def __init__(self):
tf.reset_default_graph()
pa.create_necessary_folders()
# Place Holder, batch size is 1
self.img_holder = tf.placeholder(tf.float32, [1, pa.PH, pa.PW, 3])
# PAF inference network
self.tensor_paf_pcm = PAF_network.PoseNet().inference_paf_pcm(
self.img_holder)
self.sess = tf.Session()
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state("logs/")
if ckpt:
saver.restore(self.sess, ckpt.model_checkpoint_path)
else:
raise FileNotFoundError("No PAF check point")
def main(argv=None):
print("Training with batch size: " + str(BATCH_SIZE))
# Place Holder
PH, PW = pa.PH, pa.PW
ipjc_holder = tf.placeholder(tf.float32, [BATCH_SIZE, 8, 6 + 2, 3])
img_holder = tf.placeholder(tf.float32, [BATCH_SIZE, PH, PW, 3])
# Entire network
img_tensor, i_hv_tensor = gpu_pipeline.build_training_pipeline(
ipjc_holder, img_holder)
poseNet = gpu_network.PoseNet()
loss_tensor = poseNet.build_paf_pcm_loss(img_tensor, i_hv_tensor)
lgdts_tensor = build_training_ops(loss_tensor)
# Session Saver summary_writer
sess = tf.Session()
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state("logs/")
if ckpt:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
# Global initializer
sess.run(tf.global_variables_initializer())
pa.create_necessary_folders()
summary_writer = tf.summary.FileWriter("logs/summary", sess.graph)
# Close the graph so no op can be added
tf.get_default_graph().finalize()
# Load samples from disk
samples_gen = gpu_pipeline.training_samples_gen(BATCH_SIZE)
for itr in range(1, int(1e7)):
batch_img, batch_ipjc = next(samples_gen)
feed_dict = {img_holder: batch_img, ipjc_holder: batch_ipjc}
loss_num, g_step_num, lr_num, train_op = sess.run(lgdts_tensor[0:4],
feed_dict=feed_dict)
print_log(loss_num, g_step_num, lr_num, itr)
# Summary
if itr % 100 == 0:
summary_str = sess.run(lgdts_tensor[4], feed_dict=feed_dict)
summary_writer.add_summary(summary_str, g_step_num)
saver.save(sess, "logs/ckpt")
print('Model Saved.')
sess.close()
def main(argv=None):
print("Training with batch size: " + str(BATCH_SIZE))
# Place Holder
PH, PW = pa.PH, pa.PW
HEAT_H, HEAT_W = pa.HEAT_H, pa.HEAT_W
PCM_nhwc_holder = tf.placeholder(tf.float32, [BATCH_SIZE, HEAT_H, HEAT_W, 14])
PAF_nhwc_holder = tf.placeholder(tf.float32, [BATCH_SIZE, HEAT_H, HEAT_W, 11 * 2])
img_holder = tf.placeholder(tf.float32, [BATCH_SIZE, PH, PW, 3])
# Entire network
poseNet = PAF_network.PoseNet()
loss_tensor = poseNet.build_paf_pcm_loss(img_holder, PCM_nhwc_holder, PAF_nhwc_holder)
lgdts_tensor = build_training_ops(loss_tensor) #[loss_tensor, global_step, decaying_learning_rate, train_op, summary_op]
# Session Saver summary_writer
sess = tf.Session()
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state("logs/")
if ckpt:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
# Global initializer
sess.run(tf.global_variables_initializer())
pa.create_necessary_folders()
summary_writer = tf.summary.FileWriter("logs/summary", sess.graph)
# Close the graph so no op can be added
tf.get_default_graph().finalize()
# Load samples from disk
gen_PCM_PAF_IMG = label_loader.generator_PCM_PAF_IMG(BATCH_SIZE, (512, 512), 8)
for itr in range(1, int(1e7)):
BC, BA, BI = next(gen_PCM_PAF_IMG)
feed_dict = {img_holder: BI, PCM_nhwc_holder: BC, PAF_nhwc_holder: BA}
loss_num, g_step_num, lr_num, train_op = sess.run(lgdts_tensor[0:4], feed_dict=feed_dict)
print_log(loss_num, g_step_num, lr_num, itr)
# Summary
if itr % 100 == 0:
summary_str = sess.run(lgdts_tensor[4], feed_dict=feed_dict)
summary_writer.add_summary(summary_str, g_step_num)
saver.save(sess, "logs/ckpt")
print('Model Saved.')
sess.close()
if visible(12): # Head top
set_ipjc(6, 12)
if visible(13): # Neck
set_ipjc(7, 13)
iname_arr = np.asarray(iname_list)
return [ipjc_arr, iname_arr]
target_ratio = PW / PH
files = [open(l) for l in label_collection]
json_labels_list = [json.load(l) for l in files]
[f.close() for f in files]
la_im_list = list(zip(json_labels_list, img_folder_collection))
ipjcs_inames_list = [resize_by_json(*la_im)
for la_im in la_im_list] # [3][ipjc, iname]
ipjc3, iname3 = list(zip(*ipjcs_inames_list))
ipjc3 = np.reshape(np.asarray(ipjc3), [-1])
iname3 = np.reshape(np.asarray(iname3), [-1])
ipjc_con = np.concatenate(ipjc3, 0)
iname_con = np.concatenate(iname3, 0)
np.save(AI_IPJC_FILE, ipjc_con)
np.save(AI_INAME_FILE, iname_con)
assert sys.version_info >= (3, 5)
pa.create_necessary_folders()
resize_keep_ratio(save_img=False)
def save_joints_position(v_name=None):
"""
Save joints position from a video to file
v_name: name of video, no appendix
:return:
"""
tf.reset_default_graph()
pa.create_necessary_folders()
batch_size = 15
if v_name is None:
video_path = os.path.join(pa.VIDEO_FOLDER_PATH,
pa.VIDEO_LIST[0] + ".mp4")
else:
video_path = os.path.join(pa.VIDEO_FOLDER_PATH, v_name + ".mp4")
metadata = skvideo.io.ffprobe(video_path)
total_frames = int(metadata["video"]["@nb_frames"])
v_width = int(metadata["video"]["@width"])
v_height = int(metadata["video"]["@height"])
assert (v_height == pa.PH and v_width == pa.PW)
v_gen = skvideo.io.vreader(video_path)
# Place Holder
img_holder = tf.placeholder(tf.float32, [batch_size, v_height, v_width, 3])
# Entire network
paf_pcm_tensor = gpu_network.PoseNet().inference_paf_pcm(img_holder)
# Place for argmax values
joint_ixy = list() # [i][j0~6][x,y]
# Session Saver summary_writer
with tf.Session() as sess:
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state("logs/")
if ckpt:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
raise FileNotFoundError("Tensorflow ckpt not found")
for i in range(0, total_frames - batch_size + 1, batch_size):
frames = [next(v_gen) / 255. for _ in range(batch_size)]
feed_dict = {img_holder: frames}
paf_pcm = sess.run(paf_pcm_tensor, feed_dict=feed_dict)
pcm = paf_pcm[:, :, :, 14:]
pcm = np.clip(pcm, 0., 1.)
for idx_img in range(batch_size):
# 6 joint in image
img_j6 = []
for idx_joint in range(8):
heat = pcm[idx_img, :, :, idx_joint]
c_coor_1d = np.argmax(heat)
c_coor_2d = np.unravel_index(
c_coor_1d, [pa.HEAT_SIZE[1], pa.HEAT_SIZE[0]])
c_value = heat[c_coor_2d]
j_xy = [] # x,y
if c_value > 0.15:
percent_h = c_coor_2d[0] / pa.HEAT_H
percent_w = c_coor_2d[1] / pa.HEAT_W
j_xy.append(percent_w)
j_xy.append(percent_h)
else:
j_xy.append(-1.)
j_xy.append(-1.)
img_j6.append(j_xy)
joint_ixy.append(img_j6)
print("Image: " + str(i))
# sess closed
save_path = os.path.join(pa.RNN_SAVED_JOINTS_PATH, v_name + ".npy")
np.save(save_path, joint_ixy)
print(save_path)