def detect_click(class_path,
class_name,
snr_threshold_low=5,
snr_threshold_high=20,
tar_fs=192000):
# tar_fs = 192000
# tar_fs = 400000
folder_list = find_click.list_files(class_path)
if not folder_list:
folder_list = folder_list + [class_path]
for folder in folder_list:
print(folder)
count = 0
wav_files = find_click.list_wav_files(folder)
# wav_files = shuffle_frames(wav_files)
path_name = folder.split('/')[-1]
dst_path = "./CNN_Det12_WK5/%(class)s/%(type)s" % {
'class': class_name,
'type': path_name
}
if not os.path.exists(dst_path):
mkdir(dst_path)
save_npy = True
for file_name in wav_files:
run_cnn_detection(file_name, snr_threshold_low, snr_threshold_high,
save_npy, dst_path, tar_fs)
def load_data(data_path, n_class, batch_num=20, n_total=500):
train_xs = np.empty((0, 192))
train_ys = np.empty((0, n_class))
test_xs = np.empty((0, 192))
test_ys = np.empty((0, n_class))
for c in range(0, n_class):
path = "%(path)s/%(class)d" % {'path': data_path, 'class': c}
wav_files = find_click.list_wav_files(path)
print("load data : %s, the number of files : %d" %
(path, len(wav_files)))
label = np.zeros(n_class)
label[c] = 1
# xs = np.empty((0, 256))
xs = np.empty((0, 320))
count = 0
#
for pathname in wav_files:
wave_data, frame_rate = find_click.read_wav_file(pathname)
energy = np.sqrt(np.sum(wave_data**2))
wave_data /= energy
wave_data = np.reshape(wave_data, [-1])
xs = np.vstack((xs, wave_data))
count += 1
if count >= batch_num * n_total:
break
xs0, xs1 = split_data(xs)
temp_train_xs = random_crop(xs0, batch_num, int(n_total * 4 / 5))
temp_test_xs = random_crop(xs1, batch_num, int(n_total / 5))
temp_train_ys = np.tile(label, (temp_train_xs.shape[0], 1))
temp_test_ys = np.tile(label, (temp_test_xs.shape[0], 1))
train_xs = np.vstack((train_xs, temp_train_xs))
train_ys = np.vstack((train_ys, temp_train_ys))
test_xs = np.vstack((test_xs, temp_test_xs))
test_ys = np.vstack((test_ys, temp_test_ys))
return train_xs, train_ys, test_xs, test_ys
def detect_save_click(class_path,
class_name,
snr_threshold_low=5,
snr_threshold_high=100):
tar_fs = 96000
signal_len = 320
folder_list = find_click.list_files(class_path)
if folder_list == []:
folder_list = folder_list + [class_path]
for folder in folder_list:
print(folder)
count = 0
wav_files = find_click.list_wav_files(folder)
# wav_files = shuffle_frames(wav_files)
path_name = folder.split('/')[-1]
dst_path = "./TKEO_wk3_complete/%(class)s/%(type)s" % {
'class': class_name,
'type': path_name
}
if not os.path.exists(dst_path):
mkdir(dst_path)
for pathname in wav_files:
print(pathname)
wave_data, frameRate = find_click.read_wav_file(pathname)
# wave_data = resample(wave_data, frameRate, tar_fs) #
[path, wavname_ext] = os.path.split(pathname)
wavname = wavname_ext.split('/')[-1]
wavname = wavname.split('.')[0]
fl = 5000
fwhm = 0.0004
fdr_threshold = 0.65
click_index, xn = find_click.find_click_fdr_tkeo(
wave_data, frameRate, fl, fwhm, fdr_threshold, signal_len, 8)
scale = (2**12 - 1) / max(xn)
for i in np.arange(xn.size):
xn[i] = xn[i] * scale
click_arr = []
for j in range(click_index.shape[0]):
index = click_index[j]
# click_data = wave_data[index[0]:index[1], 0]
click_data = xn[index[0]:index[1]]
# 信噪比过滤
detected_clicks_energy = calcu_click_energy(
click_data.reshape(1, -1))
noise_estimate1 = xn[index[0] - 256:index[0]]
noise_estimate2 = xn[index[1] + 1:index[1] + 257]
noise_estimate = np.hstack((noise_estimate1, noise_estimate2))
noise_energy = calcu_energy(noise_estimate)
if noise_energy <= 0 or detected_clicks_energy <= 0:
continue
snr = 10 * math.log10(detected_clicks_energy / noise_energy)
if snr < snr_threshold_low or snr > snr_threshold_high:
continue
click_data = resample(click_data, frameRate, tar_fs) # 前置TKEO前
click_data = cut_data(click_data, signal_len)
click_data = click_data.astype(np.short)
click_arr.append(click_data)
# filename = "%(path)s/%(pre)s_click_%(n)06d.wav" % {'path': dst_path, 'pre': wavname, 'n': count}
# f = wave.open(filename, "wb")
# # set wav params
# f.setnchannels(1)
# f.setsampwidth(2)
# f.setframerate(tar_fs)
# # turn the data to string
# f.writeframes(click_data.tostring())
# f.close()
count = count + 1
dst = "%(path)s/%(pre)s_N%(num)d.npy" \
% {'path': dst_path, 'pre': wavname, 'num': len(click_arr)}
print(dst)
np.save(dst, np.array(click_arr, dtype=np.short))
# if count > 20000:
# break
print("count = %(count)d" % {'count': count})
def load_lwy_data(batch_num=20, n_total=500):
dict = {
'0': '',
'1': '',
'2': '',
'3': '',
'4': '',
'5': '',
'6': '',
'7': ''
}
dict[
"0"] = "/home/fish/ROBB/CNN_click/click/WavData/BBW/Blainvilles_beaked_whale_(Mesoplodon_densirostris)"
dict[
"1"] = "/home/fish/ROBB/CNN_click/click/WavData/Gm/Pilot_whale_(Globicephala_macrorhynchus)"
dict[
"2"] = "/home/fish/ROBB/CNN_click/click/WavData/Gg/Rissos_(Grampus_grisieus)"
dict["3"] = "/home/fish/ROBB/CNN_click/click/WavData/Tt/palmyra2006"
dict["4"] = "/home/fish/ROBB/CNN_click/click/WavData/Dc/Dc"
dict["5"] = "/home/fish/ROBB/CNN_click/click/WavData/Dd/Dd"
dict["6"] = "/home/fish/ROBB/CNN_click/click/WavData/Melon/palmyra2006"
dict["7"] = "/home/fish/ROBB/CNN_click/click/WavData/Spinner/palmyra2006"
n_class = len(dict)
train_xs = np.empty((0, 192))
train_ys = np.empty((0, n_class))
test_xs = np.empty((0, 192))
test_ys = np.empty((0, n_class))
for key in dict:
# path = "%(path)s/%(class)d" % {'path': data_path, 'class': c}
path = dict[key]
c = int(key)
wav_files = find_click.list_wav_files(path)
print("load data : %s, the number of files : %d, class: %d" %
(path, len(wav_files), c))
label = np.zeros(n_class)
label[c] = 1
# xs = np.empty((0, 256))
xs = np.empty((0, 320))
count = 0
#
for pathname in wav_files:
wave_data, frame_rate = find_click.read_wav_file(pathname)
# energy = np.sqrt(np.sum(wave_data ** 2))
# wave_data /= energy
wave_data = np.reshape(wave_data, [-1])
xs = np.vstack((xs, wave_data))
count += 1
if count >= (batch_num + 10) * n_total:
break
xs0, xs1 = split_data(xs)
temp_train_xs = random_crop(xs0, batch_num, int(n_total * 4 / 5))
temp_test_xs = random_crop(xs1, batch_num, int(n_total / 5))
temp_train_ys = np.tile(label, (temp_train_xs.shape[0], 1))
temp_test_ys = np.tile(label, (temp_test_xs.shape[0], 1))
train_xs = np.vstack((train_xs, temp_train_xs))
train_ys = np.vstack((train_ys, temp_train_ys))
test_xs = np.vstack((test_xs, temp_test_xs))
test_ys = np.vstack((test_ys, temp_test_ys))
return train_xs, train_ys, test_xs, test_ys
dict[
"6"] = "/media/ywy/本地磁盘/Data/MobySound/5th_Workshop/5th_DCL_data_melon-headed"
dict[
"7"] = "/media/ywy/本地磁盘/Data/MobySound/5th_Workshop/5th_DCL_data_spinner"
dict["0"] = "/media/ywy/本地磁盘/Data/MobySound/3rd_Workshop/" \
"Training_Data/Blainvilles_beaked_whale_(Mesoplodon_densirostris)"
dict[
"1"] = "/media/ywy/本地磁盘/Data/MobySound/3rd_Workshop/Training_Data/Pilot_whale_(Globicephala_macrorhynchus)"
dict[
"2"] = "/media/ywy/本地磁盘/Data/MobySound/3rd_Workshop/Training_Data/Rissos_(Grampus_grisieus)"
for key in dict:
print(dict[key])
count = 0
wav_files = find_click.list_wav_files(dict[key])
dst_path = "./Data/ClickC8/%(class)s" % {'class': key}
mkdir(dst_path)
for pathname in wav_files:
print(pathname)
wave_data, frameRate = find_click.read_wav_file(pathname)
fl = 5000
fwhm = 0.0008
fdr_threshold = 0.62
click_index, xn = find_click.find_click_fdr_tkeo(
wave_data, frameRate, fl, fwhm, fdr_threshold, signal_len, 8)
def load_data_lstm(data_path, n_class, batch_num=20, n_total=500):
train = []
test = []
x_in = tf.placeholder("float", [None, 192])
# 输入
x_image = tf.reshape(x_in, [-1, 1, 192, 1])
# 第一个卷积层
W_conv1 = weight_variable([1, 5, 1, 32])
b_conv1 = bias_variable([32])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_1x2(h_conv1)
# 第二个卷积层
W_conv2 = weight_variable([1, 5, 32, 32])
b_conv2 = bias_variable([32])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_1x2(h_conv2)
# 密集链接层
W_fc1 = weight_variable([1 * 48 * 32, 256])
b_fc1 = bias_variable([256])
h_pool2_flat = tf.reshape(h_pool2, [-1, 1 * 48 * 32])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
# Dropout
keep_prob = tf.placeholder("float")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob=keep_prob)
# 输出层
W_fc2 = weight_variable([256, n_class])
b_fc2 = bias_variable([n_class])
y = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
init = tf.global_variables_initializer()
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(init)
saver.restore(sess, "params/cnn_net.ckpt") # 加载训练好的网络参数
test_cnn = []
for c in range(0, n_class):
# path = "./Data/Click/%(class)d" % {'class': c}
path = "%(path)s/%(class)d" % {'path': data_path, 'class': c}
wav_files = find_click.list_wav_files(path)
print("load data : %s, the number of files : %d" %
(path, len(wav_files)))
# xs = np.empty((0, 256))
xs = np.empty((0, 320))
count = 0
for pathname in wav_files:
wave_data, frame_rate = find_click.read_wav_file(pathname)
energy = np.sqrt(np.sum(wave_data**2))
wave_data /= energy
wave_data = np.reshape(wave_data, [-1])
xs = np.vstack((xs, wave_data))
count += 1
if count > batch_num * n_total:
break
xs0, xs1 = split_data(xs)
sample_num = xs0.shape[0]
for i in range(0, int(n_total * 4 / 5)):
frames = np.empty((0, 256))
for j in range(batch_num * i, batch_num * (i + 1)):
index = j % sample_num
temp_x = xs0[index]
# beg_idx = np.random.randint(0, 32)
beg_idx = np.random.randint(64, (64 + 32))
crop_x = temp_x[beg_idx:(beg_idx + 192)]
crop_x = np.reshape(crop_x, [1, 192])
ftu = sess.run(h_fc1, feed_dict={x_in:
crop_x}) # 计算CNN网络输出
frames = np.vstack((frames, ftu))
frames = np.expand_dims(np.expand_dims(frames, axis=0), axis=0)
frames = list(frames)
label = [0] * n_class
label[c] = 1
label = np.array([[label]])
label = list(label)
sample = frames + label
train.append(sample)
sample_num = xs1.shape[0]
for i in range(0, int(n_total / 5)):
frames = np.empty((0, 256))
tmp_xs = np.empty((0, 192))
for j in range(batch_num * i, batch_num * (i + 1)):
index = j % sample_num
temp_x = xs1[index]
# beg_idx = np.random.randint(0, 32)
beg_idx = np.random.randint(64, (64 + 32))
crop_x = temp_x[beg_idx:(beg_idx + 192)]
crop_x = np.reshape(crop_x, [1, 192])
ftu = sess.run(h_fc1, feed_dict={x_in:
crop_x}) # 计算CNN网络输出
frames = np.vstack((frames, ftu))
tmp_xs = np.vstack((tmp_xs, crop_x))
frames = np.expand_dims(np.expand_dims(frames, axis=0), axis=0)
frames = list(frames)
label = [0] * n_class
label[c] = 1
label = np.array([[label]])
label = list(label)
sample = frames + label
test.append(sample)
tmp_xs = np.expand_dims(np.expand_dims(tmp_xs, axis=0), axis=0)
tmp_xs = list(tmp_xs)
sample = tmp_xs + label
test_cnn.append(sample)
count = 0
for i in range(len(test_cnn)):
test_xs = test_cnn[i][0]
label = np.zeros(n_class)
for j in range(0, test_xs.shape[1]):
txs = test_xs[0, j, :]
txs = np.reshape(txs, [1, 192])
out_y = sess.run(y, feed_dict={x_in: txs, keep_prob: 1.0})
c = np.argmax(out_y, 1)
label[c] += 1
ref_y = test_cnn[i][1]
if np.equal(np.argmax(label), np.argmax(ref_y)):
count += 1
print('cnn test accuracy: ', round(count / len(test_cnn), 3))
return train, test
def test_cnn_batch_data(data_path, n_class, batch_num=20, n_total=500):
click_batch = []
for c in range(0, n_class):
path = "%(path)s/%(class)d" % {'path': data_path, 'class': c}
wav_files = find_click.list_wav_files(path)
print("load data : %s, the number of files : %d" % (path, len(wav_files)))
# 为避免训练网络用的Click用于测试, 类似于训练时区分训练和测试样本
# 利用全部样本后1/5的Click生成测试样本
xs = np.empty((0, 320))
count = 0
split_idx = int(len(wav_files) * 4 / 5)
for pathname in wav_files:
count += 1
if count < split_idx:
continue
wave_data, frame_rate = find_click.read_wav_file(pathname)
energy = np.sqrt(np.sum(wave_data ** 2))
wave_data /= energy
wave_data = np.reshape(wave_data, [-1])
xs = np.vstack((xs, wave_data))
if count >= batch_num * n_total:
break
sample_num = xs.shape[0]
for i in range(0, int(n_total / 5)):
tmp_xs = np.empty((0, 192))
for j in range(batch_num * i, batch_num * (i + 1)):
index = j % sample_num
temp_x = xs[index]
beg_idx = np.random.randint(64, (64 + 32))
crop_x = temp_x[beg_idx:(beg_idx + 192)]
crop_x = np.reshape(crop_x, [1, 192])
tmp_xs = np.vstack((tmp_xs, crop_x))
label = [0] * n_class
label[c] = 1
label = np.array([[label]])
label = list(label)
tmp_xs = np.expand_dims(np.expand_dims(tmp_xs, axis=0), axis=0)
tmp_xs = list(tmp_xs)
sample = tmp_xs + label
click_batch.append(sample)
x = tf.placeholder("float", [None, 192])
# 输入
x_image = tf.reshape(x, [-1, 1, 192, 1])
# 第一个卷积层
W_conv1 = weight_variable([1, 5, 1, 32])
b_conv1 = bias_variable([32])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_1x2(h_conv1)
# 第二个卷积层
W_conv2 = weight_variable([1, 5, 32, 32])
b_conv2 = bias_variable([32])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_1x2(h_conv2)
# 密集链接层
W_fc1 = weight_variable([1 * 48 * 32, 256])
b_fc1 = bias_variable([256])
h_pool2_flat = tf.reshape(h_pool2, [-1, 1 * 48 * 32])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
# Dropout
keep_prob = tf.placeholder("float")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob=keep_prob)
# 输出层
W_fc2 = weight_variable([256, n_class])
b_fc2 = bias_variable([n_class])
y = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
init = tf.global_variables_initializer()
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(init)
saver.restore(sess, "params/cnn_net.ckpt") # 加载训练好的网络参数
count = 0
for i in range(len(click_batch)):
temp_xs = click_batch[i][0]
label = np.zeros(n_class)
for j in range(0, temp_xs.shape[1]):
txs = temp_xs[0, j, :]
txs = np.reshape(txs, [1, 192])
out_y = sess.run(y, feed_dict={x: txs, keep_prob: 1.0})
c = np.argmax(out_y, 1)
label[c] += 1
ref_y = click_batch[i][1]
if np.equal(np.argmax(label), np.argmax(ref_y)):
count += 1
print('cnn test accuracy (majority voting): ', round(count / len(click_batch), 3))
count = 0
weight = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
for i in range(len(click_batch)):
temp_xs = click_batch[i][0]
label = np.zeros(n_class)
for j in range(0, temp_xs.shape[1]):
txs = temp_xs[0, j, :]
txs = np.reshape(txs, [1, 192])
out = sess.run(weight, feed_dict={x: txs, keep_prob: 1.0})
out = np.reshape(out, label.shape)
label = label + out
ref_y = click_batch[i][1]
if np.equal(np.argmax(label), np.argmax(ref_y)):
count += 1
print('cnn test accuracy (weight voting): ', round(count / len(click_batch), 3))
count = 0
for i in range(len(click_batch)):
temp_xs = click_batch[i][0]
label = np.zeros(n_class)
for j in range(0, temp_xs.shape[1]):
txs = temp_xs[0, j, :]
txs = np.reshape(txs, [1, 192])
out = sess.run(y, feed_dict={x: txs, keep_prob: 1.0})
out = np.reshape(out, label.shape)
label = label + out
ref_y = click_batch[i][1]
if np.equal(np.argmax(label), np.argmax(ref_y)):
count += 1
print('cnn test accuracy (sum of softmax voting): ', round(count / len(click_batch), 3))