def evaluate():
"""Run Eval once.
"""
with tf.Session() as sess:
# Make image placeholder
boundary, true_flow = flow_net.inputs_flow(batch_size=FLAGS.batch_size,
shape=shape,
dims=FLAGS.dims)
# inference model.
predicted_flow = flow_net.inference_network(boundary)
# Restore for eval
init = tf.global_variables_initializer()
sess.run(init)
# make graph def
graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
# make fake data
batch_boundary = np.zeros([FLAGS.batch_size] + shape + [1])
sess.run(predicted_flow, feed_dict={boundary: batch_boundary})
t = time.time()
for i in tqdm(xrange(run_steps)):
# calc flow
sess.run(predicted_flow, feed_dict={boundary: batch_boundary})
elapsed = time.time() - t
filename = "./figs/" + FLAGS.flow_model + "_shape_" + FLAGS.shape + "_batch_size_" + str(
FLAGS.batch_size) + ".txt"
with open(filename, "w") as f:
f.write(str(elapsed / float(FLAGS.batch_size * run_steps)))
def evaluate():
"""Run Eval once.
"""
with tf.Session() as sess:
# Make image placeholder
param_inputs, _ = flow_net.inputs_boundary(FLAGS.nr_boundary_params, 1, shape)
# Make boundary
boundary = flow_net.inference_boundary(1, FLAGS.dims*[FLAGS.obj_size], param_inputs, full_shape=shape)
boundary = tf.round(boundary)
# inference model.
predicted_flow = flow_net.inference_network(boundary)
# init graph
init = tf.global_variables_initializer()
sess.run(init)
# Restore variables
variables_to_restore = tf.all_variables()
variables_to_restore_boundary = [variable for i, variable in enumerate(variables_to_restore) if "boundary_network" in variable.name[:variable.name.index(':')]]
variables_to_restore_flow = [variable for i, variable in enumerate(variables_to_restore) if "flow_network" in variable.name[:variable.name.index(':')]]
saver_boundary = tf.train.Saver(variables_to_restore_boundary)
saver_flow = tf.train.Saver(variables_to_restore_flow)
ckpt_boundary = tf.train.get_checkpoint_state(BOUNDARY_DIR)
ckpt_flow = tf.train.get_checkpoint_state(FLOW_DIR)
saver_boundary.restore(sess, ckpt_boundary.model_checkpoint_path)
saver_flow.restore(sess, ckpt_flow.model_checkpoint_path)
global_step = 1
# make graph def
graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
for i in xrange(10):
# random params
rand_param = np.expand_dims(get_random_params(FLAGS.nr_boundary_params, 3), axis=0)
rand_param[:,0] = 0
rand_param[:,1] = 0
rand_param[:,2] = 0.5
rand_param[:,3] = 1.0
rand_param = np.load("figs/3d_wing_params_op.npy")
rand_param = np.expand_dims(rand_param, axis=0)
# calc flow
p_flow, p_boundary = sess.run([predicted_flow, boundary],feed_dict={param_inputs: rand_param})
p_flow = p_flow * ((-p_boundary) + 1.0)
dim=3
# plt boundary
#plt.imshow(p_boundary[0,:,:,72,0])
#plt.show()
# save vtk file of it
image_vtk = tvtk.ImageData(spacing=(1, 1, 1), origin=(0, 0, 0))
image_vtk.point_data.vectors = p_flow[0,:,:,:,0:3].reshape([shape[0]*shape[1]*shape[2], 3])
image_vtk.point_data.scalars = p_flow[0,:,:,:,3].reshape([shape[0]*shape[1]*shape[2]])
image_vtk.point_data.scalars.name = "pressure"
image_vtk.dimensions = shape
write_data(image_vtk, "figs/vtk_flow_wing_3d")
image_vtk = tvtk.ImageData(spacing=(1, 1, 1), origin=(0, 0, 0))
image_vtk.point_data.scalars = p_boundary[0,:,:,:,0].reshape([shape[0]*shape[1]*shape[2]])
image_vtk.point_data.scalars.name = "boundary"
image_vtk.dimensions = shape
write_data(image_vtk, "figs/vtk_boundary_wing_3d")
def evaluate():
"""Run Eval once.
"""
with tf.Session() as sess:
# Make image placeholder
boundary, true_flow = flow_net.inputs_flow(batch_size=batch_size,
shape=shape,
dims=FLAGS.dims)
# Build a Graph that computes the logits predictions from the
# inference model.
predicted_flow = flow_net.inference_network(boundary,
network_type="flow",
keep_prob=FLAGS.keep_prob)
# predict force
if FLAGS.flow_model == "xiao_network":
sharp_boundary = tf.minimum(
tf.maximum(tf.round(-boundary + .5), 0.0), 1.0)
else:
sharp_boundary = boundary
predicted_force = calc_force(sharp_boundary, predicted_flow[..., 2:3])
predicted_drag_x = tf.reduce_sum(predicted_force[..., 0], axis=[1, 2])
predicted_drag_y = tf.reduce_sum(predicted_force[..., 1], axis=[1, 2])
true_force = calc_force(sharp_boundary, true_flow[..., 2:3])
true_drag_x = tf.reduce_sum(true_force[..., 0], axis=[1, 2])
true_drag_y = tf.reduce_sum(true_force[..., 1], axis=[1, 2])
# predicted max vel
predicted_flow_norm = calc_velocity_norm(predicted_flow)
true_flow_norm = calc_velocity_norm(true_flow)
predicted_max_vel = tf.reduce_max(predicted_flow_norm, axis=[1, 2])
true_max_vel = tf.reduce_max(true_flow_norm, axis=[1, 2])
# Restore for eval
init = tf.global_variables_initializer()
sess.run(init)
variables_to_restore = tf.all_variables()
variables_to_restore_flow = [
variable for i, variable in enumerate(variables_to_restore)
if "flow_network" in variable.name[:variable.name.index(':')]
]
saver = tf.train.Saver(variables_to_restore_flow)
ckpt = tf.train.get_checkpoint_state(FLOW_DIR)
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = 1
graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
# make vtm dataset
dataset = Sailfish_data("../../data/",
size=FLAGS.obj_size,
dim=FLAGS.dims)
dataset.parse_data()
#dataset.load_data(FLAGS.dims, FLAGS.obj_size)
# store drag data
p_drag_x_data = []
t_drag_x_data = []
p_drag_y_data = []
t_drag_y_data = []
p_max_vel_data = []
t_max_vel_data = []
#for run in filenames:
#for i in tqdm(xrange(60)):
for i in tqdm(xrange(99)):
# read in boundary
batch_boundary, batch_flow = dataset.minibatch(
train=False,
batch_size=batch_size,
signed_distance_function=FLAGS.sdf)
# calc flow
p_drag_x, t_drag_x, p_drag_y, t_drag_y, p_max_vel, t_max_vel = sess.run(
[
predicted_drag_x, true_drag_x, predicted_drag_y,
true_drag_y, predicted_max_vel, true_max_vel
],
feed_dict={
boundary: batch_boundary,
true_flow: batch_flow
})
#plt.imshow(sess.run(sharp_boundary, feed_dict={boundary: batch_boundary})[0,:,:,0] - batch_boundary[0,:,:,0])
#plt.imshow(batch_boundary[0,:,:,0])
#plt.show()
p_drag_x_data.append(p_drag_x)
t_drag_x_data.append(t_drag_x)
p_drag_y_data.append(p_drag_y)
t_drag_y_data.append(t_drag_y)
p_max_vel_data.append(p_max_vel)
t_max_vel_data.append(t_max_vel)
# display it
p_drag_x_data = np.concatenate(p_drag_x_data, axis=0)
t_drag_x_data = np.concatenate(t_drag_x_data, axis=0)
p_drag_y_data = np.concatenate(p_drag_y_data, axis=0)
t_drag_y_data = np.concatenate(t_drag_y_data, axis=0)
p_max_vel_data = np.concatenate(p_max_vel_data, axis=0)
t_max_vel_data = np.concatenate(t_max_vel_data, axis=0)
# save it
np.savez("./figs/store_flow_accuracy_values/" + FLAGS.flow_model,
p_drag_x_data=p_drag_x_data,
t_drag_x_data=t_drag_x_data,
p_drag_y_data=p_drag_x_data,
t_drag_y_data=t_drag_x_data,
p_max_vel_data=p_max_vel_data,
t_max_vel_data=t_max_vel_data)
fig = plt.figure(figsize=(15, 15))
a = fig.add_subplot(2, 2, 1)
font_size_axis = 6
plt.scatter(t_drag_x_data, p_drag_x_data)
plt.plot(t_drag_x_data, t_drag_x_data, color="red")
plt.title("X Force", fontsize=38)
a = fig.add_subplot(2, 2, 2)
plt.scatter(t_drag_y_data, p_drag_y_data)
plt.plot(t_drag_y_data, t_drag_y_data, color="red")
plt.title("Y Force", fontsize=38)
a = fig.add_subplot(2, 2, 3)
plt.scatter(t_max_vel_data, p_max_vel_data)
plt.plot(t_max_vel_data, t_max_vel_data, color="red")
plt.title("Max Velocity", fontsize=38)
plt.ylabel("Predicted", fontsize=26)
plt.xlabel("True", fontsize=26)
plt.savefig("./figs/flow_accuracy_2d.pdf")
plt.show()
def evaluate():
"""Run Eval once.
Args:
saver: Saver.
summary_writer: Summary writer.
top_k_op: Top K op.
summary_op: Summary op.
"""
with tf.Graph().as_default():
# Make image placeholder
params_op, params_op_init, params_op_set, squeeze_loss = flow_net.inputs_boundary_learn(
batch_size, network_type="heat", noise_std=0.01)
# Make boundary
boundary = flow_net.inference_boundary(batch_size, 2 * [128],
params_op)
# predict steady flow on boundary
predicted_heat = flow_net.inference_network(boundary,
network_type="heat",
keep_prob=1.0)
# loss
#loss = tf.reduce_sum(predicted_heat)
loss = tf.reduce_max(predicted_heat)
loss += squeeze_loss
# train_op
variables_to_train = tf.all_variables()
variables_to_train = [
variable for i, variable in enumerate(variables_to_train)
if "params" in variable.name[:variable.name.index(':')]
]
train_step = flow_net.train(loss,
FLAGS.boundary_learn_lr,
train_type="boundary_params",
variables=variables_to_train)
# init graph
init = tf.global_variables_initializer()
# Restore the moving average version of the learned variables for eval.
variables_to_restore = tf.all_variables()
variables_to_restore_boundary = [
variable for i, variable in enumerate(variables_to_restore)
if "boundary_network" in variable.name[:variable.name.index(':')]
]
variables_to_restore_heat = [
variable for i, variable in enumerate(variables_to_restore)
if "heat_network" in variable.name[:variable.name.index(':')]
]
saver_boundary = tf.train.Saver(variables_to_restore_boundary)
saver_heat = tf.train.Saver(variables_to_restore_heat)
# start ses and init
sess = tf.Session()
sess.run(init)
ckpt_boundary = tf.train.get_checkpoint_state(BOUNDARY_DIR)
ckpt_heat = tf.train.get_checkpoint_state(FLOW_DIR)
saver_boundary.restore(sess, ckpt_boundary.model_checkpoint_path)
saver_heat.restore(sess, ckpt_heat.model_checkpoint_path)
graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
params_np = (np.random.rand(1, FLAGS.nr_boundary_params) - .5) / 2.0
sess.run(params_op_init, feed_dict={params_op_set: params_np})
run_time = FLAGS.boundary_learn_steps
# make store vectors for values
plot_error = []
# make store dir
os.system("mkdir ./figs/boundary_learn_image_store")
for i in tqdm(xrange(run_time)):
l, _, = sess.run([loss, train_step], feed_dict={})
print(l)
plot_error.append(np.sum(l))
if ((i + 1) % 1 == 0) or i == run_time - 1:
# make video with opencv
p_heat, p_boundary = sess.run([predicted_heat, boundary])
# save plot image to make video
fig = plt.figure()
fig.set_size_inches(15, 5)
a = fig.add_subplot(1, 3, 2)
plt.imshow(p_heat[0, :, :, 0])
#plt.title("Heat Dissipation", fontsize="x-large")
plt.title("Heat Dissipation", fontsize=16)
a = fig.add_subplot(1, 3, 3)
plt.imshow(p_boundary[0, :, :, 0])
plt.title("Heat Sink Geometry", fontsize=16)
a = fig.add_subplot(1, 3, 1)
plt.plot(np.array(plot_error), label="Temp at Source")
plt.xlim(-10, 510)
plt.xlabel("Step")
plt.ylabel("Temp")
plt.legend()
plt.suptitle("Heat Sink Optimization Using Gradient Descent",
fontsize=20)
plt.savefig("./figs/boundary_learn_image_store/plot_" +
str(i).zfill(5) + ".png")
if run_time - i <= 100:
plt.savefig("./figs/" + FLAGS.boundary_learn_loss +
"_plot.png")
if i == run_time - 1:
plt.savefig("./figs/heat_learn_gradient_decent.pdf")
plt.show()
plt.close(fig)
# generate video of plots
os.system("rm ./figs/heat_learn_video.mp4")
os.system(
"cat ./figs/boundary_learn_image_store/*.png | ffmpeg -f image2pipe -r 30 -vcodec png -i - -vcodec libx264 ./figs/heat_learn_video.mp4"
)
os.system("rm -r ./figs/boundary_learn_image_store")
def train():
"""Train ring_net for a number of steps."""
with tf.Graph().as_default():
# store grad and loss values
grads = []
loss_gen = []
# global step counter
global_step = tf.get_variable('global_step', [],
initializer=tf.constant_initializer(0),
trainable=False)
# do for all gpus
for i in range(FLAGS.nr_gpus):
print("Unrolling on gpu:" + str(i))
with tf.device('/gpu:%d' % i):
# make inputs
boundary, true_flow = flow_net.inputs_flow(
FLAGS.batch_size, shape, FLAGS.dims)
# create and unrap network
predicted_flow = flow_net.inference_network(
boundary, network_type="flow", keep_prob=FLAGS.keep_prob)
# if i is one then get variables to store all trainable params and
if i == 0:
all_params = tf.trainable_variables()
# calc error
error = flow_net.loss_flow(true_flow, predicted_flow, boundary)
loss_gen.append(error)
# store grads
grads.append(tf.gradients(loss_gen[i], all_params))
# exponential moving average for training
ema = tf.train.ExponentialMovingAverage(decay=.9995)
maintain_averages_op = tf.group(ema.apply(all_params))
# store up the loss and gradients on gpu:0
with tf.device('/gpu:0'):
for i in range(1, FLAGS.nr_gpus):
loss_gen[0] += loss_gen[i]
for j in range(len(grads[0])):
grads[0][j] += grads[i][j]
# train (hopefuly)
train_op = tf.group(
adam_updates(all_params,
grads[0],
lr=FLAGS.lr,
mom1=0.95,
mom2=0.9995), maintain_averages_op,
global_step.assign_add(1))
# set total loss for printing
total_loss = loss_gen[0]
tf.summary.scalar('total_loss', total_loss)
# List of all Variables
variables = tf.global_variables()
# Build a saver
saver = tf.train.Saver(tf.global_variables())
# Summary op
summary_op = tf.summary.merge_all()
# Build an initialization operation to run below.
init = tf.global_variables_initializer()
# Start running operations on the Graph.
sess = tf.Session()
# init if this is the very time training
sess.run(init)
# init from checkpoint
variables_to_restore = ema.variables_to_restore(
tf.moving_average_variables())
saver_restore = tf.train.Saver(variables_to_restore)
ckpt = tf.train.get_checkpoint_state(TRAIN_DIR)
if ckpt is not None:
print("init from " + TRAIN_DIR)
try:
saver_restore.restore(sess, ckpt.model_checkpoint_path)
except:
tf.gfile.DeleteRecursively(TRAIN_DIR)
tf.gfile.MakeDirs(TRAIN_DIR)
print(
"there was a problem using variables in checkpoint, random init will be used instead"
)
# Start que runner
tf.train.start_queue_runners(sess=sess)
# Summary op
graph_def = sess.graph.as_graph_def(add_shapes=True)
summary_writer = tf.summary.FileWriter(TRAIN_DIR, graph_def=graph_def)
# calc number of steps left to run
run_steps = FLAGS.max_steps - int(sess.run(global_step))
# make vtm dataset
dataset = Sailfish_data("../../data/",
size=FLAGS.obj_size,
dim=FLAGS.dims)
#dataset.load_data(FLAGS.dims, FLAGS.obj_size)
dataset.parse_data()
for step in xrange(run_steps):
current_step = sess.run(global_step)
t = time.time()
batch_boundary, batch_flow = dataset.minibatch(
batch_size=FLAGS.batch_size,
signed_distance_function=FLAGS.sdf)
_, loss_value = sess.run([train_op, error],
feed_dict={
boundary: batch_boundary,
true_flow: batch_flow
})
elapsed = time.time() - t
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if current_step % 10 == 0:
print("loss value at " + str(loss_value))
print("time per batch is " + str(elapsed))
if current_step % 1000 == 0:
summary_str = sess.run(summary_op,
feed_dict={
boundary: batch_boundary,
true_flow: batch_flow
})
summary_writer.add_summary(summary_str, current_step)
checkpoint_path = os.path.join(TRAIN_DIR, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=global_step)
print("saved to " + TRAIN_DIR)
def evaluate():
"""Run Eval once.
Args:
saver: Saver.
summary_writer: Summary writer.
top_k_op: Top K op.
summary_op: Summary op.
"""
with tf.Graph().as_default():
# Make image placeholder
inputs_vector, true_boundary = flow_net.inputs_boundary(
FLAGS.nr_boundary_params, batch_size, shape)
# Build a Graph that computes the logits predictions from the
# inference model.
#inputs_vector_noise = inputs_vector + tf.random_normal(shape=tf.shape(inputs_vector), mean=0.0, stddev=0.0001, dtype=tf.float32)
boundary = flow_net.inference_boundary(1,
FLAGS.dims * [FLAGS.obj_size],
inputs=inputs_vector,
full_shape=shape)
#boundary = tf.round(boundary)
predicted_flow = flow_net.inference_network(boundary, keep_prob=1.0)
# quantities to optimize
force = calc_force(boundary, predicted_flow[:, :, :, 2:3])
drag_x = tf.reduce_sum(force[:, :, :, 0])
drag_y = tf.reduce_sum(force[:, :, :, 1])
drag_ratio = (drag_y / drag_x)
# init graph
init = tf.global_variables_initializer()
# Restore the moving average version of the learned variables for eval.
variables_to_restore = tf.all_variables()
variables_to_restore_boundary = [
variable for i, variable in enumerate(variables_to_restore)
if "boundary_network" in variable.name[:variable.name.index(':')]
]
variables_to_restore_flow = [
variable for i, variable in enumerate(variables_to_restore)
if "flow_network" in variable.name[:variable.name.index(':')]
]
saver_boundary = tf.train.Saver(variables_to_restore_boundary)
saver_flow = tf.train.Saver(variables_to_restore_flow)
# start ses and init
sess = tf.Session()
sess.run(init)
ckpt_boundary = tf.train.get_checkpoint_state(BOUNDARY_DIR)
ckpt_flow = tf.train.get_checkpoint_state(FLOW_DIR)
saver_boundary.restore(sess, ckpt_boundary.model_checkpoint_path)
saver_flow.restore(sess, ckpt_flow.model_checkpoint_path)
graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
params_np = get_random_params(FLAGS.nr_boundary_params, 2)
params_np = np.expand_dims(params_np, axis=0)
params_np[0, 0] = 0.0
params_np[0, 1] = 0.5
params_np[0, 2] = 1.0
params_np[0, 4] = 0.0
# make store vectors for values
resolution = 320
loss_val = np.zeros((resolution))
max_d_ratio = np.zeros((resolution))
d_ratio_store = None
boundary_frame_store = []
store_freq = int(resolution / nr_frame_saves)
# make store dir
for i in tqdm(xrange(resolution)):
params_np[0, 4] += (0.3) / resolution
velocity_norm_g = sess.run(drag_ratio,
feed_dict={
inputs_vector:
np.concatenate(batch_size *
[params_np],
axis=0)
})
if i % store_freq == 0:
boundary_frame_store.append(
sess.run(
boundary,
feed_dict={
inputs_vector:
np.concatenate(batch_size * [params_np], axis=0)
})[0,
int(FLAGS.obj_size / 2):int(3 * FLAGS.obj_size / 2),
int(FLAGS.obj_size / 2):int(3 * FLAGS.obj_size / 2),
0])
loss_val[i] = velocity_norm_g
fig = plt.figure(figsize=(10, 5))
a = fig.add_subplot(1, 2, 1)
plt.title("Boundary from Parameter Change", fontsize=16)
boundary_frame_store = tile_frames(boundary_frame_store)
plt.imshow(boundary_frame_store)
#plt.tick_params(axis='both', top="off", bottom="off")
plt.axis('off')
a = fig.add_subplot(1, 2, 2)
#plt.imshow(np.concatenate(boundary_frame_store, axis = 0))
plt.plot(np.arange(resolution) / float(resolution) - .5, loss_val)
plt.ylabel("Lift/Drag")
plt.xlabel("Parameter Value")
plt.title("Loss vs Parameter Value", fontsize=16)
plt.savefig("./figs/boundary_space_explore.pdf")
plt.show()
def evaluate():
"""Run Eval once.
Args:
saver: Saver.
summary_writer: Summary writer.
top_k_op: Top K op.
summary_op: Summary op.
"""
num_angles = 9
max_angle = 0.10
min_angle = -0.30
set_params = np.array(num_angles * [FLAGS.nr_boundary_params * [0.0]])
set_params[:, :] = 0.0
set_params_pos = np.array(num_angles * [FLAGS.nr_boundary_params * [0.0]])
set_params_pos[:, :] = 1.0
for i in xrange(num_angles):
set_params[i, 0] = -i
set_params[:, 0] = ((max_angle - min_angle) *
(set_params[:, 0] / (num_angles - 1))) - min_angle
print(set_params[:, 0])
set_params[:, 2] = 0.5
set_params[:, 3] = 1.0
set_params[:, -1] = 0.0
set_params_pos[:, 0] = 0.0 # set angle to 0.0
set_params_pos[:, 1] = 0.0 # set angle to 0.0
set_params_pos[:, 2] = 0.0 # set n_1 to .5
set_params_pos[:, 3] = 0.0 # set n_2 to 1.0
set_params_pos[:, -1] = 0.0 # set tail hieght to 0.0
with tf.Graph().as_default():
# Make image placeholder
params_op, params_op_init, params_op_set, squeeze_loss = flow_net.inputs_boundary_learn(
batch_size,
set_params=set_params,
set_params_pos=set_params_pos,
noise_std=0.01)
# Make placeholder for flow computed by lattice boltzmann solver
solver_boundary, solver_flow = flow_net.inputs_flow(
1, shape, FLAGS.dims)
sharp_boundary, blaa = flow_net.inputs_flow(
batch_size * set_params.shape[0], shape, FLAGS.dims)
# Make boundary
params_op_0, params_op_1, params_op_2 = tf.split(params_op, 3, axis=0)
with tf.device('/gpu:0'):
boundary_0 = flow_net.inference_boundary(
batch_size * int(set_params.shape[0] / 3),
FLAGS.dims * [FLAGS.obj_size],
params_op_0,
full_shape=shape)
with tf.device('/gpu:1'):
boundary_1 = flow_net.inference_boundary(
batch_size * int(set_params.shape[0] / 3),
FLAGS.dims * [FLAGS.obj_size],
params_op_1,
full_shape=shape)
with tf.device('/gpu:2'):
boundary_2 = flow_net.inference_boundary(
batch_size * int(set_params.shape[0] / 3),
FLAGS.dims * [FLAGS.obj_size],
params_op_2,
full_shape=shape)
# predict steady flow on boundary
grads = []
loss_gen = []
with tf.device('/gpu:0'):
predicted_flow_0 = flow_net.inference_network(
boundary_0, network_type="flow", keep_prob=FLAGS.keep_prob)
force_0 = calc_force(boundary_0, predicted_flow_0[..., -1:])
drag_x_0 = tf.reduce_sum(force_0[..., 0], axis=[1, 2, 3
]) / batch_size
drag_y_0 = tf.reduce_sum(force_0[..., 1], axis=[1, 2, 3
]) / batch_size
drag_z_0 = tf.reduce_sum(force_0[..., 2], axis=[1, 2, 3
]) / batch_size
drag_lift_ratio_0 = -(drag_x_0 / drag_z_0)
loss_0 = -tf.reduce_sum(drag_lift_ratio_0)
loss_0 += squeeze_loss
variables_to_train = tf.all_variables()
variables_to_train = [
variable for i, variable in enumerate(variables_to_train)
if "params" in variable.name[:variable.name.index(':')]
]
loss_gen.append(loss_0)
# store grads
grads.append(tf.gradients(loss_gen[0], variables_to_train))
with tf.device('/gpu:1'):
predicted_flow_1 = flow_net.inference_network(
boundary_1, network_type="flow", keep_prob=FLAGS.keep_prob)
force_1 = calc_force(boundary_1, predicted_flow_1[..., -1:])
drag_x_1 = tf.reduce_sum(force_1[..., 0], axis=[1, 2, 3
]) / batch_size
drag_y_1 = tf.reduce_sum(force_1[..., 1], axis=[1, 2, 3
]) / batch_size
drag_z_1 = tf.reduce_sum(force_1[..., 2], axis=[1, 2, 3
]) / batch_size
drag_lift_ratio_1 = -(drag_x_1 / drag_z_1)
loss_1 = -tf.reduce_sum(drag_lift_ratio_1)
loss_1 += squeeze_loss
variables_to_train = tf.all_variables()
variables_to_train = [
variable for i, variable in enumerate(variables_to_train)
if "params" in variable.name[:variable.name.index(':')]
]
loss_gen.append(loss_1)
# store grads
grads.append(tf.gradients(loss_gen[1], variables_to_train))
with tf.device('/gpu:2'):
predicted_flow_2 = flow_net.inference_network(
boundary_2, network_type="flow", keep_prob=FLAGS.keep_prob)
force_2 = calc_force(boundary_2, predicted_flow_2[..., -1:])
drag_x_2 = tf.reduce_sum(force_2[..., 0], axis=[1, 2, 3
]) / batch_size
drag_y_2 = tf.reduce_sum(force_2[..., 1], axis=[1, 2, 3
]) / batch_size
drag_z_2 = tf.reduce_sum(force_2[..., 2], axis=[1, 2, 3
]) / batch_size
drag_lift_ratio_2 = -(drag_x_2 / drag_z_2)
loss_2 = -tf.reduce_sum(drag_lift_ratio_2)
loss_2 += squeeze_loss
variables_to_train = tf.all_variables()
variables_to_train = [
variable for i, variable in enumerate(variables_to_train)
if "params" in variable.name[:variable.name.index(':')]
]
loss_gen.append(loss_2)
# store grads
grads.append(tf.gradients(loss_gen[2], variables_to_train))
# store up the loss and gradients on gpu:0
with tf.device('/gpu:0'):
for i in range(1, 3):
loss_gen[0] += loss_gen[i]
for j in range(len(grads[0])):
grads[0][j] += grads[i][j]
train_step = tf.group(
adam_updates(variables_to_train,
grads[0],
lr=FLAGS.boundary_learn_lr,
mom1=0.95,
mom2=0.9995))
#with tf.device('/cpu:0'):
# sharp_predicted_flow = flow_net.inference_network(sharp_boundary, network_type="flow", keep_prob=FLAGS.keep_prob)
# quantities to optimize
#force = calc_force(boundary, predicted_flow[...,-1:])
#sharp_force = calc_force(sharp_boundary, sharp_predicted_flow[...,-1:])
#solver_force = calc_force(solver_boundary, solver_flow[...,-1:])
#drag_x = tf.reduce_sum(force[...,0], axis=[1,2,3])/batch_size
#drag_y = tf.reduce_sum(force[...,1], axis=[1,2,3])/batch_size
#drag_z = tf.reduce_sum(force[...,2], axis=[1,2,3])/batch_size
"""
sharp_drag_x = tf.reduce_sum(sharp_force[...,0], axis=[1,2,3])/batch_size
sharp_drag_y = tf.reduce_sum(sharp_force[...,1], axis=[1,2,3])/batch_size
sharp_drag_z = tf.reduce_sum(sharp_force[...,2], axis=[1,2,3])/batch_size
solver_drag_x = tf.reduce_sum(solver_force[...,0], axis=[1,2,3])/batch_size
solver_drag_y = tf.reduce_sum(solver_force[...,1], axis=[1,2,3])/batch_size
solver_drag_z = tf.reduce_sum(solver_force[...,2], axis=[1,2,3])/batch_size
"""
#drag_x = tf.concat([drag_x_0, drag_x_1, drag_x_2], axis=0)
#drag_y = tf.concat([drag_y_0, drag_y_1, drag_y_2], axis=0)
#drag_z = tf.concat([drag_z_0, drag_z_1, drag_z_2], axis=0)
#drag_lift_ratio = -(drag_x/drag_z)
#sharp_drag_lift_ratio = -(sharp_drag_x/sharp_drag_z)
#solver_drag_lift_ratio = -(solver_drag_x/solver_drag_z)
# loss
#loss = - tf.abs(tf.constant([-25.0, 100.0, 200.0]) - drag_x) - drag_z
#loss = drag_x - drag_z/2.0
#loss = -tf.reduce_sum(drag_lift_ratio)
#loss += squeeze_loss
# train_op
#variables_to_train = tf.all_variables()
#variables_to_train = [variable for i, variable in enumerate(variables_to_train) if "params" in variable.name[:variable.name.index(':')]]
#train_step = flow_net.train(loss, FLAGS.boundary_learn_lr, train_type="boundary_params", variables=variables_to_train)
# init graph
init = tf.global_variables_initializer()
# Restore the moving average version of the learned variables for eval.
variables_to_restore = tf.all_variables()
variables_to_restore_boundary = [
variable for i, variable in enumerate(variables_to_restore)
if "boundary_network" in variable.name[:variable.name.index(':')]
]
variables_to_restore_flow = [
variable for i, variable in enumerate(variables_to_restore)
if "flow_network" in variable.name[:variable.name.index(':')]
]
saver_boundary = tf.train.Saver(variables_to_restore_boundary)
saver_flow = tf.train.Saver(variables_to_restore_flow)
# start ses and init
sess = tf.Session()
sess.run(init)
ckpt_boundary = tf.train.get_checkpoint_state(BOUNDARY_DIR)
ckpt_flow = tf.train.get_checkpoint_state(FLOW_DIR)
saver_boundary.restore(sess, ckpt_boundary.model_checkpoint_path)
saver_flow.restore(sess, ckpt_flow.model_checkpoint_path)
graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
params_np = (np.random.rand(1, FLAGS.nr_boundary_params) - .5) / 2.0
#params_np = (np.random.rand(1,FLAGS.nr_boundary_params) - .5)/2.0
#params_np = np.zeros((1,FLAGS.nr_boundary_params-1))
sess.run(params_op_init, feed_dict={params_op_set: params_np})
run_time = FLAGS.boundary_learn_steps
# make store vectors for values
plot_error = np.zeros((run_time))
plot_drag_x = np.zeros((run_time))
plot_drag_y = np.zeros((run_time))
plot_drag_z = np.zeros((run_time))
# make store dir
os.system("mkdir ./figs/boundary_learn_image_store")
for i in tqdm(xrange(run_time)):
#l_0, _ = sess.run([loss_0, train_step_0], feed_dict={})
l, _, d_x, d_y, d_z = sess.run(
[loss_gen[0], train_step, drag_x_2, drag_y_2, drag_z_2],
feed_dict={})
#l_2, _ = sess.run([loss_2, train_step_2], feed_dict={})
plot_error[i] = np.sum(l)
plot_drag_x[i] = np.sum(d_x[fig_pos])
plot_drag_y[i] = np.sum(d_y[fig_pos])
plot_drag_z[i] = np.sum(d_z[fig_pos])
if ((i + 1) % 20 == 0) or i == run_time - 1:
# make video with opencv
"""
s_params = sess.run(params_op)
wing_boundary = []
for p in xrange(s_params.shape[0]):
wing_boundary.append(wing_boundary_3d(s_params[p,0], s_params[p,1], s_params[p,2],
s_params[p,3], s_params[p,4], s_params[p,5],
s_params[p,6:int((FLAGS.nr_boundary_params-7)/3+6)],
s_params[p,int((FLAGS.nr_boundary_params-7)/3+6):int(2*(FLAGS.nr_boundary_params-7)/3+6)],
s_params[p,int(2*(FLAGS.nr_boundary_params-7)/3+6):-1],
s_params[p,-1], FLAGS.dims*[FLAGS.obj_size]))
wing_boundary = np.stack(wing_boundary)
wing_boundary = np.pad(wing_boundary, [[0,0],[24,24],[24,24],[24,24],[0,0]], 'constant', constant_values=0.0)
#print(sharp_boundary.get_shape())
#print(wing_boundary.shape)
p_flow, p_boundary, d_l_ratio, sharp_d_l_ratio = sess.run([sharp_predicted_flow, boundary, drag_lift_ratio, sharp_drag_lift_ratio],feed_dict={sharp_boundary: wing_boundary})
"""
p_flow, p_boundary, d_l_ratio_0, d_l_ratio_1, d_l_ratio_2 = sess.run(
[
predicted_flow_2, boundary_2, drag_lift_ratio_0,
drag_lift_ratio_1, drag_lift_ratio_2
])
d_l_ratio = np.concatenate(
[d_l_ratio_0, d_l_ratio_1, d_l_ratio_2], axis=0)
# save plot image to make video
#p_pressure = p_flow[fig_pos,:,:,72,2]
p_boundary = np.concatenate([
p_boundary[fig_pos, :, 71, :, 0], p_boundary[fig_pos, :, :,
71, 0]
],
axis=0)
p_pressure = np.concatenate([
p_flow[fig_pos, :, 71, :, 3], p_flow[fig_pos, :, :, 71, 3]
],
axis=0)
#p_pressure = p_flow[:,:,76,:,3].reshape((p_boundary.shape[0]*p_boundary.shape[1], p_boundary.shape[2]))
#p_boundary = p_boundary[:,:,76,:,0].reshape((p_boundary.shape[0]*p_boundary.shape[1], p_boundary.shape[2]))
fig = plt.figure()
fig.set_size_inches(20, 5)
#a = fig.add_subplot(1,5,1)
#plt.imshow(p_pressure)
a = fig.add_subplot(1, 4, 1)
plt.imshow(p_boundary)
a = fig.add_subplot(1, 4, 2)
plt.plot(plot_error, label="Sum(Lift/Drag)")
plt.xlabel("Step")
plt.legend()
a = fig.add_subplot(1, 4, 3)
plt.plot(plot_drag_x, label="Lift Angle 0")
plt.plot(-plot_drag_z, label="Drag Angle 0")
plt.ylim(0.0, np.max(plot_drag_x) + 30.0)
plt.xlabel("Step")
plt.legend()
a = fig.add_subplot(1, 4, 4)
plt.plot(np.degrees(set_params[:, 0]),
d_l_ratio,
'bo',
label="Lift/Drag Network")
#plt.plot(-np.degrees(set_params[3:6,0]), sharp_d_l_ratio, 'ro', label="Lift/Drag Sharp")
#if i == run_time-1:
# solver_d_l_ratio = run_flow_solver(sess.run(params_op), solver_boundary, solver_flow, sess, solver_drag_lift_ratio)
# plt.plot(-np.degrees(set_params[:,0]), solver_d_l_ratio, 'go', label="Lift/Drag Solver")
plt.xlabel("Angle of Attack (Degrees)")
plt.xlim(
min(np.degrees(set_params[:, 0])) - 3,
max(np.degrees(set_params[:, 0])) + 3)
plt.ylim(np.min(d_l_ratio) - 1.0, np.max(d_l_ratio) + 1.0)
plt.legend()
plt.suptitle("3D Wing Optimization Using Gradient Descent",
fontsize=20)
plt.savefig("./figs/boundary_learn_image_store/plot_" +
str(i).zfill(5) + ".png")
if run_time - i <= 100:
plt.savefig("./figs/" + FLAGS.boundary_learn_loss +
"_plot.png")
if i == run_time - 1:
plt.savefig("./figs/learn_gradient_descent.pdf")
plt.show()
plt.show()
plt.close(fig)
np.save("figs/3d_wing_params_op", sess.run(params_op[6]))
print(sess.run(params_op[6]))
def evaluate():
"""Run Eval once.
"""
with tf.Session() as sess:
# Make image placeholder
boundary, true_flow = flow_net.inputs_flow(batch_size=1, shape=shape, dims=FLAGS.dims)
# inference model.
predicted_flow = flow_net.inference_network(boundary, network_type="flow", keep_prob=FLAGS.keep_prob)
# calc velocity norms
predicted_flow_norm = calc_velocity_norm(predicted_flow)
true_flow_norm = calc_velocity_norm(true_flow)
# extract out pressure field
predicted_pressure = predicted_flow[...,2]
true_pressure = true_flow[...,2]
# Restore for eval
init = tf.global_variables_initializer()
sess.run(init)
variables_to_restore = tf.all_variables()
variables_to_restore_flow = [variable for i, variable in enumerate(variables_to_restore) if "flow_network" in variable.name[:variable.name.index(':')]]
saver = tf.train.Saver(variables_to_restore_flow)
ckpt = tf.train.get_checkpoint_state(FLOW_DIR)
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = 1
# make graph def
graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
# make dataset
dataset = Sailfish_data("../../data/", size=FLAGS.obj_size, dim=FLAGS.dims)
dataset.parse_data()
for i in xrange(10):
# make plot
fig = plt.figure(figsize = (10, 5))
gs1 = gridspec.GridSpec(2, 3)
gs1.update(wspace=0.025, hspace=0.025)
# read in boundary
batch_boundary, batch_flow = dataset.minibatch(train=False, batch_size=1, signed_distance_function=FLAGS.sdf)
# calc flow
p_flow_norm, p_pressure, t_flow_norm, t_pressure = sess.run([predicted_flow_norm, predicted_pressure, true_flow_norm, true_pressure],feed_dict={boundary: batch_boundary, true_flow: batch_flow})
axarr = plt.subplot(gs1[0])
im = axarr.imshow(p_flow_norm[0])
axarr.get_xaxis().set_ticks([])
axarr.get_yaxis().set_ticks([])
axarr.set_ylabel("Velocity", y = .5, x = .5)
axarr.set_title("Generated", y = .99)
fig.colorbar(im, ax=axarr)
axarr = plt.subplot(gs1[1])
im = axarr.imshow(t_flow_norm[0])
axarr.get_xaxis().set_ticks([])
axarr.get_yaxis().set_ticks([])
axarr.set_title("True", y = .99)
fig.colorbar(im, ax=axarr)
axarr = plt.subplot(gs1[2])
im = axarr.imshow(np.abs(p_flow_norm[0] - t_flow_norm[0]))
axarr.get_xaxis().set_ticks([])
axarr.get_yaxis().set_ticks([])
axarr.set_title("Difference", y = .99)
fig.colorbar(im, ax=axarr)
axarr = plt.subplot(gs1[3])
im = axarr.imshow(p_pressure[0])
axarr.get_xaxis().set_ticks([])
axarr.get_yaxis().set_ticks([])
axarr.set_ylabel("Pressure", y = .5, x = .5)
fig.colorbar(im, ax=axarr)
axarr = plt.subplot(gs1[4])
im = axarr.imshow(t_pressure[0])
axarr.get_xaxis().set_ticks([])
axarr.get_yaxis().set_ticks([])
fig.colorbar(im, ax=axarr)
axarr = plt.subplot(gs1[5])
im = axarr.imshow(np.abs(p_pressure[0] - t_pressure[0]))
axarr.get_xaxis().set_ticks([])
axarr.get_yaxis().set_ticks([])
fig.colorbar(im, ax=axarr)
plt.suptitle("Predicted vs True Steady State Flows", fontsize=20, y=0.98)
plt.savefig("./figs/generated_flow_difference.pdf")
plt.show()
def evaluate():
"""Run Eval once.
"""
with tf.Session() as sess:
# Make image placeholder
boundary, true_flow = flow_net.inputs_flow(batch_size=batch_size,
shape=shape,
dims=FLAGS.dims)
# Build a Graph that computes the logits predictions from the
# inference model.
predicted_flow = flow_net.inference_network(boundary)
# predict force
predicted_force = calc_force(boundary, predicted_flow[..., 3:4])
predicted_drag_x = tf.reduce_sum(predicted_force[..., 0],
axis=[1, 2, 3])
predicted_drag_y = tf.reduce_sum(predicted_force[..., 1],
axis=[1, 2, 3])
predicted_drag_z = tf.reduce_sum(predicted_force[..., 2],
axis=[1, 2, 3])
true_force = calc_force(boundary, true_flow[..., 3:4])
true_drag_x = tf.reduce_sum(true_force[..., 0], axis=[1, 2, 3])
true_drag_y = tf.reduce_sum(true_force[..., 1], axis=[1, 2, 3])
true_drag_z = tf.reduce_sum(true_force[..., 2], axis=[1, 2, 3])
# predicted max vel
predicted_max_vel_x = tf.reduce_max(predicted_flow[..., 0],
axis=[1, 2, 3])
predicted_max_vel_y = tf.reduce_max(predicted_flow[..., 1],
axis=[1, 2, 3])
predicted_max_vel_z = tf.reduce_max(predicted_flow[..., 2],
axis=[1, 2, 3])
true_max_vel_x = tf.reduce_max(true_flow[..., 0], axis=[1, 2, 3])
true_max_vel_y = tf.reduce_max(true_flow[..., 1], axis=[1, 2, 3])
true_max_vel_z = tf.reduce_max(true_flow[..., 2], axis=[1, 2, 3])
# Restore for eval
init = tf.global_variables_initializer()
sess.run(init)
variables_to_restore = tf.all_variables()
variables_to_restore_flow = [
variable for i, variable in enumerate(variables_to_restore)
if "flow_network" in variable.name[:variable.name.index(':')]
]
saver = tf.train.Saver(variables_to_restore_flow)
ckpt = tf.train.get_checkpoint_state(FLOW_DIR)
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = 1
graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
# make vtm dataset
dataset = Sailfish_data("../../data/",
size=FLAGS.obj_size,
dim=FLAGS.dims)
dataset.parse_data()
# store drag data
p_drag_x_data = []
t_drag_x_data = []
p_drag_y_data = []
t_drag_y_data = []
p_drag_z_data = []
t_drag_z_data = []
p_max_vel_x_data = []
t_max_vel_x_data = []
p_max_vel_y_data = []
t_max_vel_y_data = []
p_max_vel_z_data = []
t_max_vel_z_data = []
#for run in filenames:
for i in tqdm(xrange(80)):
# read in boundary
batch_boundary, batch_flow = dataset.minibatch(
train=False,
batch_size=batch_size,
signed_distance_function=FLAGS.sdf)
# calc flow
p_drag_x, t_drag_x, p_drag_y, t_drag_y, p_drag_z, t_drag_z, p_max_vel_x, t_max_vel_x, p_max_vel_y, t_max_vel_y, p_max_vel_z, t_max_vel_z = sess.run(
[
predicted_drag_x, true_drag_x, predicted_drag_y,
true_drag_y, predicted_drag_z, true_drag_z,
predicted_max_vel_x, true_max_vel_x, predicted_max_vel_y,
true_max_vel_y, predicted_max_vel_z, true_max_vel_z
],
feed_dict={
boundary: batch_boundary,
true_flow: batch_flow
})
p_drag_x_data.append(p_drag_x)
t_drag_x_data.append(t_drag_x)
p_drag_y_data.append(p_drag_y)
t_drag_y_data.append(t_drag_y)
p_drag_z_data.append(p_drag_z)
t_drag_z_data.append(t_drag_z)
p_max_vel_x_data.append(p_max_vel_x)
t_max_vel_x_data.append(t_max_vel_x)
p_max_vel_y_data.append(p_max_vel_y)
t_max_vel_y_data.append(t_max_vel_y)
p_max_vel_z_data.append(p_max_vel_z)
t_max_vel_z_data.append(t_max_vel_z)
# display it
p_drag_x_data = np.concatenate(p_drag_x_data, axis=0)
t_drag_x_data = np.concatenate(t_drag_x_data, axis=0)
p_drag_y_data = np.concatenate(p_drag_y_data, axis=0)
t_drag_y_data = np.concatenate(t_drag_y_data, axis=0)
p_drag_z_data = np.concatenate(p_drag_z_data, axis=0)
t_drag_z_data = np.concatenate(t_drag_z_data, axis=0)
p_max_vel_x_data = np.concatenate(p_max_vel_x_data, axis=0)
t_max_vel_x_data = np.concatenate(t_max_vel_x_data, axis=0)
p_max_vel_y_data = np.concatenate(p_max_vel_y_data, axis=0)
t_max_vel_y_data = np.concatenate(t_max_vel_y_data, axis=0)
p_max_vel_z_data = np.concatenate(p_max_vel_z_data, axis=0)
t_max_vel_z_data = np.concatenate(t_max_vel_z_data, axis=0)
fig = plt.figure(figsize=(18, 3))
a = fig.add_subplot(1, 6, 1)
plt.scatter(p_drag_x_data, t_drag_x_data)
plt.plot(t_drag_x_data, t_drag_x_data, color="red")
plt.title("X Force")
a = fig.add_subplot(1, 6, 2)
plt.scatter(p_drag_y_data, t_drag_y_data)
plt.plot(t_drag_y_data, t_drag_y_data, color="red")
plt.title("Y Force")
a = fig.add_subplot(1, 6, 3)
plt.scatter(p_drag_z_data, t_drag_z_data)
plt.plot(t_drag_z_data, t_drag_z_data, color="red")
plt.title("Z Force")
a = fig.add_subplot(1, 6, 4)
plt.scatter(p_max_vel_x_data, t_max_vel_x_data)
plt.plot(t_max_vel_x_data, t_max_vel_x_data, color="red")
plt.title("Max X Velocity")
a = fig.add_subplot(1, 6, 5)
plt.scatter(p_max_vel_y_data, t_max_vel_y_data)
plt.plot(t_max_vel_y_data, t_max_vel_y_data, color="red")
plt.title("Max Y Velocity")
a = fig.add_subplot(1, 6, 6)
plt.scatter(p_max_vel_z_data, t_max_vel_z_data)
plt.plot(t_max_vel_z_data, t_max_vel_z_data, color="red")
plt.title("Max Z Velocity")
plt.savefig("./figs/flow_accuracy_3d.jpeg")
plt.show()
def evaluate():
"""Run Eval once.
Args:
saver: Saver.
summary_writer: Summary writer.
top_k_op: Top K op.
summary_op: Summary op.
"""
num_angles = 4
max_angle = 0.2
min_angle = -0.1
set_params = np.array(num_angles*[FLAGS.nr_boundary_params*[0.0]])
set_params[:,:] = 0.0
set_params_pos = np.array(num_angles*[FLAGS.nr_boundary_params*[0.0]])
set_params_pos[:,:] = 1.0
for i in xrange(num_angles):
set_params[i,0] = -i
set_params[:,0] = ((max_angle - min_angle) * (set_params[:,0]/num_angles)) - min_angle
set_params[:,1] = 0.5
set_params[:,2] = 1.0
set_params[:,-1] = 0.0
set_params_pos[:,0] = 0.0 # set angle to 0.0
set_params_pos[:,1] = 0.0 # set n_1 to .5
set_params_pos[:,2] = 0.0 # set n_2 to 1.0
set_params_pos[:,-1] = 0.0 # set tail hieght to 0.0
with tf.Graph().as_default():
# Make image placeholder
params_op, params_op_init, params_op_set, squeeze_loss = flow_net.inputs_boundary_learn(batch_size, set_params=set_params, set_params_pos=set_params_pos, noise_std=0.001)
# Make boundary
boundary = flow_net.inference_boundary(batch_size*set_params.shape[0], FLAGS.dims*[FLAGS.obj_size], params_op, full_shape=shape)
sharp_boundary = tf.round(boundary)
# predict steady flow on boundary
predicted_flow = flow_net.inference_network(boundary)
predicted_sharp_flow = flow_net.inference_network(sharp_boundary)
# quantities to optimize
force = calc_force(boundary, predicted_flow[:,:,:,2:3])
sharp_force = calc_force(sharp_boundary, predicted_sharp_flow[:,:,:,2:3])
drag_x = tf.reduce_sum(force[:,:,:,0], axis=[1,2])/batch_size
drag_y = tf.reduce_sum(force[:,:,:,1], axis=[1,2])/batch_size
sharp_drag_x = tf.reduce_sum(sharp_force[:,:,:,0], axis=[1,2])/batch_size
sharp_drag_y = tf.reduce_sum(sharp_force[:,:,:,1], axis=[1,2])/batch_size
drag_lift_ratio = (drag_x/drag_y)
sharp_drag_lift_ratio = (sharp_drag_x/sharp_drag_y)
# loss
loss = -tf.reduce_sum(drag_lift_ratio)
# init graph
init = tf.global_variables_initializer()
# Restore the moving average version of the learned variables for eval.
variables_to_restore = tf.all_variables()
variables_to_restore_boundary = [variable for i, variable in enumerate(variables_to_restore) if "boundary_network" in variable.name[:variable.name.index(':')]]
variables_to_restore_flow = [variable for i, variable in enumerate(variables_to_restore) if "flow_network" in variable.name[:variable.name.index(':')]]
saver_boundary = tf.train.Saver(variables_to_restore_boundary)
saver_flow = tf.train.Saver(variables_to_restore_flow)
# start ses and init
sess = tf.Session()
sess.run(init)
ckpt_boundary = tf.train.get_checkpoint_state(BOUNDARY_DIR)
ckpt_flow = tf.train.get_checkpoint_state(FLOW_DIR)
saver_boundary.restore(sess, ckpt_boundary.model_checkpoint_path)
saver_flow.restore(sess, ckpt_flow.model_checkpoint_path)
graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
params_np = (np.random.rand(1,FLAGS.nr_boundary_params) - .5)
sess.run(params_op_init, feed_dict={params_op_set: params_np})
run_time = FLAGS.boundary_learn_steps
# make store vectors for values
plot_error = np.zeros((run_time))
plot_drag_y = np.zeros((run_time))
plot_drag_x = np.zeros((run_time))
# make store dir
os.system("mkdir ./figs/boundary_learn_image_store")
# simulated annealing params
temp = 0.1
param_old = params_np
param_new = distort_param(params_np, std)
fittness_old = sess.run(loss)
fittness_new = 0.0
for i in tqdm(xrange(run_time)):
sess.run(params_op_init, feed_dict={params_op_set: param_new})
fittness_new = sess.run(loss)
print(fittness_new)
print(fittness_old)
param_old, fittness_old, temp = simulated_annealing_step(param_old, fittness_old, param_new, fittness_new, temp=temp)
print(temp)
param_new = distort_param(param_old, std)
l, d_y, d_x, p_o = sess.run([loss, sharp_drag_y, sharp_drag_x, params_op], feed_dict={})
plot_error[i] = np.sum(l)
plot_drag_x[i] = np.sum(d_x[2])
plot_drag_y[i] = np.sum(d_y[2])
if (i+1) % 400 == 0:
# make video with opencv
velocity_norm_g, boundary_g = sess.run([predicted_sharp_flow, sharp_boundary],feed_dict={})
d_y, d_x, l_c, p_o = sess.run([sharp_drag_y, sharp_drag_x, sharp_drag_lift_ratio, params_op], feed_dict={})
#velocity_norm_g, boundary_g = sess.run([force, boundary],feed_dict={})
#sflow_plot = np.concatenate([ 5.0*velocity_norm_g[0], boundary_g[0]], axis=1)
#sflow_plot = np.uint8(grey_to_short_rainbow(sflow_plot))
#sflow_plot = cv2.applyColorMap(sflow_plot
#video.write(sflow_plot)
# save plot image to make video
velocity_norm_g = velocity_norm_g[2,:,:,2]
boundary_g = boundary_g[2,:,:,0]
fig = plt.figure()
fig.set_size_inches(15.5, 7.5)
a = fig.add_subplot(1,5,1)
plt.imshow(velocity_norm_g)
a = fig.add_subplot(1,5,2)
plt.imshow(boundary_g)
a = fig.add_subplot(1,5,3)
plt.plot(plot_error, label="lift/drag")
plt.xlabel("step")
plt.legend()
a = fig.add_subplot(1,5,4)
plt.plot(plot_drag_x, label="drag_x")
plt.plot(plot_drag_y, label="drag_y")
plt.xlabel("step")
plt.legend()
a = fig.add_subplot(1,5,5)
plt.plot(set_params[:,0], l_c, 'bo', label="lift/drag")
plt.xlabel("angle of attack")
plt.xlim(min(set_params[:,0])-0.03, max(set_params[:,0])+0.03)
#plt.legend()
plt.suptitle("Using Gradient Decent")
plt.savefig("./figs/boundary_learn_image_store/plot_" + str(i).zfill(5) + ".png")
if run_time - i <= 100:
plt.savefig("./figs/" + FLAGS.boundary_learn_loss + "_plot.png")
#plt.show()
plt.close(fig)
# close cv video
video.release()
cv2.destroyAllWindows()
# generate video of plots
os.system("rm ./figs/" + FLAGS.boundary_learn_loss + "_plot_video.mp4")
os.system("cat ./figs/boundary_learn_image_store/*.png | ffmpeg -f image2pipe -r 30 -vcodec png -i - -vcodec libx264 ./figs/" + FLAGS.boundary_learn_loss + "_plot_video.mp4")
os.system("rm -r ./figs/boundary_learn_image_store")
def evaluate():
"""Run Eval once.
Args:
saver: Saver.
summary_writer: Summary writer.
top_k_op: Top K op.
summary_op: Summary op.
"""
num_angles = 1
max_angle = 0.0
min_angle = 0.0
set_params = np.array(num_angles * [FLAGS.nr_boundary_params * [0.0]])
set_params[:, :] = 0.0
set_params_pos = np.array(num_angles * [FLAGS.nr_boundary_params * [0.0]])
set_params_pos[:, :] = 1.0
for i in xrange(num_angles):
set_params[i, 0] = -i
set_params[:, 0] = ((max_angle - min_angle) *
(set_params[:, 0] / num_angles)) - min_angle
set_params[:, 1] = 0.5
set_params[:, 2] = 1.0
set_params[:, -1] = 0.0
set_params_pos[:, 0] = 0.0 # set angle to 0.0
set_params_pos[:, 1] = 0.0 # set n_1 to .5
set_params_pos[:, 2] = 0.0 # set n_2 to 1.0
set_params_pos[:, -1] = 0.0 # set tail hieght to 0.0
with tf.Graph().as_default():
# Make vector placeholder
params_op, params_op_init, params_op_set, squeeze_loss = flow_net.inputs_boundary_learn(
batch_size, set_params=set_params, set_params_pos=set_params_pos)
# Compute boundary
boundary = flow_net.inference_boundary(batch_size * num_angles,
FLAGS.dims * [FLAGS.obj_size],
params_op,
full_shape=shape)
boundary = tf.round(boundary)
# Build a Graph that computes the logits predictions from the
# inference model.
predicted_flow = flow_net.inference_network(boundary)
# quantities to optimize
force = calc_force(boundary, predicted_flow[:, :, :, 2:3])
drag_x = tf.reduce_sum(force[:, :, :, 0], axis=[0, 1, 2])
drag_y = tf.reduce_sum(force[:, :, :, 1], axis=[0, 1, 2])
#drag_ratio = -(drag_x/drag_y)
#drag_ratio = (drag_y/drag_x)
drag_ratio = drag_y
# init graph
init = tf.global_variables_initializer()
# Restore the moving average version of the learned variables for eval.
variables_to_restore = tf.all_variables()
variables_to_restore_boundary = [
variable for i, variable in enumerate(variables_to_restore)
if "boundary_network" in variable.name[:variable.name.index(':')]
]
variables_to_restore_flow = [
variable for i, variable in enumerate(variables_to_restore)
if "flow_network" in variable.name[:variable.name.index(':')]
]
saver_boundary = tf.train.Saver(variables_to_restore_boundary)
saver_flow = tf.train.Saver(variables_to_restore_flow)
# start ses and init
sess = tf.Session()
sess.run(init)
ckpt_boundary = tf.train.get_checkpoint_state(BOUNDARY_DIR)
ckpt_flow = tf.train.get_checkpoint_state(FLOW_DIR)
saver_boundary.restore(sess, ckpt_boundary.model_checkpoint_path)
saver_flow.restore(sess, ckpt_flow.model_checkpoint_path)
graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
run_time = FLAGS.boundary_learn_steps
# make store vectors for values
best_boundary = []
max_d_ratio = np.zeros((run_time))
iteration = np.arange(run_time) * batch_size
d_ratio_store = []
# make store dir
os.system("mkdir ./figs/boundary_random_image_store")
for i in tqdm(xrange(run_time)):
input_batch = (
np.random.rand(batch_size, FLAGS.nr_boundary_params) - .5)
sess.run(params_op_init, feed_dict={params_op_set: input_batch})
d_ratio, boundary_batch = sess.run([drag_ratio, boundary])
#plt.imshow(1.0*sess.run(force)[0,:,:,1] - boundary_batch[0,3:-3,3:-3,0])
#plt.imshow(1.0*sess.run(force)[0,:,:,1])
#plt.show()
d_ratio_store.append(d_ratio)
if np.max(np.array(d_ratio_store)) <= d_ratio:
best_boundary = boundary_batch
best_input = input_batch
max_d_ratio[i] = np.max(np.array(d_ratio_store))
if i % 100 == 0:
# make video with opencv
sess.run(params_op_init, feed_dict={params_op_set: best_input})
velocity_norm_g = sess.run(predicted_flow)
#velocity_norm_g, boundary_g = sess.run([force, boundary],feed_dict={})
#sflow_plot = np.concatenate([ 5.0*velocity_norm_g[0], boundary_g[0]], axis=1)
#sflow_plot = np.uint8(grey_to_short_rainbow(sflow_plot))
#sflow_plot = cv2.applyColorMap(sflow_plot
#video.write(sflow_plot)
# save plot image to make video
velocity_norm_g = velocity_norm_g[fig_pos, :, :, 2]
boundary_g = best_boundary[fig_pos, :, :, 0]
fig = plt.figure()
fig.set_size_inches(25.5, 7.5)
a = fig.add_subplot(1, 4, 1)
plt.imshow(velocity_norm_g)
a = fig.add_subplot(1, 4, 2)
plt.imshow(boundary_g)
a = fig.add_subplot(1, 4, 3)
plt.plot(iteration, max_d_ratio, label="best lift/drag")
plt.legend(loc=4)
a = fig.add_subplot(1, 4, 4)
# the histogram of the data
n, bins, patches = plt.hist(np.array(d_ratio_store),
50,
normed=1,
facecolor='green')
#plt.hist(d_ratio_store, 10, normed=1, facecolor='green')
plt.xlabel("lift/drag")
plt.ylabel("frequency")
plt.legend()
plt.suptitle("Using Random Search")
plt.savefig("./figs/boundary_random_image_store/plot_" +
str(i).zfill(5) + ".png")
if run_time - i <= 100:
plt.savefig("./figs/" + FLAGS.boundary_learn_loss +
"_plot.png")
#plt.show()
plt.close(fig)
# close cv video
video.release()
cv2.destroyAllWindows()
# generate video of plots
os.system("rm ./figs/random_plot_video.mp4")
os.system(
"cat ./figs/boundary_random_image_store/*.png | ffmpeg -f image2pipe -r 30 -vcodec png -i - -vcodec libx264 ./figs/random_plot_video.mp4"
)
os.system("rm -r ./figs/boundary_random_image_store")
def evaluate():
"""Run Eval once.
Args:
saver: Saver.
summary_writer: Summary writer.
top_k_op: Top K op.
summary_op: Summary op.
"""
with tf.Graph().as_default():
# Make image placeholder
params_op, params_op_init, params_op_set, squeeze_loss = flow_net.inputs_boundary_learn(
batch_size, network_type="heat", noise_std=0.01)
# Make boundary
boundary = flow_net.inference_boundary(batch_size, 2 * [128],
params_op)
# predict steady flow on boundary
predicted_heat = flow_net.inference_network(boundary,
network_type="heat")
# loss
loss = tf.reduce_max(predicted_heat)
loss += squeeze_loss
# train_op
variables_to_train = tf.all_variables()
variables_to_train = [
variable for i, variable in enumerate(variables_to_train)
if "params" in variable.name[:variable.name.index(':')]
]
train_step = flow_net.train(loss,
FLAGS.boundary_learn_lr,
train_type="boundary_params",
variables=variables_to_train)
# init graph
init = tf.global_variables_initializer()
# Restore the moving average version of the learned variables for eval.
variables_to_restore = tf.all_variables()
variables_to_restore_boundary = [
variable for i, variable in enumerate(variables_to_restore)
if "boundary_network" in variable.name[:variable.name.index(':')]
]
variables_to_restore_heat = [
variable for i, variable in enumerate(variables_to_restore)
if "heat_network" in variable.name[:variable.name.index(':')]
]
saver_boundary = tf.train.Saver(variables_to_restore_boundary)
saver_heat = tf.train.Saver(variables_to_restore_heat)
# start ses and init
sess = tf.Session()
sess.run(init)
ckpt_boundary = tf.train.get_checkpoint_state(BOUNDARY_DIR)
ckpt_heat = tf.train.get_checkpoint_state(FLOW_DIR)
saver_boundary.restore(sess, ckpt_boundary.model_checkpoint_path)
saver_heat.restore(sess, ckpt_heat.model_checkpoint_path)
# make graph
graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
# total run time
run_time = FLAGS.boundary_learn_steps
# use same start for comparison
start_params_np = (
np.random.rand(batch_size, FLAGS.nr_boundary_params) - .5) / 2.0
# gradient decent
plot_error_gradient_decent = np.zeros((num_runs, run_time))
for sim in tqdm(xrange(num_runs)):
sess.run(params_op_init,
feed_dict={params_op_set: start_params_np})
for i in tqdm(xrange(run_time)):
plot_error_gradient_decent[sim, i] = run_heat_sink_simulation(
np.minimum(np.maximum(sess.run(params_op) - 0.5, -.5),
0.5))
sess.run(train_step, feed_dict={})
gradient_descent_boundary = heat_sink_boundary_2d(
sess.run(params_op)[0], [128, 128])
# simulated annealing
plot_error_simulated_annealing = np.zeros(
(len(temps), num_runs, run_time))
for t in tqdm(xrange(len(temps))):
for sim in tqdm(xrange(num_runs)):
temp = temps[t]
param_old = start_params_np
fittness_old = run_heat_sink_simulation(param_old)
param_new = distort_param(start_params_np, std)
fittness_new = 0.0
for i in tqdm(xrange(run_time)):
plot_error_simulated_annealing[t, sim, i] = fittness_old
fittness_new = run_heat_sink_simulation(param_new)
param_old, fittness_old, temp = simulated_annealing_step(
param_old,
fittness_old,
param_new,
fittness_new,
temp=temp)
param_new = distort_param(param_old, std)
simulated_annealing_boundary = heat_sink_boundary_2d(
param_old[0] + .5, [128, 128])
x = np.arange(run_time)
plot_error_gradient_decent_mean, plot_error_gradient_decent_std = calc_mean_and_std(
plot_error_gradient_decent)
fig = plt.figure()
fig.set_size_inches(10, 5)
a = fig.add_subplot(1, 2, 1)
plt.imshow((simulated_annealing_boundary -
.5 * gradient_descent_boundary)[:, :, 0])
plt.title("Difference in Heat Sink Design", fontsize=16)
a = fig.add_subplot(1, 2, 2)
plt.errorbar(x,
plot_error_gradient_decent_mean,
yerr=plot_error_gradient_decent_std,
lw=1.0,
label="Gradient Descent")
for t in tqdm(xrange(len(temps))):
plot_error_simulated_annealing_mean, plot_error_simulated_annealing_std = calc_mean_and_std(
plot_error_simulated_annealing[t])
plt.errorbar(x,
plot_error_simulated_annealing_mean,
yerr=plot_error_simulated_annealing_std,
lw=1.0,
label="Simulated Annealing temp = " + str(temps[t]))
plt.xlabel('Step')
plt.ylabel('Temp at Source')
plt.suptitle("Gradient Descent vs Simulated Annealing", fontsize=20)
plt.legend(loc="upper_left")
plt.savefig("./figs/heat_learn_comparison.pdf")
plt.show()
def evaluate():
"""Run Eval once.
Args:
saver: Saver.
summary_writer: Summary writer.
top_k_op: Top K op.
summary_op: Summary op.
"""
num_angles = 9
max_angle = 0.30
min_angle = -0.10
set_params = np.array(num_angles * [FLAGS.nr_boundary_params * [0.0]])
set_params[:, :] = 0.0
set_params_pos = np.array(num_angles * [FLAGS.nr_boundary_params * [0.0]])
set_params_pos[:, :] = 1.0
for i in xrange(num_angles):
set_params[i, 0] = -i
set_params[:, 0] = ((max_angle - min_angle) *
(set_params[:, 0] / (num_angles - 1))) - min_angle
set_params[:, 1] = 0.5
set_params[:, 2] = 1.0
set_params[:, -1] = 0.0
set_params_pos[:, 0] = 0.0 # set angle to 0.0
set_params_pos[:, 1] = 0.0 # set n_1 to .5
set_params_pos[:, 2] = 0.0 # set n_2 to 1.0
set_params_pos[:, -1] = 0.0 # set tail hieght to 0.0
with tf.Graph().as_default():
# Make image placeholder
params_op, params_op_init, params_op_set, squeeze_loss = flow_net.inputs_boundary_learn(
batch_size,
set_params=set_params,
set_params_pos=set_params_pos,
noise_std=0.01)
# Make boundary
boundary = flow_net.inference_boundary(batch_size *
set_params.shape[0],
FLAGS.dims * [FLAGS.obj_size],
params_op,
full_shape=shape)
#sharp_boundary =
# predict steady flow on boundary
predicted_flow = flow_net.inference_network(boundary)
# quantities to optimize
force = calc_force(boundary, predicted_flow[:, :, :, 2:3])
drag_x = tf.reduce_sum(force[:, :, :, 0], axis=[1, 2]) / batch_size
drag_y = tf.reduce_sum(force[:, :, :, 1], axis=[1, 2]) / batch_size
drag_lift_ratio = -(drag_y / drag_x)
# loss
loss = -tf.reduce_sum(drag_lift_ratio)
loss += squeeze_loss
# train_op
variables_to_train = tf.all_variables()
variables_to_train = [
variable for i, variable in enumerate(variables_to_train)
if "params" in variable.name[:variable.name.index(':')]
]
train_step = flow_net.train(loss,
FLAGS.boundary_learn_lr,
train_type="boundary_params",
variables=variables_to_train)
# init graph
init = tf.global_variables_initializer()
# Restore the moving average version of the learned variables for eval.
variables_to_restore = tf.all_variables()
variables_to_restore_boundary = [
variable for i, variable in enumerate(variables_to_restore)
if "boundary_network" in variable.name[:variable.name.index(':')]
]
variables_to_restore_flow = [
variable for i, variable in enumerate(variables_to_restore)
if "flow_network" in variable.name[:variable.name.index(':')]
]
saver_boundary = tf.train.Saver(variables_to_restore_boundary)
saver_flow = tf.train.Saver(variables_to_restore_flow)
# start ses and init
sess = tf.Session()
sess.run(init)
ckpt_boundary = tf.train.get_checkpoint_state(BOUNDARY_DIR)
ckpt_flow = tf.train.get_checkpoint_state(FLOW_DIR)
saver_boundary.restore(sess, ckpt_boundary.model_checkpoint_path)
saver_flow.restore(sess, ckpt_flow.model_checkpoint_path)
# make graph
graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
# total run time
run_time = FLAGS.boundary_learn_steps
# use same start for comparison
start_params_np = (
np.random.rand(batch_size, FLAGS.nr_boundary_params) - .5)
# gradient decent
plot_error_gradient_decent = np.zeros((num_runs, run_time))
for sim in tqdm(xrange(num_runs)):
sess.run(params_op_init,
feed_dict={params_op_set: start_params_np})
for i in tqdm(xrange(run_time)):
l, _ = sess.run([loss, train_step], feed_dict={})
if i == run_time - 1:
d_l_ratio = sess.run(drag_lift_ratio)
plot_error_gradient_decent[sim, i] = np.sum(l)
# simulated annealing
plot_error_simulated_annealing = np.zeros(
(len(temps), num_runs, run_time))
for t in tqdm(xrange(len(temps))):
for sim in tqdm(xrange(num_runs)):
sess.run(params_op_init,
feed_dict={params_op_set: start_params_np})
temp = temps[t]
param_old = start_params_np
param_new = distort_param(start_params_np, std)
fittness_old = sess.run(loss)
fittness_new = 0.0
for i in tqdm(xrange(run_time)):
sess.run(params_op_init,
feed_dict={params_op_set: param_new})
fittness_new = sess.run(loss)
param_old, fittness_old, temp = simulated_annealing_step(
param_old,
fittness_old,
param_new,
fittness_new,
temp=temp)
param_new = distort_param(param_old, std)
plot_error_simulated_annealing[t, sim, i] = fittness_old
x = np.arange(run_time)
fig = plt.figure()
fig.set_size_inches(5, 5)
plot_error_gradient_decent_mean, plot_error_gradient_decent_std = calc_mean_and_std(
plot_error_gradient_decent)
plt.errorbar(x,
plot_error_gradient_decent_mean,
yerr=plot_error_gradient_decent_std,
lw=1.0,
label="Gradient Descent")
for t in tqdm(xrange(len(temps))):
plot_error_simulated_annealing_mean, plot_error_simulated_annealing_std = calc_mean_and_std(
plot_error_simulated_annealing[t])
#plt.errorbar(x, plot_error_simulated_annealing_mean, yerr=plot_error_simulated_annealing_std, c='g', lw=1.0, label="Simulated Annealing temp = " + str(temps[t]))
plt.errorbar(x,
plot_error_simulated_annealing_mean,
yerr=plot_error_simulated_annealing_std,
lw=1.0,
label="Simulated Annealing temp = " + str(temps[t]))
plt.xlabel('Step')
plt.ylabel('Loss')
plt.title("Optimization", fontsize=20)
plt.legend(loc="upper_left")
plt.savefig("./figs/learn_comparison.pdf")
plt.show()
def evaluate():
"""Run Eval once.
Args:
saver: Saver.
summary_writer: Summary writer.
top_k_op: Top K op.
summary_op: Summary op.
"""
num_angles = 9
max_angle = 0.30
min_angle = -0.10
set_params = np.array(num_angles * [FLAGS.nr_boundary_params * [0.0]])
set_params[:, :] = 0.0
set_params_pos = np.array(num_angles * [FLAGS.nr_boundary_params * [0.0]])
set_params_pos[:, :] = 1.0
for i in xrange(num_angles):
set_params[i, 0] = -i
set_params[:, 0] = ((max_angle - min_angle) *
(set_params[:, 0] / (num_angles - 1))) - min_angle
set_params[:, 1] = 0.5
set_params[:, 2] = 1.0
set_params[:, -1] = 0.0
set_params_pos[:, 0] = 0.0 # set angle to 0.0
set_params_pos[:, 1] = 0.0 # set n_1 to .5
set_params_pos[:, 2] = 0.0 # set n_2 to 1.0
set_params_pos[:, -1] = 0.0 # set tail hieght to 0.0
with tf.Graph().as_default():
# Make image placeholder
params_op, params_op_init, params_op_set, squeeze_loss = flow_net.inputs_boundary_learn(
batch_size,
set_params=set_params,
set_params_pos=set_params_pos,
noise_std=0.01)
# Make placeholder for flow computed by lattice boltzmann solver
solver_boundary, solver_flow = flow_net.inputs_flow(
1, shape, FLAGS.dims)
sharp_boundary, blaa = flow_net.inputs_flow(
batch_size * set_params.shape[0], shape, FLAGS.dims)
# Make boundary
boundary = flow_net.inference_boundary(batch_size *
set_params.shape[0],
FLAGS.dims * [FLAGS.obj_size],
params_op,
full_shape=shape)
# predict steady flow on boundary
predicted_flow = flow_net.inference_network(boundary,
network_type="flow",
keep_prob=FLAGS.keep_prob)
sharp_predicted_flow = flow_net.inference_network(
sharp_boundary, network_type="flow", keep_prob=FLAGS.keep_prob)
# quantities to optimize
force = calc_force(boundary, predicted_flow[..., -1:])
sharp_force = calc_force(sharp_boundary, sharp_predicted_flow[...,
-1:])
solver_force = calc_force(solver_boundary, solver_flow[..., -1:])
drag_x = tf.reduce_sum(force[..., 0], axis=[1, 2]) / batch_size
drag_y = tf.reduce_sum(force[..., 1], axis=[1, 2]) / batch_size
sharp_drag_x = tf.reduce_sum(sharp_force[..., 0], axis=[1, 2
]) / batch_size
sharp_drag_y = tf.reduce_sum(sharp_force[..., 1], axis=[1, 2
]) / batch_size
solver_drag_x = tf.reduce_sum(solver_force[..., 0],
axis=[1, 2]) / batch_size
solver_drag_y = tf.reduce_sum(solver_force[..., 1],
axis=[1, 2]) / batch_size
drag_lift_ratio = -(drag_y / drag_x)
sharp_drag_lift_ratio = -(sharp_drag_y / sharp_drag_x)
solver_drag_lift_ratio = -(solver_drag_y / solver_drag_x)
# loss
loss = -tf.reduce_sum(drag_lift_ratio)
#loss = -drag_y + drag_x
#loss = -tf.reduce_sum(drag_x)
loss += squeeze_loss
# train_op
variables_to_train = tf.all_variables()
variables_to_train = [
variable for i, variable in enumerate(variables_to_train)
if "params" in variable.name[:variable.name.index(':')]
]
train_step = flow_net.train(loss,
FLAGS.boundary_learn_lr,
train_type="boundary_params",
variables=variables_to_train)
# init graph
init = tf.global_variables_initializer()
# Restore the moving average version of the learned variables for eval.
variables_to_restore = tf.all_variables()
variables_to_restore_boundary = [
variable for i, variable in enumerate(variables_to_restore)
if "boundary_network" in variable.name[:variable.name.index(':')]
]
variables_to_restore_flow = [
variable for i, variable in enumerate(variables_to_restore)
if "flow_network" in variable.name[:variable.name.index(':')]
]
saver_boundary = tf.train.Saver(variables_to_restore_boundary)
saver_flow = tf.train.Saver(variables_to_restore_flow)
# start ses and init
sess = tf.Session()
sess.run(init)
ckpt_boundary = tf.train.get_checkpoint_state(BOUNDARY_DIR)
ckpt_flow = tf.train.get_checkpoint_state(FLOW_DIR)
saver_boundary.restore(sess, ckpt_boundary.model_checkpoint_path)
saver_flow.restore(sess, ckpt_flow.model_checkpoint_path)
graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
params_np = (np.random.rand(1, FLAGS.nr_boundary_params) - .5)
#params_np = np.zeros((1,FLAGS.nr_boundary_params-1))
sess.run(params_op_init, feed_dict={params_op_set: params_np})
run_time = FLAGS.boundary_learn_steps
# make store vectors for values
plot_error = np.zeros((run_time))
plot_drag_y = np.zeros((run_time))
plot_drag_x = np.zeros((run_time))
# make store dir
os.system("mkdir ./figs/boundary_learn_image_store")
for i in tqdm(xrange(run_time)):
l, _, d_y, d_x = sess.run([loss, train_step, drag_y, drag_x],
feed_dict={})
plot_error[i] = np.sum(l)
plot_drag_x[i] = np.sum(d_x[fig_pos])
plot_drag_y[i] = np.sum(d_y[fig_pos])
if ((i + 1) % 1 == 0) or i == run_time - 1:
# make video with opencv
s_params = sess.run(params_op)
wing_boundary = []
for p in xrange(s_params.shape[0]):
wing_boundary.append(
wing_boundary_2d(
s_params[p, 0], s_params[p, 1], s_params[p, 2],
s_params[p, 3:int((FLAGS.nr_boundary_params - 4) /
2)],
s_params[p,
int((FLAGS.nr_boundary_params - 4) /
2):-1], s_params[p, -1],
FLAGS.dims * [FLAGS.obj_size]))
wing_boundary = np.stack(wing_boundary)
wing_boundary = np.pad(
wing_boundary, [[0, 0], [128, 128], [128, 128], [0, 0]],
'constant',
constant_values=0.0)
#print(sharp_boundary.get_shape())
#print(wing_boundary.shape)
p_flow, p_boundary, d_l_ratio, sharp_d_l_ratio = sess.run(
[
sharp_predicted_flow, boundary, drag_lift_ratio,
sharp_drag_lift_ratio
],
feed_dict={sharp_boundary: wing_boundary})
# save plot image to make video
p_pressure = p_flow[fig_pos, :, :, 2]
p_boundary = p_boundary[fig_pos, :, :, 0]
fig = plt.figure()
fig.set_size_inches(15, 10)
a = fig.add_subplot(2, 3, 1)
plt.imshow(p_pressure)
plt.title("Pressure", fontsize=16)
a = fig.add_subplot(2, 3, 2)
plt.imshow(p_boundary)
plt.title("Boundary", fontsize=16)
a = fig.add_subplot(2, 3, 3)
plt.plot(plot_error, label="Sum(Lift/Drag)")
plt.xlabel("Step")
plt.legend()
a = fig.add_subplot(2, 3, 4)
plt.plot(-plot_drag_x, label="Drag Angle 0")
plt.plot(plot_drag_y, label="Lift Angle 0")
plt.ylim(-1.0, np.max(plot_drag_y) + 2.0)
plt.xlabel("Step")
plt.legend()
a = fig.add_subplot(2, 3, 5)
plt.plot(-np.degrees(set_params[:, 0]),
d_l_ratio,
'bo',
label="Lift/Drag Network")
#plt.plot(-np.degrees(set_params[:,0]), sharp_d_l_ratio, 'ro', label="Lift/Drag Sharp")
#if i == run_time-1:
# solver_d_l_ratio = run_flow_solver(sess.run(params_op), solver_boundary, solver_flow, sess, solver_drag_lift_ratio)
# plt.plot(-np.degrees(set_params[:,0]), solver_d_l_ratio, 'go', label="Lift/Drag Solver")
plt.xlabel("Angle of Attack (Degrees)")
plt.xlim(
min(-np.degrees(set_params[:, 0])) - 3,
max(-np.degrees(set_params[:, 0])) + 3)
plt.ylim(np.min(d_l_ratio) - 1, np.max(d_l_ratio) + 2)
plt.legend()
plt.suptitle("2D Wing Optimization Using Gradient Descent",
fontsize=20)
plt.savefig("./figs/boundary_learn_image_store/plot_" +
str(i).zfill(5) + ".png")
if run_time - i <= 100:
plt.savefig("./figs/" + FLAGS.boundary_learn_loss +
"_plot.pdf")
if i == run_time - 1:
plt.savefig("./figs/learn_gradient_descent.pdf")
plt.show()
#plt.show()
plt.close(fig)
# generate video of plots
os.system("rm ./figs/airfoil_2d_video.mp4")
os.system(
"cat ./figs/boundary_learn_image_store/*.png | ffmpeg -f image2pipe -r 30 -vcodec png -i - -vcodec libx264 ./figs/airfoil_2d_video.mp4"
)
os.system("rm -r ./figs/boundary_learn_image_store")
def evaluate():
"""Run Eval once.
"""
with tf.Session() as sess:
# Make image placeholder
boundary, true_heat = flow_net.inputs_heat(batch_size=batch_size,
shape=shape,
dims=FLAGS.dims)
# Build a Graph that computes the logits predictions from the
# inference model.
predicted_heat = flow_net.inference_network(boundary,
network_type="heat",
keep_prob=FLAGS.keep_prob)
# predict max heat
predicted_max_heat = tf.reduce_max(predicted_heat, axis=[1, 2, 3])
true_max_heat = tf.reduce_max(true_heat, axis=[1, 2, 3])
# Restore for eval
init = tf.global_variables_initializer()
sess.run(init)
variables_to_restore = tf.all_variables()
variables_to_restore_heat = [
variable for i, variable in enumerate(variables_to_restore)
if "heat_network" in variable.name[:variable.name.index(':')]
]
saver = tf.train.Saver(variables_to_restore_heat)
ckpt = tf.train.get_checkpoint_state(FLOW_DIR)
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = 1
graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
# make vtm dataset
dataset = Heat_Sink_data("../../data/",
size=FLAGS.obj_size,
dim=FLAGS.dims)
dataset.parse_data()
# store drag data
p_max_heat_data = []
t_max_heat_data = []
#for run in filenames:
for i in tqdm(xrange(99)):
# read in boundary
batch_boundary, batch_heat = dataset.minibatch(
train=False, batch_size=batch_size)
# calc heat
p_max_heat, t_max_heat = sess.run(
[predicted_max_heat, true_max_heat],
feed_dict={
boundary: batch_boundary,
true_heat: batch_heat
})
p_max_heat_data.append(p_max_heat)
t_max_heat_data.append(t_max_heat)
# display it
p_max_heat_data = np.concatenate(p_max_heat_data, axis=0)
t_max_heat_data = np.concatenate(t_max_heat_data, axis=0)
fig = plt.figure(figsize=(5, 5))
a = fig.add_subplot(1, 1, 1)
plt.scatter(p_max_heat_data, t_max_heat_data, color="blue", s=2.5)
plt.plot(t_max_heat_data, t_max_heat_data, color="red", linewidth=0.7)
plt.xlabel("True")
plt.ylabel("Predicted")
plt.title("Temp at Source")
plt.savefig("./figs/heat_accuracy.pdf")
plt.show()
# calc average error
print(p_max_heat_data)
error = np.abs(p_max_heat_data - t_max_heat_data) / t_max_heat_data
print(error)
error = np.sum(error) / error.shape[0]
print(error)