def get_loss(self):
nl = tf.nn.leaky_relu
ac = tf.one_hot(self.ac, self.ac_space.n, axis=2)
sh = tf.shape(ac)
ac = flatten_two_dims(ac)
ac_four_dim = tf.expand_dims(tf.expand_dims(ac, 1), 1)
def add_ac(x):
if x.get_shape().ndims == 2:
return tf.concat([x, ac], axis=-1)
elif x.get_shape().ndims == 4:
sh = tf.shape(x)
return tf.concat([
x, ac_four_dim + tf.zeros([
sh[0], sh[1], sh[2],
ac_four_dim.get_shape()[3].value
], tf.float32)
],
axis=-1)
with tf.variable_scope(self.scope):
x = flatten_two_dims(self.features)
x = unet(x, nl=nl, feat_dim=self.feat_dim, cond=add_ac)
x = unflatten_first_dim(x, sh)
self.prediction_pixels = x * self.ob_std + self.ob_mean
return tf.reduce_mean((x - tf.stop_gradient(self.out_features))**2,
[2, 3, 4])
def predict_next(self, reuse):
nl = tf.nn.leaky_relu
if isinstance(self.ac_space, gym.spaces.Discrete):
ac = tf.one_hot(self.ac, get_action_n(self.ac_space), axis=2)
else:
ac = self.ac
sh = tf.shape(ac)
ac = flatten_two_dims(ac)
ac_four_dim = tf.expand_dims(tf.expand_dims(ac, 1), 1)
def add_ac(x):
if x.get_shape().ndims == 2:
return tf.concat([x, ac], axis=-1)
elif x.get_shape().ndims == 4:
sh = tf.shape(x)
return tf.concat([
x, ac_four_dim + tf.zeros([
sh[0], sh[1], sh[2],
ac_four_dim.get_shape()[3].value
], tf.float32)
],
axis=-1)
with tf.variable_scope(self.scope, reuse=reuse):
x = flatten_two_dims(self.features)
x = unet(x, nl=nl, feat_dim=self.feat_dim, cond=add_ac)
x = unflatten_first_dim(x, sh)
self.prediction_pixels = x * self.ob_std + self.ob_mean
# return tf.reduce_mean((x - tf.stop_gradient(self.out_features)) ** 2, [2, 3, 4])
return x
def get_loss(self):
nl = tf.nn.leaky_relu
ac = tf.one_hot(self.ac, self.ac_space.n, axis=2)
sh = tf.shape(ac)
ac = flatten_two_dims(ac)
ac_four_dim = tf.expand_dims(tf.expand_dims(ac, 1), 1)
def add_ac(x):
if x.get_shape().ndims == 2:
return tf.concat([x, ac], axis=-1)
elif x.get_shape().ndims == 4:
sh = tf.shape(x)
return tf.concat(
[
x,
ac_four_dim + tf.zeros(
[
sh[0], sh[1], sh[2],
ac_four_dim.get_shape()[3].value
],
tf.float32,
),
],
axis=-1,
)
with tf.variable_scope(self.scope):
x = flatten_two_dims(self.features)
mu, log_sigma_squared = unet(x,
nl=nl,
feat_dim=self.feat_dim,
cond=add_ac)
mu = unflatten_first_dim(mu, sh)
log_sigma_squared = unflatten_first_dim(log_sigma_squared, sh)
prediction_pixels = mu * self.ob_std + self.ob_mean
if self.ama == "true":
mse = tf.square(mu - 2 * tf.stop_gradient(self.out_features))
dynamics_reward = tf.reduce_mean((mse - tf.exp(log_sigma_squared)),
axis=[2, 3, 4])
if self.clip_ama == "true":
dynamics_reward = tf.clip_by_value(dynamics_reward, 0, 1e6)
loss = tf.reduce_mean(
(tf.exp(-log_sigma_squared) *
(mse) + self.uncertainty_penalty * log_sigma_squared),
axis=[2, 3, 4],
)
elif self.ama == "false":
mse = tf.square(mu - tf.stop_gradient(self.out_features))
dynamics_reward = tf.reduce_mean(mse, axis=[2, 3, 4])
loss = dynamics_reward
else:
raise ValueError("Please specify whether to use AMA or not")
return (
loss,
dynamics_reward,
prediction_pixels,
log_sigma_squared,
)
patches_val,
patches_val_ref)
patches_val_ref_aug_h = tf.keras.utils.to_categorical(patches_val_ref_aug,
number_class)
#%%
start_time = time.time()
exp = 1
rows = patch_size
cols = patch_size
adam = Adam(lr=0.0001, beta_1=0.9)
batch_size = 8
weights = [0.5, 0.5, 0]
loss = weighted_categorical_crossentropy(weights)
model = unet((rows, cols, channels))
model.compile(optimizer=adam, loss=loss, metrics=['accuracy'])
# print model information
model.summary()
filepath = 'models/'
# define early stopping callback
earlystop = EarlyStopping(monitor='val_loss',
min_delta=0.0001,
patience=10,
verbose=1,
mode='min')
checkpoint = ModelCheckpoint(filepath + 'unet_exp_' + str(exp) + '.h5',
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='min')
args.num_classes, args)
model = resuneta.model
model.summary()
model.compile(optimizer=optm,
loss=loss,
metrics=[
'accuracy',
tf.keras.metrics.TruePositives(),
tf.keras.metrics.FalsePositives(),
tf.keras.metrics.TrueNegatives(),
tf.keras.metrics.FalseNegatives()
])
print('ResUnet-a compiled!')
else:
model = unet((rows, cols, channels), args.num_classes)
model.summary()
model.compile(optimizer=optm,
loss=loss,
metrics=[
'accuracy',
tf.keras.metrics.TruePositives(),
tf.keras.metrics.FalsePositives(),
tf.keras.metrics.TrueNegatives(),
tf.keras.metrics.FalseNegatives()
])
else:
# load checkpoint compiled
print(f"[INFO] loading {args.checkpoint_path}...")
model = load_model(args.checkpoint_path)
model.summary()
