def show_recons_samples(sess, ckptfile):
# visualize sample reconstructions
e2c.saver.restore(sess, ckptfile) # restore variable values
dataset=PlaneData("data/plane1.npz","data/env1.png")
dataset.initialize()
(x_val,u_val,x_next_val)=dataset.sample(e2c.batch_size)
xr,xp=sess.run([e2c.x_recons, e2c.x_predict],feed_dict={e2c.x:x_val,e2c.u:u_val,e2c.x_next:x_next_val})
#xr,xp=sess.run([e2c.x_recons0, e2c.x_predict0],feed_dict={e2c.x0:x_val,e2c.u0:u_val,e2c.x1:x_next_val})
A,B=e2c.A,e2c.B
def getimgs(x,xnext):
padsize=1
padval=.5
ph=B+2*padsize
pw=A+2*padsize
img=np.ones((10*ph,2*pw))*padval
for i in range(10):
startr=i*ph+padsize
img[startr:startr+B,padsize:padsize+A]=x[i,:].reshape((A,B))
for i in range(10):
startr=i*ph+padsize
img[startr:startr+B,pw+padsize:pw+padsize+A]=xnext[i,:].reshape((A,B))
return img
fig,arr=plt.subplots(1,2)
arr[0].matshow(getimgs(x_val,x_next_val),cmap=plt.cm.gray,vmin=0,vmax=1)
arr[0].set_title('Data')
arr[1].matshow(getimgs(xr,xp),cmap=plt.cm.gray,vmin=0,vmax=1)
arr[1].set_title('Reconstruction')
plt.show()
dataset = PlaneData("data/plane1.npz", "data/env1.png")
dataset.initialize()
# tmp
# (x_val,u_val,x_next_val)=dataset.sample(batch_size, replace=False)
# feed_dict={
# x:x_val,
# u:u_val,
# x_next:x_next_val
# }
# results=sess.run([L_x,L_x_next,L_z,L_bound,KL],feed_dict)
# pdb.set_trace()
# resume training
#saver.restore(sess, "/ltmp/e2c-plane-83000.ckpt")
train_iters = 2e5 # 5K iters
for i in range(int(train_iters)):
(x_val, u_val, x_next_val) = dataset.sample(batch_size, replace=False)
feed_dict = {x: x_val, u: u_val, x_next: x_next_val}
plt.hist(x_val[0, :])
plt.show()
results = sess.run([loss, all_summaries, train_op], feed_dict)
if i % 1000 == 0:
print("iter=%d : Loss: %f" % (i, results[0]))
if i > 2000:
writer.add_summary(results[1], i)
if (i % 100 == 0 and i < 1000) or (i % 1000 == 0):
saver.save(sess, ckpt_file + "-%05d" % (i) + ".ckpt")
sess.close()
dataset.initialize()
# tmp
# (x_val,u_val,x_next_val)=dataset.sample(batch_size, replace=False)
# feed_dict={
# x:x_val,
# u:u_val,
# x_next:x_next_val
# }
# results=sess.run([L_x,L_x_next,L_z,L_bound,KL],feed_dict)
# pdb.set_trace()
# resume training
#saver.restore(sess, "/ltmp/e2c-plane-83000.ckpt")
train_iters=2e5 # 5K iters
for i in range(int(train_iters)):
(x_val,u_val,x_next_val)=dataset.sample(batch_size, replace=False)
feed_dict={
x:x_val,
u:u_val,
x_next:x_next_val
}
plt.hist(x_val[0,:])
plt.show()
results=sess.run([loss,all_summaries,train_op],feed_dict)
if i%1000==0:
print("iter=%d : Loss: %f" % (i,results[0]))
if i>2000:
writer.add_summary(results[1], i)
if (i%100==0 and i < 1000) or (i % 1000 == 0):
saver.save(sess,ckpt_file+"-%05d"%(i)+".ckpt")