def __init__(self,model_type,model_size,learning_rate = 0.001):
self.model_type = model_type
self.constrain_op = None
self.x = tf.placeholder(dtype=tf.float32,shape=[None,None,6])
self.target_y = tf.placeholder(dtype=tf.float32,shape=[None,None,1])
self.model_size = model_size
head = self.x
if(model_type == "lstm"):
self.fused_cell = tf.nn.rnn_cell.LSTMCell(model_size)
head,_ = tf.nn.dynamic_rnn(self.fused_cell,head,dtype=tf.float32,time_major=True)
elif(model_type.startswith("ltc")):
learning_rate = 0.01 # LTC needs a higher learning rate
self.wm = ltc.LTCCell(model_size)
if(model_type.endswith("_rk")):
self.wm._solver = ltc.ODESolver.RungeKutta
elif(model_type.endswith("_ex")):
self.wm._solver = ltc.ODESolver.Explicit
else:
self.wm._solver = ltc.ODESolver.SemiImplicit
head,_ = tf.nn.dynamic_rnn(self.wm,head,dtype=tf.float32,time_major=True)
self.constrain_op = self.wm.get_param_constrain_op()
elif(model_type == "node"):
self.fused_cell = NODE(model_size,cell_clip=10)
head,_ = tf.nn.dynamic_rnn(self.fused_cell,head,dtype=tf.float32,time_major=True)
elif(model_type == "ctgru"):
self.fused_cell = CTGRU(model_size,cell_clip=-1)
head,_ = tf.nn.dynamic_rnn(self.fused_cell,head,dtype=tf.float32,time_major=True)
elif(model_type == "ctrnn"):
self.fused_cell = CTRNN(model_size,cell_clip=-1,global_feedback=True)
head,_ = tf.nn.dynamic_rnn(self.fused_cell,head,dtype=tf.float32,time_major=True)
else:
raise ValueError("Unknown model type '{}'".format(model_type))
# target_y = tf.expand_dims(self.target_y,axis=-1)
self.y = tf.layers.Dense(1,activation=None,kernel_initializer=tf.keras.initializers.TruncatedNormal())(head)
print("logit shape: ",str(self.y.shape))
self.loss = tf.reduce_mean(tf.square(self.target_y-self.y))
optimizer = tf.train.AdamOptimizer(learning_rate)
self.train_step = optimizer.minimize(self.loss)
self.accuracy = tf.reduce_mean(tf.abs(self.target_y-self.y))
self.sess = tf.InteractiveSession()
self.sess.run(tf.global_variables_initializer())
self.result_file = os.path.join("results","power","{}_{}.csv".format(model_type,model_size))
if(not os.path.exists("results/power")):
os.makedirs("results/power")
if(not os.path.isfile(self.result_file)):
with open(self.result_file,"w") as f:
f.write("best epoch, train loss, train mae, valid loss, valid mae, test loss, test mae\n")
self.checkpoint_path = os.path.join("tf_sessions","power","{}".format(model_type))
if(not os.path.exists("tf_sessions/power")):
os.makedirs("tf_sessions/power")
self.saver = tf.train.Saver()
def __init__(self,
model_type,
model_size,
sparsity_level=0.0,
learning_rate=0.001):
self.model_type = model_type
self.constrain_op = []
self.sparsity_level = sparsity_level
self.x = tf.placeholder(dtype=tf.float32, shape=[None, None, 72])
self.target_y = tf.placeholder(dtype=tf.int64, shape=[None, None])
self.model_size = model_size
head = self.x
if (model_type == "lstm"):
self.fused_cell = tf.nn.rnn_cell.LSTMCell(model_size)
head, _ = tf.nn.dynamic_rnn(self.fused_cell,
head,
dtype=tf.float32,
time_major=True)
elif (model_type.startswith("ltc")):
learning_rate = 0.01 # LTC needs a higher learning rate
self.wm = ltc.LTCCell(model_size)
if (model_type.endswith("_rk")):
self.wm._solver = ltc.ODESolver.RungeKutta
elif (model_type.endswith("_ex")):
self.wm._solver = ltc.ODESolver.Explicit
else:
self.wm._solver = ltc.ODESolver.SemiImplicit
head, _ = tf.nn.dynamic_rnn(self.wm,
head,
dtype=tf.float32,
time_major=True)
self.constrain_op.extend(self.wm.get_param_constrain_op())
elif (model_type == "node"):
self.fused_cell = NODE(model_size, cell_clip=-1)
head, _ = tf.nn.dynamic_rnn(self.fused_cell,
head,
dtype=tf.float32,
time_major=True)
elif (model_type == "ctgru"):
self.fused_cell = CTGRU(model_size, cell_clip=-1)
head, _ = tf.nn.dynamic_rnn(self.fused_cell,
head,
dtype=tf.float32,
time_major=True)
elif (model_type == "ctrnn"):
self.fused_cell = CTRNN(model_size,
cell_clip=-1,
global_feedback=True)
head, _ = tf.nn.dynamic_rnn(self.fused_cell,
head,
dtype=tf.float32,
time_major=True)
else:
raise ValueError("Unknown model type '{}'".format(model_type))
self._debug_list_sparse_vars = []
if (self.sparsity_level > 0):
self.constrain_op.extend(self.get_sparsity_ops())
self.y = tf.layers.Dense(2, activation=None)(head)
print("logit shape: ", str(self.y.shape))
weight = tf.cast(self.target_y, dtype=tf.float32) * 1.5 + 0.1
self.loss = tf.losses.sparse_softmax_cross_entropy(
labels=self.target_y, logits=self.y, weights=weight)
print("loss shape: ", str(self.loss.shape))
self.loss = tf.reduce_mean(self.loss)
optimizer = tf.train.AdamOptimizer(learning_rate)
self.train_step = optimizer.minimize(self.loss)
model_prediction = tf.argmax(input=self.y, axis=2)
lab = tf.cast(self.target_y, dtype=tf.float32)
pred = tf.cast(model_prediction, dtype=tf.float32)
# True/False positives/negatives
tp = tf.reduce_sum(lab * pred)
tn = tf.reduce_sum((1 - lab) * (1 - pred))
fp = tf.reduce_sum((1 - lab) * (pred))
fn = tf.reduce_sum((lab) * (1 - pred))
# don't divide by zero
# Precision and Recall
self.prec = tp / (tp + fp + 0.00001)
self.recall = tp / (tp + fn + 0.00001)
# F1-score (Geometric mean of precision and recall)
self.accuracy = 2 * (self.prec * self.recall) / (
self.prec + self.recall + 0.000001)
self.sess = tf.InteractiveSession()
self.sess.run(tf.global_variables_initializer())
self.result_file = os.path.join(
"results", "ozone",
"{}_{}_{:02d}.csv".format(model_type, model_size,
int(100 * self.sparsity_level)))
if (not os.path.exists("results/ozone")):
os.makedirs("results/ozone")
if (not os.path.isfile(self.result_file)):
with open(self.result_file, "w") as f:
f.write(
"best epoch, train loss, train acc, valid loss, valid acc, test loss, test acc\n"
)
self.checkpoint_path = os.path.join("tf_sessions", "ozone",
"{}".format(model_type))
if (not os.path.exists("tf_sessions/ozone")):
os.makedirs("tf_sessions/ozone")
self.saver = tf.train.Saver()
def __init__(self,model_type,model_size,sparsity_level=0.0,learning_rate = 0.001):
self.model_type = model_type
self.constrain_op = []
self.sparsity_level = sparsity_level
self.x = tf.placeholder(dtype=tf.float32,shape=[None,None,5])
self.target_y = tf.placeholder(dtype=tf.int32,shape=[None,None])
self.model_size = model_size
head = self.x
if(model_type == "lstm"):
self.fused_cell = tf.nn.rnn_cell.LSTMCell(model_size)
head,_ = tf.nn.dynamic_rnn(self.fused_cell,head,dtype=tf.float32,time_major=True)
elif(model_type.startswith("ltc")):
learning_rate = 0.005 # LTC needs a higher learning rate
self.wm = ltc.LTCCell(model_size)
if(model_type.endswith("_rk")):
self.wm._solver = ltc.ODESolver.RungeKutta
elif(model_type.endswith("_ex")):
self.wm._solver = ltc.ODESolver.Explicit
else:
self.wm._solver = ltc.ODESolver.SemiImplicit
head,_ = tf.nn.dynamic_rnn(self.wm,head,dtype=tf.float32,time_major=True)
self.constrain_op.extend(self.wm.get_param_constrain_op())
elif(model_type == "node"):
self.fused_cell = NODE(model_size,cell_clip=-1)
head,_ = tf.nn.dynamic_rnn(self.fused_cell,head,dtype=tf.float32,time_major=True)
elif(model_type == "ctgru"):
self.fused_cell = CTGRU(model_size,cell_clip=-1)
head,_ = tf.nn.dynamic_rnn(self.fused_cell,head,dtype=tf.float32,time_major=True)
elif(model_type == "ctrnn"):
self.fused_cell = CTRNN(model_size,cell_clip=-1,global_feedback=True)
head,_ = tf.nn.dynamic_rnn(self.fused_cell,head,dtype=tf.float32,time_major=True)
else:
raise ValueError("Unknown model type '{}'".format(model_type))
if(self.sparsity_level > 0):
self.constrain_op.extend(self.get_sparsity_ops())
self.y = tf.layers.Dense(2,activation=None)(head)
print("logit shape: ",str(self.y.shape))
self.loss = tf.reduce_mean(tf.losses.sparse_softmax_cross_entropy(
labels = self.target_y,
logits = self.y,
))
optimizer = tf.train.AdamOptimizer(learning_rate)
self.train_step = optimizer.minimize(self.loss)
model_prediction = tf.argmax(input=self.y, axis=2)
self.accuracy = tf.reduce_mean(tf.cast(tf.equal(model_prediction, tf.cast(self.target_y,tf.int64)), tf.float32))
self.sess = tf.InteractiveSession()
self.sess.run(tf.global_variables_initializer())
self.result_file = os.path.join("results","occupancy","{}_{}_{:02d}.csv".format(model_type,model_size,int(100*self.sparsity_level)))
if(not os.path.exists("results/occupancy")):
os.makedirs("results/occupancy")
if(not os.path.isfile(self.result_file)):
with open(self.result_file,"w") as f:
f.write("best epoch, train loss, train accuracy, valid loss, valid accuracy, test loss, test accuracy\n")
self.checkpoint_path = os.path.join("tf_sessions","occupancy","{}".format(model_type))
if(not os.path.exists("tf_sessions/occupancy")):
os.makedirs("tf_sessions/occupancy")
self.saver = tf.train.Saver()