from keras.layers import Input, Dense, Conv2D, MaxPooling2D, UpSampling2D, subtract
from keras.models import Model
from keras import backend as K
from keras.callbacks import ModelCheckpoint
from keras.layers.core import Dense,Dropout,Activation,Flatten,Lambda
MAX = 1
GAUS_KERNEL = 3
GAUS_SIGMA = 1.2
def Gaussian_Filter(kernel_size=GAUS_KERNEL, sigma=GAUS_SIGMA):
k = (kernel_size-1)//2
filter = []
sigma_2 = sigma**2
for i in range(kernel_size):
filter_row = []
for j in range(kernel_size):
Hij = np.exp(-((i+1-(k+1))**2 + (j+1-(k+1))**2)/(2*sigma_2)) /(2*np.pi*sigma_2)
filter_row.append(Hij)
filter.append(filter_row)
return np.asarray(filter).reshape(kernel_size,kernel_size,1,1)
gaussian_filter = tf.constant(Gaussian_Filter(), tf.float32)
h_filter = tf.reshape(tf.constant([[-1,0,1],[-2,0,2],[-1,0,1]], tf.float32), [3,3,1,1])
v_filter = tf.reshape(tf.constant([[1,2,1],[0,0,0],[-1,-2,-1]], tf.float32), [3,3,1,1])
np_filter_0 = np.zeros((3,3,1,2))
np_filter_0[1,0,0,0], np_filter_0[1,2,0,1] = 1,1
filter_0 = tf.constant(np_filter_0, tf.float32)
np_filter_90 = np.zeros((3,3,1,2))
np_filter_90[0,1,0,0], np_filter_90[2,1,0,1] = 1,1
filter_90 = tf.constant(np_filter_90, tf.float32)
np_filter_45 = np.zeros((3,3,1,2))
np_filter_45[0,2,0,0], np_filter_45[2,0,0,1] = 1,1
filter_45 = tf.constant(np_filter_45, tf.float32)
np_filter_135 = np.zeros((3,3,1,2))
np_filter_135[0,0,0,0], np_filter_135[2,2,0,1] = 1,1
filter_135 = tf.constant(np_filter_135, tf.float32)
np_filter_sure = np.ones([3,3,1,1]); np_filter_sure[1,1,0,0] = 0
filter_sure = tf.constant(np_filter_sure, tf.float32)
border_paddings = tf.constant([[0,0],[1,1],[1,1],[0,0]])
def Border_Padding(x, pad_width):
for _ in range(pad_width): x = tf.pad(x, border_paddings, 'SYMMETRIC')
return x
def FourAngles(d):
d0 = tf.to_float(tf.greater_equal(d,157.5))+tf.to_float(tf.less(d,22.5))
d45 = tf.to_float(tf.greater_equal(d,22.5))*tf.to_float(tf.less(d,67.5))
d90 = tf.to_float(tf.greater_equal(d,67.5))*tf.to_float(tf.less(d,112.5))
d135 = tf.to_float(tf.greater_equal(d,112.5))*tf.to_float(tf.less(d,157.5))
return (d0,d45,d90,d135)
def TF_Canny(img_tensor, minRate=0.10, maxRate=0.40,preserve_size=True, remove_high_val=False, return_raw_edges=False):
img_tensor = (img_tensor)*MAX
if preserve_size: img_tensor = Border_Padding(img_tensor, (GAUS_KERNEL-1)//2)
x_gaussian = tf.nn.convolution(img_tensor, gaussian_filter, padding='VALID')
if remove_high_val: x_gaussian = tf.clip_by_value(x_gaussian, 0, MAX/2)
if preserve_size: x_gaussian = Border_Padding(x_gaussian, 1)
Gx = tf.nn.convolution(x_gaussian, h_filter, padding='VALID')
Gy = tf.nn.convolution(x_gaussian, v_filter, padding='VALID')
G = tf.sqrt(tf.square(Gx) + tf.square(Gy))
BIG_PHI = tf.atan2(Gy,Gx)
BIG_PHI = (BIG_PHI*180/np.pi)%180
D_0,D_45,D_90,D_135 = FourAngles(BIG_PHI)
targetPixels_0 = tf.nn.convolution(G, filter_0, padding='SAME')
isGreater_0 = tf.to_float(tf.greater(G*D_0, targetPixels_0))
isMax_0 = isGreater_0[:,:,:,0:1]*isGreater_0[:,:,:,1:2]
targetPixels_90 = tf.nn.convolution(G, filter_90, padding='SAME')
isGreater_90 = tf.to_float(tf.greater(G*D_90, targetPixels_90))
isMax_90 = isGreater_90[:,:,:,0:1]*isGreater_90[:,:,:,1:2]
targetPixels_45 = tf.nn.convolution(G, filter_45, padding='SAME')
isGreater_45 = tf.to_float(tf.greater(G*D_45, targetPixels_45))
isMax_45 = isGreater_45[:,:,:,0:1]*isGreater_45[:,:,:,1:2]
targetPixels_135 = tf.nn.convolution(G, filter_135, padding='SAME')
isGreater_135 = tf.to_float(tf.greater(G*D_135, targetPixels_135))
isMax_135 = isGreater_135[:,:,:,0:1]*isGreater_135[:,:,:,1:2]
edges_raw = G*(isMax_0 + isMax_90 + isMax_45 + isMax_135)
edges_raw = tf.clip_by_value(edges_raw, 0, MAX)
if return_raw_edges: return tf.squeeze(edges_raw)
edges_sure = tf.to_float(tf.greater_equal(edges_raw, maxRate))
edges_weak = tf.to_float(tf.less(edges_raw, maxRate))*tf.to_float(tf.greater_equal(edges_raw, minRate))
edges_connected = tf.nn.convolution(edges_sure, filter_sure, padding='SAME')*edges_weak
for _ in range(10): edges_connected = tf.nn.convolution(edges_connected, filter_sure, padding='SAME')*edges_weak
edges_final = edges_sure + tf.clip_by_value(edges_connected,0,MAX)
return tf.squeeze(edges_final)
def edge_loss(y_true, y_pred):
y_true = tf.cast(y_true,dtype = tf.float32)
y_pred = tf.cast(y_pred,dtype = tf.float32)
E = TF_Canny(y_true, return_raw_edges=False)
E = 1 - E
E = tf.cast(E,dtype = tf.float32)
A = y_true - y_pred
A = tf.cast(A,dtype = tf.float32)
E = tf.reshape(E,(8,128,128,3))
LP = K.abs(A)
LP = tf.cast(LP,dtype = tf.float32)
SQ = K.mean(K.square(y_pred - y_true), axis=-1)
LE = tf.multiply(E,LP)
LE = tf.cast(LE,dtype = tf.float32)
SQ1 = K.mean(K.square(LE), axis=-1)
los = 0.7 * SQ + 0.3 * SQ1
return los
input_img = Input(shape=(128,128,3))
x = Conv2D(64, (3, 3), activation='relu', padding='same') (input_img)
x = Conv2D(64, (3, 3), strides = (2, 2), activation='relu', padding='same')(x)
# x = Conv2D(64, (3, 3), strides = (2, 2), activation='relu', padding='same')(x)
encoded = Conv2D(3, (3,3), activation='sigmoid', padding='same')(x)
inp = Lambda(bicub)(encoded)
dup = inp
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(inp)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
model = Conv2D(64, (3, 3), padding='same', kernel_initializer='he_normal')(model)
model = Activation('relu')(model)
res_img = Conv2D(3, (3, 3), padding='same', kernel_initializer='he_normal')(model)
output_img =添加（[dup，res_img]）
model = Model(input_img, res_img)
# b = keras.optimizers.Adam(lr=0.0001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)
model.summary()
model.compile(optimizer= 'adam', loss=edge_loss)
mc = keras.callbacks.ModelCheckpoint('drive/My Drive/altweights2/weights{epoch}.hdf5', save_best_only=False, period=1,verbose = 1)
model.fit(x_train2, x_train2,
validation_data = (x_train2, x_train2),
epochs=100,
batch_size=8,
shuffle=True,
callbacks=[mc])
先生，当我使用网络训练数据集但使用2次时，两次具有不同的损失函数。一个是定制的损失函数，另一个是定义的均方误差函数。在两种条件相同的情况下，当进行均方误差训练时，它的训练效果很好，但是对于另一方训练时，它显示出以下误差， 我在网络上检查了很多，但找不到任何错误。 需要帮助
NotFoundError Traceback (most recent call last)
<ipython-input-31-1b5a177291a2> in <module>()
6 batch_size=8,
7 shuffle=True,
----> 8 callbacks=[mc,TensorBoardColabCallback(tbc) ])
5 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python /framework /errors_impl.py in __exit__(self, type_arg, value_arg, traceback_arg)
526 None, None,
527 compat.as_text(c_api.TF_Message(self.status.status)),
--> 528 c_api.TF_GetCode(self.status.status))
529 # Delete the underlying status object from memory otherwise it stays alive
530 # as there is a reference to status from this from the traceback due to
NotFoundError: No algorithm worked!
[[{{node loss_6/conv2d_92_loss/convolution}}]]
[[{{node training_5/Adam/gradients/lambda_4/truediv_grad /Reshape}}]]
InvalidArgumentError Traceback (most recent call last)
<ipython-input-20-292832ca5d2b> in <module>()
6 batch_size=8,
7 shuffle=True,
----> 8 callbacks=[mc])
/home/titanxpascal/Documents/myenv/local/lib/python2.7/site-packages /keras/engine/training.pyc in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)
1037 initial_epoch=initial_epoch,
1038 steps_per_epoch=steps_per_epoch,
-> 1039验证步骤=验证步骤） 1040 1041 defvaluate（self，x = None，y = None，
/home/titanxpascal/Documents/myenv/local/lib/python2.7/site-packages /keras/engine/training_arrays.pyc in fit_loop（model，f，ins，out_labels，batch_size，历元，详细，回调， val_f，val_ins，随机播放，callback_metrics，initial_epoch，steps_per_epoch，validation_steps） 197 ins_batch [i] = ins_batch [i] .toarray（） 198 -> 199个出局= f（ins_batch） 200次= to_list（outs） zip中的l，o为201（out_labels，outs）：
/home/titanxpascal/Documents/myenv/local/lib/python2.7/site-packages /keras/backend/tensorflow_backend.pyc in __call__(self, inputs)
2713 return self._legacy_call(inputs)
2714
-> 2715返回self._call（输入） 2716其他： 2717 if py_any（is_tensor（x）for x in input）：
/home/titanxpascal/Documents/myenv/local/lib/python2.7/site-packages /keras/backend/tensorflow_backend.pyc in _call(self, inputs)
2673 fetched = self._callable_fn(*array_vals, run_metadata=self.run_metadata)
2674 else:
-> 2675 fetched = self._callable_fn(*array_vals)
2676 return fetched[:len(self.outputs)]
2677
/home/titanxpascal/Documents/myenv/local/lib/python2.7/site-packages /tensorflow/python/client/session.pyc in __call__(self, *args)
1452 else:
1453 return tf_session.TF_DeprecatedSessionRunCallable(
-> 1454 self._session._session, self._handle, args, status, None)
1455
1456 def __del__(self):
/home/titanxpascal/Documents/myenv/local/lib/python2.7/site-packages /tensorflow/python/framework/errors_impl.pyc in __exit__(self, type_arg, value_arg, traceback_arg)
517 None, None,
518 compat.as_text(c_api.TF_Message(self.status.status)),
--> 519 c_api.TF_GetCode(self.status.status))
520 # Delete the underlying status object from memory otherwise it stays alive
521 # as there is a reference to status from this from the traceback due to
InvalidArgumentError: input and filter must have the same depth: 3 vs 1
[[Node: loss_7/conv2d_23_loss/convolution = Conv2D[T=DT_FLOAT, data_format="NCHW", dilations=[1, 1, 1, 1], padding="VALID", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true, _device="/job:localhost/replica:0/task:0/device:GPU:0"](loss_7/conv2d_23_loss/convolution-0-TransposeNHWCToNCHW-LayoutOptimizer, Const_9)]]
[[Node: training_7/Adam/gradients/loss_7/conv2d_23_loss/sub_2_grad/Shape/_3001 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device_incarnation=1, tensor_name="edge_507_training_7/Adam/gradients/loss_7/conv2d_23_loss/sub_2_grad/Shape", tensor_type=DT_INT32, _device="/job:localhost/replica:0/task:0/device:CPU:0"]()]]