resampled data
# resampled_data = tck(time_new)
# scale data (center, norm)
# resampled_data = preprocessing.scale(resampled_data)
# first integral
# tck.integral = tck.antiderivative()
# resampled_data_integral = tck.integral(time_new)
# scale data (center, norm)
# resampled_data_integral = preprocessing.scale(resampled_data_integral)
# 2nd integral
# tck.integral_2 = tck.antiderivative(2)
# resampled_data_integral_2 = tck.integral_2(time_new)
# scale data (center, norm)
# resampled_data_integral_2 = preprocessing.scale(resampled_data_integral_2)
# first deriv
tck.deriv = tck.derivative()
resampled_data_deriv = tck.deriv(time_new)
# scale
resampled_data_deriv = preprocessing.scale(resampled_data_deriv)
# second deriv
tck.deriv_2 = tck.derivative(2)
resampled_data_deriv_2 = tck.deriv_2(time_new)
#scale
resampled_data_deriv_2 = preprocessing.scale(
resampled_data_deriv_2)
# concatenate into one vector
concatenated_resampled_data = np.concatenate(
(resampled_data_deriv, resampled_data_deriv_2))
# store for the correct parameter, to be used later as part of inputs to SVM
X_matrix[j] = concatenated_resampled_data
# while we're at it, square vector of resampled data to get a matrix, vectorize the matrix, and store
# for each X in list, multiply X by itself -> X_2
#- vectorize X^2 (e.g. 10 x 10 -> 100 dimensions)
# X_2_matrix = np.outer(concatenated_resampled_data, concatenated_resampled_data) # temp matrix for outer product
# X_2_vector = np.reshape(X_2_matrix, -1) # reshape into a vector
#- center and normalize X^2 by mean and standard deviation
# X_2_vector_scaled[j] = preprocessing.scale(X_2_vector)
#- concatenate with input X -> 110 dimensions
# concatenated_X_X_2[j] = np.concatenate([X_matrix[j], X_2_vector_scaled[j]])
# FOR NOW, ONLY USE X, NOT OUTER PRODUCT
concatenated_X_X_2[j] = X_matrix[j]
# NOTE, THIS SHOULD REALLY JUST BE A BIG VECTOR FOR EACH STROKE, SO RESHAPE BEFORE ADDING TO OUTPUT LIST
# ALSO, THE STROKE VALUE SHOULD BE ADDED
this_sample = np.concatenate(
(np.reshape(concatenated_X_X_2, -1), np.array([curr_stroke_val])))
concatenated_samples = np.reshape(this_sample, -1)
# ADD TO OUTPUT ARRAY
output_array[i] = concatenated_samples
print(output_array.size)
return (output_array)
def preprocess(filename, num_resamplings = 25):
# read data
#filename = "../data/MarieTherese_jul31_and_Aug07_all.pkl"
pkl_file = open(filename, 'rb')
data1 = cPickle.load(pkl_file)
num_strokes = len(data1)
# get the unique stroke labels, map to class labels (ints) for later using dictionary
stroke_dict = dict()
value_index = 0
for i in range(0,num_strokes):
current_key = data1[i][0]
if current_key not in stroke_dict:
stroke_dict[current_key] = value_index
value_index = value_index + 1
# save the dictionary to file, for later use
dict_filename = "../data/stroke_label_mapping.pkl"
dict_file = open(dict_filename, 'wb')
pickle.dump(stroke_dict, dict_file)
# - smooth data
# for each stroke, get the vector of data, smooth/interpolate it over time, store sampling from smoothed signal in vector
# - sample at regular intervals (1/30 of total time, etc.) -> input vector X
num_params = len(data1[0][1][0]) #accelx, accely, etc.
#num_params = 16 #accelx, accely, etc.
# re-sample the interpolated spline this many times (25 or so seems ok, since most letters have this many points)
# build an output array large enough to hold the vectors for each stroke and the (unicode -> int) stroke value (1 elts)
# output_array = np.zeros((num_strokes, (num_resamplings_2 + num_resamplings) * num_params + 1))
output_array = np.zeros((num_strokes, (5 * num_resamplings) * num_params + 1))
print output_array.size
print filename
print num_params
print num_resamplings_2
print
for i in range(0, num_strokes):
# how far?
if (i % 100 == 0):
print float(i)/num_strokes
X_matrix = np.zeros((num_params, num_resamplings * 5)) # the array to store in (using original data and 2 derivs, 2 integrals)
# the array to store reshaped resampled vector in
X_2_vector_scaled = np.zeros((num_params, num_resamplings_2))
# the array to store the above 2 concatenated
# concatenated_X_X_2 = np.zeros((num_params, num_resamplings_2 + num_resamplings))
concatenated_X_X_2 = np.zeros((num_params, num_resamplings * 5)) # the array to store in (using original data and 2 derivs, 2 integrals)
# for each parameter (accelX, accelY, ...)
# map the unicode character to int
curr_stroke_val = stroke_dict[data1[i][0]]
#print(len(curr_stroke))
#print(curr_stroke[0])
#print(curr_stroke[1])
curr_data = data1[i][1]
# fix if too short for interpolation - pad current data with 3 zeros
if(len(curr_data) <= 3):
curr_data = np.concatenate([curr_data, np.zeros((3,num_params))])
time = np.arange(0, len(curr_data), 1) # the sample 'times' (0 to number of samples)
time_new = np.arange(0, len(curr_data), float(len(curr_data))/num_resamplings) # the resampled time points
for j in range(0, num_params): # iterate through parameters
signal = curr_data[:,j] # one signal (accelx, etc.) to interpolate
# interpolate the signal using a spline or so, so that arbitrary points can be used
# (~30 seems reasonable based on data, for example)
#tck = interpolate.splrep(time, signal, s=0) # the interpolation represenation
tck = UnivariateSpline(time, signal, s=0)
# sample the interpolation num_resamplings times to get values
# resampled_data = interpolate.splev(time_new, tck, der=0) # the resampled data
resampled_data = tck(time_new)
# scale data (center, norm)
resampled_data = preprocessing.scale(resampled_data)
# first integral
tck.integral = tck.antiderivative()
resampled_data_integral = tck.integral(time_new)
# scale data (center, norm)
resampled_data_integral = preprocessing.scale(resampled_data_integral)
# 2nd integral
tck.integral_2 = tck.antiderivative(2)
resampled_data_integral_2 = tck.integral_2(time_new)
# scale data (center, norm)
resampled_data_integral_2 = preprocessing.scale(resampled_data_integral_2)
# first deriv
tck.deriv = tck.derivative()
resampled_data_deriv = tck.deriv(time_new)
# scale
resampled_data_deriv = preprocessing.scale(resampled_data_deriv)
# second deriv
tck.deriv_2 = tck.derivative(2)
resampled_data_deriv_2 = tck.deriv_2(time_new)
#scale
resampled_data_deriv_2 = preprocessing.scale(resampled_data_deriv_2)
# concatenate into one vector
concatenated_resampled_data = np.concatenate((resampled_data,
resampled_data_integral,
resampled_data_integral_2,
resampled_data_deriv,
resampled_data_deriv_2))
# store for the correct parameter, to be used later as part of inputs to SVM
X_matrix[j] = concatenated_resampled_data
# while we're at it, square vector of resampled data to get a matrix, vectorize the matrix, and store
# for each X in list, multiply X by itself -> X_2
#- vectorize X^2 (e.g. 10 x 10 -> 100 dimensions)
# X_2_matrix = np.outer(concatenated_resampled_data, concatenated_resampled_data) # temp matrix for outer product
# X_2_vector = np.reshape(X_2_matrix, -1) # reshape into a vector
#- center and normalize X^2 by mean and standard deviation
# X_2_vector_scaled[j] = preprocessing.scale(X_2_vector)
#- concatenate with input X -> 110 dimensions
# concatenated_X_X_2[j] = np.concatenate([X_matrix[j], X_2_vector_scaled[j]])
# FOR NOW, ONLY USE X, NOT OUTER PRODUCT
concatenated_X_X_2[j] = X_matrix[j]
# NOTE, THIS SHOULD REALLY JUST BE A BIG VECTOR FOR EACH STROKE, SO RESHAPE BEFORE ADDING TO OUTPUT LIST
# ALSO, THE STROKE VALUE SHOULD BE ADDED
this_sample = np.concatenate((np.reshape(concatenated_X_X_2, -1), np.array([curr_stroke_val])))
concatenated_samples = np.reshape(this_sample, -1)
# ADD TO OUTPUT ARRAY
output_array[i] = concatenated_samples
print(output_array.size)
return(output_array)
