def main():
basepath = "test_data"
bakpath = "bakvgg16"
mkdir(bakpath + "/" + basepath)
imgdata = ["./test_data/tiger.jpeg"]
imgdata = loadFrom(basepath) #20180330
num = len(imgdata)
if num == 0:
utils.printcolor("图像文件数量为0", mode='bold', fore='red')
return
per = 10 if (num > 10) else num
count = int(num / per) if (num % per == 0) else int(num / per) + 1
print(per, num, count, num % per)
vgg = vgg16.Vgg16()
images = tf.placeholder("float", [per, 224, 224, 3])
with tf.name_scope("content_vgg"):
vgg.build(images)
for x in range(0, count):
xdata = imgdata[x * per:x * per + per]
#print(len(xdata))
if len(xdata) == num % per:
vggx = vgg16.Vgg16()
images = tf.placeholder("float", [len(xdata), 224, 224, 3])
with tf.name_scope("content_vgg"):
vggx.build(images)
tensor_imgdata(xdata, images, vggx, bakpath)
else:
tensor_imgdata(xdata, images, vgg, bakpath)
def VGG16_Synthesis_Network(input_img, input_Semantic): # img encoder vgg = vgg16.Vgg16() vgg.build(input_img) img_layer = tf.contrib.layers.flatten(vgg.pool5) img_layer = tf.contrib.layers.fully_connected(img_layer, 512) # Semantic encoder Semantic_layer = tf.contrib.layers.fully_connected(input_Semantic, 512) # Semantic synthesis z = tf.concat([img_layer, Semantic_layer], -1) # img decoder z = tf.expand_dims(z,1) z = tf.expand_dims(z,1) # output_1 = tf.layers.conv2d_transpose(inputs = z, filters = 512, kernel_size = 3, strides = (2,2), padding='same', data_format='channels_last', activation = tf.nn.relu) # output_2 = tf.layers.conv2d_transpose(inputs = output_1, filters = 256, kernel_size = 3, strides = (2,2), padding='same', data_format='channels_last', activation = tf.nn.relu) # output_3 = tf.layers.conv2d_transpose(inputs = output_2, filters = 256, kernel_size = 3, strides = (2,2), padding='same', data_format='channels_last', activation = tf.nn.relu) # output_4 = tf.layers.conv2d_transpose(inputs = output_3, filters = 128, kernel_size = 3, strides = (2,2), padding='same', data_format='channels_last', activation = tf.nn.relu) # output_5 = tf.layers.conv2d_transpose(inputs = output_4, filters = 64, kernel_size = 3, strides = (2,2), padding='same', data_format='channels_last', activation = tf.nn.relu) # output_6 = tf.layers.conv2d_transpose(inputs = output_5, filters = 32, kernel_size = 3, strides = (2,2), padding='same', data_format='channels_last', activation = tf.nn.relu) # output_7 = tf.layers.conv2d_transpose(inputs = output_6, filters = 3, kernel_size = 3, strides = (2,2), padding='same', data_format='channels_last', activation = tf.nn.relu) output_1 = residual_block(inputs_layer = z, nb_blocks = 1, out_filters = 512, strides = 2) output_2 = residual_block(inputs_layer = output_1, nb_blocks = 1, out_filters = 256, strides = 2) output_3 = residual_block(inputs_layer = output_2, nb_blocks = 1, out_filters = 256, strides = 2) output_4 = residual_block(inputs_layer = output_3, nb_blocks = 1, out_filters = 128, strides = 2) output_5 = residual_block(inputs_layer = output_4, nb_blocks = 1, out_filters = 64, strides = 2) output_6 = residual_block(inputs_layer = output_5, nb_blocks = 1, out_filters = 32, strides = 2) output_7 = residual_block(inputs_layer = output_6, nb_blocks = 1, out_filters = 3, strides = 2) # output output_7 = tf.image.resize_images(output_7, [224,224]) return output_7
def run(self):
#set enviornment
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = str(self._gpuid)
#load models
import wingdbstub
wingdbstub.Ensure()
wingdbstub.debugger.StartDebug()
import vgg16
#download the vgg16 weights from
#https://github.com/fchollet/deep-learning-models/releases/download/v0.1/vgg16_weights_tf_dim_ordering_tf_kernels.h5
xnet = vgg16.Vgg16(
'/media/zhaoke/b0685ee4-63e3-4691-ae02-feceacff6996/models/vgg16_weights_tf_dim_ordering_tf_kernels.h5'
)
print 'vggnet init done', self._gpuid
while True:
xfile = self._queue.get()
if xfile == None:
self._queue.put(None)
break
label = self.predict(xnet, xfile)
print 'woker', self._gpuid, ' xfile ', xfile, " predicted as label", label
print 'vggnet done ', self._gpuid
wingdbstub.debugger.StopDebug()
def detect_image_vgg16(request):
if request.method == "POST": # 请求方法为POST时，进行处理
myFile = request.FILES.get("myfile", None) # 获取上传的文件，如果没有文件，则默认为None
if not myFile:
return HttpResponse("no files for upload!")
img_ready = utils.load_image(myFile)
with tf.Session() as sess:
# 给输入图片占位
images = tf.placeholder(tf.float32, [1, 224, 224, 3])
# 初始化vgg,读出保存的文件参数
vgg = vgg16.Vgg16()
# 调用前向传播函数,把待识别图片喂入神经网络
vgg.forward(images)
# 运行会话,预测分类结果
probability = sess.run(vgg.prob, feed_dict={images: img_ready})
top5 = np.argsort(probability[0])[-1:-6:-1]
data = ""
data += "top5:"
data += str(top5)
data += "<br/>"
for n, i in enumerate(top5):
data += str(i)
data += ":"
data += labels[i]
data += "----"
data += str(utils.percent(probability[0][i]))
data += "<br/>"
return HttpResponse(
"<html><center><h1>VGG16物体分类结果展示</h1><body>%s</body></center></html>"
% data)
def __init__(self, model_vgg_path = config.MODEL_VGG_PATH,\
model_dnn_path = config.MODEL_DNN_PATH, \
uid_path = config.UID_PATH):
lr = 1e-4
feature_columns = [
tf.contrib.layers.real_valued_column("", dimension=4)
]
self.classifier = learn.DNNClassifier(
model_dir=model_dnn_path,
feature_columns=feature_columns,
hidden_units=[4096, 2048],
n_classes=69,
activation_fn=tf.nn.sigmoid,
optimizer=tf.train.AdamOptimizer(learning_rate=lr))
self.sess = tf.Session()
self.vgg = vgg16.Vgg16(model_vgg_path)
self.images = tf.placeholder("float", [None, 224, 224, 3])
with tf.name_scope("content_vgg"):
self.vgg.build(self.images)
self.uid = []
with open(uid_path, 'rt') as f:
for it in f:
self.uid.append(it.replace('\n', ''))
def main():
with tf.Session() as sess:
input_tensor = tf.placeholder("float", [32, 224, 224, 3])
vgg = vgg16.Vgg16()
with tf.name_scope("content_vgg"):
vgg.build(input_tensor)
# a loop for all folders
folders = os.listdir(root_folder)
folders.sort()
try:
start = int(sys.argv[1])
end = int(sys.argv[2])
except:
start = 0
end = len(folders)
for folder in folders[start:end]:
folder_path = os.path.join(root_folder, folder)
prob_path = os.path.join(probs_folder, folder + '.txt')
if os.path.exists(prob_path):
continue
probs_for_folder = get_1365_vec(vgg, input_tensor, sess,
folder_path)
try:
np.savetxt(prob_path, probs_for_folder)
except Exception as e:
print(e)
file = open('error.txt', 'a')
file.write(folder_path)
file.close()
def __init__(self):
self.vgg = vgg16.Vgg16()
self.input_holder = tf.placeholder(tf.float32,
[1, img_size, img_size, 3])
self.label_holder = tf.placeholder(tf.float32,
[label_size * label_size, 2])
self.S11_holder = tf.placeholder(
tf.float32,
[3 * 3 * int(np.floor(cfg.B / 2)),
int(np.floor(cfg.C / 2))])
self.S12_holder = tf.placeholder(
tf.float32,
[3 * 3 * int(np.floor(cfg.B / 2)),
int(np.floor(cfg.C / 2))])
self.S21_holder = tf.placeholder(
tf.float32,
[3 * 3 * int(np.floor(cfg.C / 2)),
int(np.floor(cfg.D / 2))])
self.S22_holder = tf.placeholder(
tf.float32,
[3 * 3 * int(np.floor(cfg.C / 2)),
int(np.floor(cfg.D / 2))])
self.sobel_fx, self.sobel_fy = self.sobel_filter()
self.contour_th = 1.5
self.contour_weight = 0.0001
def main():
num_frames = 80
video_path = './data'
video_save_path = './data_features'
videos = os.listdir(video_path)
videos = filter(lambda x: x.endswith('avi'), videos)
# Load the CNN model trained on images to recognize objects.
with tf.device('/gpu:0'):
with tf.Session() as sess:
vgg = vgg16.Vgg16()
for idx, video in enumerate(videos):
print idx, video
if os.path.exists(os.path.join(video_save_path, video)):
print "Already processed ... "
continue
video_fullpath = os.path.join(video_path, video)
try:
cap = cv2.VideoCapture(video_fullpath)
except:
print "cv2 not working"
pass
frame_count = 0
frame_list = []
while True:
ret, frame = cap.read()
if ret is False:
break
frame_list.append(frame)
frame_count += 1
frame_list = np.array(frame_list)
if frame_count > 80:
frame_indices = np.linspace(0,
frame_count,
num=num_frames,
endpoint=False).astype(int)
frame_list = frame_list[frame_indices]
processed_frame_list = []
for x in frame_list:
processed_frame_list.append(preprocess_frame(x))
# Generate the frame list of equally spaced 80 frames.
cropped_frame_list = np.array(processed_frame_list)
print "Cropped Frame List: ", cropped_frame_list.shape
images = tf.placeholder("float",
[cropped_frame_list.shape[0], 224, 224, 3])
with tf.name_scope("content_vgg"):
vgg.build(images)
feats = sess.run(vgg.fc6, feed_dict=feed_dict)
print "Feature Length: ", len(feats)
save_full_path = os.path.join(video_save_path, video + '.npy')
np.save(save_full_path, feats)
def predict(models_path,image_dir,labels_filename,labels_nums, data_format):
[batch_size, resize_height, resize_width, depths] = data_format
labels = np.loadtxt(labels_filename, str, delimiter='\t')
input_images = tf.placeholder(dtype=tf.float32, shape=[None, resize_height, resize_width, depths], name='input')
#其他模型预测请修改这里
# with slim.arg_scope(inception_v3.inception_v3_arg_scope()):
# out, end_points = inception_v3.inception_v3(inputs=input_images, num_classes=labels_nums, dropout_keep_prob=1.0, is_training=False)
vgg = vgg16.Vgg16(vgg16_npy_path='./vgg16.npy',input=input_images)
out = vgg.out
# 将输出结果进行softmax分布,再求最大概率所属类别
score = tf.nn.softmax(out,name='pre')
class_id = tf.argmax(score, 1)
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
saver.restore(sess, models_path)
images_list=glob.glob(os.path.join(image_dir,'*.jpg'))
for image_path in images_list:
im=read_image(image_path,resize_height,resize_width,normalization=True)
im=im[np.newaxis,:]
#pred = sess.run(f_cls, feed_dict={x:im, keep_prob:1.0})
pre_score,pre_label = sess.run([score,class_id], feed_dict={input_images:im})
max_score=pre_score[0,pre_label]
img = cv2.imread(image_path)
plt.imshow(img)
plt.title(os.path.basename(image_path))
print("{} is: pre labels:{},name:{} score: {}".format(image_path,pre_label,labels[pre_label], max_score))
plt.show() #自己加的
sess.close()
def get_feats(image_dict):
vgg16_feats = np.zeros((len(image_dict), 4096))
with tf.Session() as sess:
images = tf.placeholder(dtype=tf.float32, shape=[None, 224, 224, 3])
vgg = vgg16.Vgg16()
vgg.build(images)
for i in range(len(image_dict)):
print(i)
img_list = load_image(image_dict[i])
batch = img_list.reshape((1, 224, 224, 3))
print(batch.shape)
feature = sess.run(vgg.fc7, feed_dict={images: batch}) #提取fc7层的特征
print(feature.shape)
feature = np.reshape(feature, [4096])
feature = feature.reshape(-1)
print(feature.shape)
feature /= norm(feature) # 特征归一化
# print(feature)
vgg16_feats[i, :] = feature #每张图片的特征向量为1行
# print(vgg16_feats.shape)
# print(len(vgg16_feats))
# print(len(vgg16_feats[0]))
# vgg16_feats = np.save('X:/github/vgg-tensorflow/vgg16_feats', vgg16_feats)
return vgg16_feats
def get_activations(layer_name, batch):
#uncomment for gpu
# with tf.Session(config=tf.ConfigProto(gpu_options=(tf.GPUOptions(per_process_gpu_memory_fraction=0.7)))) as sess:
with tf.device('/cpu:0'):
with tf.Session() as sess:
images = tf.placeholder("float", [len(batch), 224, 224, 3])
feed_dict = {images: batch}
vgg = vgg16.Vgg16()
with tf.name_scope("content_vgg"):
vgg.build(images)
map_layer_name = {}
map_layer_name['conv1_1'] = vgg.conv1_1
map_layer_name['conv1_2'] = vgg.conv1_2
map_layer_name['conv2_1'] = vgg.conv2_1
map_layer_name['conv2_2'] = vgg.conv2_2
map_layer_name['conv3_1'] = vgg.conv3_1
map_layer_name['conv3_2'] = vgg.conv3_2
map_layer_name['conv3_3'] = vgg.conv3_3
map_layer_name['conv4_1'] = vgg.conv4_1
map_layer_name['conv4_2'] = vgg.conv4_2
map_layer_name['conv4_3'] = vgg.conv4_3
map_layer_name['conv5_1'] = vgg.conv5_1
map_layer_name['conv5_2'] = vgg.conv5_2
map_layer_name['conv5_3'] = vgg.conv5_3
map_layer_name['fc6'] = vgg.fc6
map_layer_name['fc7'] = vgg.fc7
map_layer_name['fc8'] = vgg.fc8
activations = sess.run(map_layer_name[layer_name], feed_dict=feed_dict)
return activations
def main(_):
input_ = tf.placeholder(tf.float32, [None, 224, 224, 3])
vgg = vgg16.Vgg16()
vgg.build(input_)
with tf.Session() as sess:
img = load_image(FLAGS.image_path)
img = img.reshape((1, 224, 224, 3))
feed_dict = {input_: img}
codes = sess.run(vgg.relu6, feed_dict=feed_dict)
with tf.Session() as sess:
feature_columns = [
tf.feature_column.numeric_column('code', shape=4096)
]
classifier = tf.estimator.DNNClassifier(
feature_columns=feature_columns,
hidden_units=[512],
n_classes=5,
model_dir='./estimator_model_dir')
# train_input_fun = create_input_fun(FLAGS.train_tfrecords, repeat_count=2)
# classifier.train(input_fn=train_input_fun)
# test_input_fun = create_input_fun(FLAGS.test_tfrecords, repeat_count=1, perform_shuffle=False)
validate_input_fun = create_validate_input_fun(codes)
# evaluate_result = classifier.evaluate(input_fn=test_input_fun)
# print("Evaluation results:")
# for key in evaluate_result:
# print(" {}, was: {}".format(key, evaluate_result[key]))
predict_results = classifier.predict(input_fn=validate_input_fun)
print("Predictions:")
print(
list(
map(lambda x: INDEX_LABEL_MAP[x["class_ids"][0]],
predict_results)))
def __init__(self):
self.vgg_t = vgg16.Vgg16()
self.input_holder_t = tf.placeholder(tf.float32, [1, img_size, img_size, 3])
self.label_holder = tf.placeholder(tf.float32, [label_size*label_size, 2])
def compute_gram_matrices(
images,
vgg_layers=['conv1_2', 'conv2_2', 'conv3_2', 'conv4_2', 'conv5_2']):
"""Computes the gram matrix representation of a batch of images"""
vgg_net = vgg16.Vgg16(opts.vgg16_path)
vgg_acts = vgg_net.get_vgg_activations(images, vgg_layers)
grams = [gram_matrix(layer) for layer in vgg_acts]
return grams
def G_Net(self, x, keep_prob):
vgg = vgg16.Vgg16()
print('making g-network')
with tf.variable_scope('G_Net'):
#x = norm_image(x)
A_input, S_input = vgg.build(x, self.batch_size, keep_prob)
print('finishing g-network')
return A_input, S_input
def perc_loss_cal(self, input_tensor):
"""
得到输入图像在预训练的VGG16模型中的中间层，用于计算感知损失
:param input_tensor: 输入人像
:return: 输入人像的conv4_1参数
"""
vgg = vgg16.Vgg16("./preTrainedModel/vgg16.npy")
vgg.build(input_tensor)
return vgg.conv4_1
def __init__(self):
self.vgg = vgg16.Vgg16()
self.input_holder = tf.placeholder(tf.float32, [1, img_size, img_size, 3])
self.label_holder = tf.placeholder(tf.float32, [label_size*label_size, 2])
self.sobel_fx, self.sobel_fy = self.sobel_filter()
self.contour_th = 1.5
self.contour_weight = 0.0001
def __init__(self):
self.vgg = vgg16.Vgg16()
self.input_holder = tf.placeholder(tf.float32,
[5, img_size, img_size, 3])
self.label_holder = tf.placeholder(tf.float32,
[5, label_size, label_size, 2])
self.contour_th = 1.5
self.contour_weight = 0.0001
def compute_vgg_loss(images, layer_names, batch_size, ch, vgg_weight):
# images = [ba, B , ab, A]
""" Preprocess images """
processed = [vgg_preprocess(i) for i in images]
#ba = vgg_preprocess( ba )
#ab = vgg_preprocess( ab )
#B = vgg_preprocess( B )
#A = vgg_preprocess( A )
img_shape = processed[-1].get_shape()
h = img_shape[1]
w = img_shape[2]
layers_ = list()
with tf.name_scope("VGG16"):
# Create first instance of the VGG16-model
vgg1 = vgg16.Vgg16(vgg_weight, h, w, ch)
for enum, i in enumerate(processed):
with tf.variable_scope(tf.get_variable_scope(), reuse=True):
with tf.name_scope("p" + str(enum)):
vgg1.build(i)
layers_.append(vgg1.get_layer_tensors(layer_names, enum))
# vgg1.build( ba )
# layers_ba = vgg1.get_layer_tensors( layer_names )
# vgg1.build( A )
# layers_A = vgg1.get_layer_tensors( layer_names )
# vgg1.build( ab )
# layers_ab = vgg1.get_layer_tensors( layer_names )
vgg1.data_dict = None
# Create second instance of the VGG16-model
# vgg1 = vgg16.Vgg16( vgg_weight, h, w, ch )
# vgg1.build( ba )
# layers_ba = vgg1.get_layer_tensors( layer_names )
# compute perceptual loss
with tf.variable_scope('loss_a', reuse=tf.AUTO_REUSE):
loss_a = tf.zeros(batch_size, tf.float32)
for f, f_ in zip(layers_[0], layers_[1]):
loss_a += tf.reduce_mean(
tf.subtract(instance_norm(f), instance_norm(f_))**2,
[1, 2, 3])
with tf.variable_scope('loss_b', reuse=tf.AUTO_REUSE):
loss_b = tf.zeros(batch_size, tf.float32)
for f, f_ in zip(layers_[2], layers_[3]):
loss_b += tf.reduce_mean(
tf.subtract(instance_norm(f), instance_norm(f_))**2,
[1, 2, 3])
return tf.reduce_mean(loss_a), tf.reduce_mean(loss_b)
def __init__(self):
self.vgg = vgg16.Vgg16()
self.input_holder = tf.placeholder(tf.float32,
[1, img_size, img_size, 3])
self.label_holder = tf.placeholder(tf.float32,
[label_size * label_size, 2])
self.label_noise = tf.placeholder(tf.float32,
[1, label_size, label_size, 1])
self.training = tf.placeholder(tf.bool, shape=[])
def load_model(name, path):
if (name == "vgg19"):
model = vgg19.Vgg19(get_data_mean(),
path,
trainable=False,
skippable=False)
if (name == "vgg16"):
model = vgg16.Vgg16(get_data_mean(),
path,
trainable=False,
skippable=False)
return model
def test():
with tf.Graph().as_default() as g:
x = tf.placeholder(
tf.float32,
[TEST_NUM, vgg16.IMAGE_SIZE, vgg16.IMAGE_SIZE, vgg16.NUM_CHANNELS])
y_ = tf.placeholder(tf.float32, [None, 2])
vgg = vgg16.Vgg16()
y = vgg.forward(x, False, None)
ema = tf.train.ExponentialMovingAverage(backward.MOVING_AVERAGE_DECAY)
ema_restore = ema.variables_to_restore()
saver = tf.train.Saver(ema_restore)
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
img_batch, label_batch = tfrecords_generateds.get_tfrecord(
TEST_NUM, isTrain=False)
while True:
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(backward.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split(
'/')[-1].split('-')[-1]
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess,
coord=coord)
score = []
for i in range(10):
xs, ys = sess.run([img_batch, label_batch])
reshaped_xs = np.reshape(
xs, (TEST_NUM, vgg16.IMAGE_SIZE, vgg16.IMAGE_SIZE,
vgg16.NUM_CHANNELS))
accuracy_score = sess.run(accuracy,
feed_dict={
x: reshaped_xs,
y_: ys
})
score.append(accuracy_score)
print("After %s training step(s), test accuracy = %g" %
(global_step, np.mean(score)))
# print("After %s training step(s), test accuracy = %g" % (global_step, accuracy_score))
coord.request_stop()
coord.join(threads)
else:
print('No checkpoint file found')
return
time.sleep(TEST_INTERVAL_SECS)
def main(args):
train_data = 'expression/fer2013-train'
test_data = 'expression/fer2013-test'
train_img, train_label = load_image(train_data)
test_img, test_label = load_image(test_data)
images = tf.placeholder(tf.float32, [None, 224, 224, 3])
label = tf.placeholder(tf.float32, [None, 7])
if args.network_model == "vgg16":
vgg = vgg16.Vgg16(args.fine_tuning)
else:
vgg = vgg19.Vgg19(args.fine_tuning)
predict = vgg.build(images)
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=label, logits=predict))
#if fine_tuning='all',we will use GradientDescent to optimize all variables
#if fine_tuning!='all',we will only use Adam to optimize the fully connected layers
if args.fine_tuning == "all":
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=0.001).minimize(cross_entropy)
else:
optimizer = tf.train.AdamOptimizer(
learning_rate=0.001).minimize(cross_entropy)
correct = tf.equal(tf.argmax(tf.nn.softmax(predict), 1),
tf.argmax(label, 1))
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(tf.all_variables())
train = next_batch(train_img, train_label, 32)
for i in range(args.train_step):
x_batch, y_batch = train.next()
loss, _, acc, = sess.run([cross_entropy, optimizer, accuracy],
feed_dict={
images: x_batch,
label: y_batch
})
if i % 10 == 0:
saver.save(sess, 'save_variables/vgg.module', global_step=i)
print('number %d loss is %f' % (i, loss))
print('number %d accuracy is %f' % (i, acc))
test_accuracy = 0
test = next_batch(test_img, test_label, 32)
for j in range(100):
x_batch, y_batch = test.next()
acc = sess.run(accuracy,
feed_dict={
images: x_batch,
label: y_batch
})
test_accuracy += acc
print('test accuracy is %f' % (test_accuracy / 100))
def run_train():
fout = open('inf.txt', 'w+')
test_config = ModelConfig()
test_config.keep_prob = 1.0
test_config.batch_size = 1
Session_config = tf.ConfigProto(allow_soft_placement=True)
Session_config.gpu_options.allow_growth = True
with tf.Graph().as_default(), tf.Session(config=Session_config) as sess:
with tf.device('/gpu:0'):
#if True:
initializer = tf.random_uniform_initializer(
-test_config.init_scale, test_config.init_scale)
train_model = vgg16.Vgg16(FLAGS.vgg16_file_path)
train_model.build(initializer)
data_test = dataset.DataSet(FLAGS.file_path_test,
FLAGS.data_root_dir,
TEST_SIZE,
is_train_set=False)
test_writer = tf.summary.FileWriter(FLAGS.log_dir + '/test')
saver = tf.train.Saver(max_to_keep=100)
last_epoch = load_model(sess, saver, FLAGS.saveModelPath,
train_model)
print('start: ', last_epoch + 1)
test_accury_1, test_accury_5, test_loss = run_epoch(
sess,
test_config.keep_prob,
fout,
test_config.batch_size,
train_model,
data_test,
tf.no_op(),
2,
test_writer,
istraining=False)
info = "Final: Test accury(top 1): %.4f Test accury(top 5): %.4f Loss %.4f" % (
test_accury_1, test_accury_5, test_loss)
print(info)
fout.write(info + '\n')
fout.flush()
test_writer.close()
print("Training step is compeleted!")
fout.close()
def set_global_xnet(gpuids):
#from wingdebugger import wingdbstub
#wingdbstub.Ensure()
#wingdbstub.debugger.StartDebug()
global xnet
gpuid = gpuids.get()
print('#############################', gpuids)
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = str(gpuid)
import vgg16
xnet = vgg16.Vgg16(
'/media/zhaoke/b0685ee4-63e3-4691-ae02-feceacff6996/models/vgg16_weights_tf_dim_ordering_tf_kernels.h5'
)
def get_feats():
vgg16_feats = np.zeros((len(unique_images), 4096))
with tf.Session() as sess:
images = tf.placeholder(dtype=tf.float32, shape=[None, 224, 224, 3])
vgg = vgg16.Vgg16()
vgg.build(images)
for i in range(len(unique_images)):
img_list = utils.load_image(unique_images[i])
batch = img_list.reshape((1, 224, 224, 3))
feature = sess.run(vgg.fc7, feed_dict={images: batch})#提取fc7层的特征
feature = np.reshape(feature, [4096])
feature /= norm(feature) # 特征归一化
vgg16_feats[i, :] = feature #每张图片的特征向量为1行
def test(batch):
with tf.Session(
config=tf.ConfigProto(gpu_options=(
tf.GPUOptions(per_process_gpu_memory_fraction=0.7)))) as sess:
images = tf.placeholder("float", [None, 224, 224, 3])
feed_dict = {images: batch}
vgg = vgg16.Vgg16()
with tf.name_scope("content_vgg"):
vgg.build(images)
prob = sess.run(vgg.prob, feed_dict=feed_dict)
print(prob)
utils.print_prob(prob[0], './synset.txt')
def app():
csvfile = open("./npy/submission.csv", "w")
Writer = csv.writer(csvfile, delimiter=',', lineterminator='\r\n')
Writer.writerow(['id', 'label'])
with tf.Graph().as_default() as g:
images = tf.placeholder(tf.float32, [1, 224, 224, 3])
vgg = vgg16.Vgg16()
y = vgg.forward(images, False, None)
y = tf.nn.softmax(y)
ema = tf.train.ExponentialMovingAverage(backward.MOVING_AVERAGE_DECAY)
ema_restore = ema.variables_to_restore()
saver = tf.train.Saver(ema_restore)
with tf.Session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
ckpt = tf.train.get_checkpoint_state(backward.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split('/')[-1].split(
'-')[-1]
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
for num in range(1, 5):
img_ready = utils.load_image(path + str(num) +
'.jpg') / 255.0
img_ready = img_ready.reshape((1, 224, 224, 3))
probability = sess.run(y, feed_dict={images: img_ready})
# print(probability, probability[0][0]+probability[0][1])
top5 = np.argsort(probability[0])[-1:-3:-1]
row = [num]
row.append(probability[0][1])
print(num, probability[0][1])
Writer.writerow(row)
csvfile.close()
print("write finished")
coord.request_stop()
coord.join(threads)
else:
print('No checkpoint file found')
def batch_vgg(training_tuple):
batch = np.concatenate(tuple(training_tuple), 0)
tf.reset_default_graph()
with tf.Session() as sess:
images = tf.placeholder("float", [training_size, 224, 224, 3])
feed_dict = {images: batch}
vgg = vgg16.Vgg16()
vgg.build(images)
prob = sess.run(vgg.prob, feed_dict=feed_dict)
print 'assign prob'
for x, y in zip(training_pic_name, prob):
pic_dict[x] = y
def main():
# load an img file, preprocess and reshape it
img_original, img_reshaped = load_img(
'./Black_Footed_Albatross_0001_796111.jpg')
# open a session
sess = open_sess("0")
# build a model
images = tf.placeholder("float", [1, config["img_W"], config["img_H"], 3])
vgg = vgg16.Vgg16()
vgg.build(images)
feed_dict = {images: img_reshaped}
prob = sess.run(vgg.prob, feed_dict=feed_dict)
#print(prob)
# get the class map value
target_class = [np.argmax(prob)] # the class that we're interested in
output = vgg.fc8 # the output(score) layer before the softmax
last_conv = vgg.conv5_3 # the last feature maps
label = tf.placeholder(tf.int64, [None], name='label')
classmap = get_classmap(label, output, last_conv, config["img_W"],
config["img_H"])
gradcam = sess.run(classmap,
feed_dict={
images: img_reshaped,
label: target_class
})
# draw the class map
plt.imshow(img_original.resize([config["img_W"], config["img_H"]]),
cmap='gray')
#plt.imshow(gradcam[0], cmap=plt.cm.seismic, alpha=0.5, interpolation='nearest') # normalized
plt.imshow(gradcam[0],
cmap=plt.cm.seismic,
alpha=0.5,
interpolation='nearest',
vmin=-0.4,
vmax=0.4)
plt.savefig('test.png', dpi=256)
plt.close()