def __init__(self,
state_size=128,
correct=1,
gamma=1,
ewma=0.7,
input_size=(3, 70, 70),
teacher=None):
NN.Module.__init__(self)
self.correct = correct
self.critic = cuda(
build_mlp(input_size=state_size, layer_sizes=[state_size, 1]))
self.opt = T.optim.Adam(self.critic.parameters())
self.input_size = input_size
self.teacher = teacher
self.gamma = gamma
for p in teacher.parameters():
p.requires_grad = False
vf_args = (observation_space.shape[0] + 1, vf_hidden_dims, 1)
# Initialize the policy
if type(action_space) is Box:
policy_args = (observation_space.shape[0], policy_hidden_dims,
action_space.shape[0])
policy = build_diag_gauss_policy(*policy_args)
elif type(action_space) is Discrete:
policy_args = (observation_space.shape[0], policy_hidden_dims,
action_space.n)
policy = build_multinomial_policy(*policy_args)
else:
raise NotImplementedError
# Initalize the value function
value_fun = build_mlp(*vf_args)
policy.to(device)
value_fun.to(device)
# Initialize the state transformation
z_filter = ZFilter()
state_bound = Bound(-5, 5)
state_filter = Transform(state_bound, z_filter)
# Initialize the simulator
n_trajectories = config['n_trajectories']
max_timesteps = config['max_timesteps']
# try:
# env_args = config['env_args']
# except:
env_args = {}
width=args.width,
height=args.height,
max_=args.max,
verbose=True) # args.verbose)
if args.model == 'cnn':
label = '10cnn_%d_%d_%d' % (32, 128, args.epochs)
model = build_cnn(training_data,
width=args.width,
height=args.height,
verbose=True) # args.verbose)
train(model, label, training_data, epochs=args.epochs)
elif args.model == 'mlp':
label = '10mlp_%d_%d' % (32, args.epochs)
model = build_mlp(training_data,
width=args.width,
height=args.height,
verbose=True) # args.verbose)
train(model, label, training_data, epochs=args.epochs)
elif args.model == 'student':
model = build_mlp(training_data,
width=args.width,
height=args.height,
verbose=True) # args.verbose)
temp = 2.0
lamb = 0.5
label = '10student_mlp_%d_%d_lambda%s_temp%s' % (32, args.epochs,
str(lamb), str(temp))
train_student(model,
label,
training_data,
images = images / 255.0
# Split the data
(trainAttrX, testAttrX, trainImagesX,
testImagesX) = train_test_split(df, images, test_size=0.25, random_state=42)
# Normalization
max_price = trainAttrX["price"].max()
trainY = trainAttrX["price"] / max_price
testY = testAttrX["price"] / max_price
(trainAttrX, testAttrX) = datasets.process_attributes(df, trainAttrX,
testAttrX)
# Create network
mlp = models.build_mlp(trainAttrX.shape[1], False)
cnn = models.build_cnn(64, 64, 3, regress=False)
concatenated = concatenate([mlp.output, cnn.output])
x = Dense(4, activation="relu")(concatenated)
x = Dense(1, activation="linear")(x)
model = Model(inputs=[mlp.input, cnn.input], outputs=x)
optimizer = Adam(lr=1e-3, decay=1e-3 / 200)
model.compile(loss="mean_absolute_percentage_error", optimizer=optimizer)
model.fit([trainAttrX, trainImagesX],
trainY,
validation_data=([testAttrX, testImagesX], testY),
def exp( args ):
'''
alg -- elu, qelu, relu, qrelu
lrate -- learning rate
dropout -- whether to dropout
batch_size -- number of samples in a mini batch
epochs -- how many epochs to train
'''
counter = Counter()
callbacks = [ counter ]
if args.arch == 'mlp':
model = models.build_mlp( args, counter )
elif args.arch == 'cnn':
model = models.build_cnn( args, counter )
elif args.arch == 'resnet':
model = models.build_resnet( args, counter )
elif args.arch == 'siamese':
model = models.build_siamese( args, counter )
elif args.arch == 'ae':
model = models.build_ae( args, counter )
else:
raise RuntimeError( 'unknown architecture' )
if args.arch == 'resnet':
if args.alg[0] == 'n':
optimizer = add_gradient_noise( keras.optimizers.Adam )( lr=lr_schedule(0), noise_eta=args.init_v, counter=counter )
else:
optimizer = keras.optimizers.Adam( lr=lr_schedule(0) )
lr_scheduler = LearningRateScheduler( lr_schedule )
lr_reducer = ReduceLROnPlateau( factor=np.sqrt(0.1),
cooldown=0,
patience=5,
min_lr=0.5e-6 )
callbacks += [ lr_reducer, lr_scheduler ]
elif args.alg[0] == 'n':
optimizer = add_gradient_noise( keras.optimizers.SGD )( lr=args.lrate, decay=1e-6, momentum=0.5, nesterov=False, noise_eta=args.init_v, counter=counter )
else:
optimizer = keras.optimizers.SGD( lr=args.lrate, decay=1e-6, momentum=0.5, nesterov=False )
if args.arch == 'siamese':
model.compile( loss=contrastive_loss, optimizer=optimizer, metrics=[siamese_accuracy] )
tr_pairs, tr_y = data.train_generator
te_pairs, te_y = data.test_generator
history = model.fit( [tr_pairs[:, 0], tr_pairs[:, 1]], tr_y,
batch_size=args.batch_size,
epochs=args.epochs,
callbacks=callbacks,
validation_data=( [te_pairs[:, 0], te_pairs[:, 1]], te_y),
verbose=args.verbose )
return history.history['loss'], history.history['val_siamese_accuracy']
elif args.arch == 'ae':
model.compile( loss=keras.losses.mse,
optimizer=optimizer )
history = model.fit_generator(
data.train_generator,
steps_per_epoch=(data.x_train.shape[0]//args.batch_size+1),
epochs=args.epochs,
callbacks=callbacks,
validation_data=data.test_generator,
validation_steps=data.x_test.shape[0]//args.test_batch_size,
verbose=args.verbose )
return history.history['loss'], history.history['val_loss']
else:
model.compile( loss=keras.losses.categorical_crossentropy,
optimizer=optimizer,
metrics=['accuracy'] )
history = model.fit_generator(
data.train_generator,
steps_per_epoch=(data.x_train.shape[0]//args.batch_size+1),
epochs=args.epochs,
callbacks=callbacks,
validation_data=data.test_generator,
validation_steps=data.x_test.shape[0]//args.test_batch_size,
verbose=args.verbose )
return history.history['loss'], history.history['val_acc']
action_dim = config['action_dim']
n_episodes = config['n_episodes']
env_name = config['env_name']
n_episodes = config['n_episodes']
n_trajectories = config['n_trajectories']
trajectory_len = config['max_timesteps']
policy_dims = config['policy_hidden_dims']
vf_dims = config['vf_hidden_dims']
cf_dims = config['cf_hidden_dims']
max_constraint_val = config['max_constraint_val']
bias_red_cost = config['bias_red_cost']
device = get_device()
policy = build_diag_gauss_policy(state_dim, policy_dims, action_dim)
value_fun = build_mlp(state_dim + 1, vf_dims, 1)
cost_fun = build_mlp(state_dim + 1, cf_dims, 1)
policy.to(device)
value_fun.to(device)
cost_fun.to(device)
simulator = SinglePathSimulator(env_name, policy, n_trajectories,
trajectory_len)
cpo = CPO(policy,
value_fun,
cost_fun,
simulator,
model_name=model_name,
bias_red_cost=bias_red_cost,
max_constraint_val=max_constraint_val)
# Load the initialized multi-layer perceptron
mlp_weight_path = "./trained_models/approximate_plane_case10/cp.ckpt"
trainingSet_delgrs = pd.read_csv('./data/approximate_plane_set.csv')
inputsMLPTrain = trainingSet_delgrs[[
'Dkappa', 'dynamicLoads', 'bearingTemp'
]]
inputsMLPTrain_min = inputsMLPTrain.min(axis=0)
inputsMLPTrain_range = inputsMLPTrain.max(axis=0) - inputsMLPTrain_min
dLInputScaling = getScalingDenseLayer(inputsMLPTrain_min,
inputsMLPTrain_range)
lowBounds_delgrs = np.asarray([0.05 / 25920])
upBounds_delgrs = np.asarray([np.max(trainingSet_delgrs['delDkappa'])])
mlp_model = build_mlp(dLInputScaling)
mlp_model.load_weights(mlp_weight_path)
mlp_model.trainable = True
# Get the discrete ordinal classifier
cl_model = discrete_ordinal_classifier(
np.expand_dims(np.transpose(multipleInspections), -1).shape, myDtype)
cl_model.build(np.expand_dims(np.transpose(multipleInspections), -1).shape)
cl_layer = cl_model.layers[0]
cl_layer.trainable = True
# Build stacked RNN model
RNNmodel = create_stacked_rnn(cl_layer,
inspectionArray,
mlp_model,
d0RNN,