def RLDeconvolution(self, RLiterations, PSF, threads=multiprocessing.cpu_count(), PSF_pad=0): '''Input: RLiterations=number of iterations to perform PSF=point spread function (an EELS spectrum object) Optional argument: threads=number of computer's CPUs to use while deconvolving, default is all of them PSF_pad=value to pad PSF with (or None to not pad PSF)''' PSF_sym = PSF.SymmetrizeAroundZLP() if PSF_pad is not None: data_length = np.size(self.SpectrumRange) PSF_length = np.size(PSF_sym.intensity) pad_length = data_length/2 - (1 + data_length) % 2 - (PSF_length-(PSF_length % 2))/2 if PSF_sym.ZLP < data_length/2: PSF_sym = PSF.PadSpectrum(pad_length, pad_value=PSF_pad, pad_side='left').SymmetrizeAroundZLP() elif PSF_sym.ZLP > data_length/2: PSF_sym = PSF_sym.PadSpectrum(pad_length, pad_value=PSF_pad, pad_side='right') print 'Beginning deconvolution...' loopyP = partial(loopy, iterations=RLiterations, PSF=PSF_sym.Normalize().intensity) x_deconv = np.array(handythread.parallel_map(loopyP, abs(self.Normalize()), threads = threads)) # x_deconv = np.array(handythread.parallel_map(loopyP, self.Normalize(), # threads = threads)) x_deconv = np.ma.array(x_deconv, mask = self.data.mask) print 'Done %s iterations!' %RLiterations return EELSSpectrumImage(x_deconv, self.dispersion)
def RLDeconvolution_Adaptive(self,
RLiterations,
PSF,
threads=multiprocessing.cpu_count(),
PSF_pad=0):
PSF_sym = PSF.SymmetrizeAroundZLP()
data_sym = self.SymmetrizeAroundZLP()
if PSF_pad is not None:
data_length = np.size(
self.SpectrumRange) ##replace w/ self.size[2]
PSF_length = np.shape(PSF_sym.data)[2]
pad_length = int(data_length / 2 - (1 + data_length) % 2 -
PSF_length // 2)
if PSF_sym.ZLP < data_length / 2:
PSF_sym = PSF.PadSpectrum(
pad_length, pad_value=PSF_pad,
pad_side='left').SymmetrizeAroundZLP()
elif PSF_sym.ZLP > data_length / 2:
PSF_sym = PSF_sym.PadSpectrum(pad_length,
pad_value=PSF_pad,
pad_side='right')
if data_sym.ZLP < data_length / 2:
data_sym = self.PadSpectrum(
pad_length, pad_value=PSF_pad + 1,
pad_side='left').SymmetrizeAroundZLP()
elif PSF_sym.ZLP > data_length / 2:
data_sym = self.PadSpectrum(pad_length,
pad_value=PSF_pad + 1,
pad_side='right')
print('Beginning deconvolution...')
loopyP_adapt = partial(loopy_adapt, iterations=RLiterations)
deconvolution_arrays = np.append(np.expand_dims(data_sym.Normalize(),
axis=-1),
np.expand_dims(PSF_sym.Normalize(),
axis=-1),
axis=-1)
x_deconv = np.array(
handythread.parallel_map(loopyP_adapt,
deconvolution_arrays,
threads=threads))
x_deconv = np.ma.array(x_deconv, mask=self.data.mask)
print('Done %s iterations!' % RLiterations)
return EELSSpectrumImage(x_deconv, dispersion=self.dispersion)
def litekmeans(X, k, max_iter=50):
X = X.T
n = X.shape[1]
ndim = X.shape[0]
last = 0
label = np.random.randint(k, size=(n, ))
iteration = 0
batchsize = 100000
nbatches = int(np.ceil(n / batchsize))
center = np.zeros((ndim, k), dtype=np.float32)
while np.any(label != last):
start = time.time()
iteration += 1
print 'iteration: {0}'.format(iteration)
E = scipy.sparse.coo_matrix((np.ones(
(n, ), dtype=np.int), (np.arange(n), label)),
shape=(n, k),
dtype=np.float64).tocsr()
# E = one hot assignments
# spdiags... = counts
print 'max of E.sum(0): %s' % (E.sum(0).max(), )
print 'max of (1.0/E.sum(0)): %s' % ((1.0 / E.sum(0)).max(), )
print 'min of E.sum(0): %s' % (E.sum(0).min(), )
print 'min of (1.0/E.sum(0)): %s' % ((1.0 / E.sum(0)).min(), )
print 'np.all(1.0/E.sum(0) == np.inf): %r' % (np.all(
1.0 / E.sum(0) == np.inf), )
center = X * E * scipy.sparse.spdiags(1.0 / (E.sum(0) + 0.0000000001),
0, k, k)
c2 = 0.5 * np.sum(center**2, 0).T[:, None]
last = label
label = np.zeros((n, ), dtype=np.int)
def get_labels(batchidx):
return np.argmax(np.dot(
center.T, X[:, j * batchsize + batchidx *
1000:min(n, j * batchsize +
(batchidx + 1) * 1000)]) - c2,
axis=0)
for j in range(nbatches):
print 'processing batch {0:d} / {1:d}'.format(j + 1, nbatches)
tmp = handythread.parallel_map(get_labels,
range(int(np.ceil(batchsize /
1000))),
threads=8)
label[j * batchsize:min(n, int((j + 1) *
batchsize))] = np.concatenate(tmp)
if iteration >= max_iter:
break
print 'iteration took {0:d} seconds'.format(int(time.time() - start))
obj = 0
Xsq = 0.5 * np.sum(X**2, 0)
batchsize = 10000
nbatches = int(np.ceil(n / batchsize))
csq = 0.5 * np.sum(center**2, 0)
# TODO: do this stuff in parallel as well (takes longer than expected)
def compute_sqd(batchidx):
tempX = X[:, j * batchsize +
batchidx * 100:min(n, j * batchsize + (batchidx + 1) * 100)]
temp = np.dot(-center.T, tempX) + csq[:, None]
tmp = Xsq[j * batchsize +
batchidx * 100:min(n, j * batchsize +
(batchidx + 1) * 100)] + temp
temp_mindist = np.min(
Xsq[j * batchsize +
batchidx * 100:min(n, j * batchsize +
(batchidx + 1) * 100)] + temp,
axis=0)
return np.sum(temp_mindist)
for j in range(nbatches):
tmp = handythread.parallel_map(compute_sqd,
range(int(np.ceil(batchsize / 100))),
threads=8)
obj += np.sum(tmp)
print 'obj: %r' % (obj, )
#print obj
#for j in range(nbatches):
# tempX = X[:, j * batchsize:min(n, (j + 1) * batchsize)]
# temp = np.dot(-center.T, tempX) + csq[:, None]
# print 'Xsq[j * batchsize:min(n, (j + 1) * batchsize)].mean(): %r' % (
# Xsq[j * batchsize:min(n, (j + 1) * batchsize)].mean(), )
# print 'mean of temp: %s' % (temp.mean(), )
# tmp = Xsq[j * batchsize:min(n, (j + 1) * batchsize)] + temp
#
# temp_mindist = np.min(
# Xsq[j * batchsize:min(n, (j + 1) * batchsize)] + temp,
# axis=0
# )
# obj = obj + np.sum(temp_mindist)
print 'obj: %r' % (obj, )
center = center.T
return (label, center, obj)
def test_parallel_map(self):
l = range(100)
r = handythread.parallel_map(lambda x: x**2, l)
for i in range(len(l)):
self.assertEqual(l[i]**2,r[i])
def run(doc_topics_filename, topic_keys_filename, state_filename, max_dict, meta_filename,
solr_url, output_dir, date_format):
if not os.path.exists(doc_topics_filename):
raise FileNotFoundError(doc_topics_filename)
if not os.path.exists(topic_keys_filename):
raise FileNotFoundError(topic_keys_filename)
if not os.path.exists(state_filename):
raise FileNotFoundError(state_filename)
if meta_filename is not None and not os.path.exists(meta_filename):
raise FileNotFoundError(meta_filename)
start = time()
# read the mallet output
print("Reading {}...".format(doc_topics_filename), end='', flush=True)
doc_topics, num_topics = read_doc_topics(doc_topics_filename, return_num_topics=True)
print("done, {:,} topics, {:,} documents".format(num_topics, len(doc_topics)))
print("Reading {}...".format(topic_keys_filename), end='', flush=True)
topic_keys, num_topwords = read_topic_keys(topic_keys_filename, return_num_topwords=True)
print("done, {:,} top words per topic".format(num_topwords))
print("Reading {}...".format(state_filename), end='', flush=True)
state = read_state(state_filename)
print("done")
if meta_filename is not None:
print("Reading metadata from {}...".format(meta_filename), end='', flush=True)
doc_meta = read_meta(meta_filename, date_format)
print("done, {:,} entries found".format(len(doc_meta)))
else:
print("Retrieving metadata from {}...".format(solr_url), end='', flush=True)
doc_meta = retrieve_meta(doc_topics, solr_url, date_format)
print("done, {:,} entries retrieved".format(len(doc_meta)))
topicids = range(num_topics)
# check if pruning is desired
if max_dict > 0:
print("Pruning tokens, keeping only the top {:,} tokens by frequency...".format(max_dict), end='', flush=True)
state = prune_state(state, max_dict)
print("done")
print("Re-indexing tokens...", end='', flush=True)
state['typeindex'] = pd.factorize(state['typeindex'])[0]
print("done")
print("Processing state data...", end='', flush=True)
# extract mapping of token id to token
tokenid_map = state[['typeindex', 'type']].drop_duplicates()
tokenid_map.columns = ['tokenid', 'token']
# extract mapping of token -> document -> topic
state = state[['doc', 'typeindex', 'topic']]
state.columns = ['docid', 'tokenid', 'topic']
# compute token, topic -> count mapping
token_topic_count = state[['tokenid', 'topic']].groupby(['tokenid', 'topic'], sort=False).size().reset_index()
token_topic_count.columns = ['tokenid', 'topic', 'count']
num_tokens = len(tokenid_map)
doc_topics_complete = pd.merge(doc_meta, doc_topics, on='source')
doc_topics_complete = doc_topics_complete.rename(columns={'id': 'docid', 'id_x': 'volid', 'id_y': 'docid'},
copy=False)
docid_publishdate = doc_topics_complete[['docid', 'year']]
print("done, {:,} tokens".format(num_tokens))
# Create aggregate state object
print("Calculating aggregate stats...", end='', flush=True)
agg = state.groupby(['docid', 'tokenid', 'topic']).size().reset_index()
agg.columns = ['docid', 'tokenid', 'topic', 'count']
agg_token = pd.merge(agg, tokenid_map, on='tokenid')
full_state = pd.merge(agg_token, docid_publishdate, on='docid')
full_state.drop('docid', axis=1, inplace=True)
corpus_token_counts_by_year = full_state[['year', 'count']].groupby('year').sum().reset_index()
topic_keywords = pd.melt(topic_keys,
id_vars='id',
value_vars=['key.' + str(i) for i in range(num_topwords)],
value_name='token')
topic_keywords = topic_keywords.rename(columns={'id': 'topic'}, copy=False)
topic_keywords.drop('variable', axis=1, inplace=True)
topic_token_counts_by_year = \
pd.merge(full_state[['topic', 'year', 'token', 'count']],
topic_keywords, on=['topic', 'token']).groupby(['topic', 'year', 'token']).sum().reset_index()
print("done")
print("Creating topics...", end='', flush=True)
def create_topic(topicid):
token_count = token_topic_count[token_topic_count['topic'] == topicid]
vector = np.repeat(np.float64(0), num_tokens)
for tid, _, cnt in token_count.itertuples(index=False):
vector[tid] = np.float64(cnt)
return Topic(vector)
topics = list(parallel_map(create_topic, topicids))
print("done")
# calculate trend
print("Calculating topic trend...", end='', flush=True)
state_trend = pd.merge(state[['docid', 'topic']],
docid_publishdate,
on='docid')[['topic', 'year']].groupby(['topic', 'year']).size().reset_index()
state_trend.columns = ['topic', 'year', 'count']
def slope(topic):
state_small = state_trend[(state_trend['topic'] == topic) & (state_trend['year'] != -1)]
dates = state_small['year'].values
values = state_small['count'].values
lm = stats.linregress(dates, values)
return lm[0]
topic_keys['trend'] = list(parallel_map(slope, topicids))
print("done")
# Insert dists
print("Calculating topic distance from center...", end='', flush=True)
topic_keys['dist'] = list(parallel_map(lambda x: x.length, topics))
print("done")
print("Calculating topic means...", end='', flush=True)
topic_keys['mean'] = doc_topics.ix[:, 'topic.0':].mean().values
print("done")
# Create Distance Matrix
dist = pd.DataFrame(data=0, dtype='float64', index=topicids, columns=topicids)
print("Calculating inter-topic distances...", end='', flush=True)
def calc_dist(tuple):
x, y = tuple
d = topics[x].distance(topics[y])
dist.iloc[x, y] = d
dist.iloc[y, x] = d
parallel_for(calc_dist, itertools.combinations(topicids, 2))
print("done")
# write results
print("Writing out results...", end='', flush=True)
doc_topics.to_csv(os.path.join(output_dir, 'documents.csv'), index=False, encoding='utf-8')
topic_keys.to_csv(os.path.join(output_dir, 'topics.csv'), index=False, encoding='utf-8')
tokenid_map.to_csv(os.path.join(output_dir, 'tokens.csv'), index=False, encoding='utf-8')
agg.to_csv(os.path.join(output_dir, 'state.csv'), index=False, encoding='utf-8')
dist.to_csv(os.path.join(output_dir, 'distance.csv'), encoding='utf-8')
corpus_token_counts_by_year.to_csv(os.path.join(output_dir, 'counts_by_year.csv'), index=False, encoding='utf-8')
topic_token_counts_by_year.to_csv(os.path.join(output_dir, 'counts_by_topic_year.csv'),
index=False, encoding='utf-8')
if solr_url is not None:
doc_meta.drop('year', axis=1, inplace=True)
doc_meta.to_csv(os.path.join(output_dir, 'docmeta.csv'), index=False, encoding='utf-8')
print("done")
elapsed = int(time() - start)
print("All done. Time elapsed: {}".format(timedelta(seconds=elapsed)))
