def on_train_begin(self, logs={}):
self.nlayerinput = lambda x: K.function([self.model.layers[0].input], [self.kdelayer.input])([x])[0]
N, dims = self.entropy_train_data.shape
Kdists = K.placeholder(ndim=2)
Klogvar = K.placeholder(ndim=0)
def obj(logvar, dists):
#print 'here', logvar # lossfunc([dists, logvar[0]])[0]
return lossfunc([dists, logvar.flat[0]])[0]
def jac(logvar, dists):
#print logvar, lossfunc([dists, logvar[0]]), jacfunc([dists, logvar[0]])
return np.atleast_2d(np.array(jacfunc([dists, logvar.flat[0]])))[0]
lossfunc = K.function([Kdists, Klogvar,], [kde_entropy_from_dists_loo(Kdists, N, dims, K.exp(Klogvar))])
jacfunc = K.function([Kdists, Klogvar,], K.gradients(kde_entropy_from_dists_loo(Kdists, N, dims, K.exp(Klogvar)), Klogvar))
self.obj =obj # lambda logvar, dists: np.array([lossfunc([dists, logvar[0]]),]) # [0]
self.jac =jac # lambda logvar, dists: jacfunc([dists, np.array([logvar]).flat[0]])[0]
def on_train_begin(self, logs={}):
modelobj = self.model.model
inputs = modelobj.inputs + modelobj.targets + modelobj.sample_weights + [ K.learning_phase(),]
lossfunc = K.function(inputs, [modelobj.total_loss])
jacfunc = K.function(inputs, K.gradients(modelobj.total_loss, self.noiselayer.logvar))
sampleweights = np.ones(len(self.traindata.X))
def obj(logvar):
v = K.get_value(self.noiselayer.logvar)
K.set_value(self.noiselayer.logvar, logvar.flat[0])
r = lossfunc([self.traindata.X, self.traindata.Y, sampleweights, 1])[0]
K.set_value(self.noiselayer.logvar, v)
return r
def jac(logvar):
v = K.get_value(self.noiselayer.logvar)
K.set_value(self.noiselayer.logvar, logvar.flat[0])
r = np.atleast_2d(np.array(jacfunc([self.traindata.X, self.traindata.Y, sampleweights, 1])))[0]
K.set_value(self.noiselayer.logvar, v)
return r
self.obj = obj # lambda logvar: lossfunc([self.traindata.X_train, self.traindata.Y_train, self.sampleweights, logvar[0], 1])[0]
self.jac = jac # lambda logvar: np.array(jacfunc([self.traindata.X_train, self.traindata.Y_train, self.sampleweights, logvar[0], 1]))
def get_logs(model, data, kdelayer, noiselayer, max_entropy_calc_N=None):
logs = {}
modelobj = model.model
inputs = modelobj.inputs + modelobj.targets + modelobj.sample_weights + [ K.learning_phase(),]
lossfunc = K.function(inputs, [modelobj.total_loss])
sampleweightstrn = np.ones(len(data.train.X))
sampleweightstst = np.ones(len(data.test.X))
noreglosstrn = lambda: lossfunc([data.train.X, data.train.Y, sampleweightstrn, 0])[0]
noreglosstst = lambda: lossfunc([data.test.X , data.test.Y , sampleweightstst, 0])[0]
if kdelayer is not None:
lv1 = K.get_value(kdelayer.logvar)
logs['kdeLV'] = lv1
print 'kdeLV=%.5f,' % lv1,
if noiselayer is not None:
lv2 = K.get_value(noiselayer.logvar)
logs['noiseLV'] = lv2
print 'noiseLV=%.5f' % lv2
if kdelayer is not None and noiselayer is not None:
if max_entropy_calc_N is None:
mitrn = data.train.X
mitst = data.test.X
else:
mitrn = randsample(data.train.X, max_entropy_calc_N)
mitst = randsample(data.test.X, max_entropy_calc_N)
mi_obj_trn = MIComputer(noiselayer.get_noise_input_func(mitrn), kdelayer=kdelayer, noiselayer=noiselayer)
mi_obj_tst = MIComputer(noiselayer.get_noise_input_func(mitst), kdelayer=kdelayer, noiselayer=noiselayer)
if True:
mivals_trn = map(lambda x: float(K.eval(x)), [mi_obj_trn.get_mi(), mi_obj_trn.get_h(), mi_obj_trn.get_hcond()]) # [data.train.X,]))
logs['mi_trn'] = mivals_trn[0]
mivals_tst = map(lambda x: float(K.eval(x)), [mi_obj_tst.get_mi(), mi_obj_tst.get_h(), mi_obj_tst.get_hcond()]) # [data.train.X,]))
logs['mi_tst'] = mivals_tst[0]
logs['kl_trn'] = noreglosstrn()
logs['kl_tst'] = noreglosstst()
print ', mitrn=%s, mitst=%s, kltrn=%.3f, kltst=%.3f' % (mivals_trn, mivals_tst, logs['kl_trn'], logs['kl_tst'])
else:
print
return logs
#logs['tstloss'] = self.totalloss([self.xX_test,0])
def extractFeatures(X, model):
"""
Extract the features of X using the activation layer of the model
Inputs:
- X: data sample to extract features for
- model: model to use to get the features
Returns: the np array of features (output from the last layer of the model)
"""
# https://keras.io/getting-started/faq/#how-can-i-visualize-the-output-of-an-intermediate-layer
# https://github.com/fchollet/keras/issues/1641
# extract layer
get_last_layer_output = K.function(
[model.layers[0].input, K.learning_phase()], [model.layers[-4].output])
layer_output = get_last_layer_output([X, 0])[0]
return layer_output
def plotCropImg(img,simg):
from keras.layers.core import K
model = Sequential()
#model.add(Lambda(lambda x: x, input_shape=(160,320,3)))
model.add(Lambda(lambda x: x / 127.5 - 1., input_shape=(160,320,3)))
model.add(Cropping2D(cropping=((50,25),(0,0))))
output = K.function([model.layers[0].input], [model.layers[1].output])
crop_img = output([img[None,...]])[0]
plt.figure()
plt.imshow(img, cmap='gray')
plt.savefig(simg+"_org.png")
plt.imshow(np.uint8(crop_img[0,...]), cmap='gray')
plt.savefig(simg+"_crop.png")
def extractLearnedFeatures(model, newData):
"""
Using the previously trained model, extract a learned set of features for the new
data from the second to last layer of the model.
Inputs:
- model: the trained model
- newData: the new data to extract features from
Returns:
- learnedFeats: features extracted from the model
"""
# https://keras.io/getting-started/faq/#how-can-i-visualize-the-output-of-an-intermediate-layer
# https://github.com/fchollet/keras/issues/1641
# extract layer
get_last_layer_output = K.function(
[model.layers[0].input, K.learning_phase()], [model.layers[-4].output])
learnedFeats = get_last_layer_output([newData, 0])[0]
return learnedFeats
def train(epoch_num=None, name=MODEL_NAME):
input_tensor = Input(name='the_input',
shape=(width, height, 3),
dtype='float32') #Input((width, height, 1))
x = input_tensor
for i in range(2):
# x = Conv2D(filters=2 ** (3+i), kernel_size=(3, 3), padding="same", activation='relu', kernel_initializer='he_normal')(x)
x = Conv2D(filters=16 * (i + 1),
kernel_size=(3, 3),
padding="same",
activation='relu',
kernel_initializer='he_normal')(x)
x = Conv2D(filters=16 * (i + 1),
kernel_size=(3, 3),
padding="same",
activation='relu',
kernel_initializer='he_normal')(x)
# x = Conv2D(filters=32, kernel_size=(3, 3), activation='relu')(x)
x = MaxPool2D(pool_size=(2, 2))(x)
conv_shape = x.get_shape()
# conv_to_rnn_dims = (width // (2 ** 3),
# (height // (2 ** 3)) * 32)
x = Reshape(target_shape=(int(conv_shape[1]),
int(conv_shape[2] * conv_shape[3])))(x)
x = Dense(dense_size, activation='relu')(x)
# (batch_size, 20, 8 )
gru_1 = GRU(rnn_size,
return_sequences=True,
kernel_initializer='he_normal',
name='gru1')(x)
gru_1b = GRU(rnn_size,
return_sequences=True,
go_backwards=True,
kernel_initializer='he_normal',
name='gru1_b')(x)
gru1_merged = Add()([gru_1, gru_1b]) #sum
gru_2 = GRU(rnn_size,
return_sequences=True,
kernel_initializer='he_normal',
name='gru2')(gru1_merged)
gru_2b = GRU(rnn_size,
return_sequences=True,
go_backwards=True,
kernel_initializer='he_normal',
name='gru2_b')(gru1_merged)
gru_2 = TimeDistributed(BatchNormalization())(gru_2)
gru_2b = TimeDistributed(BatchNormalization())(gru_2b)
x = Concatenate()([gru_2, gru_2b]) #concat
# x = Dropout(0.25)(x)
"""
最后结果是[batch_size, 最大时间序列, 分类总数+1位空白符+1位CTC校验位]，使用softmax函数，将所有结果的概率分布在（0，1）之间，激活用在每一帧时间序列上，求最大概率的分类，得出该帧的预测结果。
因此，此处dense层设置 分类总数的数量为结果，并采用softmax多分类激活函数
"""
x = Dense(n_class, kernel_initializer='he_normal', activation='softmax')(x)
# Model(inputs=input_tensor, outputs=x).summary()
# base_model = Model(inputs=input_tensor, outputs=x)
# 评估回调函数
evaluator_func = K.function([input_tensor, K.learning_phase()], [x])
# evaluator_func.
# base_model.summary()
evaluator = Evaluate(validation_func=evaluator_func,
val_seq=val_obj,
name="keras_cnn_gru_add_batch")
labels = Input(name='the_labels', shape=[n_len], dtype='float32')
input_length = Input(name='input_length', shape=[1], dtype='int64')
label_length = Input(name='label_length', shape=[1], dtype='int64')
loss_out = Lambda(ctc_lambda_func, output_shape=(1, ),
name='ctc')([x, labels, input_length, label_length])
model = Model(inputs=[input_tensor, labels, input_length, label_length],
outputs=[loss_out]) #.summary()
model.summary()
model.compile(loss={
'ctc': lambda y_true, y_pred: y_pred
},
optimizer='adadelta')
if epoch_num is not None:
weight_file = os.path.join(
OUTPUT_DIR, os.path.join(name, 'epoch_%02d.h5' % (epoch_num)))
model.load_weights(weight_file)
# print(base_model == model)
# model.fit_generator(train_gen.gen_batch(n_len, 200), steps_per_epoch=100, epochs=100, max_queue_size=1, workers=1, callbacks=[evaluator])
# model.fit_generator(image_gen.next_val(), steps_per_epoch=1, epochs=100, max_queue_size=1, workers=1, callbacks=[evaluator]) #单线程,易调试
model.fit_generator(image_gen,
steps_per_epoch=200,
epochs=100,
callbacks=[evaluator],
use_multiprocessing=True,
workers=2) #多线程
def get_activations(model, layer, X_batch):
get_activations = K.function([model.layers[0].input, K.learning_phase()], [model.layers[layer].output,])
activations = get_activations([X_batch,0])
return activations
model.add(Dense(512, input_shape=X_train.shape[1:]))
model.add(Activation('relu'))
model.add(Dropout(0.2))
model.add(Dense(512))
model.add(Activation('relu'))
model.add(Dropout(0.2))
model.add(Dense(10))
model.add(Activation('softmax'))
model.summary()
model.compile(loss='categorical_crossentropy',
optimizer=RMSprop(),
metrics=['accuracy'])
history = model.fit(X_train,
Y_train,
batch_size=batch_size,
nb_epoch=nb_epoch,
verbose=1,
validation_data=(X_test, Y_test))
score = model.evaluate(X_test, Y_test, verbose=0)
print('Test score:', score[0])
print('Test accuracy:', score[1])
# TESTING
get_last_layer_output = K.function(
[model.layers[0].input, K.learning_phase()], [model.layers[-2].output])
layer_output = get_last_layer_output([X_test, 0])[0]
output2 = get_last_layer_output([X_test, 0])
output=[loss_out])
# load weights into new model
print("Loading weights...")
model.load_weights(os.path.join(file_path_weigths, "weights275.h5")) #TODO
print("Loaded weights to model")
# the loss calc occurs elsewhere, so use a dummy lambda func for the loss
model.compile(optimizer=optimizer,
loss={
'ctc': lambda y_true, y_pred: y_pred
}) # , metrics=[self.tim_metric]
# Reporter captures output of softmax so we can decode the output during visualization
print("Init Reporter")
test_func = K.function([input_data], [y_pred])
# print("Loading model...")
# loaded_model = model_from_json(loaded_model_json)
# print("Loaded model from disk")
plot(model, to_file=os.path.join(file_path_model, 'model_eval.png'))
input_tuple = [[
('../media/nas/01_Datasets/IAM/words/a01/a01-011/a01-011-03-08.png')
]]
X = preprocessor.prep_run(input_tuple)
if K.image_dim_ordering() == 'th':
in1 = np.ones([1, 1, img_h, img_w])
else:
def evaluate_activations(model, X, layer):
from keras.layers.core import K
get_layer_output = K.function(
[model.layers[0].input, K.learning_phase()],
[model.layers[layer].output])
return get_layer_output([X, 0])[0]
# show example data
ex_image = images[random.randint(0, len(measurements)), :, :, :]
plt.figure()
plt.imshow(ex_image, cmap='gray')
plt.show()
from keras.models import Sequential
from keras.layers import Cropping2D
from keras.layers.core import Reshape
### crop test
model_crop = Sequential()
model_crop.add(
Cropping2D(cropping=((40, 25), (0, 0)), input_shape=(160, 320, 3)))
cropping_output = K.function([model_crop.layers[0].input],
[model_crop.layers[0].output])
cropped_image = cropping_output([ex_image[None, ...]])[0]
plt.imshow(cropped_image[0, ...] / 255, cmap='gray')
plt.show()
print(cropped_image.shape)
# plot data
plt.figure(figsize=(5, 2))
plt.plot(measurements)
plt.xlabel('frame')
plt.ylabel('steering angle (deg)')
plt.xlim([0, len(measurements)])
# plot histogram
plt.figure(figsize=(5, 3))
