def build_deepvo(inputs,
is_training,
seq_length,
height,
width,
scope='deepvo'):
# print (inputs)
# inputs_list = tf.unstack(inputs)
# for inputs in inputs_list:
# pred = build_flownet(inputs, is_training=is_training)
#
with tf.variable_scope('timedistruted_cnn', reuse=tf.AUTO_REUSE):
print(inputs)
inputs = tf.reshape(inputs, (1, seq_length, height, width, 2))
net = core.input_data(shape=(None, seq_length, height, width, 2),
placeholder=inputs)
net = core.time_distributed(net, build_flownet, [inputs])
# net = core.time_distributed(net, conv.conv_2d, [32, 7, 2, 'same', 'relu'])
# net = core.time_distributed(net, conv.conv_2d, [64, 5, 2, 'same', 'relu'])
# net = core.time_distributed(net, conv.conv_2d, [128, 3, 2, 'same', 'relu'])
# net = core.time_distributed(net, conv.conv_2d, [256, 3, 2, 'same', 'relu'])
# net = core.time_distributed(net, conv.conv_2d, [256, 3, 2, 'same', 'relu'])
# net = core.time_distributed(net, conv.conv_2d, [256, 3, 2, 'same', 'relu'])
net = core.time_distributed(net, conv.global_max_pool)
with tf.variable_scope('rnn'):
net = recurrent.lstm(net,
n_units=124,
activation=tf.nn.relu,
return_seq=True,
name='lstm1')
net = recurrent.lstm(net,
n_units=124,
activation=tf.nn.relu,
return_seq=True,
name='lstm2')
net = core.time_distributed(net, core.fully_connected, [128, 'relu'])
net = core.time_distributed(net, core.fully_connected, [12])
pose, uncertainty = tf.split(net, 2, axis=2)
pose = tf.cast(pose, tf.float64)
print("pose output shape")
print(pose)
#pose = core.fully_connected(net, activation=tf.nn.relu, n_units=128)
#pose = core.fully_connected(net, n_units=6)
#pose = tf.cast(pose, tf.float64)
#print (pose)
#print (tf.trainable_variables())
return pose, uncertainty
def cnnLSTM_model():
global model
filter_size_conv1 = 11
num_filters_conv1 = 5
filter_size_conv2 = 6
num_filters_conv2 = 10
filter_size_conv3 = 5
num_filters_conv3 = 5
filter_size_conv4 = 2
num_filters_conv4 = 2
lstm_units = 500
learning_rate = 1e-4
y_true = tf.placeholder(tf.float32,
shape=[None, num_classes],
name='y_true')
net = tflearn.input_data([None, 22, 64, 48, 1], name="input")
net = time_distributed(
net,
conv_2d,
args=[num_filters_conv1, filter_size_conv1, 1, 'same', 'tanh'])
net = time_distributed(net, max_pool_2d, args=[2])
net = time_distributed(
net,
conv_2d,
args=[num_filters_conv2, filter_size_conv2, 1, 'same', 'tanh'])
net = time_distributed(net, max_pool_2d, args=[2])
net = time_distributed(
net,
conv_2d,
args=[num_filters_conv3, filter_size_conv3, 1, 'same', 'tanh'])
net = time_distributed(net, max_pool_2d, args=[2])
net = time_distributed(net, flatten, args=['flat'])
net = lstm(net, lstm_units)
fc_layer = tflearn.fully_connected(net, num_classes, activation='softmax')
loss = tflearn.objectives.categorical_crossentropy(fc_layer, y_true)
network = regression(fc_layer,
optimizer='adam',
loss='categorical_crossentropy',
learning_rate=0.001)
# Training
model = tflearn.DNN(network,
tensorboard_verbose=0,
checkpoint_path='gestureCNNLSTM.tfl.ckpt')
cnnlstmmodel = conv_2d(cnnlstmmodel, 512, 2, activation='relu')
cnnlstmmodel = max_pool_2d(cnnlstmmodel, 2)
cnnlstmmodel = conv_2d(cnnlstmmodel, 256, 2, activation='relu')
cnnlstmmodel = max_pool_2d(cnnlstmmodel, 2)
cnnlstmmodel = conv_2d(cnnlstmmodel, 128, 2, activation='relu')
cnnlstmmodel = max_pool_2d(cnnlstmmodel, 2)
cnnlstmmodel = conv_2d(cnnlstmmodel, 64, 2, activation='relu')
cnnlstmmodel = max_pool_2d(cnnlstmmodel, 2)
cnnlstmmodel = conv_2d(cnnlstmmodel, 32, 2, activation='relu')
cnnlstmmodel = max_pool_2d(cnnlstmmodel, 2)
cnnlstmmodel = time_distributed(cnnlstmmodel, flatten, args=['flat'])
cnnlstmmodel = lstm(cnnlstmmodel, 1024)
cnnlstmmodel = fully_connected(cnnlstmmodel, 1000, activation='relu')
cnnlstmmodel = dropout(cnnlstmmodel, 0.7)
cnnlstmmodel = fully_connected(cnnlstmmodel, 10, activation='softmax')
cnnlstmmodel = regression(cnnlstmmodel,
optimizer='adam',
learning_rate=0.001,
loss='categorical_crossentropy',
name='regression')
model = tflearn.DNN(cnnlstmmodel, tensorboard_verbose=0)
# Load Saved Model
elif predictedClass == 9:
className = "Umur"
cv2.putText(textImage, "Pedicted Class : " + className, (30, 30),
cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 255, 255), 2)
cv2.putText(textImage, "Confidence : " + str(confidence * 100) + '%',
(30, 100), cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 255, 255), 2)
cv2.imshow("Statistics", textImage)
tf.reset_default_graph()
lstmmodel = input_data(shape=[None, 89, 100, 1], name='input')
lstmmodel = time_distributed(lstmmodel, dropout, args=[0.7])
lstmmodel = time_distributed(lstmmodel, flatten, args=['flat'])
lstmmodel = lstm(lstmmodel, 512)
lstmmodel = fully_connected(lstmmodel, 1000, activation='relu')
lstmmodel = dropout(lstmmodel, 0.7)
lstmmodel = fully_connected(lstmmodel, 10, activation='softmax')
lstmmodel = regression(lstmmodel,
optimizer='adam',
learning_rate=0.001,
loss='categorical_crossentropy',
name='regression')