def prepare_darknet_params():
inputs = Input(shape=(None, None, 3))
model = Model(inputs, darknet_body(inputs))
# det_body = DarkNetDet((None, None), 85)
# model = det_body.get_detection_body()
with h5py.File('model_data/yolo3_weights.h5', 'r') as f:
for idx, layer in enumerate(model.layers):
print(idx, layer.name)
if len(layer.weights) != 0:
splited = layer.name.split("_")
if splited[-1].isdigit():
splited[-1] = str(int(splited[-1]) + 1)
else:
splited.append("1")
h5_name = "_".join(splited)
weight_names = load_attributes_from_hdf5_group(f[h5_name], 'weight_names')
weight_values = [np.asarray(f[h5_name][weight_name]) for weight_name in weight_names]
layer.set_weights(weight_values)
model.save_weights("yolo_pretrained.h5")
class BaseKerasModel(BaseModel):
model = None
tensorboard = None
train_names = ['train_loss', 'train_mse', 'train_mae']
val_names = ['val_loss', 'val_mse', 'val_mae']
counter = 0
inputs = None
hidden_layer = None
outputs = None
def __init__(self,
use_default_dense=True,
activation='relu',
kernel_regularizer=tf.keras.regularizers.l1(0.001)):
super().__init__()
if use_default_dense:
self.activation = activation
self.kernel_regularizer = kernel_regularizer
def create_input_layer(self, input_placeholder: BaseInputFormatter):
"""Creates keras model"""
self.inputs = tf.keras.layers.InputLayer(
input_shape=input_placeholder.get_input_state_dimension())
return self.inputs
def create_hidden_layers(self, input_layer=None):
if input_layer is None:
input_layer = self.inputs
hidden_layer = tf.keras.layers.Dropout(0.3)(input_layer)
hidden_layer = tf.keras.layers.Dense(
128,
kernel_regularizer=self.kernel_regularizer,
activation=self.activation)(hidden_layer)
hidden_layer = tf.keras.layers.Dropout(0.4)(hidden_layer)
hidden_layer = tf.keras.layers.Dense(
64,
kernel_regularizer=self.kernel_regularizer,
activation=self.activation)(hidden_layer)
hidden_layer = tf.keras.layers.Dropout(0.3)(hidden_layer)
hidden_layer = tf.keras.layers.Dense(
32,
kernel_regularizer=self.kernel_regularizer,
activation=self.activation)(hidden_layer)
hidden_layer = tf.keras.layers.Dropout(0.1)(hidden_layer)
self.hidden_layer = hidden_layer
return self.hidden_layer
def create_output_layer(self,
output_formatter: BaseOutputFormatter,
hidden_layer=None):
# sigmoid/tanh all you want on self.model
if hidden_layer is None:
hidden_layer = self.hidden_layer
self.outputs = tf.keras.layers.Dense(
output_formatter.get_model_output_dimension()[0],
activation='tanh')(hidden_layer)
self.model = Model(inputs=self.inputs, outputs=self.outputs)
return self.outputs
def write_log(self, callback, names, logs, batch_no, eval=False):
for name, value in zip(names, logs):
summary = tf.Summary()
summary_value = summary.value.add()
summary_value.simple_value = value
tag_name = name
if eval:
tag_name = 'eval_' + tag_name
summary_value.tag = tag_name
callback.writer.add_summary(summary, batch_no)
callback.writer.flush()
def finalize_model(self, logname=str(int(random() * 1000))):
loss, loss_weights = self.create_loss()
self.model.compile(tf.keras.optimizers.Nadam(lr=0.001),
loss=loss,
loss_weights=loss_weights,
metrics=[
tf.keras.metrics.mean_absolute_error,
tf.keras.metrics.binary_accuracy
])
log_name = './logs/' + logname
self.logger.info("log_name: " + log_name)
self.tensorboard = tf.keras.callbacks.TensorBoard(
log_dir=log_name,
histogram_freq=1,
write_images=False,
batch_size=1000,
)
self.tensorboard.set_model(self.model)
self.logger.info("Model has been finalized")
def fit(self, x, y, batch_size=1):
if self.counter % 200 == 0:
logs = self.model.evaluate(x, y, batch_size=batch_size, verbose=1)
self.write_log(self.tensorboard,
self.model.metrics_names,
logs,
self.counter,
eval=True)
print('step:', self.counter)
else:
logs = self.model.train_on_batch(x, y)
self.write_log(self.tensorboard, self.model.metrics_names, logs,
self.counter)
self.counter += 1
def predict(self, arr):
return self.model.predict(arr)
def save(self, file_path):
self.model.save_weights(filepath=file_path, overwrite=True)
def load(self, file_path):
path = os.path.abspath(file_path)
self.model.load_weights(filepath=os.path.abspath(file_path))
def create_loss(self):
return 'mean_absolute_error', None
class jyHEDModelV2_2_SGD_GradientTape_L1(jyModelBase):
def __init__(self):
super(jyHEDModelV2_2_SGD_GradientTape_L1, self).__init__()
self.__listLayerName = []
self.__pVisualModel = None
self.__bLoadModel = False
self.__pTrainFW = tf.summary.create_file_writer(self._strLogPath +
'/train')
self.__pValidFW = tf.summary.create_file_writer(self._strLogPath +
'/valid')
self.__pMetricsFW = tf.summary.create_file_writer(self._strLogPath +
'/metrics')
def structureModel(self):
weightDecay = 0.00001
Inputs = layers.Input(shape=self._inputShape,
batch_size=self._iBatchSize)
Con1 = layers.Conv2D(64, (3, 3),
name='Con1',
activation='relu',
padding='SAME',
input_shape=self._inputShape,
strides=1,
kernel_regularizer=l2(weightDecay))(Inputs)
Con2 = layers.Conv2D(64, (3, 3),
name='Con2',
activation='relu',
padding='SAME',
strides=1,
kernel_regularizer=l2(weightDecay))(Con1)
Side1 = sideBranch(Con2, 1)
MaxPooling1 = layers.MaxPooling2D((2, 2),
name='MaxPooling1',
strides=2,
padding='SAME')(Con2)
# outputs1
Con3 = layers.Conv2D(128, (3, 3),
name='Con3',
activation='relu',
padding='SAME',
strides=1,
kernel_regularizer=l2(weightDecay))(MaxPooling1)
Con4 = layers.Conv2D(128, (3, 3),
name='Con4',
activation='relu',
padding='SAME',
strides=1,
kernel_regularizer=l2(weightDecay))(Con3)
Side2 = sideBranch(Con4, 2)
MaxPooling2 = layers.MaxPooling2D((2, 2),
name='MaxPooling2',
strides=2,
padding='SAME')(Con4)
# outputs2
Con5 = layers.Conv2D(256, (3, 3),
name='Con5',
activation='relu',
padding='SAME',
strides=1,
kernel_regularizer=l2(weightDecay))(MaxPooling2)
Con6 = layers.Conv2D(256, (3, 3),
name='Con6',
activation='relu',
padding='SAME',
strides=1,
kernel_regularizer=l2(weightDecay))(Con5)
Con7 = layers.Conv2D(256, (3, 3),
name='Con7',
activation='relu',
padding='SAME',
strides=1,
kernel_regularizer=l2(weightDecay))(Con6)
Side3 = sideBranch(Con7, 4)
MaxPooling3 = layers.MaxPooling2D((2, 2),
name='MaxPooling3',
strides=2,
padding='SAME')(Con7)
# outputs3
Con8 = layers.Conv2D(512, (3, 3),
name='Con8',
activation='relu',
padding='SAME',
strides=1,
kernel_regularizer=l2(weightDecay))(MaxPooling3)
Con9 = layers.Conv2D(512, (3, 3),
name='Con9',
activation='relu',
padding='SAME',
strides=1,
kernel_regularizer=l2(weightDecay))(Con8)
Con10 = layers.Conv2D(512, (3, 3),
name='Con10',
activation='relu',
padding='SAME',
strides=1,
kernel_regularizer=l2(weightDecay))(Con9)
Side4 = sideBranch(Con10, 8)
MaxPooling4 = layers.MaxPooling2D((2, 2),
name='MaxPooling4',
strides=2,
padding='SAME')(Con10)
# outputs4
Con11 = layers.Conv2D(512, (3, 3),
name='Con11',
activation='relu',
padding='SAME',
strides=1,
kernel_regularizer=l2(weightDecay))(MaxPooling4)
Con12 = layers.Conv2D(512, (3, 3),
name='Con12',
activation='relu',
padding='SAME',
strides=1,
kernel_regularizer=l2(weightDecay))(Con11)
Con13 = layers.Conv2D(512, (3, 3),
name='Con13',
activation='relu',
padding='SAME',
strides=1,
kernel_regularizer=l2(weightDecay))(Con12)
Side5 = sideBranch(Con13, 16)
Fuse = layers.Concatenate(axis=-1)([Side1, Side2, Side3, Side4, Side5])
# learn fusion weight
fuseInitWeight = initializers.constant(0.2)
Fuse = layers.Conv2D(1, (1, 1),
name='Fuse',
padding='SAME',
use_bias=False,
activation=None,
kernel_initializer=fuseInitWeight,
kernel_regularizer=l1(weightDecay))(Fuse)
# output1 = layers.Activation('sigmoid', name='output1')(Side1)
# output2 = layers.Activation('sigmoid', name='output2')(Side2)
# output3 = layers.Activation('sigmoid', name='output3')(Side3)
# output4 = layers.Activation('sigmoid', name='output4')(Side4)
# output5 = layers.Activation('sigmoid', name='output5')(Side5)
output6 = layers.Activation('sigmoid', name='output6')(Fuse)
outputs = [output6
] # [output1, output2, output3, output4, output5, output6]
self._pModel = Model(inputs=Inputs, outputs=outputs)
pOptimizer = optimizers.adam(lr=0.0001)
pOptimizer = optimizers.SGD(lr=0.000001, decay=0., momentum=0.9)
pOptimizer = tf.optimizers.SGD(lr=0.5, decay=0., momentum=0.9)
# pOptimizer = monitorSGD(lr=0.000001, decay=0., momentum=0.9)
# grads = tf.gradients(classBalancedSigmoidCrossEntropy, self._pModel.trainable_weights)
# pSGD = optimizers.SGD()
self._pModel.compile(
loss={
# 'output1': classBalancedSigmoidCrossEntropy,
# 'output2': classBalancedSigmoidCrossEntropy,
# 'output3': classBalancedSigmoidCrossEntropy,
# 'output4': classBalancedSigmoidCrossEntropy,
# 'output5': classBalancedSigmoidCrossEntropy,
'output6': classBalancedSigmoidCrossEntropy
},
optimizer=pOptimizer)
# self._pModel.summary()
def startTrain(self, listDS, iMaxLen, iBatchSize):
'''
itrTrain = tf.compat.v1.data.make_one_shot_iterator(listDS[0])
itrValid = tf.compat.v1.data.make_one_shot_iterator(listDS[1])
iStepsPerEpochTrain = int(iMaxLen[0] / iBatchSize[0])
iStepsPerEpochValid = int(iMaxLen[1] / iBatchSize[1])
pBack = myCallback(self._strLogPath)
self._pModel.fit(itrTrain, validation_data=itrValid, epochs=self._iEpochs,
callbacks=[self._pSaveModel, self._pTensorboard, pBack], steps_per_epoch=iStepsPerEpochTrain,
validation_steps=iStepsPerEpochValid)
'''
itrTrain = tf.compat.v1.data.make_one_shot_iterator(listDS[0])
itrValid = tf.compat.v1.data.make_one_shot_iterator(listDS[1])
iStepsPerEpochTrain = int(iMaxLen[0] / iBatchSize[0])
iStepsPerEpochValid = int(iMaxLen[1] / iBatchSize[1])
# trainLoss = tf.keras.metrics.Mean(name='train_loss')
dictLossGroup = self._pModel.loss
# t = self._pModel.layers[23].losses
# t = self._pModel.weights[0].name
# p = self._pModel.loss
iTick = 0
# epoch
for epoch in range(self._iEpochs):
# save model
if iTick > self._iPeriod:
strModelFileName = self._strModelFileName.format(epoch=epoch +
1)
filepath = self._strSavePath + strModelFileName
print(self._strFormat %
('Epoch: %s/%s, SaveModel: %s' %
(str(epoch), str(self._iEpochs), strModelFileName)))
self._pModel.save_weights(filepath, overwrite=True)
iTick = 0
iTick += 1
# stepsPerEpoch
for stepsPerEpoch in range(iStepsPerEpochTrain):
with tf.GradientTape() as tape:
itr = itrTrain.next()
# output define as [out1, out2, ....., out6]
listPredict = [self._pModel(itr[0])]
t = self._pModel.weights
listLabel = [itr[1]]
listLoss = []
fAllLoss = 0.
template = 'Per: {}/{}, TrainLoss: {} -- '
i = 0
# multiple output, calculate loss
for key in dictLossGroup:
# loss function
pLoss = dictLossGroup[key]
# add regularize
regularization_loss = tf.math.add_n(
self._pModel.losses)
# pLoss += tf.add_n
# loss value
outputLoss = pLoss(
listLabel[i], listPredict[i]) + regularization_loss
listLoss.append(outputLoss)
# sum of loss
fAllLoss += outputLoss
# print format
template += 'train_loss_%s: {} -- ' % key
i += 1
# calculate gradient
gradient = tape.gradient(fAllLoss,
self._pModel.trainable_weights)
# trainLoss(fAllLoss)
template += '\n'
print(
template.format(stepsPerEpoch + 1, iStepsPerEpochTrain,
fAllLoss, listLoss[0]))
# backprop
self._pModel.optimizer.apply_gradients(
zip(gradient, self._pModel.trainable_weights))
# 每执行完一个train epoch 进行validcross 因此valid计算不能与train同步进行要在train epoch结束后进行
fValidAllLoss = 0.
listValidLoss = list(0 for n in range(len(dictLossGroup)))
for stepsPerEpochValid in range(iStepsPerEpochValid):
itr2 = itrValid.next()
listPreValid = [self._pModel(itr2[0])]
listValidLabel = [itr2[1]]
i = 0
for key in dictLossGroup:
# loss function
pLoss = dictLossGroup[key]
# loss value
outputValidLoss = pLoss(listValidLabel[i], listPreValid[i])
listValidLoss[i] += outputValidLoss
# sum of loss
fValidAllLoss += outputValidLoss
# print format
# template += ' --train_loss_%s: {}-- ' % key
i += 1
# mean of val_loss
fValidAllLoss /= iStepsPerEpochValid
validTemplate = 'Epoch {}, val_loss: {} -- '.format(
epoch + 1, fValidAllLoss)
for k in range(len(listValidLoss)):
listValidLoss[k] /= iStepsPerEpochValid
validTemplate += 'val_loss_{}: {} -- '.format(
k + 1, listValidLoss[k])
print(
'\n-----------------------------------------------------------------------\n'
)
print(validTemplate)
print(
'\n-----------------------------------------------------------------------\n'
)
# per epoch output
with self.__pTrainFW.as_default():
i = 0
tf.summary.scalar('loss: ', fAllLoss, step=epoch)
# tf.summary.scalar('val_loss: ', fValidAllLoss, step=epoch)
for key in dictLossGroup:
tf.summary.scalar('loss_' + key, listLoss[i], step=epoch)
# tf.summary.scalar('val_loss_' + key, listValidLoss[i], step=epoch)
i += 1
with self.__pMetricsFW.as_default():
# save gradient each layer
pLayerWeight = self._pModel.trainable_weights
for i in range(len(pLayerWeight)):
strName = pLayerWeight[i].name + '/Grad'
tf.summary.histogram(strName, gradient[i], step=epoch)
# mean grad
meanGrad = tf.reduce_mean(gradient[i])
tf.summary.scalar(strName + '/Mean', meanGrad, step=epoch)
# model grad
tensorNorm = tf.norm(gradient[i])
tf.summary.scalar(strName + '/Norm',
tensorNorm,
step=epoch)
with self.__pValidFW.as_default():
i = 0
tf.summary.scalar('loss: ', fValidAllLoss, step=epoch)
for key in dictLossGroup:
tf.summary.scalar('loss_' + key,
listValidLoss[i],
step=epoch)
i += 1
def loadWeights(self, strPath):
# last = tf.train.latest_checkpoint(strPath)
# checkPoint = tf.train.load_checkpoint(strPath)
self._pModel.load_weights(strPath)
# w = self._pModel.weights
# visual model
self.__bLoadModel = True
def generateVisualModel(self):
outputs = []
for myLayer in self._pModel.layers:
self.__listLayerName.append(myLayer.name)
outputs.append(myLayer.output)
# print(self.__pModel.layers[0])
# self.__pVisualModel = Model(self.__pModel.inputs, outputs=outputs)
self.__pVisualModel = Model(self._pModel.inputs,
outputs=self._pModel.outputs)
return self.__pVisualModel
def predict(self, IMG):
# pImage = open(IMG, 'rb').read()
# tensorIMG = tf.image.decode_jpeg(pImage)
pIMG = image.array_to_img(IMG) # .resize((256, 144))
tensorIMG = image.img_to_array(pIMG)
x = np.array(tensorIMG / 255.0)
# show image
iColumn = 4
# generate window
plt.figure(num='Input')
# plt.subplot(1, 1, 1)
plt.imshow(x)
# imagetest = x
x = np.expand_dims(x, axis=0)
# pyplot.imshow(x)
time1 = datetime.datetime.now()
outputs = self.__pVisualModel.predict(x)
time2 = datetime.datetime.now()
print(time2 - time1)
i = 100
listOutput = []
for i in range(len(outputs)):
outputShape = outputs[i].shape
singleOut = outputs[i].reshape(outputShape[0], outputShape[1],
outputShape[2])
# singleOut *= 255
listOutput.append(singleOut)
singleOut = listOutput[-1]
singleOut[singleOut > 0.5] = 1
listOutput[-1] = singleOut
return listOutput
'''
for output in outputs:
# plt.figure(num='%s' % str(i))
outputShape = output.shape
singleOut = output.reshape(outputShape[1], outputShape[2], outputShape[3])
singleOut *= 255
if outputShape[3] == 1:
# test = x - output
# test = np.abs(test)
# return mysum
# plt.subplot(1, 1, 1)
# plt.imshow(singleOut, camp='gray')
# cv2.imwrite('D:\wyc\Projects\TrainDataSet\HED\Result/%s.jpg' % str(i), singleOut)
return singleOut
# i += 1
# plt.show()
'''
def getModelConfig(self):
return self._iBatchSize
best_metric = float("inf")
wait = 0
for epoch in range(args.epoches):
printf("training epoch %s..." % (epoch, ))
train_loss, train_ci = loop_dataset(train_inds,
optimizer=optimizer)
printf("train epoch %d loss %.4f ci %.4f \n" %
(epoch, train_loss, train_ci))
printf("validating epoch %s..." % (epoch, ))
val_loss, val_ci = loop_dataset(val_inds, optimizer=None)
printf("validating epoch %d loss %.4f ci %.4f \n" %
(epoch, val_loss, val_ci))
if val_loss <= best_metric:
best_metric = val_loss
DTAModel.save_weights(os.path.join(chkpt_subdir, "DTA"), )
wait = 0
else:
wait += 1
if wait > args.patience:
break
DTAModel.load_weights(os.path.join(chkpt_subdir, "DTA"))
printf("start testing...")
test_loss, test_ci = loop_dataset(test_inds, optimizer=None)
if test_ci > best_ci:
best_ci = test_ci
best_loss = test_loss
best_it = it
kl_loss *= -0.5
vae_loss = K.mean(reconstruction_loss + kl_loss)
vae.add_loss(vae_loss)
vae.compile(optimizer='adam')
vae.summary()
plot_model(vae, to_file='vae_mlp.png', show_shapes=True)
if args.weights:
vae.load_weights(args.weights)
else:
# train the autoencoder
vae.fit(x_train,
epochs=epochs,
batch_size=batch_size,
validation_data=(x_test, None))
vae.save_weights('vae_mlp_mnist.h5')
plot_results(models, data, batch_size=batch_size, model_name="vae_mlp")
# # plot the latent space of the encoded digits #########################################
# x_test_encoded = encoder.predict(x_test, batch_size=batch_size)
# plt.figure(figsize=(6, 6))
# plt.scatter(x_test_encoded[:, 0], x_test_encoded[:, 1], c=y_test)
# plt.colorbar()
# plt.show()
#
# # scan the latent plane and generate new digits #########################################
# # display a 2D manifold of the digits
# n = 15 # figure with 15x15 digits
# digit_size = 28
# figure = np.zeros((digit_size * n, digit_size * n))
