def experiment_50_50_split(ndims, avg_variant, date_of_wordvecs,
subtract_temporal_average_bool):
if avg_variant == 'avg':
fmris = load_avg(1) # 1 because there's no averaging (see load_fmri)
elif avg_variant == 'pca':
fmris = load_avg_pca(ndims, 1)
elif avg_variant == 'srm' or avg_variant == 'srmica':
# here, srm means we don't do the testing over T iterations
# we just use the average of all of them without error bars
if avg_variant == 'srm':
srms = load_srm(ndims, 1)
elif avg_variant == 'srmica':
srms = load_srmica(ndims, 1)
# calculate the average for each mask, and replace with that
fmris = {}
for mask in srms.keys():
#if mask != 'erez_dmna_network':
# continue
srm = srms[mask]
Ws = srm[0] # the map
S = srm[1] # averaged SRM-projected training data
num_training_time_steps = S.shape[1]
tst_data_list = srm[2] # test portion of the data, pre-selected
#print "tst_data_list length = " + str(len(tst_data_list))
num_subjs = len(tst_data_list)
avg_srm_tst_data = None
for i in xrange(num_subjs):
#print "Subject #"+ str(i)
tst_data = tst_data_list[i]
#print "Test data shape = " + str(tst_data.shape)
transformed = np.dot(Ws[i, :, :].T, tst_data)
if i == 0:
avg_srm_tst_data = transformed / (0. + num_subjs)
else:
avg_srm_tst_data += transformed / (0. + num_subjs)
fmri_mask = np.c_[S, avg_srm_tst_data]
fmris[mask] = fmri_mask
else:
print 'avg_variant is wrong'
return
if date_of_wordvecs == 'may15':
semantic_vecs = load_may15_annotation_vecs()
elif date_of_wordvecs == 'sep12weighted':
semantic_vecs = load_sep12_weighted_word_vecs_annotations()
elif date_of_wordvecs == 'sep12unweighted':
semantic_vecs = load_sep12_unweighted_word_vecs_annotations()
else:
# default
semantic_vecs = load_sep12_weighted_word_vecs_annotations()
print "Semantic vector shape = " + str(semantic_vecs.shape)
num_sem_vecs = semantic_vecs.shape[1] # assume time is # cols
# don't subtract out average of all semantic vectors for all time points
# only calculate average for training, and subtract that average out of the test
# CHANGE MADE: sep 19 12:40 AM
'''
# subtract out average if we're doing this case
if subtract_temporal_average_bool == True:
semantic_vecs, avg_semantic_vec = subtract_column_mean(semantic_vecs)
'''
#print "num semantic vectors = " + str(num_sem_vecs)
word_tr = semantic_vecs[:, xrange(num_sem_vecs / 2 + 1)]
#print "word_tr.shape = " + str(word_tr.shape)
word_tst = semantic_vecs[:, xrange(num_sem_vecs / 2 + 1, num_sem_vecs)]
#print "word_tst.shape = " + str(word_tst.shape)
#print "word_tr.shape = " + str(word_tr.shape)
#print "word_tst.shape = " + str(word_tst.shape)
# don't subtract out average of all semantic vectors for all time points
# only calculate average for training, and subtract that average out of the test
# CHANGE MADE: sep 19 12:40 AM
if subtract_temporal_average_bool == True:
word_tr, avg_tr_word_vec = subtract_column_mean(word_tr)
word_tst = word_tst - avg_tr_word_vec[:, np.newaxis]
mask_results = {}
for mask in fmris.keys():
fmri = fmris[mask]
num_total_time_steps = fmri.shape[1]
#print "fmri.shape = " + str(fmri.shape)
fmri_tr = fmri[:, xrange(num_sem_vecs / 2 + 1)]
fmri_tst = fmri[:, xrange(num_sem_vecs / 2 + 1, num_sem_vecs)]
#print "fmri_tr.shape = " + str(fmri_tr.shape)
#print "fmri_tst.shape = " + str(fmri_tst.shape)
#print "Training linear maps..."
Wridge_ft, Wridge_tf, Wpro_ft, Wpro_tf = learn_linear_maps(
xrange(num_sem_vecs / 2 + 1), [], fmri_tr, mask + avg_variant,
word_tr, date_of_wordvecs)
#print "Done training linear maps..."
TF_class_scores = []
TF_rank_scores = []
FT_class_scores = []
FT_rank_scores = []
num_time_steps = fmri_tst.shape[
1] # assume time steps are # of columns
chunk_array = []
index = 0
num_chunks = 25 ### CHANGED JAN 11
while index < num_time_steps:
chunk_array.append(index)
index += num_time_steps / num_chunks
if index > num_time_steps:
index = num_time_steps
k = 1 #-> 4% chance rate ## CHANGED FEB 4
# comparisons in fMRI space (i.e. text -> fMRI) (best was ridge, use that map first)
# truth is fmri_tst
# prediction is Wridge_tf*word_tst
# flipped is using procrustes instead
TF_truth = fmri_tst
TF_prediction = np.dot(Wridge_tf, word_tst)
TF_prediction_flipped = np.dot(Wpro_tf, word_tst)
TF_prediction = TF_prediction[0, :, :]
TF_prediction_flipped = TF_prediction_flipped[0, :, :]
#print "TF pred shape: " + str(TF_prediction.shape)
#print "TF pred flipped shape: " + str(TF_prediction_flipped.shape)
#print "TF truth shape: " + str(TF_truth.shape)
assert (TF_truth.shape == TF_prediction.shape)
assert (TF_truth.shape == TF_prediction_flipped.shape)
TF_classification_score = scene_classification(TF_truth, TF_prediction,
chunk_array, k)
TF_classification_score_flipped = scene_classification(
TF_truth, TF_prediction_flipped, chunk_array, k)
TF_rank_score, TF_ranks = scene_ranking(TF_truth, TF_prediction,
chunk_array)
TF_rank_score_flipped, TF_ranks_flipped = scene_ranking(
TF_truth, TF_prediction_flipped, chunk_array)
TF_rank_score = TF_rank_score / (num_chunks + 0.)
TF_rank_score_flipped = TF_rank_score_flipped / (num_chunks + 0.)
# comparisons in Semantic space (i.e. fMRI -> text) (best was procrustes, use that map first)
# truth is word_tst
# prediction is Wpro_ft*fmri_tst
# flipped is using ridge instead
FT_truth = word_tst
FT_prediction = np.dot(Wpro_ft, fmri_tst)
FT_prediction_flipped = np.dot(Wridge_ft, fmri_tst)
FT_prediction = FT_prediction[0, :, :]
FT_prediction_flipped = FT_prediction_flipped[0, :, :]
#print "FT pred shape: " + str(FT_prediction.shape)
#print "FT pred flipped shape: " + str(FT_prediction_flipped.shape)
#print "FT truth shape: " + str(FT_truth.shape)
assert (FT_truth.shape == FT_prediction.shape)
assert (FT_truth.shape == FT_prediction_flipped.shape)
FT_classification_score = scene_classification(FT_truth, FT_prediction,
chunk_array, k)
FT_classification_score_flipped = scene_classification(
FT_truth, FT_prediction_flipped, chunk_array, k)
FT_rank_score, FT_ranks = scene_ranking(FT_truth, FT_prediction,
chunk_array)
FT_rank_score_flipped, FT_ranks_flipped = scene_ranking(
FT_truth, FT_prediction_flipped, chunk_array)
FT_rank_score = FT_rank_score / (num_chunks + 0.)
FT_rank_score_flipped = FT_rank_score_flipped / (num_chunks + 0.)
print "Mask: " + mask
print "Using Ridge for Text -> fMRI and Procrustes for fMRI -> Text"
print "Text -> fMRI (Ridge) scene classification (" + str(
float(k) /
num_chunks) + "% chance) avg = " + str(TF_classification_score)
print "Text -> fMRI (Ridge) scene ranking (50% chance) avg = " + str(
TF_rank_score)
print "fMRI -> Text (Procrustes) scene classification (" + str(
float(k) /
num_chunks) + "% chance) avg = " + str(FT_classification_score)
print "fMRI -> Text (Procrustes) scene ranking (50% chance) avg = " + str(
FT_rank_score)
print "---------------------------------------------------------------------------"
print "Using Procrustes for Text -> fMRI and Ridge for fMRI -> Text"
print "Text -> fMRI (Procrustes) scene classification (" + str(
float(k) / num_chunks) + "% chance) avg = " + str(
TF_classification_score_flipped)
print "Text -> fMRI (Procrustes) scene ranking (50% chance) avg = " + str(
TF_rank_score_flipped)
print "fMRI -> Text (Ridge) scene classification (" + str(
float(k) / num_chunks) + "% chance) avg = " + str(
FT_classification_score_flipped)
print "fMRI -> Text (Ridge) scene ranking (50% chance) avg = " + str(
FT_rank_score_flipped)
print "///////////////////////////////////////////////////////////////////////////"
mask_results[mask] = (TF_classification_score, TF_rank_score,
FT_classification_score, FT_rank_score,
TF_classification_score_flipped,
TF_rank_score_flipped,
FT_classification_score_flipped,
FT_rank_score_flipped)
return mask_results
def past_timesteps_experiment_50_50_split(ndims, improper_classification,
num_previous, avg_variant,
date_of_wordvecs,
subtract_temporal_average_bool):
if avg_variant == 'avg':
print "AVG version is not supported for previous timestep mode"
return
elif avg_variant == 'pca':
fmris = load_avg_pca(ndims, 1)
for mask in fmris.keys():
fmri = fmris[mask]
print "size of fmri PCA: " + str(fmri.shape)
# assume that time = #cols
num_time_steps = fmri.shape[1]
fmri_tr = fmri[:, xrange(num_time_steps / 2 + 1)]
fmri_tst = fmri[:, xrange(num_time_steps / 2 + 1, num_time_steps)]
fmri_tr = add_prev_time_steps(fmri_tr, num_previous)
fmri_tst = add_prev_time_steps(fmri_tst, num_previous)
new_fmri = np.c_[fmri_tr, fmri_tst]
fmris[mask] = new_fmri
elif avg_variant == 'srm' or avg_variant == 'srmica':
# here, srm means we don't do the testing over T iterations
# we just use the average of all of them without error bars
if avg_variant == 'srm':
srms = load_srm(ndims, 1)
elif avg_variant == 'srmica':
srms = load_srmica(ndims, 1)
# calculate the average for each mask, and replace with that
fmris = {}
for mask in srms.keys():
#if mask != 'erez_dmna_network':
# continue
srm = srms[mask]
Ws = srm[0] # the map
S = srm[1] # averaged SRM-projected training data
num_training_time_steps = S.shape[1]
tst_data_list = srm[2] # test portion of the data, pre-selected
#print "tst_data_list length = " + str(len(tst_data_list))
num_subjs = len(tst_data_list)
avg_srm_tst_data = None
for i in xrange(num_subjs):
#print "Subject #"+ str(i)
tst_data = tst_data_list[i]
#print "Test data shape = " + str(tst_data.shape)
transformed = np.dot(Ws[i, :, :].T, tst_data)
if i == 0:
avg_srm_tst_data = transformed / (0. + num_subjs)
else:
avg_srm_tst_data += transformed / (0. + num_subjs)
# add previous time steps
# note that we add previous time steps SEPARATELY to training and test
# this means that there is NO OVERLAP between training and test in the previous time steps.
S_prev_times = add_prev_time_steps(S, num_previous)
avg_srm_prev_times = add_prev_time_steps(avg_srm_tst_data,
num_previous)
#print "shape of S_prev_times = " + str(S_prev_times.shape)
#print "shape of avg_srm_prev_times = " + str(avg_srm_prev_times.shape)
fmri_mask = np.c_[S_prev_times, avg_srm_prev_times]
fmris[mask] = fmri_mask
else:
print 'avg_variant is wrong'
return
if date_of_wordvecs == 'may15':
semantic_vecs = load_may15_annotation_vecs()
elif date_of_wordvecs == 'sep12weighted':
semantic_vecs = load_sep12_weighted_word_vecs_annotations()
elif date_of_wordvecs == 'sep12unweighted':
semantic_vecs = load_sep12_unweighted_word_vecs_annotations()
else:
# default
semantic_vecs = load_sep12_weighted_word_vecs_annotations()
print "Semantic vector shape = " + str(semantic_vecs.shape)
num_sem_vecs = semantic_vecs.shape[1] # assume time is # cols
# don't subtract out average of all semantic vectors for all time points
# only calculate average for training, and subtract that average out of the test
# CHANGE MADE: sep 19 12:40 AM
'''
# subtract out average if we're doing this case
if subtract_temporal_average_bool == True:
semantic_vecs, avg_semantic_vec = subtract_column_mean(semantic_vecs)
'''
#print "num semantic vectors = " + str(num_sem_vecs)
word_tr = semantic_vecs[:, xrange(num_sem_vecs / 2 + 1)]
#print "word_tr.shape = " + str(word_tr.shape)
word_tst = semantic_vecs[:, xrange(num_sem_vecs / 2 + 1, num_sem_vecs)]
#print "word_tst.shape = " + str(word_tst.shape)
#print "word_tr.shape = " + str(word_tr.shape)
#print "word_tst.shape = " + str(word_tst.shape)
# don't subtract out average of all semantic vectors for all time points
# only calculate average for training, and subtract that average out of the test
# CHANGE MADE: sep 19 12:40 AM
if subtract_temporal_average_bool == True:
word_tr, avg_tr_word_vec = subtract_column_mean(word_tr)
word_tst = word_tst - avg_tr_word_vec[:, np.newaxis]
# add previous time steps to semantic stuff
word_tr = add_prev_time_steps(word_tr, num_previous)
word_tst = add_prev_time_steps(word_tst, num_previous)
mask_results = {}
for mask in fmris.keys():
fmri = fmris[mask]
num_total_time_steps = fmri.shape[1]
#print "fmri.shape = " + str(fmri.shape)
fmri_tr = fmri[:, xrange(num_sem_vecs / 2 + 1)]
fmri_tst = fmri[:, xrange(num_sem_vecs / 2 + 1, num_sem_vecs)]
print "fmri_tr.shape = " + str(fmri_tr.shape)
print "fmri_tst.shape = " + str(fmri_tst.shape)
print "Training linear maps..."
if improper_classification == None:
# truncate one side before learning linear map (for each direction)
# modify fmri_tr, word_tr (aren't used again)
long_fmri_tr = fmri_tr.copy()
long_word_tr = word_tr.copy()
short_fmri_tr = keep_only_first_time_step(fmri_tr, num_previous)
short_word_tr = keep_only_first_time_step(word_tr, num_previous)
# learn long fmri -> short text
Wridge_ft, useless1, Wpro_ft, useless2 = learn_linear_maps(
xrange(num_sem_vecs / 2 + 1), [], long_fmri_tr,
mask + avg_variant, short_word_tr, date_of_wordvecs)
# learn long text -> short fmri
useless1, Wridge_tf, useless2, Wpro_tf = learn_linear_maps(
xrange(num_sem_vecs / 2 + 1), [], short_fmri_tr,
mask + avg_variant, long_word_tr, date_of_wordvecs)
#clear memory
useless1 = None
useless2 = None
long_fmri_tr = None
long_word_tr = None
short_fmri_tr = None
short_word_tr = None
else:
Wridge_ft, Wridge_tf, Wpro_ft, Wpro_tf = learn_linear_maps(
xrange(num_sem_vecs / 2 + 1), [], fmri_tr, mask + avg_variant,
word_tr, date_of_wordvecs)
print "Done training linear maps..."
print "Wridge_ft.shape = " + str(Wridge_ft[0].shape)
print "Wridge_tf.shape = " + str(Wridge_tf[0].shape)
print "Wpro_ft.shape = " + str(Wpro_ft[0].shape)
print "Wpro_tf.shape = " + str(Wpro_tf[0].shape)
TF_class_scores = []
TF_rank_scores = []
FT_class_scores = []
FT_rank_scores = []
num_time_steps = fmri_tst.shape[
1] # assume time steps are # of columns
chunk_array = []
index = 0
num_chunks = 25 # CHANGED JAN 11
while index < num_time_steps:
chunk_array.append(index)
index += num_time_steps / num_chunks
if index > num_time_steps:
index = num_time_steps
k = 1 #-> 4% chance rate
if improper_classification == True or improper_classification == False:
# comparisons in fMRI space (i.e. text -> fMRI) (best was ridge, use that map first)
# truth is fmri_tst
# prediction is Wridge_tf*word_tst
# flipped is using procrustes instead
TF_prediction = np.dot(Wridge_tf, word_tst)
TF_prediction_flipped = np.dot(Wpro_tf, word_tst)
# comparisons in Semantic space (i.e. fMRI -> text) (best was procrustes, use that map first)
# truth is word_tst
# prediction is Wpro_ft*fmri_tst
# flipped is using ridge instead
FT_prediction = np.dot(Wpro_ft, fmri_tst)
FT_prediction_flipped = np.dot(Wridge_ft, fmri_tst)
# fix shape
TF_prediction = TF_prediction[0, :, :]
TF_prediction_flipped = TF_prediction_flipped[0, :, :]
FT_prediction = FT_prediction[0, :, :]
FT_prediction_flipped = FT_prediction_flipped[0, :, :]
elif improper_classification == None:
long_fmri_tst = fmri_tst
long_word_tst = word_tst
# these linear maps will already be in the short space
# so the predictions don't need to be truncated, they're already the right size
TF_prediction = np.dot(Wridge_tf, long_word_tst)
TF_prediction_flipped = np.dot(Wpro_tf, long_word_tst)
FT_prediction = np.dot(Wpro_ft, long_fmri_tst)
FT_prediction_flipped = np.dot(Wridge_ft, long_fmri_tst)
# fix shape
TF_prediction = TF_prediction[0, :, :]
TF_prediction_flipped = TF_prediction_flipped[0, :, :]
FT_prediction = FT_prediction[0, :, :]
FT_prediction_flipped = FT_prediction_flipped[0, :, :]
TF_truth = keep_only_first_time_step(fmri_tst, num_previous)
FT_truth = keep_only_first_time_step(word_tst, num_previous)
Wridge_tf = None
Wpro_tf = None
Wpro_ft = None
Wridge_ft = None
if improper_classification == True:
# no truncation
TF_truth = fmri_tst.copy()
FT_truth = word_tst.copy()
# Truncate after learning linear maps btwn long and long
elif improper_classification == False:
# evaluate in fMRI space
TF_prediction = keep_only_first_time_step(TF_prediction,
num_previous)
TF_prediction_flipped = keep_only_first_time_step(
TF_prediction_flipped, num_previous)
# evaluate in text space
FT_prediction = keep_only_first_time_step(FT_prediction,
num_previous)
FT_prediction_flipped = keep_only_first_time_step(
FT_prediction_flipped, num_previous)
TF_truth = keep_only_first_time_step(fmri_tst, num_previous)
FT_truth = keep_only_first_time_step(word_tst, num_previous)
# check everything is the right size
print "TF_prediction.shape = " + str(TF_prediction.shape)
print "TF_prediction_flipped.shape = " + str(
TF_prediction_flipped.shape)
print "TF truth shape: " + str(TF_truth.shape)
assert (TF_truth.shape == TF_prediction.shape)
assert (TF_truth.shape == TF_prediction_flipped.shape)
print "FT_prediction.shape = " + str(FT_prediction.shape)
print "FT_prediction_flipped.shape = " + str(
FT_prediction_flipped.shape)
print "FT truth shape: " + str(FT_truth.shape)
assert (FT_truth.shape == FT_prediction.shape)
assert (FT_truth.shape == FT_prediction_flipped.shape)
# CALCULATE SCORES
TF_classification_score = scene_classification(TF_truth, TF_prediction,
chunk_array, k)
TF_classification_score_flipped = scene_classification(
TF_truth, TF_prediction_flipped, chunk_array, k)
TF_rank_score, TF_ranks = scene_ranking(TF_truth, TF_prediction,
chunk_array)
TF_rank_score_flipped, TF_ranks_flipped = scene_ranking(
TF_truth, TF_prediction_flipped, chunk_array)
TF_rank_score = TF_rank_score / (num_chunks + 0.)
TF_rank_score_flipped = TF_rank_score_flipped / (num_chunks + 0.)
FT_classification_score = scene_classification(FT_truth, FT_prediction,
chunk_array, k)
FT_classification_score_flipped = scene_classification(
FT_truth, FT_prediction_flipped, chunk_array, k)
FT_rank_score, FT_ranks = scene_ranking(FT_truth, FT_prediction,
chunk_array)
FT_rank_score_flipped, FT_ranks_flipped = scene_ranking(
FT_truth, FT_prediction_flipped, chunk_array)
FT_rank_score = FT_rank_score / (num_chunks + 0.)
FT_rank_score_flipped = FT_rank_score_flipped / (num_chunks + 0.)
# DISPLAY SCORES
print "---------------------------------------------------------------------------"
print "Mask: " + mask
print "Using Ridge for Text -> fMRI and Procrustes for fMRI -> Text"
print "Text -> fMRI (Ridge) scene classification (" + str(
float(k) * 100 /
num_chunks) + "% chance) avg = " + str(TF_classification_score)
print "Text -> fMRI (Ridge) scene ranking (50% chance) avg = " + str(
TF_rank_score)
print "fMRI -> Text (Procrustes) scene classification (" + str(
float(k) * 100 /
num_chunks) + "% chance) avg = " + str(FT_classification_score)
print "fMRI -> Text (Procrustes) scene ranking (50% chance) avg = " + str(
FT_rank_score)
print "---------------------------------------------------------------------------"
print "Using Procrustes for Text -> fMRI and Ridge for fMRI -> Text"
print "Text -> fMRI (Procrustes) scene classification (" + str(
float(k) * 100 / num_chunks) + "% chance) avg = " + str(
TF_classification_score_flipped)
print "Text -> fMRI (Procrustes) scene ranking (50% chance) avg = " + str(
TF_rank_score_flipped)
print "fMRI -> Text (Ridge) scene classification (" + str(
float(k) * 100 / num_chunks) + "% chance) avg = " + str(
FT_classification_score_flipped)
print "fMRI -> Text (Ridge) scene ranking (50% chance) avg = " + str(
FT_rank_score_flipped)
print "---------------------------------------------------------------------------"
print "///////////////////////////////////////////////////////////////////////////"
mask_results[mask] = (TF_classification_score, TF_rank_score,
FT_classification_score, FT_rank_score,
TF_classification_score_flipped,
TF_rank_score_flipped,
FT_classification_score_flipped,
FT_rank_score_flipped)
return mask_results
import numpy as np from path_to_RESULTS import PATH_TO_RESULTS from load_text_data import load_sep12_weighted_word_vecs_annotations, subtract_column_mean from load_fMRI_data import load_srmica, load_srm from performance_metrics import pearson_r, cosdist import matplotlib.pyplot as plt # pearson_r basically subtracts the mean out, so you could technically use it without calculating # the subtracted annotation mean. # what this means is that cosine distance for mean-subtracted vectors = pearson_r for either mean-subtracted # or not mean-subtracted vectors calculated_fmri_text_correlations = True if not calculated_fmri_text_correlations: # subtract mean sub_mean_annotation_vecs, avg_vec = subtract_column_mean(load_sep12_weighted_word_vecs_annotations()) #print "avg vec= "+ str(avg_vec) #print "avg vec norm = " + str(np.linalg.norm(avg_vec)) annotation_vecs = sub_mean_annotation_vecs print "Annotation vector shape: " + str(annotation_vecs.shape) #SRM = load_srm(20, 0) SRMICA = load_srmica(20, 0) fMRI_vecs = SRMICA["erez_dmna_network"] Ws = fMRI_vecs[0] S = fMRI_vecs[1] tst = fMRI_vecs[2] for subj in range(16): if subj == 0:
def semantic_time_weighting_experiment_50_50_split(
fmri_or_text_time_weights, timestep_weighting, learn_weights,
avg_variant, date_of_wordvecs, subtract_temporal_average_bool):
# if we're learning the weights, then timestep_weighting is irrelevant
if learn_weights == True:
assert (timestep_weighting == None)
assert ((fmri_or_text_time_weights == "fmri")
or (fmri_or_text_time_weights == "text"))
if avg_variant == 'avg':
fmris = load_avg(1) # 1 because there's no averaging (see load_fmri)
elif avg_variant == 'pca':
fmris = load_avg_pca(1)
elif avg_variant == 'srm' or avg_variant == 'srmica':
# here, srm means we don't do the testing over T iterations
# we just use the average of all of them without error bars
if avg_variant == 'srm':
srms = load_srm(1)
elif avg_variant == 'srmica':
srms = load_srmica(1)
# calculate the average for each mask, and replace with that
fmris = {}
for mask in srms.keys():
#if mask != 'erez_dmna_network':
# continue
srm = srms[mask]
Ws = srm[0] # the map
S = srm[1] # averaged SRM-projected training data
num_training_time_steps = S.shape[1]
tst_data_list = srm[2] # test portion of the data, pre-selected
#print "tst_data_list length = " + str(len(tst_data_list))
num_subjs = len(tst_data_list)
avg_srm_tst_data = None
for i in xrange(num_subjs):
#print "Subject #"+ str(i)
tst_data = tst_data_list[i]
#print "Test data shape = " + str(tst_data.shape)
transformed = np.dot(Ws[i, :, :].T, tst_data)
if i == 0:
avg_srm_tst_data = transformed / (0. + num_subjs)
else:
avg_srm_tst_data += transformed / (0. + num_subjs)
fmri_mask = np.c_[S, avg_srm_tst_data]
fmris[mask] = fmri_mask
else:
print 'avg_variant is wrong'
return
if date_of_wordvecs == 'may15':
semantic_vecs = load_may15_annotation_vecs()
elif date_of_wordvecs == 'sep12weighted':
semantic_vecs = load_sep12_weighted_word_vecs_annotations()
elif date_of_wordvecs == 'sep12unweighted':
semantic_vecs = load_sep12_unweighted_word_vecs_annotations()
else:
# default
semantic_vecs = load_sep12_weighted_word_vecs_annotations()
print "Semantic vector shape = " + str(semantic_vecs.shape)
num_sem_vecs = semantic_vecs.shape[1] # assume time is # cols
# don't subtract out average of all semantic vectors for all time points
# only calculate average for training, and subtract that average out of the test
# CHANGE MADE: sep 19 12:40 AM
'''
# subtract out average if we're doing this case
if subtract_temporal_average_bool == True:
semantic_vecs, avg_semantic_vec = subtract_column_mean(semantic_vecs)
'''
#print "num semantic vectors = " + str(num_sem_vecs)
word_tr = semantic_vecs[:, xrange(num_sem_vecs / 2 + 1)]
#print "word_tr.shape = " + str(word_tr.shape)
word_tst = semantic_vecs[:, xrange(num_sem_vecs / 2 + 1, num_sem_vecs)]
#print "word_tst.shape = " + str(word_tst.shape)
#print "word_tr.shape = " + str(word_tr.shape)
#print "word_tst.shape = " + str(word_tst.shape)
# don't subtract out average of all semantic vectors for all time points
# only calculate average for training, and subtract that average out of the test
# CHANGE MADE: sep 19 12:40 AM
if subtract_temporal_average_bool == True:
word_tr, avg_tr_word_vec = subtract_column_mean(word_tr)
word_tst = word_tst - avg_tr_word_vec[:, np.newaxis]
mask_results = {}
for mask in fmris.keys():
fmri = fmris[mask]
num_total_time_steps = fmri.shape[1]
#print "fmri.shape = " + str(fmri.shape)
fmri_tr = fmri[:, xrange(num_sem_vecs / 2 + 1)]
fmri_tst = fmri[:, xrange(num_sem_vecs / 2 + 1, num_sem_vecs)]
#print "fmri_tr.shape = " + str(fmri_tr.shape)
#print "fmri_tst.shape = " + str(fmri_tst.shape)
print "Training linear maps..."
# all procrustes
if learn_weights == True:
if fmri_or_text_time_weights == "fmri":
W_tf, W_ft, C_weights = learn_matrix_factorization_for_time_text_weights(
xrange(num_sem_vecs / 2 + 1), [], fmri_tr,
mask + avg_variant, word_tr, date_of_wordvecs)
elif fmri_or_text_time_weights == "text":
W_tf, W_ft, C_weights = learn_matrix_factorization_for_time_text_weights(
xrange(num_sem_vecs / 2 + 1), [], word_tr,
date_of_wordvecs, fmri_tr, mask + avg_variant)
else:
C_weights = timestep_weighting
num_times = len(xrange(num_sem_vecs / 2 + 1))
C = create_k_diag_conv_mat_from_weights(
np.zeros((num_times, num_times)), C_weights, num_times)
if fmri_or_text_time_weights == "fmri":
X = fmri_tr[:, xrange(num_sem_vecs / 2 + 1)]
Y = word_tr[:, xrange(num_sem_vecs / 2 + 1)]
YC = np.dot(Y, C)
# text -> fmri
W_tf = procrustes_fit([X, YC])
W_tf = W_tf[0]
W_ft = W_tf.T
elif fmri_or_text_time_weights == "text":
Y = fmri_tr[:, xrange(num_sem_vecs / 2 + 1)]
X = word_tr[:, xrange(num_sem_vecs / 2 + 1)]
YC = np.dot(Y, C)
# text -> fmri
W_tf = procrustes_fit([YC, X])
W_tf = W_tf[0]
W_ft = W_tf.T
print "Done training linear maps..."
TF_class_scores = []
TF_rank_scores = []
FT_class_scores = []
FT_rank_scores = []
num_time_steps = fmri_tst.shape[
1] # assume time steps are # of columns
chunk_array = []
index = 0
num_chunks = 25 ### CHANGED JAN 11
while index < num_time_steps:
chunk_array.append(index)
index += num_time_steps / num_chunks
if index > num_time_steps:
index = num_time_steps
k = 1 #-> 4% chance rate ## CHANGED FEB 4
# weighting matrices for semantic vectors
C = create_k_diag_conv_mat_from_weights(
np.zeros((num_time_steps, num_time_steps)), C_weights,
num_time_steps)
time_weighted_word_tst = np.dot(word_tst, C)
# comparisons in fMRI space (i.e. text -> fMRI)
# truth is fmri_tst
# prediction is W_tf*time_weighted_word_tst
TF_truth = fmri_tst
TF_prediction = np.dot(W_tf, time_weighted_word_tst)
#TF_prediction = TF_prediction[0, :, :]
#print "TF pred shape: " + str(TF_prediction.shape)
#print "TF truth shape: " + str(TF_truth.shape)
assert (TF_truth.shape == TF_prediction.shape)
TF_classification_score = scene_classification(TF_truth, TF_prediction,
chunk_array, k)
TF_rank_score, TF_ranks = scene_ranking(TF_truth, TF_prediction,
chunk_array)
TF_rank_score = TF_rank_score / (num_chunks + 0.)
# comparisons in Semantic space (i.e. fMRI -> text) (time-weighted_averages)
# truth is time_weighted_word_tst
# prediction is W_ft*fmri_tst
FT_truth = time_weighted_word_tst
FT_prediction = np.dot(W_ft, fmri_tst)
#FT_prediction = FT_prediction[0, :, :]
#print "FT pred shape: " + str(FT_prediction.shape)
#print "FT truth shape: " + str(FT_truth.shape)
assert (FT_truth.shape == FT_prediction.shape)
FT_classification_score = scene_classification(FT_truth, FT_prediction,
chunk_array, k)
FT_rank_score, FT_ranks = scene_ranking(FT_truth, FT_prediction,
chunk_array)
FT_rank_score = FT_rank_score / (num_chunks + 0.)
print "Mask: " + mask
print "---------------------------------------------------------------------------"
print "fMRI -> Text (Procrustes) scene classification (" + str(
float(k) /
num_chunks) + "% chance) avg = " + str(FT_classification_score)
print "fMRI -> Text (Procrustes) scene ranking (50% chance) avg = " + str(
FT_rank_score)
print "---------------------------------------------------------------------------"
print "Text -> fMRI (Procrustes) scene classification (" + str(
float(k) /
num_chunks) + "% chance) avg = " + str(TF_classification_score)
print "Text -> fMRI (Procrustes) scene ranking (50% chance) avg = " + str(
TF_rank_score)
print "---------------------------------------------------------------------------"
print "///////////////////////////////////////////////////////////////////////////"
mask_results[mask] = (C_weights, TF_classification_score,
TF_rank_score, FT_classification_score,
FT_rank_score)
return mask_results
def make_weighted_average_word_vecs_over_multiple_TRs(
weighted_bool, number_TRs_we_average_over,
subtract_out_temporal_mean_bool):
alpha = 0.0001 # as per yingyu
avg_wordvec_per_multiple_TRs = []
freq_dict = p.load(open(main_path + 'SEMVECS/enwiki_freq_dict.p', 'rb'))
# implement the avoidance of the last 3 TRs here
num_TRs = len(txt_TRs) - 3
# just ensuring that it'll never go past that TR number, which is what we want
for i in range(0, num_TRs, number_TRs_we_average_over):
# make sure we don't go past num_TRs
endpoint = min(num_TRs, i + number_TRs_we_average_over)
text_array = [txt_TRs[j][0] for j in range(i, endpoint)]
#print "============= START OF TR " + str(i) + " ==============="
txt_str = ''
num_words_TR = 0
wordvec_avg = None
#print "Length of text array = " + str(len(text_array))
#print text_array
# calculate number of words in all the TRs as well as the text description
for txt in text_array:
for word in txt:
num_words_TR += 1
txt_str += word + ' '
for txt in text_array:
for word in txt:
#print "Word: " + word
good_return, wordvec = get_word_vector(word)
if good_return != -1:
if weighted_bool:
yingyu_coeff = alpha / (0. + freq_dict[word] + alpha)
else:
yingyu_coeff = 1.
if wordvec_avg == None:
wordvec_avg = yingyu_coeff * wordvec / (0. +
num_words_TR)
else:
wordvec_avg += yingyu_coeff * wordvec / (0. +
num_words_TR)
#print "-----"
avg_wordvec_per_multiple_TRs.append(wordvec_avg)
#print "============= END OF TR " + str(i) + " ==============="
avg_word_vec_over_multiple_TRs_mat = np.matrix(
avg_wordvec_per_multiple_TRs)
avg_word_vec_over_multiple_TRs_mat = avg_word_vec_over_multiple_TRs_mat.T
print "avg_word_vec_over_multiple_TRs_mat.shape = " + str(
avg_word_vec_over_multiple_TRs_mat.shape)
if subtract_out_temporal_mean_bool:
avg_word_vec_over_multiple_TRs_mat = subtract_column_mean(
avg_word_vec_over_multiple_TRs_mat)
print "after subtracting: avg_word_vec_over_multiple_TRs_mat.shape = " + str(
avg_word_vec_over_multiple_TRs_mat.shape)
#print "shape of wordvec mat = " + str(avg_word_vec_over_multiple_TRs_mat.shape)
# save as file for learning the Dictionary representation
filename = main_path + 'SEMVECS/'
if weighted_bool:
filename += 'weighted_'
else:
filename += 'unweighted_'
filename += 'word_vector_average_over_' + str(
number_TRs_we_average_over) + '_TRs'
if subtract_out_temporal_mean_bool:
filename += '_subtract_out_temporal_mean'
filename += '.mat'
savemat(filename, {'vecs': avg_word_vec_over_multiple_TRs_mat})
return avg_word_vec_over_multiple_TRs_mat # return transpose here to avoid flipping