with tf.device("/device:GPU:0"):
ordinal_model.build_model(input_images)
ordinal_model.build_evaluation(eval_batch_size=configs.batch_size,
flip_array=preprocessor.flip_array)
print("Network built!")
# log_writer = tf.summary.FileWriter(logdir=log_dir, graph=sess.graph)
model_saver = tf.train.Saver()
net_init = tf.global_variables_initializer()
sess.run([net_init])
# reload the model
for cur_model_iterations in evaluation_models:
mean_coords_eval = evaluators.mEvaluatorPose3D(
nJoints=configs.nJoints)
opt_mean_coords_eval = evaluators.mEvaluatorPose3D(
nJoints=configs.nJoints)
raw_coords_eval = evaluators.mEvaluatorPose3D(
nJoints=configs.nJoints)
opt_raw_coords_eval = evaluators.mEvaluatorPose3D(
nJoints=configs.nJoints)
data_index = my_utils.mRangeVariable(min_val=data_from,
max_val=data_to - 1,
initial_val=data_from)
################# Restore the model ################
if os.path.exists(
configs.restore_model_path_fn(cur_model_iterations) +
".index"):
cur_model_depth_scale.add(cur_scale)
scale_for_show.append(cur_scale)
print("Iter: {:07d} Loss : {:07f} Scales: {}\n\n".format(
scale_data_index.val, scale_loss, scale_for_show))
print("Cur Scale: {:07f}\n\n".format(
cur_model_depth_scale.cur_average[0]))
################################################################################################
##### Then evaluate it #####
cur_depth_scale = cur_model_depth_scale.cur_average[0]
print("Scale used to evaluate: {:07f}".format(cur_depth_scale))
depth_eval = evaluators.mEvaluatorDepth(nJoints=configs.nJoints)
coords_eval = evaluators.mEvaluatorPose3D(nJoints=configs.nJoints)
data_index = my_utils.mRangeVariable(min_val=data_from,
max_val=data_to - 1,
initial_val=data_from)
while not data_index.isEnd():
global_steps = sess.run(ordinal_model.global_steps)
batch_images_np = np.zeros([
configs.batch_size, configs.img_size, configs.img_size, 3
],
dtype=np.float32)
batch_relation_table_np = np.zeros(
[configs.batch_size, configs.nJoints, configs.nJoints],
dtype=np.float32)
batch_loss_table_log_np = np.zeros(
import numpy as np
import os
import sys
sys.path.append("../")
from utils.evaluate_utils import evaluators
eval_iterations = 100000
nJoints = 17
if __name__ == "__main__":
eval_depth = evaluators.mEvaluatorDepth(nJoints)
eval_coords = evaluators.mEvaluatorPose3D(nJoints)
gt_depth = np.random.random([eval_iterations, nJoints]) * 100
pd_depth = np.random.random([eval_iterations, nJoints]) * 100
mean_depth = np.mean(np.abs(gt_depth - pd_depth))
gt_coords = np.random.random([eval_iterations, nJoints, 3]) * 100
pd_coords = np.random.random([eval_iterations, nJoints, 3]) * 100
mean_coords = np.mean(np.sqrt(np.sum((gt_coords - pd_coords)**2, axis=2)))
for i in range(eval_iterations):
eval_depth.add(gt_depth[i], pd_depth[i])
eval_coords.add(gt_coords[i], pd_coords[i])
print("mean_depth({} | {}), mean_coords({} | {})".format(mean_depth, eval_depth.mean(), mean_coords, eval_coords.mean()))