def get_CNN(shape=[100,176,256, 256],optimizer='adam', learning_rate=0.001, loss='categorical_crossentropy'):
if platform == 'win32':
root = 'C:/Users/douce/Desktop/MIT Fall 2018/6.869 Machine Vision/Final Project/'
else:
root = '/home/ubuntu'
tf.reset_default_graph()
net = tflearn.input_data(shape)
net = tflearn.conv_3d(net, 16, 5, strides=2, activation='leaky_relu', padding='VALID', weights_init='xavier',
regularizer='L2', weight_decay=0.01)
net = tflearn.max_pool_3d(net, kernel_size=3, strides=2, padding='VALID')
net = tflearn.conv_3d(net, 32, 3, strides=2, padding='VALID', weights_init='xavier', regularizer='L2',
weight_decay=0.01)
net = tflearn.normalization.batch_normalization(net)
net = tflearn.activations.leaky_relu(net)
net = tflearn.max_pool_3d(net, kernel_size=2, strides=2, padding='VALID')
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, 1024, weights_init='xavier', regularizer='L2')
net = tflearn.normalization.batch_normalization(net, gamma=1.1, beta=0.1)
net = tflearn.activations.leaky_relu(net)
net = tflearn.dropout(net, 0.6)
net = tflearn.fully_connected(net, 512, weights_init='xavier', regularizer='L2')
net = tflearn.normalization.batch_normalization(net, gamma=1.2, beta=0.2)
net = tflearn.activations.leaky_relu(net)
net = tflearn.dropout(net, 0.7)
net = tflearn.fully_connected(net, 128, weights_init='xavier', regularizer='L2')
net = tflearn.normalization.batch_normalization(net, gamma=1.4, beta=0.4)
net = tflearn.activations.leaky_relu(net)
net = tflearn.dropout(net, 0.7)
net = tflearn.fully_connected(net, 3, weights_init='xavier', regularizer='L2')
net = tflearn.normalization.batch_normalization(net, gamma=1.3, beta=0.3)
net = tflearn.activations.softmax(net)
net = tflearn.regression(net, optimizer='adam', learning_rate=0.001, loss='categorical_crossentropy')
model = tflearn.DNN(net, checkpoint_path=os.path.join(root, 'model.tfl.ckpt'), max_checkpoints=3)
def main():
'''
Establishes architecture for CNN-fully-connected-RNN neural net.
'''
# Create the input layer: None: batch size, 120: frames, 80: height, 80: width, 3: RGB
network = tflearn.input_data(shape=[None, 120, 80, 80, 3], name='input')
# Convolutional network
network = tflearn.conv_3d(
network, 32, (3, 3, 3), activation='relu'
) # 32 conv layers of 3x3x3 (3x3x3 convolves for each 3 frames, 3px height, and 3px width)
network = tflearn.max_pool_3d(
network, (1, 2, 2),
strides=(1, 2, 2)) # Pools results of the conv_3d layer
network = tflearn.conv_3d(network, 64, (3, 3, 3),
activation='relu') # 64 layers of 3x3x3
network = tflearn.max_pool_3d(network, (1, 2, 2), strides=(1, 2, 2))
network = tflearn.conv_3d(network, 128, (3, 3, 3),
activation='relu') # 128 layers of 3x3x3
network = tflearn.conv_3d(network, 128, (3, 3, 3),
activation='relu') # another one?
network = tflearn.max_pool_3d(network, (1, 2, 2), strides=(1, 2, 2))
network = tflearn.conv_3d(network, 256, (2, 2, 2),
activation='relu') # 256 layers of 2x2x2
network = tflearn.conv_3d(network, 256, (2, 2, 2), activation='relu')
network = tflearn.max_pool_3d(network, (1, 2, 2), strides=(1, 2, 2))
network = tflearn.conv_2d(network,
64,
4,
activation='relu',
regularizer="L2") # 64 layers of 4x4
network = tflearn.max_pool_2d(network, 2) # and then max pool
# Normalize activations of the previous layer at each batch.
network = tflearn.local_response_normalization(network)
# And now the fully-connected neural net (128 & 256 neurons + dropout)
network = tflearn.fully_connected(network, 128, activation='tanh')
network = tflearn.dropout(network, 0.8)
network = tflearn.fully_connected(network, 256, activation='tanh')
network = tflearn.dropout(network, 0.8)
network = tflearn.reshape(network, [-1, 1, 256]) # Why 256?
# LSTM layers
network = tflearn.lstm(network, 128, return_seq=True) # LSTM of 128 units
network = tflearn.lstm(network, 128)
network = tflearn.fully_connected(
network, 4, activation='softmax') # Just four neurons... okay?
network = tflearn.regression(network,
optimizer='adam',
loss='categorical_crossentropy',
name='target')
model = tflearn.DNN(network, tensorboard_verbose=0)
X, Y = getXY()
model.fit(X,
Y,
n_epoch=1,
validation_set=0.1,
show_metric=True,
snapshot_step=100)
def inference(images, seismic, dropout_keep):
with tf.variable_scope("inference"):
# Image Model
image_feats = image_features(images)
# Seismic Model
seismic_feats = seismic_features(seismic)
# Fusion
with tf.variable_scope("Fusion"):
i_shape = image_feats.get_shape()
s_shape = seismic_feats.get_shape()
output_shape = [
FLAGS.batch_size,
int(i_shape[1]),
int(i_shape[2]),
int(s_shape[1]),
int(i_shape[3] + s_shape[3])
] # Bx7x7x39x3072
# ipdb.set_trace()
i_arranged = tf.expand_dims(image_feats, 3)
F1 = tf.tile(i_arranged, [1, 1, 1, output_shape[3], 1])
s_arranged = tf.transpose(tf.expand_dims(
seismic_feats, 1), [0, 1, 3, 2, 4]) # has dims 256,1,1,19,20
F2 = tf.tile(s_arranged,
[1, output_shape[1], output_shape[2], 1, 1])
# F1_norm = slim.batch_norm
# bil_feats = tf.batch_matmul(F1, F2, name="Tensor_Product")
# ipdb.set_trace()
spatio_temporal_feats = tf.concat([F1, F2], 4)
tf.add_to_collection(
'feats',
[seismic_feats, image_feats, F2, F1, spatio_temporal_feats])
# ipdb.set_trace()
print('asf')
c3_1 = tflearn.conv_3d(spatio_temporal_feats,
128,
3,
activation='relu',
weights_init='normal')
c3_p = tflearn.max_pool_3d(c3_1,
kernel_size=[1, 2, 2, 5, 1],
strides=[1, 2, 2, 5, 1])
tf.add_to_collection('WtoRegu', c3_1.W)
# ipdb.set_trace()
# c3_2 = tflearn.conv_3d(c3_p, 64, 3, activation='relu', weights_init='normal')
# c3_p2 = tflearn.max_pool_3d(c3_2, kernel_size=[1, 2, 2, 5, 1], strides=[1, 2, 2, 5, 1])
# ipdb.set_trace()
final_feats = tf.reshape(c3_p, [FLAGS.batch_size, -1])
# Classifier
logits = classification(final_feats, dropout_keep)
return logits
datanp=[] #images
truenp=[] #labels
for file in os.listdir(data_dir):
data=np.load(os.path.join(data_dir,file))
datanp.append((data[0][0]))
truenp.append(data[0][1])
sh=datanp.shape
tf.reset_default_graph()
net = tflearn.input_data(shape=[None, sh[1], sh[2], sh[3], sh[4]])
net = tflearn.conv_3d(net, 16,5,strides=2,activation='leaky_relu', padding='VALID',weights_init='xavier',regularizer='L2',weight_decay=0.01)
net = tflearn.max_pool_3d(net, kernel_size = 3, strides=2, padding='VALID')
net = tflearn.conv_3d(net, 32,3,strides=2, padding='VALID',weights_init='xavier',regularizer='L2',weight_decay=0.01)
net = tflearn.normalization.batch_normalization(net)
net = tflearn.activations.leaky_relu (net)
net = tflearn.max_pool_3d(net, kernel_size = 2, strides=2, padding='VALID')
net = tflearn.dropout(net,0.5)
net = tflearn.fully_connected(net, 1024,weights_init='xavier',regularizer='L2')
net = tflearn.normalization.batch_normalization(net,gamma=1.1,beta=0.1)
net = tflearn.activations.leaky_relu (net)
net = tflearn.dropout(net,0.6)
net = tflearn.fully_connected(net, 512,weights_init='xavier',regularizer='L2')
net = tflearn.normalization.batch_normalization(net,gamma=1.2,beta=0.2)
net = tflearn.activations.leaky_relu (net)
net = tflearn.dropout(net,0.7)
net = tflearn.fully_connected(net, 128,weights_init='xavier',regularizer='L2')
net = tflearn.normalization.batch_normalization(net,gamma=1.4,beta=0.4)
test = 'test'
train = 'training'
def getXY():
X, Y = image_preloader(train,
image_shape=(80, 80),
mode='folder',
categorical_labels='True',
normalize=True)
return X, Y
network = tflearn.input_data(shape=[None, 120, 80, 80, 3], name='input')
network = tflearn.conv_3d(network, 32, (3, 3, 3), activation='relu')
network = tflearn.max_pool_3d(network, (1, 2, 2), strides=(1, 2, 2))
network = tflearn.conv_3d(network, 64, (3, 3, 3), activation='relu')
network = tflearn.max_pool_3d(network, (1, 2, 2), strides=(1, 2, 2))
network = tflearn.conv_3d(network, 128, (3, 3, 3), activation='relu')
network = tflearn.conv_3d(network, 128, (3, 3, 3), activation='relu')
network = tflearn.max_pool_3d(network, (1, 2, 2), strides=(1, 2, 2))
network = tflearn.conv_3d(network, 256, (2, 2, 2), activation='relu')
network = tflearn.conv_3d(network, 256, (2, 2, 2), activation='relu')
network = tflearn.max_pool_3d(network, (1, 2, 2), strides=(1, 2, 2))
network = tflearn.conv_2d(network, 64, 4, activation='relu', regularizer="L2")
network = tflearn.max_pool_2d(network, 2)
network = tflearn.local_response_normalization(network)
network = fully_connected(network, 128, activation='tanh')
network = dropout(network, 0.8)
network = fully_connected(network, 256, activation='tanh')
