def main():
"""Entry point function"""
adder = tf.Variable(10.)
life = tfh.Life(
tfh.NeuralNetwork(
name="Inner_NN",
layers=[
# Even though for ValidationLayer's name isn't required to not include space
# but it is good to stay consistance since other might require it
tfh.ValidationLayer(name="Input_Validation", shape=[None], dtype=tf.float32),
tfh.OpLayer(lambda input: input + adder, [adder]),
tfh.ValidationLayer(shape=[None], dtype=tf.float32)]),
optimizer=tf.train.AdamOptimizer(0.3))
input_x = np.array([
1, 2, 3, 4
], dtype=np.float32)
expect_y = np.array([
3, 4, 5, 6
], dtype=np.float32)
life.connect_neural_network(sample_input=input_x, sample_output=expect_y, will_train=True)
life.init_var()
hypo = life.feed(input_x)
print(hypo)
for counter in range(200):
cost = life.train(input_x, expect_y, process_list=[adder])
if counter%10 == 0:
print(cost)
def main():
"""Entry point function"""
sess = tf.Session()
x_input = tf.placeholder(tf.float32, shape=[None])
my_nn = tfh.NeuralNetwork(
name="Main_Layer",
layers=[
tfh.ValidationLayer(shape=[None], dtype=tf.float32),
tfh.ReshapeLayer(shape=[-1, 1]),
tfh.FeedForwardLayer(features_out=2, name="FFLayer"),
#features_out is the number of neuron in the layer
tfh.FeedForwardLayer(features_in=2, features_out=5, name="FFLayer"),
#features_in can also be set to verify
tfh.ValidationLayer(shape=[None, 5], dtype=tf.float32),
]
)
y_hypothesis = my_nn.connect(x_input)
sess.run(tf.initialize_all_variables())
batch = [1, 2, 3, 4]
print(sess.run(y_hypothesis, feed_dict={x_input: batch}))
def main():
"""Entry point function"""
life = tfh.Life(tfh.NeuralNetwork(
name="Inner_NN",
layers=[
tfh.ValidationLayer(name="Input_Validation",
shape=[None, 10, 10],
dtype=tf.float32),
tfh.ReshapeLayer(shape=[None, 10, 10, 1]),
tfh.ConvLayer(depth_out=2, kernel_width=2),
tfh.ConvLayer(depth_out=1, kernel_width=3, padding=False),
tfh.ValidationLayer(shape=[None, 8, 8, 1], dtype=tf.float32),
tfh.ReshapeLayer(shape=[None, 8, 8])
]),
optimizer=tf.train.AdamOptimizer(0.3))
input_x = np.random.randn(1, 10, 10).astype(np.float32)
expect_y = np.random.randn(1, 8, 8).astype(np.float32)
life.connect_neural_network(sample_input=input_x,
sample_output=expect_y,
will_train=True)
life.init_var()
hypo = life.feed(input_x)
print(hypo)
for counter in range(200):
cost = life.train(input_x, expect_y, process_list=[])
if counter % 10 == 0:
print(cost)
def main():
"""Entry point function"""
model_name = "model2-d"
life = tfh.Life(tfh.NeuralNetwork(layers=[
tfh.ValidationLayer(shape=[None, 96 + 8, 120 + 8, 3], dtype=tf.uint8),
tfh.OpLayer(tf.to_float),
tfh.OpLayer(lambda x: x / 255.),
tfh.ConvLayer(kernel_width=3, depth_out=30, depth_in=3, padding=False),
tfh.ConvLayer(kernel_width=3, depth_out=50, padding=False),
tfh.ConvLayer(kernel_width=3, depth_out=40, padding=False),
tfh.ConvLayer(kernel_width=3, depth_out=9, padding=False),
tfh.OpLayer(tf.sigmoid),
tfh.ValidationLayer(shape=[None, 96, 120, 9], dtype=tf.float32),
]),
cost_function=conv_cross_entropy,
optimizer=tf.train.AdamOptimizer(0.001))
data = Data.DataFeeder("data/",
dynamic_load=False,
filename="TrainCache",
data_padding=4,
label_height=96,
label_width=120)
batch = data.get_batch(60)
print(batch[0].dtype, batch[0].shape)
print(batch[1].dtype, batch[1].shape)
life.connect_neural_network(sample_input=batch[0],
sample_output=batch[1],
will_train=True)
# life.load_saved_model("auto-save/"+model_name+"-save-data3.0")
# life.load_saved_model(model_name+"-save-data")
life.init_var()
for counter in range(100):
batch = data.get_batch(60, shuffle=True)
result = life.train(input_layer_value=batch[0],
output_layer_value=batch[1])
if counter % 27 == 0:
print("{} iteration".format(counter), result)
data_in = batch[0]
feed_result = life.feed(input_layer_value=data_in)
hypo = feed_result[0]
# print("HYPO SHAPE", hypo.shape)
expect = batch[1][0]
cv2.imshow("image-in", data_in[0])
cv2.imshow("hypo", ObjClass.combine_label(hypo))
cv2.imshow("expect", ObjClass.combine_label(expect))
cv2.waitKey(20)
if counter % 100 == 0:
life.save_current_model("auto-save/" + model_name + "-save-data" +
str(counter / 100))
life.save_current_model(model_name + "-save-data")
def main():
"""Entry point function"""
model_name = "model4-a"
padding_size = 0
train_data_height = 512
train_data_width = 640
batch_size = 5
life = tfh.Life(
tfh.NeuralNetwork(
layers=[
tfh.ValidationLayer(shape=[None, train_data_height + padding_size*2, train_data_width + padding_size*2, 3], dtype=tf.uint8),
tfh.OpLayer(tf.to_float),
tfh.OpLayer(lambda x: x/255. - 0.5),
tfh.ConvLayer(kernel_width=3, depth_out=10, depth_in=3, padding=True),
tfh.ConvLayer(kernel_width=3, depth_out=10, padding=True),
tfh.ConvLayer(kernel_width=3, depth_out=20, padding=True),
tfh.ConvLayer(kernel_width=3, depth_out=10, padding=True),
tfh.ConvLayer(kernel_width=3, depth_out=2, padding=True),
tfh.ValidationLayer(shape=[None, train_data_height, train_data_width, 2], dtype=tf.float32),
# tfh.ReshapeLayer(shape=[None, 2]),
]
)
)
data = Data.DataFeeder("data/", dynamic_load=True,
filename=model_name+"-FeedCache", data_padding=0,
label_height=512, label_width=640)
batch = data.get_batch(5)
life.connect_neural_network(sample_input=batch[0], will_train=False)
# life.load_saved_model(model_name+"-save-data")
# life.load_saved_model("auto-save/"+model_name+"-save-data1.0")
life.load_saved_model("auto-save/model4-a-sava-data20.0")
# life.init_var()
print(batch[0].dtype, batch[0].shape)
print(batch[1].dtype, batch[1].shape)
##
# data.get_batch(500)
for _ in range(100):
batch = data.get_batch(5, shuffle=True)
data_in = batch[0]
feed_result = life.feed(input_layer_value=data_in)
hypo = feed_result[0]
print("HYPO SHAPE", hypo.shape)
for index, (img, expect_label) in enumerate(zip(batch[0], batch[1])):
cv2.imshow("image-in", img)
cv2.imshow("hypo", ObjClass.combine_label(feed_result[index]))
# print(feed_result[index])
cv2.imshow("expect", ObjClass.combine_label(expect_label))
cv2.waitKey(0)
def main():
"""Entry point function"""
adder = tf.Variable(10.)
adder2 = tf.Variable(7.)
nn1 = tfh.NeuralNetwork(
name="Inner_NN",
layers=[
tfh.ValidationLayer(name="Input_Validation", shape=[None], dtype=tf.float32),
tfh.OpLayer(lambda input: input + adder, [adder]),
tfh.OpLayer(tf.nn.relu),
tfh.ValidationLayer(shape=[None], dtype=tf.float32)])
nn2 = tfh.NeuralNetwork(
name="Inner2_NN",
layers=[
tfh.ValidationLayer(name="Input_Validation", shape=[None], dtype=tf.float32),
tfh.OpLayer(lambda input: input + adder2, [adder2]),
tfh.ValidationLayer(shape=[None], dtype=tf.float32)])
life = tfh.Life(
tfh.NeuralNetwork(
name="Main_NN",
layers=[nn1, nn2]),
optimizer=tf.train.AdamOptimizer(0.3))
input_x = np.array([
0, 1, 2, 3, 4
], dtype=np.float32)
expect_y = np.array([
2, 3, 4, 5, 6
], dtype=np.float32)
life.connect_neural_network(sample_input=input_x, sample_output=expect_y, will_train=True)
life.init_var()
# life.init_network([nn1, nn2])
# life.load_saved_model("test.ckpt", nn2)
hypo = life.feed(input_x)
print(hypo)
life.save_current_model("test.ckpt", nn1)
def main():
"""Entry point function"""
life = tfh.Life(
tfh.NeuralNetwork(
layers=[
tfh.ValidationLayer(shape=[None, 64 + 4, 80 + 4, 3], dtype=tf.uint8),
tfh.OpLayer(tf.to_float),
tfh.OpLayer(lambda x: x/255.),
tfh.ConvLayer(kernel_width=3, depth_out=30, depth_in=3, padding=False),
tfh.ConvLayer(kernel_width=3, depth_out=9, padding=False),
tfh.ValidationLayer(shape=[None, 64, 80, 9], dtype=tf.float32),
]
)
)
data = Data.DataFeeder("data/",
filename="TrainCache", data_padding=2,
label_height=64, label_width=80)
batch = data.get_batch(100)
print(batch[0].dtype, batch[0].shape)
print(batch[1].dtype, batch[1].shape)
life.connect_neural_network(sample_input=batch[0], sample_output=batch[1], will_train=True)
life.load_saved_model("model1")
# life.init_var()
for counter in range(500):
batch = data.get_batch(100)
result = life.train(input_layer_value=batch[0], output_layer_value=batch[1])
if counter%27 == 0:
print("{} iteration".format(counter), result)
data_in = batch[0]
feed_result = life.feed(input_layer_value=data_in)
hypo = feed_result[0]
# print("HYPO SHAPE", hypo.shape)
expect = batch[1][0]
cv2.imshow("image-in", data_in[0])
cv2.imshow("hypo", ObjClass.combine_label(hypo))
cv2.imshow("expect", ObjClass.combine_label(expect))
cv2.waitKey(20)
life.save_current_model("model1")
def main():
"""Entry point function"""
life = tfh.Life(
tfh.NeuralNetwork(
layers=[
tfh.ValidationLayer(shape=[None, 512, 640, 3], dtype=tf.uint8),
tfh.OpLayer(tf.to_float),
tfh.OpLayer(lambda x: x/255.),
tfh.ConvLayer(kernel_width=3, depth_out=30, depth_in=3),
tfh.ConvLayer(kernel_width=3, depth_out=9, ),
tfh.ValidationLayer(shape=[None, 512, 640, 9], dtype=tf.float32),
]
)
)
data = Data.DataFeeder("data/",
filename="FeedCache", data_padding=0,
label_height=512, label_width=640)
batch = data.get_batch(5)
life.connect_neural_network(sample_input=batch[0], will_train=False)
life.load_saved_model("model1")
# life.init_var()
print(batch[0].dtype, batch[0].shape)
print(batch[1].dtype, batch[1].shape)
##
data_in = batch[0]
feed_result = life.feed(input_layer_value=data_in)
hypo = feed_result[0]
print("HYPO SHAPE", hypo.shape)
for _ in range(2):
for index, (img, expect_label) in enumerate(zip(batch[0], batch[1])):
cv2.imshow("image-in", img)
cv2.imshow("hypo", ObjClass.combine_label(feed_result[index]))
cv2.imshow("expect", ObjClass.combine_label(expect_label))
cv2.waitKey(0)
def main():
"""Entry point function"""
model_name = "model6-a"
padding_size = 5
train_data_height = 160
train_data_width = 160
batch_size = 15
total_number_of_iteration = 3000
life = tfh.Life(
tfh.NeuralNetwork(layers=[
tfh.ValidationLayer(shape=[
None, train_data_height + padding_size * 2, train_data_width +
padding_size * 2, 3
],
dtype=tf.uint8),
tfh.OpLayer(tf.to_float),
tfh.OpLayer(lambda x: x / 255. - 0.5),
tfh.ConvLayer(
kernel_width=3, depth_out=10, depth_in=3, padding=False),
tfh.OpLayer(half_sigmoid),
tfh.ConvLayer(kernel_width=3, depth_out=10, padding=False),
tfh.OpLayer(half_sigmoid),
tfh.ConvLayer(kernel_width=3, depth_out=20, padding=False),
tfh.OpLayer(half_sigmoid),
tfh.ConvLayer(kernel_width=3, depth_out=10, padding=False),
tfh.OpLayer(half_sigmoid),
tfh.ConvLayer(kernel_width=3, depth_out=2, padding=False),
tfh.ValidationLayer(
shape=[None, train_data_height, train_data_width, 2],
dtype=tf.float32),
# tfh.ReshapeLayer(shape=[None, 2]),
]))
data = Data.DataFeeder("data/",
dynamic_load=True,
raw_preprocess=raw_preprocess,
cache_name=model_name + "-TrainCache",
data_padding=padding_size,
label_height=train_data_height,
label_width=train_data_width)
batch = data.get_batch(batch_size, shuffle=True, slient=False)
print(batch[0].dtype, batch[0].shape)
print(batch[1].dtype, batch[1].shape)
print(np.amax(batch[0], axis=(0, 1, 2)), np.amax(batch[1], axis=(0, 1, 2)))
life.connect_neural_network(sample_input=batch[0])
tf_flatten_label = tf.placeholder(dtype=tf.float32, shape=[None, 2])
tf_flatten_result = tf.reshape(life.tfvResult_pipe, shape=[-1, 2])
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(tf_flatten_result,
tf_flatten_label))
trainer = tf.train.AdamOptimizer(0.001).minimize(cross_entropy)
# life.load_saved_model("auto-save/"+model_name+"-save-data3.0")
life.session.run(tf.initialize_all_variables())
# life.load_saved_model(model_name+"-save-data")
# life.init_var()
for counter in range(total_number_of_iteration):
batch = data.get_batch(batch_size, shuffle=True)
preview = counter % 20 == 0
process_list = [cross_entropy, trainer, life.tfvResult_pipe
] if preview else [cross_entropy, trainer]
result = life.feed(
batch[0],
process_list=process_list,
feed_dict={tf_flatten_label: batch[1].reshape((-1, 2))})
print(counter, "loss", result[0])
if preview:
# print(result[2][0].sum())
# print(result[2][0])
# for index in range(batch_size):
for index in range(2):
cv2.imshow("image-in", batch[0][index])
cv2.imshow("hypo", ObjClass.combine_label(result[2][index]))
cv2.imshow("expecr", ObjClass.combine_label(batch[1][index]))
cv2.waitKey(30)
if counter % 100 == 0:
life.save_current_model("auto-save/" + model_name + "-save-data" +
str(counter / 100))
#for counter in range(100):
# batch = data.get_batch(60, shuffle=True)
# result = life.train(input_layer_value=batch[0], output_layer_value=batch[1])
# if counter%27 == 0:
# print("{} iteration".format(counter), result)
# data_in = batch[0]
# feed_result = life.feed(input_layer_value=data_in)
# hypo = feed_result[0]
# # print("HYPO SHAPE", hypo.shape)
# expect = batch[1][0]
# if counter%100 == 0:
# life.save_current_model("auto-save/"+model_name+"-save-data"+str(counter/100))
life.save_current_model(model_name + "-save-data")