def do_predict(image_ids):
image_datas = []
for image_id in image_ids:
image_test = Image.open('testset/' + image_id.replace('.xml', '.' + cfg.image_format))
resized_image = image_test.resize((cfg.sample_size, cfg.sample_size), Image.BICUBIC)
image_data = np.array(resized_image, dtype='float32')
image_datas.append((image_id, image_data, image_test))
imgs_holder = tf.placeholder(tf.float32, shape=[1, cfg.sample_size, cfg.sample_size, 3])
istraining = tf.constant(False, tf.bool)
cfg.batch_size = 1
cfg.scratch = True
model = yolov3(imgs_holder, None, istraining)
img_hw = tf.placeholder(dtype=tf.float32, shape=[2])
boxes, scores, classes = model.pedict(img_hw, iou_threshold=0.5, score_threshold=0.001)
saver = tf.train.Saver()
ckpt_dir = 'ckpt/'
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(ckpt_dir)
saver.restore(sess, ckpt.model_checkpoint_path)
gs = int(ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1])
print('Restore batch: ', gs)
for image_id, image_data, image_test in image_datas[:5]:
boxes_, scores_, classes_ = sess.run([boxes, scores, classes],
feed_dict={
img_hw: [image_test.size[1], image_test.size[0]],
imgs_holder: np.reshape(image_data / 255, [1, cfg.sample_size, cfg.sample_size, 3])})
try:
draw_xml(np.array(image_test, dtype=np.float32) / 255, boxes_, classes_, cfg.names, scores=scores_, image_id=image_id)
print('predict:', image_id)
except Exception as e:
print(e)
from draw_boxes import draw_boxes
import matplotlib.pyplot as plt
# IMG_ID ='008957'
# image_test = Image.open('/home/raytroop/Dataset4ML/VOC2007/VOCdevkit/VOC2007/JPEGImages/{}.jpg'.format(IMG_ID))
image_test = Image.open('image/dog.jpg')
resized_image = image_test.resize((416, 416), Image.BICUBIC)
image_data = np.array(resized_image, dtype='float32')
imgs_holder = tf.placeholder(tf.float32, shape=[1, 416, 416, 3])
istraining = tf.constant(False, tf.bool)
cfg.batch_size = 1
cfg.scratch = True
model = yolov3(imgs_holder, None, istraining)
img_hw = tf.placeholder(dtype=tf.float32, shape=[2])
boxes, scores, classes = model.pedict(img_hw, iou_threshold=0.5, score_threshold=0.5)
saver = tf.train.Saver()
ckpt_dir = './ckpt/'
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(ckpt_dir)
saver.restore(sess, ckpt.model_checkpoint_path)
boxes_, scores_, classes_ = sess.run([boxes, scores, classes],
feed_dict={
img_hw: [image_test.size[1], image_test.size[0]],
imgs_holder: np.reshape(image_data / 255, [1, 416, 416, 3])})
image_draw = draw_boxes(np.array(image_test, dtype=np.float32) / 255, boxes_, classes_, cfg.names, scores=scores_)
from config import cfg
from PIL import Image, ImageDraw, ImageFont
from draw_boxes import draw_boxes
import matplotlib.pyplot as plt
# IMG_ID ='008957'
# image_test = Image.open('/home/raytroop/Dataset4ML/VOC2007/VOCdevkit/VOC2007/JPEGImages/{}.jpg'.format(IMG_ID))
imgs_holder = tf.placeholder(tf.float32, shape=[1, 416, 416, 3])
istraining = tf.constant(False, tf.bool)
cfg.batch_size = 1
cfg.scratch = True
model = yolov3(imgs_holder, None, istraining)
img_hw = tf.placeholder(dtype=tf.float32, shape=[2])
boxes, scores, classes = model.pedict(img_hw, iou_threshold=0.5, score_threshold=0.5)
saver = tf.train.Saver()
ckpt_dir = './ckpt/'
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(ckpt_dir)
print(ckpt.model_checkpoint_path)
saver.restore(sess, ckpt.model_checkpoint_path)
import datetime
import os
path = r"E:\github\darknet_windows\build\darknet\x64\data\voc\VOCdevkit\VOC2007\JPEGImages"
files = os.listdir(path)
# files.sort(key=lambda x: int(x[:-4]))
def do_predict(image_ids):
image_datas = []
for image_id in image_ids[:]:
image_test = Image.open(testset + '\\' + image_id)
resized_image = image_test.resize((cfg.sample_size, cfg.sample_size),
Image.BICUBIC)
image_data = np.array(resized_image, dtype='float32')
image_datas.append((image_id, image_data, image_test))
imgs_holder = tf.placeholder(
tf.float32, shape=[1, cfg.sample_size, cfg.sample_size, 3])
istraining = tf.constant(False, tf.bool)
cfg.batch_size = 1
cfg.scratch = True
model = yolov3(imgs_holder, None, istraining)
img_hw = tf.placeholder(dtype=tf.float32, shape=[2])
boxes, scores, classes = model.pedict(img_hw,
iou_threshold=0.5,
score_threshold=score_threshold)
saver = tf.train.Saver()
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(ckpt_dir)
saver.restore(sess, ckpt.model_checkpoint_path)
gs = int(ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1])
print('Restore batch: ', gs)
for image_id, image_data, image_test in image_datas:
boxes_, scores_, classes_ = sess.run(
[boxes, scores, classes],
feed_dict={
img_hw: [image_test.size[1], image_test.size[0]],
imgs_holder:
np.reshape(image_data / 255,
[1, cfg.sample_size, cfg.sample_size, 3])
})
try:
image_draw = draw_boxes(
np.array(image_test, dtype=np.float32) / 255,
boxes_,
classes_,
cfg.names,
scores=scores_)
# cv2.imshow("prediction.png", cv2.cvtColor(image_draw, cv2.COLOR_RGB2BGR))
print('predict:', image_id)
# tree = draw_xml(np.array(image_test, dtype=np.float32) / 255, boxes_, classes_, cfg.names, scores=scores_, image_id=image_id)
# tree.write(result_dir + '\\predicted_xmls\\{}.xml'.format(image_id.split('.')[0]))
cv2.imwrite(result_dir + '\\' + image_id,
cv2.cvtColor(image_draw, cv2.COLOR_RGB2BGR),
[int(cv2.IMWRITE_PNG_COMPRESSION), 9])
# fig = plt.figure(frameon=False)
# ax = plt.Axes(fig, [0, 0, 1, 1])
# ax.set_axis_off()
# fig.add_axes(ax)
# plt.imshow(image_draw, interpolation='none')
# plt.savefig('C:\\Users\\P900\\Desktop\\myWork\\YOLOv3_tf\\prediction.jpg', dpi='figure', interpolation='none')
# # fig.savefig('prediction.jpg')
# plt.show()
except Exception as e:
print(e)
from yolo_top import yolov3 import numpy as np import tensorflow as tf img = tf.constant(np.random.normal(0, 1, [8, 416, 416, 3]), tf.float32) truth = tf.constant(np.random.randint(0, 2, size=[8, 30, 5]), tf.float32) istraining = tf.constant(False, tf.bool) model = yolov3(img, truth, istraining) loss = model.compute_loss() sess = tf.Session() sess.run(tf.global_variables_initializer()) print(sess.run(loss))
from yolo_top import yolov3
import numpy as np
import tensorflow as tf
from data_pipeline import data_pipeline
import config as cfg
import os
os.environ["CUDA_VISIBLE_DEVICES"]="3"
file_path = 'E:/Python/tensorflow/YOLO/people count/yuncong_data/our/trainval_2014.tfrecord'
log_dir='E:/Python/tensorflow/YOLO/people count/yuncong_data/our/log/'
imgs, true_boxes = data_pipeline(file_path, cfg.batch_size)
istraining = tf.constant(True, tf.bool)
model = yolov3(imgs, true_boxes, istraining)
with tf.name_scope('loss'):
loss,AVG_IOU,coordinates_loss_sum,objectness_loss,no_objects_loss_mean = model.compute_loss()
tf.summary.scalar('loss', loss)
tf.summary.scalar('avg', AVG_IOU)
tf.summary.scalar('coord', coordinates_loss_sum)
tf.summary.scalar('obj', objectness_loss)
tf.summary.scalar('no_obj', no_objects_loss_mean)
global_step = tf.Variable(0, trainable=False)
#lr = tf.train.exponential_decay(0.0001, global_step=global_step, decay_steps=2e4, decay_rate=0.1)
lr = tf.train.piecewise_constant(global_step, [30000, 45000], [1e-4, 5e-5, 1e-5]) ##作用在不同步长时更改学习率
optimizer = tf.train.AdamOptimizer(learning_rate=lr)
#optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.00001)
update_op = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
vars_det = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="Head")
# for var in vars_det:
# print(var)
with tf.control_dependencies(update_op):
from yolo_top import yolov3 import numpy as np import tensorflow as tf img = tf.constant(np.random.normal(0, 1, [8, 416, 416, 3]), tf.float32) truth = tf.constant(np.random.randint(0, 2, size=[8, 30, 5]), tf.float32) istraining = tf.constant(True, tf.bool) model = yolov3(img, truth, istraining) loss = model.compute_loss() sess = tf.Session() sess.run(tf.global_variables_initializer()) print(sess.run(loss))
# print(img,truth)
# tfconfig = tf.ConfigProto()
# tfconfig.gpu_options.allow_growth = True
# tfconfig.gpu_options.per_process_gpu_memory_fraction = 0.9
img_holder = tf.placeholder(tf.float32,
shape=(cfg.batch_size, cfg.train.image_resized,
cfg.train.image_resized, 3),
name='img_holder')
truth_holder = tf.placeholder(tf.float32,
shape=(cfg.batch_size, 30, 5),
name='truth_holder')
istraining = tf.placeholder(tf.bool, shape=(), name='istraining')
model = yolov3(img_holder, truth_holder, istraining, trainable_head=True)
loss = model.compute_loss()
# optimizer
global_step = tf.Variable(0, trainable=False)
# lr = tf.train.exponential_decay(0.0001, global_step=global_step, decay_steps=2e4, decay_rate=0.1)
lr1 = tf.train.piecewise_constant(global_step, [2215 * 3, 2215 * 5],
[5e-5, 1e-4, 1e-5])
# lr2 = tf.train.piecewise_constant(global_step, [100, 1000], [1e-5, 5e-6, 1e-6])
# optimizer1 = tf.train.AdamOptimizer(learning_rate=lr1)
optimizer2 = tf.train.AdamOptimizer(learning_rate=lr1)
# optimizer = tf.train.MomentumOptimizer(learning_rate=lr, momentum=0.9)
update_op = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
from yolo_top import yolov3
import numpy as np
import tensorflow as tf
from data_pipeline import data_pipeline
from config import cfg
file_path = 'trainval0712.tfrecords'
imgs, true_boxes = data_pipeline(file_path, cfg.batch_size)
istraining = tf.constant(True, tf.bool)
model = yolov3(imgs, true_boxes, istraining)
loss = model.compute_loss()
global_step = tf.Variable(0, trainable=False)
# lr = tf.train.exponential_decay(0.0001, global_step=global_step, decay_steps=2e4, decay_rate=0.1)
lr = tf.train.piecewise_constant(global_step, [40000, 45000], [1e-3, 1e-4, 1e-5])
optimizer = tf.train.AdamOptimizer(learning_rate=lr)
# optimizer = tf.train.MomentumOptimizer(learning_rate=lr, momentum=0.9)
update_op = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
vars_det = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="Head")
# for var in vars_det:
# print(var)
with tf.control_dependencies(update_op):
train_op = optimizer.minimize(loss, global_step=global_step, var_list=vars_det)
saver = tf.train.Saver()
ckpt_dir = './ckpt/'
gs = 0
batch_per_epoch = 2000
cfg.train.max_batches = int(batch_per_epoch * 10)
cfg.train.image_resized = 608