def __call__(self, inp):
with tf.variable_scope("discriminator",
reuse=tf.AUTO_REUSE) as model_scope:
x = tfl.Flatten()(inp)
x = tfl.Dense(256)(x)
x = self.batchnorm()(x)
x = tf.nn.leaky_relu(x)
x = self.dropout(0.25)(x)
x = tfl.Dense(256)(x)
x = self.batchnorm()(x)
x = tf.nn.leaky_relu(x)
x = self.dropout(0.25)(x)
x = tfl.Dense(64)(x)
x = self.batchnorm()(x)
x = tf.nn.leaky_relu(x)
x = self.dropout(0.25)(x)
x = tfl.Dense(1, activation=None)(x)
self.variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope=model_scope.name)
self.post_call(model_scope)
return x
def __init__(self, num_filters, kernel_size):
# inputs
self.inputs = tf.placeholder(tf.float32, shape=[None, 28, 28, 1])
# first convolutional layer, default values: strides=1, use_bias=True
conv1 = layers.Conv2D(filters=num_filters,
kernel_size=kernel_size,
padding="same",
activation="relu")
pooling = layers.MaxPooling2D(pool_size=2, strides=2)
conv2 = layers.Conv2D(filters=num_filters,
kernel_size=kernel_size,
padding="same",
activation="relu")
# flatten layer before the dense layers
flatten = layers.Flatten()
# dense layers -> first one with relu?
linear1 = layers.Dense(units=128, activation="relu")
# second dense layer only computes logits
linear2 = layers.Dense(units=10, activation=None)
# define the graph
self.logits = conv1(self.inputs)
for layer in [pooling, conv2, pooling, flatten, linear1, linear2]:
self.logits = layer(self.logits)
self.out_soft = tf.nn.softmax(self.logits)
# intialize the variables
init = tf.global_variables_initializer()
self.sess = tf.Session()
self.sess.run(init)
def conv_net(inputs, conv1_kernels, kernel_size1, pool_size1, conv2_kernels,
kernel_size2, pool_size2):
# *** WARNING: HARDCODED SHAPES, CHANGE ACCORDINGLY ***
in_layer = tf.reshape(inputs, [-1, 28, 28, 1])
# 1st conv layer
conv1 = layers.conv2d(inputs=in_layer,
filters=conv1_kernels,
kernel_size=kernel_size1,
padding='same',
activation=tf.nn.relu)
# # 1st pooling layer
# pool1 = layers.max_pooling2d(inputs=conv1,
# pool_size=pool_size1,
# strides=2)
# 2nd conv layer
conv2 = layers.conv2d(inputs=conv1,
filters=conv2_kernels,
kernel_size=kernel_size2,
padding='same',
activation=tf.nn.relu)
# # 2nd pooling layer
# pool2 = layers.max_pooling2d(inputs=conv2,
# pool_size=pool_size2,
# strides=2)
fl = layers.Flatten()(conv2)
return fl
def __init__(self):
self.conv_1 = layers.Conv2D(filters=512,
kernel_size=[3, 3],
strides=[2, 2],
padding='same',
activation=nn.sigmoid,
name="gfn_conv_1")
self.conv_2 = layers.Conv2D(filters=512,
kernel_size=[3, 3],
strides=[1, 1],
padding='same',
activation=nn.sigmoid,
name="gfn_conv_2")
self.conv_3 = layers.Conv2D(filters=512,
kernel_size=[3, 3],
strides=[2, 2],
padding='same',
activation=nn.sigmoid,
name="gfn_conv_3")
self.conv_4 = layers.Conv2D(filters=512,
kernel_size=[3, 3],
strides=[1, 1],
padding='same',
activation=nn.sigmoid,
name="gfn_conv_4")
self.flatten = layers.Flatten(name="flatten")
self.dense_5 = layers.Dense(1024,
activation=nn.sigmoid,
name="gfn_dense_1")
self.dense_6 = layers.Dense(512,
activation=nn.sigmoid,
name="gfn_dense_1")
self.dense_7 = layers.Dense(256,
activation=nn.sigmoid,
name="gfn_desne_3")
return
def get_model(feature, label):
# Reshape the input vector into a 28x28 image
input_layer = tf.reshape(feature, [-1, 28, 28, 1])
conv1 = tf.layers.conv2d(inputs=input_layer,
filters=32,
kernel_size=[5, 5],
padding="same",
activation=tf.nn.relu,
name="conv1")
pool1 = tf.layers.max_pooling2d(inputs=conv1,
pool_size=[2, 2],
strides=2,
name="pool1")
conv2 = tf.layers.conv2d(inputs=pool1,
filters=64,
kernel_size=[5, 5],
padding="same",
activation=tf.nn.relu,
name="conv2")
pool2 = tf.layers.max_pooling2d(inputs=conv2,
pool_size=[2, 2],
strides=2,
name="pool2")
# Flatten
pool2_flat = layers.Flatten()(pool2)
# Add a Dense layer
dense = tf.layers.dense(inputs=pool2_flat,
units=1024,
activation=tf.nn.relu,
name="Dense1")
# Add a dropout layer
dropout = tf.layers.dropout(
inputs=dense,
rate=0.4, # Dropout rate
training=True,
name="dropout")
# Logits Layer
logits = tf.layers.dense(inputs=dropout, units=10, name="Logits_layer")
onehot_labels = tf.one_hot(tf.cast(label, tf.int32), 10, 1, 0)
loss = tf.losses.softmax_cross_entropy(onehot_labels=onehot_labels,
logits=logits)
correct_prediction = tf.equal(tf.cast(tf.argmax(logits, 1), tf.float32),
label)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
"""
Print out TensorBoard Summaries
"""
tf.summary.scalar("loss", loss)
tf.summary.histogram("loss", loss)
tf.summary.scalar("training_accuracy", accuracy)
tf.summary.image("images", input_layer, max_outputs=3)
# summary_op = tf.summary.merge_all()
return tf.argmax(logits, 1), loss, accuracy
X = []
y = []
for feature, label in training_dataset():
X.append(feature)
y.append(label)
X_train = np.array(X).reshape(-1, 150, 150, 1)
train_label = np.array(y).reshape(-1, 1)
label_encoded = OneHotEncoder().fit_transform(train_label)
y_train = label_encoded.A
X_train = X_train / 255.0
model = keras.Sequential([
layers.Conv2D(64, (3, 3),
activation=tf.nn.relu,
input_shape=X_train.shape[1:]),
layers.MaxPooling2D(pool_size=(2, 2), strides=(1, 1)),
layers.Conv2D(64, (3, 3), activation=tf.nn.relu),
layers.MaxPooling2D(pool_size=(2, 2), strides=(1, 1)),
layers.Flatten(),
layers.Dense(64, activation=tf.nn.relu),
layers.Dense(8, activation=tf.nn.softmax)
])
model.compile(loss='categorical_crossentropy',
optimizer=tf.keras.optimizers.Adam(),
metrics=['accuracy'])
model.fit(X_train, y_train, batch_size=4, validation_split=0.1, epoch=5)
def __init__(self,
name,
state_size,
action_size,
opt,
feature_layers=None,
critic_layers=None,
actor_layers=None):
self.name = name
self.state_size = state_size
self.action_size = action_size
self.optimizer = opt
self.feature_layers = [
layers.Conv2D(filters=16,
kernel_size=(8, 8),
strides=(4, 4),
activation=tf.nn.leaky_relu),
layers.Conv2D(filters=32,
kernel_size=(4, 4),
strides=(2, 2),
activation=tf.nn.leaky_relu),
layers.Flatten(),
layers.Dense(256, activation=tf.nn.leaky_relu, name="features"),
] if (feature_layers is None
or not isinstance(feature_layers, Iterable)) else feature_layers
critic_layers = [layers.Dense(1, name='value')] if (
critic_layers is None
or not isinstance(critic_layers, Iterable)) else critic_layers
actor_layers = [layers.Dense(action_size, name='logits')] if (
actor_layers is None
or not isinstance(actor_layers, Iterable)) else actor_layers
self.selected_action = tf.placeholder(tf.uint8, [None], name="labels")
self.actions_onehot = tf.one_hot(self.selected_action,
self.action_size,
dtype=tf.float32)
self.advantages = tf.placeholder(tf.float32, [None])
self.discounted_reward = tf.placeholder(tf.float32, [None])
self.state = tf.placeholder(tf.float32,
shape=[None, *state_size],
name="states")
with tf.variable_scope(self.name):
self.feature = self._layers_output(self.feature_layers, self.state)
self.value = self._layers_output(critic_layers, self.feature)
self.logits = self._layers_output(actor_layers, self.feature)
self.policy = tf.nn.softmax(self.logits, name='policy')
# self.value_loss, self.policy_loss, self.entropy, self.total_loss = self._compute_loss()
# self.target = tf.placeholder(tf.float32, [None])
responsible_outputs = tf.reduce_sum(
self.policy * self.actions_onehot, 1)
self.entropy = 0.005 * tf.reduce_sum(
-self.policy * tf.log(self.policy + 1e-7), 1)
self.policy_loss = -tf.reduce_mean(
(tf.log(responsible_outputs + 1e-7)) * self.advantages +
self.entropy)
self.value_loss = tf.losses.mean_squared_error(
self.advantages, tf.squeeze(self.value))
self.total_loss = 0.5 * self.value_loss + self.policy_loss # - entropy * 0.005
self.optimizer = tf.train.RMSPropOptimizer(learning_rate=0.01,
decay=.99)
# if name != 'global':
var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope=self.name)
self.gradients = self.optimizer.compute_gradients(
self.total_loss, var_list)
self.gradients_placeholders = []
for grad, var in self.gradients:
self.gradients_placeholders.append(
(tf.placeholder(var.dtype, shape=var.get_shape()), var))
self.apply_gradients = self.optimizer.apply_gradients(
self.gradients_placeholders)
# self.gradients = tf.gradients(self.total_loss,
# tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=self.name))
# self.grads, self.grad_norms = tf.clip_by_global_norm(self.gradients, 40.0)
# self.apply_grads = opt.apply_gradients(
# zip(self.grads, tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, 'global')))
# else:
self.reward_summary_ph = tf.placeholder(tf.float32,
name="reward_summary")
self.reward_summary = tf.summary.scalar(name='reward_summary',
tensor=self.reward_summary_ph)
self.merged_summary_op = tf.summary.merge_all()
self.writer = tf.summary.FileWriter('./graphs', tf.get_default_graph())
self.test = tf.get_default_graph().get_tensor_by_name(
os.path.split(self.value.name)[0] + '/kernel:0')
def __call__(self, inputs, castFromUint8=True):
# self.is_training
pr_shape = lambda var: print(var.shape)
# CNN
# with tf.name_scope('')
if self.data_format == 'NCHW':
inputs = tf.transpose(inputs, [0, 3, 1, 2])
# print(inputs.shape.dims)
if castFromUint8:
inputs = tf.cast(inputs, self.dtype)
with tf.variable_scope("init_conv"):
out = self.Conv2D(32, 3, strides=2)(inputs)
# pr_shape(out)
out = self.BN()(out, training=self.is_training)
out = tf.nn.relu6(out)
with tf.variable_scope("body"):
out = self._inverted_bottleneck(out, 6, 16, stride=1)
# pr_shape(out)
# out = tf.nn.relu6(out)
out = self._inverted_bottleneck(out, 6, 24, stride=2)
# pr_shape(out)
# out = tf.nn.relu6(out)
out = self._inverted_bottleneck(out, 6, 32, stride=2)
# out = tf.nn.relu6(out)
out = self._inverted_bottleneck(out, 6, 64, stride=2)
# out = tf.nn.relu6(out)
out = self.Pool2D(4, 3, 'max')(out)
pr_shape(out)
# residualParam = []
# param = {'filters': 32, 'kernel_sz': 5, 'strides': 2}
# residualParam.append(param)
# param = {'filters': 48, 'kernel_sz': 3, 'strides': 1}
# residualParam.append(param)
# with tf.variable_scope("res1"):
# out = self.residual(inputs, residualParam)
# out = tf.nn.relu6(out)
# with tf.variable_scope("conv1_relu"):
# out0 = self.Conv2D(48, 5)(inputs)
# out0 = self.BN()(out0, training=self.is_training)
# out0 = tf.nn.relu6(out0)
# with tf.variable_scope("conv1_relu"):
# with tf.variable_scope("pool1"):
# out = self.Pool2D(3, 2, 'max')(out)
# with tf.variable_scope("conv2_relu"):
# out = self.Conv2D(48, 5)(out)
# out = self.BN()(out, training=self.is_training)
# out = tf.nn.relu6(out)
# with tf.variable_scope("pool2"):
# out = self.Pool2D(3, 3, 'max')(out)
# with tf.variable_scope("conv3_relu"):
# out = self.Conv2D(48, 5)(out)
# out = self.BN()(out, training=self.is_training)
# out = tf.nn.relu6(out)
# with tf.variable_scope("pool3"):
# out = self.Pool2D(3, 3, 'max')(out)
with tf.variable_scope('fc1'):
out = tl.Flatten()(out)
out = self.FC(128)(out)
self.feature2rnn = out
with tf.variable_scope("dropout"):
out = tl.Dropout(0.5)(out, training=self.is_training)
with tf.variable_scope('fc2'):
out = self.FC(2)(out)
# pr_shape(out)
return out
def __init__(self,
name,
state_size,
action_size,
opt,
feature_layers=None,
critic_layers=None,
actor_layers=None):
self.name = name
self.state_size = state_size
self.action_size = action_size
self.optimizer = opt
self.feature_layers = [
# layers.Dense(100, activation='relu', name="features"),
layers.Conv2D(filters=16,
kernel_size=(8, 8),
strides=(4, 4),
activation=tf.nn.leaky_relu),
layers.Conv2D(filters=32,
kernel_size=(4, 4),
strides=(2, 2),
activation=tf.nn.leaky_relu),
layers.Flatten(),
layers.Dense(256, activation=tf.nn.leaky_relu, name="features"),
] if (feature_layers is None
or not isinstance(feature_layers, Iterable)) else feature_layers
critic_layers = [layers.Dense(1, name='value')] if (
critic_layers is None
or not isinstance(critic_layers, Iterable)) else critic_layers
# actor_layers = [
# layers.Dense(action_size, activation='sigmoid', name='logits')
# ] if (actor_layers is None or not isinstance(actor_layers, Iterable)) else actor_layers
self.state = tf.placeholder(tf.float32,
shape=[None, *state_size],
name="states")
with tf.variable_scope(self.name):
self.feature = self._layers_output(self.feature_layers, self.state)
self.value = self._layers_output(critic_layers, self.feature)
# self.logits = self._layers_output(actor_layers, self.feature)
# self.policy = tf.nn.softmax(self.logits, name='policy')
# self.selected_action = tf.placeholder(tf.float32, [None], name="labels")
# self.actions_onehot = tf.one_hot(tf.cast(self.selected_action, tf.int32), self.action_size, dtype=tf.float32)
self.advantages = tf.placeholder(tf.float32, [None])
if name != 'global':
# self.value_loss, self.policy_loss, self.entropy_loss, self.total_loss = self._compute_loss()
# self.responsible_outputs = tf.reduce_sum(self.policy * self.actions_onehot, 1)
# responsible_outputs = tf.reduce_sum(self.logits * actions_onehot, 1)
self.value_loss = (self.advantages - self.value)**2
# self.entropy = -tf.reduce_sum(self.policy * tf.log(self.policy), 1)
# self.policy_loss = -tf.reduce_sum(tf.log(self.responsible_outputs)*self.advantages)
self.total_loss = tf.reduce_mean(
0.5 * self.value_loss) # + self.policy_loss)
self.gradients = tf.gradients(
self.total_loss,
tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope=self.name))
def flatten(x):
return ly.Flatten()(x)
