def calc_loss_and_print(y_data, print_dir, step, N):
val_loss, orig_loss, n_batch = 0, 0, 0
for b in range(int(y_data.shape[0] / FLAGS.batch_size)):
y_batch = y_data[b * FLAGS.batch_size:(b + 1) *
FLAGS.batch_size, :, :, :]
feed_dict = {y_: y_batch}
if FLAGS.regularization > 0:
err1, err2, xeval, ypred, xtpred, ytpred = sess.run(
[cost, cost_input_output, y_l, y, y_s, y_t],
feed_dict=feed_dict)
else:
err1, err2, xeval, ypred = sess.run(
[cost, cost_input_output, y_l, y], feed_dict=feed_dict)
val_loss += err1
orig_loss += err2
n_batch += 1
batch_dir = print_dir
if y_data.shape[0] > y_batch.shape[0]:
batch_dir = '%s/batch_%03d' % (print_dir, n_batch)
if n_batch <= N or N < 0:
if not os.path.exists(batch_dir):
os.makedirs(batch_dir)
for i in range(0, y_batch.shape[0]):
# Print LDR samples
if FLAGS.print_im:
img_io.writeLDR(
np.squeeze(xeval[i]),
"%s/%06d_%03d_in.png" % (batch_dir, step, i + 1))
img_io.writeLDR(
np.squeeze(y_batch[i]),
"%s/%06d_%03d_gt.png" % (batch_dir, step, i + 1))
img_io.writeLDR(
np.squeeze(ypred[i]),
"%s/%06d_%03d_out.png" % (batch_dir, step, i + 1))
if FLAGS.regularization > 0:
img_io.writeLDR(
np.squeeze(xtpred[i]), "%s/%06d_%03d_out_R.png" %
(batch_dir, step, i + 1))
img_io.writeLDR(
np.squeeze(ytpred[i]), "%s/%06d_%03d_out_T.png" %
(batch_dir, step, i + 1))
return (val_loss / n_batch, orig_loss / n_batch)
y_predict = sess.run([y], feed_dict=feed_dict)
y_predict = np.power(np.maximum(y_predict, 0.0), FLAGS.gamma)
print_("\tdone\n")
# Gamma corrected output 用伽马校正代替色调映射
y_gamma = np.power(np.maximum(y_predict, 0.0), 0.5)
# Write to disc
print_("\tWriting...")
k += 1
# 仅仅对过曝光区域进行了细节增强
image_name = frames[i]
image_name_out = image_name[image_name.find("/")+1:image_name.find(".")]
print("image_name_out:%s" % image_name_out)
img_io.writeLDR(x_buffer, '%s/%s_in_%d.png' % (FLAGS.out_dir,image_name_out, k), -3) # -3的曝光度
img_io.writeLDR(y_gamma, '%s/%s_out_%d.png' % (FLAGS.out_dir,image_name_out, k), -3)
# 对曝光度增强没有效果-》欠曝光细节没有增强
img_io.writeEXR(y_predict, '%s/%s_out_%d.exr' % (FLAGS.out_dir,image_name_out, k))
print_("\tdone\n")
except img_io.IOException as e:
print_("\n\t\tWarning! ", 'w', True)
print_("%s\n" % e, 'w')
except Exception as e:
print_("\n\t\tError: ", 'e', True)
print_("%s\n" % e, 'e')
print_("Done!\n")
# g alters this according to y = f(x^(1/g))^g
print_("\tInference...")
feed_dict = {y_: np.power(np.maximum(y_buffer, 0.0), 1.0/FLAGS.gamma)}
if FLAGS.input:
x_buffer = sess.run(x, feed_dict=feed_dict)
else:
x_buffer, y_predict = sess.run([x,y], feed_dict=feed_dict)
print_("\tdone\n")
# Write to disc
print_("\tWriting...")
for b in range(FLAGS.batch_size):
k += 1;
#img_io.writeLDR(np.squeeze(y_buffer[b,:,:,:]), '%s/%06d_gt.png' % (out_dir, k), -3)
if FLAGS.input:
img_io.writeLDR(np.squeeze(x_buffer[b,:,:]), '%s/%06d.png' % (out_dir, k), -3)
else:
img_io.writeLDR(np.squeeze(y_predict[b,:,:,:]), '%s/%06d.png' % (out_dir, k), -3)
print_("\tdone\n")
except img_io.IOException as e:
print_("\n\t\tWarning! ", 'w', True)
print_("%s\n"%e, 'w')
except Exception as e:
print_("\n\t\tError: ", 'e', True)
print_("%s\n"%e, 'e')
print_("Done!\n")
def calc_loss_and_print(x_data, y_data, print_dir, step, N):
val_loss, orig_loss, n_batch = 0, 0, 0
for b in range(int(x_data.shape[0] / FLAGS.batch_size)):
x_batch = x_data[b * FLAGS.batch_size:(b + 1) *
FLAGS.batch_size, :, :, :]
y_batch = y_data[b * FLAGS.batch_size:(b + 1) *
FLAGS.batch_size, :, :, :]
feed_dict = {x: x_batch, y_: y_batch}
err1, err2, y_predict, y_gt, M = sess.run(
[cost, cost_input_output, y, y_, msk], feed_dict=feed_dict)
val_loss += err1
orig_loss += err2
n_batch += 1
batch_dir = print_dir
if x_data.shape[0] > x_batch.shape[0]:
batch_dir = '%s/batch_%03d' % (print_dir, n_batch)
if n_batch <= N or N < 0:
if not os.path.exists(batch_dir):
os.makedirs(batch_dir)
for i in range(0, x_batch.shape[0]):
yy_p = np.squeeze(y_predict[i])
xx = np.squeeze(x_batch[i])
yy = np.squeeze(y_gt[i])
mm = np.squeeze(M[i])
# Apply inverse camera curve
x_lin = np.power(
np.divide(0.6 * xx, np.maximum(1.6 - xx, 1e-10)),
1.0 / 0.9)
# Transform log predictions to linear domain
yy_p = np.exp(yy_p) - eps
# Masking
y_final = (1 - mm) * x_lin + mm * yy_p
# Gamma correction
yy_p = np.power(np.maximum(yy_p, 0.0), 0.5)
y_final = np.power(np.maximum(y_final, 0.0), 0.5)
yy = np.power(np.maximum(yy, 0.0), 0.5)
xx = np.power(np.maximum(x_lin, 0.0), 0.5)
# Print LDR samples
if FLAGS.print_im:
img_io.writeLDR(
xx, "%s/%06d_%03d_in.png" % (batch_dir, step, i + 1),
-3)
img_io.writeLDR(
yy, "%s/%06d_%03d_gt.png" % (batch_dir, step, i + 1),
-3)
img_io.writeLDR(
y_final,
"%s/%06d_%03d_out.png" % (batch_dir, step, i + 1), -3)
# Print HDR samples
if FLAGS.print_hdr:
img_io.writeEXR(
xx, "%s/%06d_%03d_in.exr" % (batch_dir, step, i + 1))
img_io.writeEXR(
yy, "%s/%06d_%03d_gt.exr" % (batch_dir, step, i + 1))
img_io.writeEXR(
y_final,
"%s/%06d_%03d_out.exr" % (batch_dir, step, i + 1))
return (val_loss / n_batch, orig_loss / n_batch)
# the reconstructed highlights. If y = f(x) is the reconstruction, the gamma
# g alters this according to y = f(x^(1/g))^g
print_("\tInference...")
feed_dict = {x: np.power(np.maximum(x_buffer, 0.0), 1.0 / FLAGS.gamma)}
y_predict = sess.run([y], feed_dict=feed_dict)
y_predict = np.power(np.maximum(y_predict, 0.0), FLAGS.gamma)
print_("\tdone\n")
# Gamma corrected output
y_gamma = np.power(np.maximum(y_predict, 0.0), 0.5)
# Write to disc
print_("\tWriting...")
k += 1
# img_io.writeLDR(x_buffer, '%s/%06d_in.jpg' % (FLAGS.out_dir, k), -3)
# img_io.writeLDR(y_gamma, '%s/%06d_out.jpg' % (FLAGS.out_dir, k), -3)
# img_io.writeEXR(y_predict, '%s/%06d_out.exr' % (FLAGS.out_dir, k))
img_io.writeLDR(y_predict, '%s/%06d_predict.jpg' % (FLAGS.out_dir, k))
print_("\tdone\n")
except img_io.IOException as e:
print_("\n\t\tWarning! ", 'w', True)
print_("%s\n" % e, 'w')
except Exception as e:
print_("\n\t\tError: ", 'e', True)
print_("%s\n" % e, 'e')
print_("Done!\n")
sess.close()
print_("\t(Saturation: %0.2f%%)\n" % (100.0*(x_buffer>=1).sum()/x_buffer.size), 'm')
# Run prediction
print_("\tInference...")
feed_dict = {x: x_buffer}
y_predict = sess.run([y], feed_dict=feed_dict)
print_("\tdone\n")
# Gamma corrected output
y_gamma = np.power(np.maximum(y_predict, 0.0), 0.5)
# Write to disc
print_("\tWriting...")
k += 1;
img_io.writeLDR(x_buffer, '%s/%06d_in.png' % (FLAGS.out_dir, k), -3)
img_io.writeLDR(y_gamma, '%s/%06d_out.png' % (FLAGS.out_dir, k), -3)
img_io.writeEXR(y_predict, '%s/%06d_out.exr' % (FLAGS.out_dir, k))
print_("\tdone\n")
except img_io.IOException as e:
print_("\n\t\tWarning! ", 'w', True)
print_("%s\n"%e, 'w')
except Exception as e:
print_("\n\t\tError: ", 'e', True)
print_("%s\n"%e, 'e')
print_("Done!\n")
sess.close()
