def __init__(self, config):
self.config = config
# load data here
self.h5_reader = H5DataReader(config.h5_data_path, mode='r', data_key=config.h5_data_key, label_key=config.h5_label_key, seed=config.h5_shuffle_seed)
if config.get("h5_condition_args"):
self.h5_reader.set_condition_idx(*config.h5_condition_args)
self.train_batch_generator = None
self.test_batch_generator = None
self.batch_generator = None
#数据长度
sig_len = 10000
#训练集数据量比测试集数据量
train_test_rate = 4
#设置训练模式
mode = 'f'
acc_rate = 0.90 # 预设准确率
batch_size = 20
LR = 1e-4
labels = ['dt1', 'dt2', 'dt3', 'dt4', 'dt5', 'dt6', 'dt7', 'dt8', 'dt9']
tick_marks = np.array(range(len(labels))) + 0.5
row = 2000
data = H5DataReader(dir_path, 'r', 'signals', 'labels')
kind_num = 9
shape1 = 128
shape2 = 128
x = tf.placeholder(tf.float32,
shape=(None, shape1, shape2, 4),
name='x_features')
y_ = tf.placeholder(tf.int64, shape=[None, kind_num], name='y_')
keep_prob = tf.placeholder("float", name='keep_prob')
training = tf.placeholder(tf.bool, name='training')
n_batch = len(data.train_indices) // batch_size
# prelogits, end_points = network_incep.inference(x, keep_prob,
ckpt = tf.train.get_checkpoint_state(model_path)
load_model.load_model(model_path)
# files = os.listdir(model_path)
# meta_file = [s for s in files if s.endswith('.meta')][0]
# saver = tf.train.import_meta_graph(os.path.join(model_path, meta_file))
# Get input and output tensors
input = tf.get_default_graph().get_tensor_by_name("input:0")
is_training = tf.get_default_graph().get_tensor_by_name("is_training:0")
decode = tf.get_default_graph().get_tensor_by_name("decode:0")
#回复模型
# saver.restore(sess, ckpt.model_checkpoint_path)
loss = tf.get_default_graph().get_tensor_by_name("loss/bce:0")
# stft_reader = H5DataReader(stft_path,)
sig_reader = H5DataReader(signal_path, 'r', 'signals',seed=303)
# sig_reader = H5DataReader(stft_path, 'r',seed=30)
batch, _ = sig_reader.get_shuffle_data(20)
batch = np.asarray([signal_regulation(x) for x in batch])
decode_arr ,los_array= sess.run([decode,loss],feed_dict={input:batch,is_training:False})
diff = np.abs(decode_arr-batch)
for i in range(len(decode_arr)):
save_sig_fig(diff[i,:1000,0], '../figures/sv2_diff_{:d}'.format(i))
save_sig_fig(batch[i,:1000,0], '../figures/sv2_input_{:d}'.format(i))
save_sig_fig(decode_arr[i,:1000,0], '../figures/sv2_output_{:d}'.format(i))
# np.exp(decode_arr) / sum(np.exp(decode_arr))
# plot(np.decode_arr, 0)
net = tf.layers.dense(flatten, 1024, activation='relu')
logits = tf.layers.dense(net, N)
softmax = tf.nn.softmax(logits)
loss = tf.losses.sparse_softmax_cross_entropy(labels, logits)
predict = tf.arg_max(softmax, -1)
accuracy = tf.reduce_mean(
tf.cast(tf.equal(predict, labels), tf.float32))
train_varlist = tf.trainable_variables("Finetune")
train_op = tf.train.AdamOptimizer(init_lr).minimize(
loss, var_list=train_varlist, global_step=finetune_step)
#began training
sess.run(tf.global_variables_initializer())
saver.restore(sess, ckpt)
h5_reader = H5DataReader(stft_path, seed=678)
i = 0
los_list = []
acc_list = []
while (iters > i):
i += 1
batch_x, batch_y = h5_reader.get_shuffle_data(batch_size)
# preprocessing
batch_x = [myfft1_norm(x) for x in batch_x]
feed_dict = {is_training: False, input: batch_x, labels: batch_y}
_, prd, los, acc = sess.run([train_op, predict, loss, accuracy],
feed_dict)
los_list.append(los)
acc_list.append(acc)
plt.clf()
def savefig(features, i, name):
k = 0
plt.figure(figsize=(20, 6))
for feature in features:
k += 1
plt.plot(feature[:1000, i], linewidth=1)
plt.savefig(name + str(k) + '.png')
plt.clf()
def plot2d(features, i):
for feature in features:
plt.imshow(feature[:, :, i])
plt.show()
plt.clf()
if __name__ == '__main__':
stft_path = '../../dataset/LTE_dataset_stft_256x256x4_3c_1216.h5'
signal_path = '../../dataset/LTE_origin_3240_dataset_5c_10s_1202.h5'
print(os.path.abspath(signal_path))
# stft_reader = H5DataReader(stft_path,)
# sig_reader = H5DataReader(signal_path, 'r', 'signals')
sig_reader = H5DataReader(stft_path, 'r')
batch, _ = sig_reader.get_shuffle_data(1)
print(batch[0].shape)
plot2d(batch, 2)
with tf.variable_scope("Finetune"):
flatten = tf.layers.flatten(code)
net=tf.layers.dense(flatten, 128, activation='relu')
logits = tf.layers.dense(net,N)
softmax = tf.nn.softmax(logits)
loss = tf.losses.sparse_softmax_cross_entropy(labels,logits)
predict = tf.arg_max(softmax,-1)
accuracy = tf.reduce_mean(tf.cast(tf.equal(predict,labels),tf.float32))
train_varlist = tf.trainable_variables("Finetune")
train_op = tf.train.AdamOptimizer(init_lr).minimize(loss,var_list=train_varlist)
#began training
sess.run(tf.global_variables_initializer())
saver.restore(sess, ckpt)
h5_reader = H5DataReader(signal_path,data_key='signals')
batch_x,batch_y = h5_reader.get_shuffle_data(batch_size)
feed_dict = {is_training:False, input:batch_x, labels:batch_y}
i = 0
los_list= []
acc_list =[]
while(iters > i ):
i+= 1
_, prd, los, acc =sess.run([train_op,predict,loss,accuracy],feed_dict)
los_list.append(los)
acc_list.append(acc)
if i%20 ==0:
los = np.mean(los_list)
acc = np.mean(acc_list)
softmax = tf.nn.softmax(logits)
loss = tf.losses.sparse_softmax_cross_entropy(labels, logits)
predict = tf.arg_max(softmax, -1)
accuracy = tf.reduce_mean(
tf.cast(tf.equal(predict, labels), tf.float32))
train_varlist = tf.trainable_variables("Finetune")
train_op = tf.train.AdamOptimizer(init_lr).minimize(
loss, var_list=train_varlist, global_step=finetune_step)
#began training
new_saver = tf.train.Saver(max_to_keep=1)
sess.run(tf.global_variables_initializer())
import_saver.restore(sess, ckpt)
h5_reader = H5DataReader(stft_path,
seg_set_method='txt',
txt_path=train_idx_txt)
unknown_reader = H5DataReader(stft_path)
unknown_reader.set_condition_idx(condition_keys=['labels', 'fc'],
include_conditions=[(6, 225),
(7, 225)],
exclude_conditions=None)
i = 0
los_list = []
acc_list = []
e_los_list = []
e_acc_list = []
u_los_list = []
u_acc_list = []
while (iters > i):
i += 1