def upload() -> dict:
cli_uuid = unquote(get_arg("uuid"))
print("post from:", cli_uuid)
if len(cli_uuid) == 2:
if os.path.exists(user_dir + cli_uuid):
length = int(request.headers.get('content-length'))
print("准备接收:", length, "bytes")
if length < MAXBUFFSZ:
return save_img(request.get_data(), cli_uuid, image_dir, info_json_path)
else:
return save_img(request.stream.read(length), cli_uuid, image_dir, info_json_path)
else: return {"stat":"noid"}
else: return {"stat":"invid"}
def run(self) -> None:
while True:
r = self.p.request('GET', RANDOM_IMG_API, preload_content=False)
data = r.read()
print(save_img(data, "涩酱", image_dir, json_dir))
r.release_conn()
sleep(DELAY * THREAD_NUM)
def run_infer(weights_file, labels_file, image_path, out_filename):
model = YOLOv4Model()
model.load_weights(weights_file)
img, input = read_img(image_path, 608)
cls_names = open(labels_file, "r").read().split("\n")
boxes, scores, labels = inference.infer(model, cls_names, input)
pixels = add_bboxes(img, boxes, scores, labels)
if out_filename:
save_img(out_filename, pixels)
else:
draw_img(pixels)
async def upload() -> dict:
cli_uuid = unquote(get_arg("uuid"))
print("post from:", cli_uuid)
if len(cli_uuid) == 2:
if os.path.exists(user_dir + cli_uuid):
return save_img(await request.get_data(), cli_uuid, image_dir, info_json_path)
else: return {"stat":"noid"}
else: return {"stat":"invid"}
def run_infer(weights_file, labels_file, image_path, out_filename):
cls_names = open(labels_file, "r").read().split("\n")
model = YOLOv4Model()
model.load_weights(weights_file)
img, input = read_img(image_path, 608)
prediction = model.predict(input)
boxes, scores, labels = inference.decode_prediction(prediction, len(cls_names))
labels = [cls_names[cls] for cls in labels]
pixels = add_bboxes(img, boxes, scores, labels)
if out_filename:
save_img(out_filename, pixels)
else:
draw_img(pixels)
def upform() -> dict:
cli_uuid = unquote(get_arg("uuid"))
print("post from:", cli_uuid)
if len(cli_uuid) == 2:
if os.path.exists(user_dir + cli_uuid):
re = []
for f in request.files.getlist("img"):
re.append({"name":f.filename, **save_img(f.read(), cli_uuid, image_dir, info_json_path)})
return {"result": re}
else: return {"stat":"noid"}
else: return {"stat":"invid"}
def load_and_train(options):
# unpack parameters
sty_imgs = options.sty_imgs
sty_weights = np.array(options.sty_weights)
cont_img = options.cont_img
output_file = options.output_file
output_scale = options.output_scale
learn_rate = options.learn_rate
alpha = np.float32(options.alpha)
beta = np.float32(options.beta)
num_epoch = options.num_epoch
vgg19_loc = options.vgg19_loc
# load images and vgg19 coefficients
sty_features = load_sty_features()
cont_feature = load_cont_feature()
cont = load_cont_img(cont_img, output_scale)
vgg_obj = vgg.VGG19(vgg19_loc)
cont_ten = comp_cont_ten(cont, cont_feature, vgg_obj)
gram, gram_coef = comp_sty_gram(load_sty_imgs(sty_imgs), sty_weights,
sty_features, vgg_obj)
# model
cont_remix = tf.Variable(cont)
vgg_obj.build_upto(cont_remix, 'pool5', False)
# style loss function
gamma = np.float32(1.0 / len(sty_features))
gram_style = {}
for style in sty_features:
this_shape = vgg_obj.layers[style].get_shape().as_list()
this_Ml = this_shape[1] * this_shape[2]
reshaped = tf.reshape(vgg_obj.layers[style], (-1, this_shape[3]))
gram_style[style] = tf.matmul(tf.transpose(reshaped),
reshaped) / (this_Ml**2)
loss_style = tf.constant(np.float32(0.0))
for style in sty_features:
loss_style += tf.reduce_sum(
tf.square(gram_style[style] - gram[style])) * gram_coef[style]
# content loss function
loss_content = tf.reduce_mean(
tf.square(vgg_obj.layers[cont_feature] - cont_ten))
# punish local pixel noise
loss_noise = tf.reduce_mean(
tf.abs(
tf.nn.max_pool(cont_remix,
ksize=[1, 3, 3, 1],
strides=[1, 1, 1, 1],
padding='VALID') -
tf.nn.max_pool(-cont_remix,
ksize=[1, 3, 3, 1],
strides=[1, 1, 1, 1],
padding='VALID')))
# train step
loss = gamma * loss_style + alpha * loss_content + beta * loss_noise
err = float('inf')
train_step = tf.train.AdamOptimizer(learn_rate).minimize(loss)
with tf.Session() as sess:
tf.initialize_all_variables().run()
for idx in range(num_epoch):
sess.run(train_step)
# list all errors
this_loss_content = alpha * loss_content.eval()
this_loss_style = gamma * loss_style.eval()
this_loss_noise = beta * loss_noise.eval()
this_err = this_loss_content + this_loss_style + this_loss_noise
print('epoch', idx, ': content loss', this_loss_content,
'style loss', this_loss_style, 'noise loss', this_loss_noise)
if this_err < err:
err = this_err
output = cont_remix.eval()[0, :, :, :]
# save image
img.save_img(output_file, img.revert_img(output))