def Process_690(args):
path = args.file_path
all_folders = os.listdir(path)
count = 0
for folder in all_folders:
# load data
train_seq_path = os.path.join(args.file_path, folder, "train",
"sequences_alph.npy")
test_seq_path = os.path.join(args.file_path, folder, "test",
"sequences_alph.npy")
train_lab_path = os.path.join(args.file_path, folder, "train",
"targets.npy")
test_lab_path = os.path.join(args.file_path, folder, "test",
"targets.npy")
train_sequences = np.load(train_seq_path)
test_sequences = np.load(test_seq_path)
train_labels = np.load(train_lab_path)
test_labels = np.load(test_lab_path)
train_sequences = train_sequences.reshape(train_sequences.shape[0], 1)
test_sequences = test_sequences.reshape(test_sequences.shape[0], 1)
train_labels = train_labels.reshape(train_labels.shape[0], 1)
test_labels = test_labels.reshape(test_labels.shape[0], 1)
# concat sequence and labels together
trains = list(np.concatenate((train_sequences, train_labels), axis=1))
tests = list(np.concatenate((test_sequences, test_labels), axis=1))
random.seed(args.seed)
random.shuffle(trains)
random.shuffle(trains)
random.shuffle(tests)
random.shuffle(tests)
# make output path
output_path = os.path.join(args.output_path, str(args.kmer), folder)
if not os.path.exists(output_path):
os.makedirs(output_path)
# write files
f_train = open(os.path.join(output_path, "train.tsv"), 'wt')
tsv_train = csv.writer(f_train, delimiter='\t')
tsv_train.writerow(["sequence", "label"])
for i in range(len(trains)):
sentence = get_kmer_sentence(trains[i][0].decode("utf-8"),
args.kmer)
tsv_train.writerow([sentence, int(trains[i][1])])
f_dev = open(os.path.join(output_path, "dev.tsv"), 'wt')
tsv_dev = csv.writer(f_dev, delimiter='\t')
tsv_dev.writerow(["sequence", "label"])
for i in range(len(tests)):
sentence = get_kmer_sentence(tests[i][0].decode("utf-8"),
args.kmer)
tsv_dev.writerow([sentence, int(tests[i][1])])
count += 1
print("Finish %s folders" % (count))
def write_file_pair(lines,
writer,
seq1_index=0,
seq2_index=1,
label_index=2):
for line in lines:
seq1 = get_kmer_sentence(line[seq1_index], args.kmer)
seq2 = get_kmer_sentence(line[seq2_index], args.kmer)
writer.writerow([seq1, seq2, str(int(line[label_index]))])
def write_file(lines, writer, seq_index=2, label_index=3):
global max_length
for line in lines:
sentence = get_kmer_sentence(line[seq_index], args.kmer)
if len(sentence.split()) > max_length:
max_length = len(sentence.split())
writer.writerow([sentence, str(line[label_index])])
def Process(args):
SCAN_LIST = [int(500 / (args.slide - 1)) * i for i in range(args.slide)]
old_file = open(args.file_path, "r", encoding="utf-8-sig")
old_lines = list(csv.reader(old_file, delimiter=",", quotechar=None))[1:]
if args.output_path:
root_path = args.output_path + "/"
else:
root_path = "/".join(args.file_path.split("/")[:-1]) + "/" + str(
args.kmer) + "/"
if not os.path.exists(root_path):
os.makedirs(root_path)
labels = np.array([])
new_file = open(root_path + "dev.tsv", 'wt')
tsv_w = csv.writer(new_file, delimiter='\t')
tsv_w.writerow(["setence", "label"])
for line in old_lines:
label = line[6]
labels = np.append(labels, int(label))
for index in SCAN_LIST:
sub_sequence = line[8][index:index + 500]
sub_sentence = get_kmer_sentence(sub_sequence, kmer=args.kmer)
tsv_w.writerow([sub_sentence, label])
np.save(root_path + "label.npy", labels)
def write_file(lines, path, kmer, head=True, seq_index=0, label_index=1):
with open(path, 'wt') as f:
tsv_w = csv.writer(f, delimiter='\t')
if head:
tsv_w.writerow(["setence", "label"])
for line in lines:
if kmer == 0:
sentence = str(line[seq_index])
else:
sentence = str(
get_kmer_sentence("".join(line[seq_index].split()), kmer))
if label_index == None:
label = "0"
else:
label = str(line[label_index])
tsv_w.writerow([sentence, label])
def Process_splice(args):
# X_train = np.load(os.path.join(args.file_path, "x_train.npy"))
# X_dev = np.load(os.path.join(args.file_path, "x_dev.npy"))
# Y_train = np.load(os.path.join(args.file_path, "y_train.npy"))
# Y_dev = np.load(os.path.join(args.file_path, "y_dev.npy"))
# assert len(X_train) == len(Y_train)
# assert len(X_dev) == len(Y_dev)
# for kmer in range(3,7):
# root_path = os.path.join(args.file_path, str(kmer))
# os.makedirs(root_path)
# f_train = open(os.path.join(root_path, "train.tsv"), "wt")
# f_dev = open(os.path.join(root_path, "dev.tsv"), "wt")
# tsv_train = csv.writer(f_train, delimiter='\t')
# tsv_dev = csv.writer(f_dev, delimiter='\t')
# tsv_train.writerow(["seq", "label"])
# tsv_dev.writerow(["seq", "label"])
# for i, seq in enumerate(X_train):
# sequence = get_kmer_sentence(str(seq), kmer)
# tsv_train.writerow([sequence, int(Y_train[i])])
# for j, seq in enumerate(X_dev):
# sequence = get_kmer_sentence(str(seq), kmer)
# tsv_dev.writerow([sequence, int(Y_dev[j])])
X_test = np.load(os.path.join(args.file_path, "x_test.npy"))
Y_test = np.load(os.path.join(args.file_path, "y_test.npy"))
assert len(X_test) == len(Y_test)
for kmer in range(3, 7):
root_path = os.path.join(args.file_path, str(kmer))
os.makedirs(root_path)
f_test = open(os.path.join(root_path, "dev.tsv"), "wt")
tsv_test = csv.writer(f_test, delimiter='\t')
tsv_test.writerow(["seq", "label"])
for i, seq in enumerate(X_test):
sequence = get_kmer_sentence(str(seq), kmer)
label = int(np.where(Y_test[i] == 1)[0])
tsv_test.writerow([sequence, label])
def Visualize(args):
if args.kmer == 0:
KMER_LIST = [3, 4, 5, 6]
for kmer in KMER_LIST:
tokenizer_name = 'dna' + str(kmer)
model_path = os.path.join(args.model_path, str(kmer))
model = BertModel.from_pretrained(model_path,
output_attentions=True)
tokenizer = DNATokenizer.from_pretrained(tokenizer_name,
do_lower_case=False)
raw_sentence = args.sequence if args.sequence else SEQUENCE
sentence_a = get_kmer_sentence(raw_sentence, kmer)
tokens = sentence_a.split()
attention = get_attention_dna(model,
tokenizer,
sentence_a,
start=args.start_layer,
end=args.end_layer)
attention_scores = np.array(attention).reshape(
np.array(attention).shape[0], 1)
# attention_scores[0] = 0
real_scores = get_real_score(attention_scores, kmer, args.metric)
real_scores = real_scores / np.linalg.norm(real_scores)
if kmer != KMER_LIST[0]:
scores += real_scores.reshape(1, real_scores.shape[0])
else:
scores = real_scores.reshape(1, real_scores.shape[0])
else:
# load model and calculate attention
tokenizer_name = 'dna' + str(args.kmer)
model_path = args.model_path
model = BertModel.from_pretrained(model_path, output_attentions=True)
tokenizer = DNATokenizer.from_pretrained(tokenizer_name,
do_lower_case=False)
raw_sentence = args.sequence if args.sequence else SEQUENCE
sentence_a = get_kmer_sentence(raw_sentence, args.kmer)
tokens = sentence_a.split()
attention = get_attention_dna(model,
tokenizer,
sentence_a,
start=args.start_layer,
end=args.end_layer)
attention_scores = np.array(attention).reshape(
np.array(attention).shape[0], 1)
# attention_scores[0] = 0
real_scores = get_real_score(attention_scores, args.kmer, args.metric)
scores = real_scores.reshape(1, real_scores.shape[0])
ave = np.sum(scores) / scores.shape[1]
print(ave)
print(scores)
# plot
sns.set()
ax = sns.heatmap(scores, cmap='YlGnBu', vmin=0)
plt.show()
