def load_data():
r = csv.reader(open(predict_file, 'r', encoding='utf-8'),
delimiter=',',
quotechar='"')
raw_data = np.array(list(r))
s.print_data("raw", raw_data)
str_data = np.delete(raw_data[1:, 1:], [6], axis=1)
po_nums = raw_data[1:, 0]
return s.normalize(str_data.astype('float32')), po_nums
def parse(tokens):
"""
parse tokenized query to simplified executable list structure based on Shunting-yard algorithm
:param tokens: tokens of a query, including words and strings of 'AND', 'OR', 'NOT', '(', and ')'
:return: parsed tokens
"""
# Parse
special_tokens = ['AND', 'OR', 'NOT', '(', ')']
ops = {
'AND': 3,
'OR': 2,
} # precedence
output = []
op_stack = []
# Adapted from pseudo code provided in the Shunting-yard algorithm wikipedia page
for i in range(len(tokens)):
if not tokens[i] in special_tokens:
word = normalize(tokens[i])
output.append(word)
elif tokens[i] == 'NOT':
op_stack.append('NOT')
elif tokens[i] in ops:
while op_stack and op_stack[-1] != '(' and ops[
op_stack[-1]] >= ops[tokens[i]]:
output.append(op_stack.pop())
op_stack.append(tokens[i])
elif tokens[i] == '(':
op_stack.append('(')
elif tokens[i] == ')':
while op_stack and op_stack[-1] != '(':
output.append(op_stack.pop())
if op_stack and op_stack[-1] == '(':
op_stack.pop()
if op_stack and op_stack[-1] == 'NOT':
output.append(op_stack.pop())
while op_stack:
output.append(op_stack.pop())
# Simplify
i = 0
while i < len(output):
# Double negation is canceled
if output[i] == 'NOT' and output[i - 1] == 'NOT':
output[(i - 1):] = output[(i + 1):]
# Use De Morgan's Law to take OR in NOT out
elif output[i] == 'NOT' and output[i - 1] == 'OR':
first_operand = get_operand(output, i - 1)
second_operand = get_operand(output, i - 1 - len(first_operand))
remaining = i + 1
i = i - 2 - len(first_operand) - len(second_operand)
output[(i + 1):] = second_operand + ['NOT'] + first_operand + [
'NOT', 'AND'
] + output[remaining:]
i = i + 1
return output
def resize_save_preview(self, outpath, target_shape=[4000, 4000]):
h_, w_, *_ = self.shape
rrate_ = min(target_shape[0]/h_, target_shape[1]/w_)
target_shape = (np.array([h_, w_], dtype=Config.TYPE_FLOAT)*rrate_).astype(dtype=np.int)
tmppath = path.join(Config.TEMP_DIR, str(uuid.uuid4()))
resized = DiskMap(tmppath, dtype=self.dtype, mode='w+', shape=self.shape)
resized[:] = self[:]
resized = imresize(resized, target_shape)
normalized = (normalize(resized)*255.).astype(np.uint8)
imsave(outpath, exposure.rescale_intensity(normalized))
del normalized
print('Resized preview saved', outpath, target_shape)
DiskMap.remove_file(resized)
def bigram_inverted_index(collection):
bigram_index = defaultdict(set)
bigram_iindex = defaultdict(set)
for doc in collection:
# preprocess without lemmatization
new_text = normalize(doc.full_text())
tokens = tokenize(new_text)
text_tokens = remove_stop_word(tokens)
for word in set(text_tokens):
if word in bigram_index:
continue
bigrams = get_bigrams(word)
bigram_index[word] = set(bigrams)
for bigram in bigrams:
bigram_iindex[bigram].add(word)
return bigram_index, bigram_iindex
def load_process_data_b():
f1 = open(pos_file, 'r' ,encoding='utf-8')
f2 = open(neg_file, 'r' ,encoding='utf-8')
r1 = csv.reader(f1, delimiter=',', quotechar='"')
r2 = csv.reader(f2, delimiter=',', quotechar='"')
pos_data = process_b(np.array(list(r1)))
neg_data = process_b(np.array(list(r2)))
data = np.concatenate((pos_data, neg_data), axis=0)
np.random.shuffle(data)
#x_data = np.delete(data, [5], axis=1)#第5列是 violation 其实也就是train 中的y_data 因此删掉
x_data = s.normalize(np.delete(data, [5], axis=1)) #第5列是 violation 其实也就是train 中的y_data，因此删掉
#s.print_data("normalized data", x_data)
y_data = data[:, 5]
#len_data = data.shape[0] // 1280 * 1280
#print("25016 trimmed is %d" %(25016 // 1280 * 1280))
#len_data = data.shape[0]#hongwei 的改动，因为不明白为何需要 除以1280 取整后再 乘回来1280，数字少的时候结果就直接变0了，无法debug了
len_data = data.shape[0] // const_folds * const_folds #20156 --> 20150
f1.close(); f2.close()
return x_data[:len_data], y_data[:len_data]
def load_process_data_m():
f1 = open(neg_file, 'r', encoding='utf-8')
r1 = csv.reader(f1, delimiter=',', quotechar='"')
#s.print_data("r1", r1)
data_list=np.array(list(r1))
#s.print_data("data_list",data_list)
data = process_m(data_list)
#s.print_data("str_data3",data)
#s=data.shape
#s.print_data("s the shape",s)
#len_data = data.shape[0] // 1280 * 1280
#len_data = data.shape[0]
len_data = data.shape[0] // const_folds * const_folds #by hongwei: 20156 --> 20150
#print("=========================> len_data is %d" %(len_data))
x_data = s.normalize(np.delete(data, [5], axis=1))
#s.print_data("x_data after normalize",x_data)
y_data = data[:, 5] #只取二维数组的第五列
#s.print_data("data[:, 5]",y_data)
#s.print_data("y_data after takiing data[:, 5]", y_data)
f1.close()
#s.print_data("x_data[:len_data]",x_data[:len_data])
return x_data[:len_data], y_data[:len_data]
def build_index(in_dir, out_dict, out_postings):
"""
build index from documents stored in the input directory,
then output the dictionary file and postings file
"""
print("indexing...")
# Initializations
dictionary = {}
doc_freq = {}
# Iterate through each file
# Not directly using files.sorted() because this sorts
# alphabetically instead of by index number
files = os.listdir(in_dir)
file_index = 0
block_count = 0
file_count = 0
while file_index < max_doc_id:
# For each file, tokenize and normalize the words
file_index = file_index + 1
if not str(file_index) in files:
continue
file_name = in_dir + "/" + str(file_index)
reader = open(file_name, 'r')
# Get the unprocessed words
content = reader.read()
reader.close()
file_count += 1
words_in_doc = []
words = nltk.word_tokenize(content)
for w in words:
ws = w.split('/')
for word in ws:
word = normalize(word)
if not word == "" and word not in words_in_doc:
words_in_doc.append(word)
for word in words_in_doc:
if word not in dictionary:
dictionary[word] = []
doc_freq[word] = 0
dictionary[word].append(file_index)
doc_freq[word] += 1
# Divide the files into blocks, each block with block_size files
# index every block by applying BSBI
if file_count >= block_size:
block_count += 1
temp_dict_path = "./temp_dict" + str(block_count) + ".txt"
temp_posting_path = "./temp_post" + str(block_count) + ".txt"
bsbi_invert(dictionary, doc_freq, temp_dict_path,
temp_posting_path)
dictionary = {}
doc_freq = {}
file_count = 0
if len(dictionary) >= 1: # Construct last block using BSBI
block_count += 1
temp_dict_path = "./temp_dict" + str(block_count) + ".txt"
temp_posting_path = "./temp_post" + str(block_count) + ".txt"
bsbi_invert(dictionary, doc_freq, temp_dict_path, temp_posting_path)
merge_block(block_count, out_dict, out_postings)
def slice_normalize(self, roi):
roi = [min(roi[0], roi[2]), min(roi[1], roi[3]), max(roi[0], roi[2]), max(roi[1], roi[3])]
w_ = roi[2] - roi[0]
h_ = roi[3] - roi[1]
sliced = np.array(self[roi[1]:roi[1]+h_, roi[0]:roi[0]+w_]).astype(dtype=Config.TYPE_FLOAT)
return (normalize(sliced)*255.).astype(np.uint8)
def copy_normalize(self, signed=False):
tmppath = path.join(Config.TEMP_DIR, str(uuid.uuid4()))
copy = DiskMap(tmppath, dtype=Config.TYPE_FLOAT, mode='w+', shape=self.shape)
copy[:] = self[:]
normalize(copy, signed)
return copy
