def main():
print("reading data...")
data = ngsim_manipulation.Data()
x = tf.placeholder(tf.float32, [None, rnn_rbm.n_visible],
name="x") #The placeholder variable that holds our data
W = tf.Variable(tf.random_normal([rnn_rbm.n_visible, rnn_rbm.n_hidden],
0.01),
name="W") #The weight matrix of the RBM
Wuh = tf.Variable(tf.random_normal(
[rnn_rbm.n_hidden_recurrent, rnn_rbm.n_hidden], 0.0001),
name="Wuh") #The RNN -> RBM hidden weight matrix
bh = tf.Variable(tf.zeros([1, rnn_rbm.n_hidden], tf.float32),
name="bh") #The RNN -> RBM hidden bias vector
Wuv = tf.Variable(tf.random_normal(
[rnn_rbm.n_hidden_recurrent, rnn_rbm.n_visible], 0.0001),
name="Wuv") #The RNN -> RBM visible weight matrix
bv = tf.Variable(tf.zeros([1, rnn_rbm.n_visible], tf.float32),
name="bv") #The RNN -> RBM visible bias vector
Wvu = tf.Variable(tf.random_normal(
[rnn_rbm.n_visible, rnn_rbm.n_hidden_recurrent], 0.0001),
name="Wvu") #The data -> RNN weight matrix
Wuu = tf.Variable(tf.random_normal(
[rnn_rbm.n_hidden_recurrent, rnn_rbm.n_hidden_recurrent], 0.0001),
name="Wuu") #The RNN hidden unit weight matrix
bu = tf.Variable(tf.zeros([1, rnn_rbm.n_hidden_recurrent], tf.float32),
name="bu") #The RNN hidden unit bias vector
u0 = tf.Variable(tf.zeros([1, rnn_rbm.n_hidden_recurrent], tf.float32),
name="u0") #The initial state of the RNN
#The RBM bias vectors. These matrices will get populated during rnn-rbm training and generation
BH_t = tf.Variable(tf.ones([1, rnn_rbm.n_hidden], tf.float32), name="BH_t")
BV_t = tf.Variable(tf.ones([1, rnn_rbm.n_visible], tf.float32),
name="BV_t")
#Build the RBM optimization
saver = tf.train.Saver()
#Note that we initialize the RNN->RBM bias vectors with the bias vectors of the trained RBM. These vectors will form the templates for the bv_t and
#bh_t of each RBM that we create when we run the RNN-RBM
updt = RBM.get_cd_update(x, W, bv, bh, k=1, lr=lr) # CD-k
print("training RBM for weight initialization...")
#Run the session
with tf.Session() as sess:
#Initialize the variables of the model
init = tf.global_variables_initializer()
sess.run(init)
#Run over each trajectory num_epoch times
for epoch in tqdm(range(num_epochs)):
#for idx in range(100):
for idx in range(data.num_trajectories):
sess.run(
updt, feed_dict={x: data.next_traj()}
) # even if traj is comprised of data in different time, we treat it like a batch of data in the same time, because we can only initialize one RBM.
#Save the initialized model here
save_path = saver.save(sess, "parameter_checkpoints/initialized.ckpt")
def main(num_epochs):
print("reading data...")
data = ngsim_manipulation.Data()
#First, we build the model and get pointers to the model parameters
x, cost, reconstruction, W, Wuh, Wuv, Wvu, Wuu, bh, bv, bu, lr, u0 = rnn_rbm.rnnrbm(
)
#The trainable parameters, as well as the initial state of the RNN
params = [W, Wuh, Wuv, Wvu, Wuu, bh, bv, bu, u0]
opt_func = tf.train.AdamOptimizer(learning_rate=lr)
grad_and_params = opt_func.compute_gradients(cost, params)
grad_and_params = [(tf.clip_by_value(grad, -10., 10.), var)
for grad, var in grad_and_params]
updt = opt_func.apply_gradients(grad_and_params)
#The learning rate of the optimizer is a parameter that we set on a schedule during training
#opt_func = tf.train.GradientDescentOptimizer(learning_rate=lr)
#grad_and_params = opt_func.compute_gradients(cost, params)
#grad_and_params = [(tf.clip_by_value(grad, -10., 10.), var) for grad, var in grad_and_params] #We use gradient clipping to prevent gradients from blowing up during training
#updt = opt_func.apply_gradients(grad_and_params)
saver = tf.train.Saver(
params, max_to_keep=1
) #We use this saver object to restore the weights of the model and save the weights every few epochs
with tf.Session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
saver.restore(
sess, saved_initial_weights_path
) #Here we load the initial weights of the model that we created with weight_initializations.py
##这边，可以考虑batch，也可以一次就只处理一个traj 随机梯度下降
print("training in progress...")
for epoch in range(num_epochs):
costs = []
start = time.time()
#for j in tqdm(range(100)):
for j in tqdm(range(data.num_trajectories)):
_, C = sess.run([updt, cost],
feed_dict={
x: data.next_traj(),
lr: learningRate
})
costs.append(C)
#Print the progress at epoch
print("epoch: {} cost: {} time: {}".format(epoch, np.mean(costs),
time.time() - start))
print("\n")
#Here we save the weights of the model every few epochs
saver.save(sess,
"parameter_checkpoints/epoch_{}.ckpt".format(epoch))
def draw_all_trajectories():
if not os.path.isdir("./picture_folder"):
os.makedirs("./picture_folder")
data = ngsim_manipulation.Data()
trajectories = data.dataset
plt.figure(figsize=(30, 15))
for traj in trajectories:
traj = data.decontraction(traj)
plt.plot(traj[:, 0], traj[:, 1])
plt.xlabel("x")
plt.ylabel("y")
plt.title("All trajectories")
plt.legend()
plt.savefig("./picture_folder/All_trajectories")
mng = plt.get_current_fig_manager()
mng.window.showMaximized()
plt.show()
def main(saved_weights_path):
print("reading data...")
data = ngsim_manipulation.Data()
x, cost, reconstruction, W, Wuh, Wuv, Wvu, Wuu, bh, bv, bu, lr, u0 = rnn_rbm.rnnrbm(
) # First we build and get the parameters of the network
params = [W, Wuh, Wuv, Wvu, Wuu, bh, bv, bu, u0]
saver = tf.train.Saver(
params) #We use this saver object to restore the weights of the model
#get many trajectories
#trajectories_primer = data.get_trajectories(start=0, end=5)
idx = [
0, 3, 4
] # three representative traj: get_traj(0) get_traj(3) get_traj(4)
trajectories_primer = data.get_trajs(idx)
trajectories_primer = data.add_noise_gaussian(trajectories_primer)
#trajectories_primer = data.add_noise_zero(idx)
#get one traj
#idx=0
#trajectories_primer = [data.get_traj(idx)]
#trajectories_primer = data.add_noise_gaussian(trajectories_primer)
#trajectories_primer = data.add_noise_zero([idx])
#xytraj=data.xyset[1]
#xytraj = xytraj*[data.max[0],1]
#x_sample=RBM.gibbs_sample(x, W, bv, bh, 1)
reconstructed_trajectories = []
decontr_trajectories_primer = []
print("reconstruction...")
with tf.Session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
saver.restore(
sess, saved_weights_path) #load the saved weights of the network
'''
# without time dependence, reconstruction based on a single RBM(only one RBM for all time steps)
# actually, this performs not too badly, even just one RBM for different time(because we choose rbm_steps=100m it include time dependence already)
for j in tqdm( range(len(trajectories_primer)) ):
decontract_traj = data.decontraction( sess.run(x_sample, feed_dict={x: trajectories_primer[j]}) )
reconstructed_trajectories.append( decontract_traj )
decontr_trajectories_primer.append( data.decontraction(trajectories_primer[j]) )
'''
# reconstruction based on RNN-RBM, time dependance. Expectation RBM
for j in tqdm(range(len(trajectories_primer))):
for i in range(num_sample):
decontract_traj = data.decontraction(
sess.run(reconstruction(),
feed_dict={x: trajectories_primer[j]}))
reconstructed_trajectories.append(
decontract_traj
) #Prime the network with primer and reconstruct this trajectory
decontr_trajectories_primer.append(
data.decontraction(trajectories_primer[j]))
trajectories = decontr_trajectories_primer + reconstructed_trajectories #+ [xytraj]
draw.draw_trajectories(trajectories,
num_trajs=len(trajectories_primer))
'''