def check_quality(result, log_path, global_step):
img_path = log_path + "imgs_{}/".format(global_step)
make_dir(img_path)
np.savetxt(img_path + "param_ground_truth.csv", result[2], delimiter=',')
# np.savetxt(img_path+"param_guessed.csv",result[3],delimiter=',')
np.savetxt(img_path + "position_ground_truth.csv",
result[3],
delimiter=',')
np.savetxt(img_path + "n_dot_view.csv",
result[4].reshape([-1, 1]),
delimiter=',')
np.savetxt(img_path + "view_dir.csv",
result[5].reshape([-1, 3]),
delimiter=',')
np.savetxt(img_path + "n.csv", result[6].reshape([-1, 3]), delimiter=',')
np.savetxt(img_path + "n_local.csv",
result[7].reshape([-1, 3]),
delimiter=',')
# np.savetxt(img_path+"param_origin.csv",result[5],delimiter=',')
# np.savetxt(img_path+"position_guessed.csv",result[5],delimiter=',')
for idx, a_gt_lumi in enumerate(result[0]):
gt_lumi_img = visualize_new(a_gt_lumi.reshape([-1]), scalerf=255.0)
guessed_lumi_img = visualize_new(np.exp(result[1][idx].reshape([-1])) -
1,
scalerf=255.0)
cv2.imwrite(img_path + "{}_gt.png".format(idx), gt_lumi_img)
cv2.imwrite(img_path + "{}_guessed.png".format(idx), guessed_lumi_img)
np.savetxt(img_path + "{}_mm.csv".format(idx),
result[2][idx].reshape([-1, 1]),
delimiter=',')
def saveWLog(self,globalStep,tmp_log_path):
W = self.sess.run(self.end_points['projection_matrix'])
W_vector = W.reshape([-1])
f = open(tmp_log_path+"W_{}.csv".format(globalStep),"w")
[f.write("{},\n".format(W_vector[i])) for i in range(W_vector.shape[0])]
f.close()
W = W.T
for i in range(W.shape[0]):
one_pattern = visualize_new(W[i]/W[i].max(),scalerf=255)#cook10240(W[i])#visualize(W[i]/W[i].max(),scalerf=255)
one_pattern = np.expand_dims(one_pattern,axis = -1)
one_pattern_pos = np.maximum(one_pattern,0.0)
one_pattern_neg = np.minimum(one_pattern,0.0)*-1
one_pattern = np.concatenate([one_pattern_neg,np.zeros(one_pattern.shape),one_pattern_pos],axis = -1)
# one_pattern /= one_pattern.max()
cv2.imwrite(tmp_log_path+"{}.png".format(i),one_pattern)
def check_quality(self,val_data):
results = self.sess.run([
self.end_points["ground_truth_lumitexels_direct"],
self.end_points["guessed_lumitexels_pd"],
self.end_points["ground_truth_slice_pd"],
self.end_points["guessed_lumitexels_ps"],
self.end_points["ground_truth_slice_ps"],
self.end_points["ground_truth_slice_length"],
self.end_points["ps_predicted"],
self.end_points["light_dir"],
self.end_points["view_dir"],
self.end_points["normal"],
self.end_points["pos"],#12
self.end_points['projection_matrix']
# self.end_points["n_local_gt"],
# self.end_points["guessed_normal"],
# self.end_points["pd_gt"],
# self.end_points["guessed_pd"],
],feed_dict={
self.end_points["input_params"]:val_data[0],
self.end_points["input_positions"]:val_data[1],
self.end_points["istraining"]:False,
self.end_points["input_rotate_angles"]:val_data[2]
})
_,img_height,img_width,_ = self.check_images.shape.as_list()
img_width = img_width//5
check_img_buffer = np.zeros([self.batch_size*5,img_height,img_width],np.float32)
for idx,a_gt_lumi in enumerate(results[0]):
# pos = results[12][idx]
# view_dir = results[10][idx] #[1,3]
# light_dir = results[9][idx] #[lumi_size, 3]
# n = results[11][idx]
tmp = visualize_new(a_gt_lumi.reshape([-1]),scalerf=255.0)
check_img_buffer[idx*5+0] = tmp
check_img_buffer[idx*5+1] = expand_img(visualize_cube_slice(results[1][idx].reshape([-1]),sample_num=8,scalerf=255.0),8,"copy_only")
check_img_buffer[idx*5+2] = expand_img(visualize_cube_slice(results[2][idx].reshape([-1]),sample_num=8,scalerf=255.0),8,"copy_only")
a = results[4][idx].reshape([-1])
tmp = visualize_cube_slice(a/a.max(),sample_num=64,scalerf=2550.0)
cv2.putText(tmp, '{:.2f}'.format(results[5][idx][0]), (128,64), cv2.FONT_HERSHEY_SIMPLEX, 1.2, 255, 2)
check_img_buffer[idx*5+4] = tmp
a = results[3][idx].reshape([-1])/a.max()
tmp = visualize_cube_slice(a,sample_num=64,scalerf=2550.0)
cv2.putText(tmp, '{:.2f}'.format(results[6][idx][0]), (128,64), cv2.FONT_HERSHEY_SIMPLEX, 1.2, 255, 2)
check_img_buffer[idx*5+3] = tmp
check_img_buffer = np.transpose(check_img_buffer,axes=[1,0,2])
check_img_buffer = np.reshape(check_img_buffer,[check_img_buffer.shape[0],self.batch_size,-1])
check_img_buffer = np.transpose(check_img_buffer,axes=[1,0,2])
check_img_buffer = np.expand_dims(check_img_buffer,axis=-1)
check_img_buffer = np.minimum(check_img_buffer,255.0)
check_img_buffer = np.maximum(check_img_buffer,0.0)
self.check_images.load(check_img_buffer.astype(np.uint8),self.sess)
# self.check_formfactor.load(check_img_ff_buffer.astype(np.uint8),self.sess)
W = results[-1]
W = W.T
pattern_buffer = np.zeros([self.measurements_length,img_height,img_width,3],np.float32)
for i in range(W.shape[0]):
one_pattern = visualize_new(W[i]/W[i].max(),scalerf=255)#cook10240(W[i])#visualize(W[i]/W[i].max(),scalerf=255)
one_pattern = np.expand_dims(one_pattern,axis = -1)
one_pattern_pos = np.maximum(one_pattern,0.0)
one_pattern_neg = np.minimum(one_pattern,0.0)*-1
pattern_buffer[i] = np.concatenate([one_pattern_neg,np.zeros(one_pattern.shape),one_pattern_pos],axis = -1)
pattern_buffer = np.minimum(pattern_buffer,255.0)
pattern_buffer = np.maximum(pattern_buffer,0.0)
self.pattern_check.load(pattern_buffer.astype(np.uint8),self.sess)
summary,global_step = self.sess.run([
self.end_points["check_summary"],
self.end_points["global_step"]
])
self.end_points["writer"].add_summary(summary,global_step)
return results
def draw_train_net(self,net_name,projectionMatrix_trainable):
'''
net_name: a string,the name of this net
'''
with tf.variable_scope(net_name,reuse=tf.AUTO_REUSE):
#[1]define input parameters
input_params = tf.placeholder(tf.float32,shape=[self.batch_size,self.parameter_len],name="input_params")
self.end_points["input_params"] = input_params
input_positions = tf.placeholder(tf.float32,shape=[self.batch_size,3],name="render_positions")
self.end_points["input_positions"] = input_positions
input_rotate_angles = tf.placeholder(tf.float32,shape=[self.batch_size,1],name="rotate_angles")
self.end_points["input_rotate_angles"] = input_rotate_angles
isTraining = tf.placeholder(tf.bool,name = "isTraining_ph")
self.end_points["istraining"] = isTraining
#[2]draw rendering net,render ground truth
renderer = tf_ggx_render(self.end_points["standard_rendering_parameters"])
self.renderer = renderer
slice_rendering_parameters = self.end_points["standard_rendering_parameters"].copy()
slice_rendering_parameters["config_dir"] = "auxiliary/tf_ggx_render_newparam/tf_ggx_render_configs_cube_slice_64x64/"
slice_rendering_parameters["lumitexel_size"] = self.slice_length_ps
renderer_slice_ps = tf_ggx_render(slice_rendering_parameters)
slice_rendering_parameters = self.end_points["standard_rendering_parameters"].copy()
slice_rendering_parameters["config_dir"] = "auxiliary/tf_ggx_render_newparam/tf_ggx_render_configs_cube_slice_8x8/"
slice_rendering_parameters["lumitexel_size"] = self.slice_length_pd
renderer_slice_pd = tf_ggx_render(slice_rendering_parameters)
rotate_theta_zero = tf.zeros([self.batch_size,1],tf.float32)
ground_truth_lumitexels_direct = renderer.draw_rendering_net(input_params,input_positions,rotate_theta_zero,"ground_truth_renderer_direct")#[batch,lightnum,1]
ground_truth_lumitexels_direct *= self.RENDER_SCALAR
self.end_points["ground_truth_lumitexels_direct"] = ground_truth_lumitexels_direct
ground_truth_slice_ps = renderer_slice_ps.draw_rendering_net(input_params,input_positions,rotate_theta_zero,"ground_truth_renderer_cube_slice_ps",pd_ps_wanted="ps_only")
ground_truth_slice_ps = ground_truth_slice_ps*self.SLICE_SCALAR
self.end_points["ground_truth_slice_ps"] = ground_truth_slice_ps
ground_truth_slice_pd = renderer_slice_pd.draw_rendering_net(input_params,input_positions,rotate_theta_zero,"ground_truth_renderer_cube_slice_pd",pd_ps_wanted="pd_only")
ground_truth_slice_pd = ground_truth_slice_pd*self.SLICE_SCALAR
self.end_points["ground_truth_slice_pd"] = ground_truth_slice_pd
# n_local,_ = renderer.get_n("ground_truth_renderer_direct")
# self.end_points["n_local_gt"] = n_local
# pd = renderer.get_pd("ground_truth_renderer_direct")
# self.end_points["pd_gt"] = pd
#[3]draw linear projection layer
with tf.variable_scope('linear_projection',reuse = tf.AUTO_REUSE):
kernel = tf.get_variable(name = 'projection_matrix',
trainable=projectionMatrix_trainable,
shape = [self.lumitexel_length,self.measurements_length],
initializer=tf.contrib.layers.xavier_initializer()
)#[lightnum,measure_ments]
self.end_points['projection_matrix_origin'] = kernel
kernel = tf.transpose(kernel,perm=[1,0])
kernel = tf.nn.l2_normalize(kernel,axis=1)
kernel = tf.transpose(kernel,perm=[1,0])
self.end_points['projection_matrix'] = kernel
measure_ments = tf.matmul(tf.squeeze(ground_truth_lumitexels_direct,axis=2),kernel)#[batch,measure_ments]
#[3.1]noiser
with tf.variable_scope("noiser"):
noise = tf.random_normal(tf.shape(measure_ments),mean = 0.0,stddev = 0.01,name="noise_generator")+1.
measure_ments_noised = tf.multiply(measure_ments,noise)
# measure_ments_noised_theta = tf.concat([measure_ments_noised,tf.sin(input_rotate_angles),tf.cos(input_rotate_angles)],axis=-1)
#[4]draw guesser net
guessed_lumitexels_pd,guessed_lumitexels_ps,ps_predicted,p_predicted = self.draw_param_guesser(measure_ments_noised,"param_guesser",renderer,isTraining)#[batch,param_length]
self.end_points["ps_predicted"] = ps_predicted
guessed_lumitexels_pd = tf.expand_dims(guessed_lumitexels_pd,axis=-1)
self.end_points["guessed_lumitexels_pd"] = guessed_lumitexels_pd
guessed_lumitexels_ps = tf.expand_dims(guessed_lumitexels_ps,axis=-1)
self.end_points["guessed_lumitexels_ps"] = tf.exp(guessed_lumitexels_ps)-1.0
predicted_normal = self.draw_normal_net(measure_ments_noised,"normal_net",renderer,isTraining)
_,label_normal = renderer.get_n("ground_truth_renderer_direct")
self.end_points["wi"] = renderer.get_wi("ground_truth_renderer_direct")
self.end_points["light_dir"] = self.end_points["wi"]
self.end_points["pos"] = input_positions
self.end_points["view_dir"] = renderer.get_compute_node("ground_truth_renderer_direct","view_dir_rotated")
self.end_points["normal"] = renderer.get_compute_node("ground_truth_renderer_direct","normal")
# self.end_points["guessed_normal"] = normal_guessed
# self.end_points["guessed_pd"] = pd_guessed
# guessed_n,guessed_t,guessed_axay,guessed_pd,guessed_ps = tf.split(guessed_params,[3,3,2,1,1],axis=-1)
# guessed_axay = tf.clip_by_value(guessed_axay,0.006,0.503)
# guessed_pd = tf.clip_by_value(guessed_pd,0.0,1.0)
# guessed_ps = tf.clip_by_value(guessed_ps,0.0,10.0)
# guessed_params = tf.concat([guessed_n,guessed_t,guessed_axay,guessed_pd,guessed_ps],axis=-1)
# self.end_points["guessed_params"] = guessed_params
# # self.end_points["guessed_position"] = guessed_position
# #[5]render guessed lumitexel
# guessed_lumitexels = renderer.draw_rendering_net(guessed_params,input_positions,"guessed_renderer")#[batch,lightnum,1]
# # guessed_lumitexels = renderer.draw_rendering_net(guessed_params,input_positions,"guessed_renderer")#[batch,lightnum,1]
# guessed_lumitexels = guessed_lumitexels*self.RENDER_SCALAR
# self.end_points["guessed_lumitexels"] = guessed_lumitexels
# _,self.end_points["n_dot_view"],self.end_points["view_dir"],self.end_points["n"],self.end_points["n_local"] = renderer.get_n_dot_view_penalty("ground_truth_renderer")
#[6]define loss
##########################
######pd loss
#########################
l2_loss_pd = tf.nn.l2_loss(guessed_lumitexels_pd-ground_truth_slice_pd,"l2_loss_pd")
self.end_points["l2_loss_pd"] = l2_loss_pd
##########################
######ps loss
#########################
#----------part 1
if self.loss_configs["use_weight"]:
print("nonono")
exit(-1)
assert self.loss_configs["use_log"] == False
ground_truth_slice_ps_tmp = tf.reshape(ground_truth_slice_ps,[self.batch_size,self.slice_length_ps])
weights = 1.0/tf.reduce_sum(tf.square(ground_truth_slice_ps_tmp),axis=-1,keepdims=True)#[batch,1]
subed = ground_truth_slice_ps-guessed_lumitexels_ps
subed = tf.reshape(subed,[self.batch_size,self.slice_length_ps])
loss_tmp = tf.reduce_sum(tf.square(subed),axis=-1,keepdims=True)*weights
l2_loss_ps = tf.reduce_sum(loss_tmp)/2.0
elif self.loss_configs["use_dot"]:
print("nonono")
exit(-1)
ground_truth_slice_ps_tmp = tf.reshape(ground_truth_slice_ps,[self.batch_size,self.slice_length_ps])
guessed_slice_ps_tmp = tf.reshape(guessed_lumitexels_ps,[self.batch_size,self.slice_length_ps])
ground_truth_slice_ps_tmp = tf.nn.l2_normalize(ground_truth_slice_ps_tmp,axis=-1)
guessed_slice_ps_tmp = tf.nn.l2_normalize(guessed_slice_ps_tmp,axis=-1)
l2_loss_ps = tf.reduce_sum(1.0-tf.reduce_sum(tf.multiply(ground_truth_slice_ps_tmp,guessed_slice_ps_tmp),axis=-1))
elif self.loss_configs["use_log"]:
print("nonono")
exit(-1)
l2_loss_ps = tf.nn.l2_loss(guessed_lumitexels_ps-tf.maximum(tf.log(1.0+ground_truth_slice_ps),0.0),"l2_loss_ps")
elif self.loss_configs["use_l2"]:
# ground_truth_slice_ps_tmp = tf.reshape(ground_truth_slice_ps,[self.batch_size,self.slice_length_ps])
# guessed_slice_ps_tmp = tf.reshape(guessed_lumitexels_ps,[self.batch_size,self.slice_length_ps])
# ground_truth_slice_ps_tmp = tf.nn.l2_normalize(ground_truth_slice_ps_tmp,axis=-1)
# alpha = 1e3
ground_truth_slice_ps_tmp = tf.log(1.0+ground_truth_slice_ps)
guessed_slice_ps_tmp = guessed_lumitexels_ps
self.end_points["guessed_lumitexels_ps_normalized"] = guessed_slice_ps_tmp
self.end_points["ground_truth_slice_ps_normalized"] = ground_truth_slice_ps_tmp
l2_loss_ps = tf.nn.l2_loss(ground_truth_slice_ps_tmp-guessed_slice_ps_tmp,"l2_loss_ps")
else:
print("[TAMER]undefined l2 loss of ps")
self.end_points["l2_loss_ps"] = l2_loss_ps
#------------part 2
tmp = tf.squeeze(ground_truth_slice_ps_tmp,axis=2)
ground_truth_slice_length = tf.sqrt(tf.reduce_sum(tf.square(tmp),axis=-1,keepdims=True))
ground_truth_slice_length = tf.log(1+ground_truth_slice_length)
self.end_points["ground_truth_slice_length"] = ground_truth_slice_length
if self.with_length_predict:
l2_loss_length = tf.nn.l2_loss(ps_predicted-ground_truth_slice_length,"l2_loss_normalized")
else:
l2_loss_length = 0.0
self.end_points["l2_loss_length"] = l2_loss_length
#-------------part 3
l2_loss_p = tf.nn.l2_loss(p_predicted-input_positions,"l2_loss_p")
self.end_points["p_loss"] = l2_loss_p
##########################
######normal loss
#########################
l2_loss_normal = tf.nn.l2_loss(predicted_normal - label_normal)
self.end_points["l2_loss_normal"] = l2_loss_normal
l2_loss = self.lambdas["pd"]*l2_loss_pd+self.lambdas["ps"]*l2_loss_ps+self.lambdas["ps_length"]*l2_loss_length+self.lambdas["normal"]*l2_loss_normal+self.lambdas["p"]*l2_loss_p
# l2_loss_normal = tf.nn.l2_loss(n_local-normal_guessed,"l2_loss_normal")
# self.end_points["l2_loss_normal"] = l2_loss_normal
# l2_loss_pd = tf.nn.l2_loss(pd-pd_guessed,"l2_loss_pd")
# self.end_points["l2_loss_pd"]=l2_loss_pd
# l2_loss_param = tf.nn.l2_loss((input_params-guessed_params)*np.array([1,1,1,0,0,0,0,0,1,0.1]),"l2_loss_param")*1e3
# l2_loss_pos = tf.nn.l2_loss((input_positions-guessed_position)*np.array([0.02,0.02,0.02]),"l2_loss_position")*1e3
# l2_loss = tf.nn.l2_loss(tf.maximum(tf.log(1.0+guessed_lumitexels),0.0)-tf.maximum(tf.log(1.0+ground_truth_lumitexels),0.0),"l2_loss")
# print(l2_loss)
# n_3d,tangent_3d,ax,ay,pd,ps = tf.split(guessed_params,[3,3,1,1,1,1],axis=1)
loss_k = 1e5
if projectionMatrix_trainable:
regularization_loss = renderer.regularizer_relu(kernel,-1.0,1.0,loss_k)
# renderer.regularizer_relu(guessed_position,-50.0,50.0,loss_k)+\
# n_dot_view_penalty+\
# renderer.regularizer_relu(theta,0.0,math.pi,loss_k)+\
else:
regularization_loss = 0
total_loss = l2_loss+regularization_loss#+l2_loss_normal+l2_loss_pd#+l2_loss_param+l2_loss_pos
# all_trainable_variables = tf.get_collection_ref(tf.GraphKeys.TRAINABLE_VARIABLES)
# for a_variable in all_trainable_variables:
# print(a_variable)
# exit()
#[7]define optimizer
global_step = tf.Variable(0, name='global_step',trainable=False)
self.end_points["global_step"] = global_step
learning_rate =tf.Variable(self.learning_rate, name='learning_rate',trainable=False)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
optimizer = tf.train.AdamOptimizer
optimizer = optimizer(learning_rate)
self.trainHandler = optimizer.minimize(total_loss,global_step=global_step)
self.end_points["optimizer"] = optimizer
#[8]define others
#[8.1] writer
train_penalty_summary = tf.summary.scalar("loss_penalty_"+self.tamer_name,regularization_loss)
train_l2_loss_summary = tf.summary.scalar("loss_l2_total_train_"+self.tamer_name,l2_loss)
train_l2_loss_pd_summary = tf.summary.scalar("loss_pd_train_"+self.tamer_name,l2_loss_pd)
train_l2_loss_pslength_summary = tf.summary.scalar("loss_pslength_train_"+self.tamer_name,l2_loss_length)
train_l2_loss_ps_summary = tf.summary.scalar("loss_ps_train_"+self.tamer_name,l2_loss_ps)
train_l2_loss_normal_summary = tf.summary.scalar("loss_normal_train_"+self.tamer_name,l2_loss_normal)
train_l2_loss_p_summary = tf.summary.scalar("loss_p_train_"+self.tamer_name,l2_loss_p)
val_l2_loss_summary = tf.summary.scalar("loss_l2_total_val_"+self.tamer_name,l2_loss)
val_l2_loss_pd_summary = tf.summary.scalar("loss_pd_val_"+self.tamer_name,l2_loss_pd)
val_l2_loss_pslength_summary = tf.summary.scalar("loss_pslength_val_"+self.tamer_name,l2_loss_length)
val_l2_loss_ps_summary = tf.summary.scalar("loss_ps_val_"+self.tamer_name,l2_loss_ps)
val_l2_loss_normal_summary = tf.summary.scalar("loss_normal_val_"+self.tamer_name,l2_loss_normal)
val_l2_loss_p_summary = tf.summary.scalar("loss_p_val_"+self.tamer_name,l2_loss_p)
# val_normal_l2_loss_summary = tf.summary.scalar("loss_normal_val_"+self.tamer_name,l2_loss_normal)
# val_pd_l2_loss_summary = tf.summary.scalar("loss_pd_val_"+self.tamer_name,l2_loss_pd)
tmp_zero_lumi = np.zeros(self.lumitexel_length,np.float32)
tmp_zero_lumi_img = visualize_new(tmp_zero_lumi.reshape([-1]))
zeros_lumis_init = tf.zeros_initializer()#tf.constant_initializer(zeros_lumis)
self.check_images = tf.get_variable(name = 'check_pic',
trainable=False,
shape = [self.batch_size,tmp_zero_lumi_img.shape[0],tmp_zero_lumi_img.shape[1]*5,1],
initializer=zeros_lumis_init,
dtype=tf.uint8
)#[lightnum,measure_ments]
self.pattern_check = tf.get_variable(name = 'patterns',
trainable=False,
shape = [self.measurements_length,tmp_zero_lumi_img.shape[0],tmp_zero_lumi_img.shape[1],3],
initializer=zeros_lumis_init,
dtype=tf.uint8
)
# self.check_formfactor = tf.get_variable(name = 'check_pic_ff',
# trainable=False,
# shape = [self.batch_size,tmp_zero_lumi_img.shape[0],tmp_zero_lumi_img.shape[1],1],
# initializer=tf.zeros_initializer(),
# dtype=tf.uint8
# )#[lightnum,measure_ments]
check_pic_summary = tf.summary.image("lumitexels",self.check_images,max_outputs=self.batch_size)
pattern_summary = tf.summary.image("patterns",self.pattern_check,max_outputs=self.measurements_length)
# check_pic_ff_summary = tf.summary.image("ff",self.check_formfactor,max_outputs=self.batch_size)
#loss of matcher
train_summary = tf.summary.merge([
train_l2_loss_summary,train_penalty_summary,train_l2_loss_pslength_summary,train_l2_loss_pd_summary,train_l2_loss_ps_summary,train_l2_loss_normal_summary,train_l2_loss_p_summary
])
val_summary = tf.summary.merge([
val_l2_loss_summary,val_l2_loss_pd_summary,val_l2_loss_pslength_summary,val_l2_loss_ps_summary,val_l2_loss_normal_summary,val_l2_loss_p_summary
])
check_summary = tf.summary.merge([
check_pic_summary,pattern_summary#,check_pic_ff_summary
])
self.end_points["train_summary"] = train_summary
self.end_points["val_summary"] = val_summary
self.end_points["check_summary"] = check_summary
#[8.2] initializer
init = tf.global_variables_initializer()
self.end_points['init'] = init
#[8.3] saver
saver = tf.train.Saver()
self.end_points['saver'] = saver
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
self.sess = tf.Session(config=config)
writer = tf.summary.FileWriter(self.logPath, tf.get_default_graph())
self.end_points['writer'] = writer
make_dir(log_measurments_dir)
make_dir(img_root)
pf = open(log_root + "ae_output_data.bin", "wb")
counter = 0
while True:
tmpData, ifValid, num = myMine.gen_batch()
if not ifValid:
break
result = myTamer.check_quality(tmpData)
result[1][:num].astype(np.float32).tofile(pf)
tmp_m = result[2].reshape(
[-1, 3, train_configs["measurements_length"]])
tmp_lumis = result[1][:num].reshape(
[-1, 3, train_configs["lumitexel_length"]])
num = tmp_m.shape[0]
for idx in range(num):
np.savetxt(log_measurments_dir + "{}_mm.csv".format(counter),
tmp_m[idx].T,
delimiter=',')
img = []
for lumi in tmp_lumis[idx]:
img.append(visualize_new(lumi))
img = np.array(img)
img = np.transpose(img, axes=[1, 2, 0])
img = img[:, :, ::-1]
cv2.imwrite(img_root + "{}.png".format(counter), img * 255.0 * 0.5)
counter += 1
# break
pf.close()
print("[TAME]training done.")
if need_dump:
tmp_guessed_params, tmp_rendered = sess.run(
[guessed_params_node, rendered_lumi_node],
feed_dict={
input_measurements:
tmp_measurements,
input_positions:
np.zeros([test_configs["batch_size"], 3],
np.float32)
})
tmp_guessed_params = tmp_guessed_params[:valids]
tmp_guessed_params.tofile(presf)
np.savetxt(f_test, tmp_guessed_params, delimiter=',')
tmp_rendered = tmp_rendered[:valids] * 5 * 1e3 / math.pi
for idx_tmp, a_lumi in enumerate(tmp_rendered):
img = visualize_new(a_lumi, scalerf=255.0)
cv2.imwrite(
log_root + "img_{}.png".format(data_ptr + idx_tmp),
img)
else:
tmp_guessed_params = sess.run(
guessed_params_node,
feed_dict={
input_measurements:
tmp_measurements,
input_positions:
np.zeros([test_configs["batch_size"], 3],
np.float32)
})
np.savetxt(f_test, tmp_guessed_params, delimiter=',')
import numpy as np
import cv2
import math
import sys
sys.path.append("../utils/")
from lumitexel_related import visualize_init, visualize_new
lumitexel_size = 24576
scaler = 5 * 1e3 / math.pi
UTILS_CONFIG_PATH = "G:/current_work/utils/"
if __name__ == "__main__":
visualize_init(UTILS_CONFIG_PATH)
thread_id = 1
parameters = {}
parameters["config_dir"] = "../tf_ggx_render/tf_ggx_render_configs_1x1/"
root = "G:/no_where/test_rendering/"
data = np.fromfile(root + "cpu_lumi.bin",
np.float32).reshape([-1, 2, lumitexel_size])
data2 = np.fromfile(root + "tf_lumi.bin",
np.float32).reshape([-1, 2, lumitexel_size])
for idx, a_lumi in enumerate(data):
cv2.imshow("img_direct", visualize_new(a_lumi[0], scalerf=scaler))
cv2.imshow("img_omega", visualize_new(a_lumi[1], scalerf=scaler))
cv2.imshow("img2_direct", visualize_new(data2[idx][0], scalerf=1))
cv2.imshow("img2_omega", visualize_new(data2[idx][1], scalerf=1))
cv2.waitKey(0)
import sys
sys.path.append("../")
from lumitexel_related import visualize_init, shrink, expand_img, visualize_new, get_visualize_idxs
import numpy as np
import cv2
lumitexel_size = 24576
block_size = 64
UITLS_PATH = "../"
if __name__ == "__main__":
visualize_init(UITLS_PATH)
idxes = get_visualize_idxs() #[24576,2]
origin_lumitexel = np.array(range(lumitexel_size), np.int32) + 1
lumitexel_img = visualize_new(origin_lumitexel).astype(np.int32)
itr_count = 0
block_num = 64 // block_size
idx_collector = []
idx_invert_collector = np.zeros(lumitexel_size, np.int32)
for x in range(block_num * 4):
for y in range(block_num * 3):
real_x = x * block_size
real_y = y * block_size
hit_point = (idxes[:, 0] == real_x) & (idxes[:, 1] == real_y)
if np.count_nonzero(hit_point) != 0:
for i in range(block_size):
import numpy as np
import cv2
import sys
sys.path.append("../utils/")
from dir_folder_and_files import make_dir
from lumitexel_related import visualize_init,visualize_new
UTILS_CONFIG_PATH = "G:/current_work/utils/"
if __name__ == "__main__":
data_root=sys.argv[1]
data_name=sys.argv[2]
visualize_init(UTILS_CONFIG_PATH)
data = np.fromfile(data_root+data_name,np.float32).reshape([-1,3,24576])
img_root = data_root+"img/"
make_dir(img_root)
for idx,apixel in enumerate(data):
r = np.expand_dims(visualize_new(apixel[0]),axis=2)
g = np.expand_dims(visualize_new(apixel[1]),axis=2)
b = np.expand_dims(visualize_new(apixel[2]),axis=2)
img = np.concatenate([b,g,r],axis=-1)
cv2.imwrite(img_root+"{}.png".format(idx),img*255)