def data_pipeline(param):
# load data
X_train, y_train = utils.load_data('./data/train.p')
X_valid, y_valid = utils.load_data('./data/valid.p')
X_test, y_test = utils.load_data('./data/test.p')
n_train = len(X_train)
n_test = len(X_test)
print("Number of training examples =", n_train)
print("Number of testing examples =", n_test)
image_shape = X_train.shape[1:]
print("Image data shape =", image_shape)
n_classes = np.max(y_train) + 1
print("Number of classes =", n_classes)
# data augmentation
X_train, y_train = utils.augment_data(X_train, y_train, param)
print("Number of augmented training examples =", len(X_train))
print("Number of validation examples =", len(X_valid))
# pre-process
X_train = np.array(
[utils.pre_process(X_train[i]) for i in range(len(X_train))],
dtype=np.float32)
X_valid = np.array(
[utils.pre_process(X_valid[i]) for i in range(len(X_valid))],
dtype=np.float32)
X_test = np.array(
[utils.pre_process(X_test[i]) for i in range(len(X_test))],
dtype=np.float32)
return X_train, y_train, X_valid, y_valid, X_test, y_test
def _train_episode(self, episode):
"""
training phase
"""
state = self.env.reset()
state = pre_process(state).repeat(self.config.in_c, 1, 1)
loss, reward_sum, done = 0, 0, False
while not done:
action = random.randint(0, self.num_actions - 1) if random.random() >= 1 - self.epsilon \
else self._make_action(state, False)
next_state, reward, done, _ = self.env.step(action)
next_state = pre_process(next_state)
next_state = torch.cat([state[:3], next_state], dim=0)
reward_sum += self.reward_func(reward)
self.step += 1
self.memory.append((state, next_state, torch.LongTensor([action]),
torch.FloatTensor([reward]), torch.FloatTensor([done])))
state = next_state
if self.step >= self.config.observate_time:
loss = self._update_param()
self.update_step += 1
if self.update_step % self.config.update_target:
self.target.load_state_dict(self.model.state_dict())
if self.step <= self.eps_step:
self.epsilon -= (self.init_eps - self.final_eps) / self.eps_step
self.reward_list.append(reward_sum)
self.reward_mean = reward_sum if self.reward_mean is None else self.reward_mean * 0.99 + reward_sum * 0.01
if self.config.visdom:
self.visual.update_vis_line(episode - 1, [self.reward_mean], 'train', 'append')
log = {'Episode': episode, 'Reward_cur': reward_sum, 'Reward_mean': self.reward_mean, 'Loss': loss,
'Reward_{}'.format(self.config.display_interval): sum(self.reward_list)}
return log
def video_upload():
target = os.path.join(APP_ROOT, 'files/')
clean_data(target)
print(target)
if not os.path.isdir(target):
os.mkdir(target)
print(request.files.getlist("file"))
for upload in request.files.getlist("file"):
print(upload)
print("{} is the file name".format(upload.filename))
filename = upload.filename
print(filename + "ana henaaa")
ext = os.path.splitext(filename)[1]
if (ext == ".mp4"):
print("File supported moving on...")
else:
return render_template("Error.html",
message="""The application supports only mp4
videos, this format is not supported""")
destination = "".join([target, filename])
print("Accept incoming file:", filename)
print("Save it to:", destination)
upload.save(destination)
pre_process(target, destination, filename)
generate_video(target, filename)
return render_template("complete_video.html", value=filename)
def track(self, current_frame):
# for idx in range(len(frame_list)):
frame_gray = cv2.cvtColor(current_frame, cv2.COLOR_BGR2GRAY)
frame_gray = frame_gray.astype(np.float32)
# import ipdb;ipdb.set_trace()
Hi = self.Ai / self.Bi
fi = frame_gray[self.pos[1]:self.pos[1] + self.pos[3],
self.pos[0]:self.pos[0] + self.pos[2]]
fi = pre_process(fi)
Gi = Hi * np.fft.fft2(fi)
gi = linear_mapping(np.fft.ifft2(Gi))
# find the max self.pos...
max_pos = np.unravel_index(np.argmax(gi, axis=None), gi.shape)
# update the position...
self.pos[1] += max_pos[0] - gi.shape[0] // 2
self.pos[0] += max_pos[1] - gi.shape[1] // 2
# get the current fi..
fi = frame_gray[self.pos[1]:self.pos[1] + self.pos[3],
self.pos[0]:self.pos[0] + self.pos[2]]
fi = pre_process(fi)
# online update...
self.Ai = self.learning_rate * (self.G * np.conjugate(
np.fft.fft2(fi))) + (1 - self.learning_rate) * self.Ai
self.Bi = self.learning_rate * (np.fft.fft2(fi) * np.conjugate(
np.fft.fft2(fi))) + (1 - self.learning_rate) * self.Bi
return self.pos
def main():
env = gym.make(args.env_name)
env.seed(500)
torch.manual_seed(500)
img_shape = env.observation_space.shape
num_actions = 3
print('image size:', img_shape)
print('action size:', num_actions)
net = QNet(num_actions)
net.load_state_dict(torch.load(args.save_path + 'model.pth'))
net.to(device)
net.eval()
epsilon = 0
for e in range(5):
done = False
score = 0
state = env.reset()
state = pre_process(state)
state = torch.Tensor(state).to(device)
history = torch.stack((state, state, state, state))
for i in range(3):
action = env.action_space.sample()
state, reward, done, info = env.step(action)
state = pre_process(state)
state = torch.Tensor(state).to(device)
state = state.unsqueeze(0)
history = torch.cat((state, history[:-1]), dim=0)
while not done:
if args.render:
env.render()
steps += 1
qvalue = net(history.unsqueeze(0))
action = get_action(0, qvalue, num_actions)
next_state, reward, done, info = env.step(action + 1)
next_state = pre_process(next_state)
next_state = torch.Tensor(next_state).to(device)
next_state = next_state.unsqueeze(0)
next_history = torch.cat((next_state, history[:-1]), dim=0)
score += reward
history = next_history
print('{} episode | score: {:.2f}'.format(e, score))
def _pre_training(self, init_frame, G):
height, width = G.shape
fi = cv2.resize(init_frame, (width, height))
# pre-process img..
fi = pre_process(fi)
Ai = G * np.conjugate(np.fft.fft2(fi))
Bi = np.fft.fft2(init_frame) * np.conjugate(np.fft.fft2(init_frame))
for _ in range(self.args.num_pretrain):
if self.args.rotate:
fi = pre_process(random_warp(init_frame))
else:
fi = pre_process(init_frame)
Ai = Ai + G * np.conjugate(np.fft.fft2(fi))
Bi = Bi + np.fft.fft2(fi) * np.conjugate(np.fft.fft2(fi))
return Ai, Bi
def pre_train(self, training_img):
init_frame = cv2.cvtColor(training_img, cv2.COLOR_BGR2GRAY)
init_frame = init_frame.astype(np.float32)
# Select Object to Track [x, y, width, height]
init_gt = cv2.selectROI('initial_img', training_img, False, False)
init_gt = np.array(init_gt).astype(np.int64)
# Compute Gaussian Response
g = np.zeros((init_frame.shape[0:2])).astype(np.float32)
g[(init_gt[1] + init_gt[3] // 2), (init_gt[0] + init_gt[2] // 2)] = 1.0
gaussian_response = cv2.GaussianBlur(g,
(0, 0), self.sigma) * init_frame
# start to create the training set ...
# get the goal..
g = gaussian_response[init_gt[1]:init_gt[1] + init_gt[3],
init_gt[0]:init_gt[0] + init_gt[2]]
init_frame = init_frame[init_gt[1]:init_gt[1] + init_gt[3],
init_gt[0]:init_gt[0] + init_gt[2]]
self.G = np.fft.fft2(g)
# start to do the pre-training...
Ai = np.zeros(self.G.shape)
Bi = np.zeros(self.G.shape)
for _ in range(self.num_pretrain):
fi = pre_process(init_frame)
Ai = Ai + self.G * np.conjugate(np.fft.fft2(fi))
Bi = Bi + np.fft.fft2(fi) * np.conjugate(np.fft.fft2(fi))
self.Ai = Ai * self.learning_rate
self.Bi = Bi * self.learning_rate
self.pos = init_gt.copy()
def _pre_training(self, init_frame, G):
height, width = G.shape
fi = cv2.resize(init_frame, (width, height))
# pre-process img..
fi = pre_process(fi)
Ai = G * np.conjugate(np.fft.fft2(fi))
Bi = np.fft.fft2(init_frame) * np.conjugate(np.fft.fft2(init_frame))
for _ in range(self.args.num_pretrain):
if self.args.rotate:
fi = pre_process(random_warp(init_frame))
else:
fi = pre_process(init_frame)
Ai = Ai + G * np.conjugate(np.fft.fft2(fi))
Bi = Bi + np.fft.fft2(fi) * np.conjugate(np.fft.fft2(fi))
return Ai, Bi
def exp_web_images():
# parameters
param = ExperimentParam(n_rows=32, n_cols=32, n_channels=3, n_classes=43)
# load data
folder = './from-web'
X, y = load_web_images(folder, param)
X = np.array([utils.pre_process(X[i]) for i in range(len(X))],
dtype=np.float32)
# load model
model_fname = param._model_fname
net, sess = load_model(model_fname, param)
preds, softmax = sess.run([net._preds, net._softmax], {
net._X: X,
net._is_training: False
})
accuracy = utils.classification_accuracy(y, preds)
print('Accuracy on web images: ', accuracy)
print('labels: ', y)
print('predictions: ', preds)
# top softmax
topk = sess.run(tf.nn.top_k(tf.constant(softmax), k=3))
print(topk)
def telemetry(sid, data):
if data:
# The current steering angle of the car
steering_angle = data["steering_angle"]
# The current throttle of the car
throttle = data["throttle"]
# The current speed of the car
speed = data["speed"]
# The current image from the center camera of the car
imgString = data["image"]
image = Image.open(BytesIO(base64.b64decode(imgString)))
image_array = utils.pre_process(image)
image_array = utils.to_keras(image_array)
steering_angle = float(
model.predict(image_array[None, :, :, :], batch_size=1))
throttle = controller.update(float(speed))
print(steering_angle, throttle)
send_control(steering_angle, throttle)
# save frame
if args.image_folder != '':
timestamp = datetime.utcnow().strftime('%Y_%m_%d_%H_%M_%S_%f')[:-3]
image_filename = os.path.join(args.image_folder, timestamp)
image.save('{}.jpg'.format(image_filename))
else:
# NOTE: DON'T EDIT THIS.
sio.emit('manual', data={}, skip_sid=True)
def test_pre_process(self):
tokens = nltk.word_tokenize(text)
res = utils.pre_process(tokens)
self.assertNotIn("the", res, msg="stopwords not excluded!!!")
self.assertIn("student", res, msg="student in text but not in result!!!")
def main(out_video_path):
# init camera
print("[INFO] starting video stream...")
vs = VideoStream(src=1).start()
time.sleep(2.0)
# use VideoWriter object to save video
fourcc = cv2.VideoWriter_fourcc(*'XVID')
out = cv2.VideoWriter(out_video_path, fourcc, 20.0, (1920, 1080))
# begin detect
while True:
frame = vs.read()
# get all 6 poly
opened = pre_process(frame)
poly_6_contours = find_poly(frame, opened, side_num=6)
# get all circle
circle_contours = find_circle(frame, opened)
# draw contours
cv2.drawContours(frame, poly_6_contours, -1, (255, 0, 0), 2)
cv2.drawContours(frame, circle_contours, -1, (0, 255, 0), 2)
# find momentum
for cnt in poly_6_contours:
center_x, center_y = get_momentum(cnt)
cv2.circle(frame, (center_x, center_y), 3, 128, -1) # 绘制中心点
for cnt in circle_contours:
center_x, center_y = get_momentum(cnt)
cv2.circle(frame, (center_x, center_y), 3, 128, -1)
# perspective trans
if len(poly_6_contours) >= 4: # enough
# find 4 max area poly 6
# perspective trans
# find point
pass
else: # not enough
pass
# imshow
cv2.imshow("frame", frame)
out.write(frame)
# break
key = cv2.waitKey(1)
if key == 27:
break
out.release()
cv2.destroyAllWindows()
vs.stop()
def forward(self, y, ret_edge=False):
yp, params = utils.pre_process(y, self.tkargs[1], self.eval())
z = ST(self.A[0](yp), self.tau[0])
for i in range(1, self.iters):
if ((i - 1) % self.edge_freq) == 0:
edge = self.topK(z)
r = self.B[i](z, edge) - yp
z = ST(z - self.A[i](r, edge), self.tau[i])
edge = self.topK(z)
xphat = self.D(z, edge)
xhat = utils.post_process(xphat, params)
if ret_edge:
return xhat, edge
return xhat
def train(dataset, output):
raw_text = open(dataset).read()
pre_text = utils.pre_process(raw_text)
char_map = utils.map_chars_to_int(pre_text)
params = {
"seq_length": 80,
"n_chars": len(pre_text),
"n_vocab": len(char_map)
}
X = []
y = []
X, y = utils.prepare_dset(pre_text, char_map, params)
params["n_patterns"] = len(X)
X = np.reshape(X, (params["n_patterns"], params["seq_length"], 1))
y = encode(y)
model_params = {
"LSTM-1": 512,
"LSTM-2": 256,
"Dropout-1": 0.3,
"Dropout-2": 0.2,
"activation": "softmax",
"loss": "categorical_crossentropy",
"optimizer": "adam",
"epochs": 100,
"batch_size": 32
}
model = create_model(X, y, model_params)
filepath = os.path.join(
output, "weights-improvement-{epoch:02d}-{loss:.4f}-bigger.hdf5")
checkpoint = ModelCheckpoint(filepath,
monitor='loss',
verbose=1,
save_best_only=True,
mode='min')
callbacks_list = [checkpoint]
model.compile(loss=model_params["loss"],
optimizer=model_params["optimizer"])
model.fit(X,
y,
epochs=model_params["epochs"],
batch_size=model_params["batch_size"],
callbacks=callbacks_list)
with open("model-opm.json", "w") as json_file:
json_file.write(model.to_json())
def _pre_training(self, init_frame, G):
# G 的大小就是选中的目标区域的大小
height, width = G.shape
fi = cv2.resize(init_frame, (width, height))
# pre-process img..
fi = pre_process(fi)
# np.fft.fft2 表示求 fi 的傅立叶变换
# np.conjugate 表示求矩阵的共轭
# 比如 g = np.matrix('[1+2j, 2+3j; 3-2j, 1-4j]')
# g.conjugate 为 matrix([[1-2j, 2-3j],[3+2j, 1+4j]])
Ai = G * np.conjugate(np.fft.fft2(fi))
Bi = np.fft.fft2(init_frame) * np.conjugate(np.fft.fft2(init_frame))
# 对 fi 进行多次刚性形变，增强检测的鲁棒性，计算出 Ai 和 Bi 的初始值
for _ in range(self.args.num_pretrain):
if self.args.rotate:
fi = pre_process(random_warp(init_frame))
else:
fi = pre_process(init_frame)
Ai = Ai + G * np.conjugate(np.fft.fft2(fi))
Bi = Bi + np.fft.fft2(fi) * np.conjugate(np.fft.fft2(fi))
return Ai, Bi
def gen_new_train(param):
# load data
X_train, y_train = utils.load_data('./data/train.p')
# data augmentation
X_train, y_train = utils.augment_data(X_train, y_train, param)
# pre-process
X_train = np.array(
[utils.pre_process(X_train[i]) for i in range(len(X_train))],
dtype=np.float32)
# one hot
oh_y_train = utils.one_hot_encode(y_train)
return X_train, y_train, oh_y_train
def exp_test_data():
# parameters
param = ExperimentParam(n_rows=32, n_cols=32, n_channels=3, n_classes=43)
# load data
X_test, y_test = utils.load_data('./data/test.p')
X_test = np.array(
[utils.pre_process(X_test[i]) for i in range(len(X_test))],
dtype=np.float32)
# load model
model_fname = param._model_fname
#preds = sess.run(net._preds, {net._X:X_test, net._is_training:False})
accuracy, _ = test_model(model_fname, param, X_test, y_test)
print('Test accuracy: ', accuracy)
def test(valid_queue, model, num_samples, args, logging):
if args.distributed:
dist.barrier()
nelbo_avg = utils.AvgrageMeter()
neg_log_p_avg = utils.AvgrageMeter()
model.eval()
for step, x in enumerate(valid_queue):
x = x[0] if len(x) > 1 else x
x = x.float().cuda()
# change bit length
x = utils.pre_process(x, args.num_x_bits)
with torch.no_grad():
nelbo, log_iw = [], []
for k in range(num_samples):
logits, log_q, log_p, kl_all, _ = model(x)
output = model.decoder_output(logits)
recon_loss = utils.reconstruction_loss(output,
x,
crop=model.crop_output)
balanced_kl, _, _ = utils.kl_balancer(kl_all, kl_balance=False)
nelbo_batch = recon_loss + balanced_kl
nelbo.append(nelbo_batch)
log_iw.append(
utils.log_iw(output,
x,
log_q,
log_p,
crop=model.crop_output))
nelbo = torch.mean(torch.stack(nelbo, dim=1))
log_p = torch.mean(
torch.logsumexp(torch.stack(log_iw, dim=1), dim=1) -
np.log(num_samples))
nelbo_avg.update(nelbo.data, x.size(0))
neg_log_p_avg.update(-log_p.data, x.size(0))
utils.average_tensor(nelbo_avg.avg, args.distributed)
utils.average_tensor(neg_log_p_avg.avg, args.distributed)
if args.distributed:
# block to sync
dist.barrier()
logging.info('val, step: %d, NELBO: %f, neg Log p %f', step, nelbo_avg.avg,
neg_log_p_avg.avg)
return neg_log_p_avg.avg, nelbo_avg.avg
def test_document_features(self):
tokens = nltk.word_tokenize(text)
all_processed_words = utils.pre_process(tokens)
all_word_freq = nltk.FreqDist(all_processed_words)
word_features = list(all_word_freq)
document_features = utils.wrap_document_features(word_features)
res = document_features(text)
for name, val in res.iteritems():
if val:
s = name.replace('contains(', '').replace(')', '')
self.assertIn(s, all_processed_words, msg="Does not contain the word feature:'{word}' from text given".format(word=s))
def run(self):
super(EnvWorker, self).run()
episode = 0
steps = 0
score = 0
life = 5
dead = False
while True:
if self.render:
self.env.render()
action = self.child_conn.recv()
next_state, reward, done, info = self.env.step(action + 1)
if life > info['ale.lives']:
dead = True
life = info['ale.lives']
next_state = pre_process(next_state)
self.history = np.moveaxis(next_state, -1, 0)
steps += 1
score += reward
self.child_conn.send([deepcopy(self.history), reward, dead, done])
if done and dead:
# print('{} episode | score: {:2f} | steps: {}'.format(
# episode, score, steps
# ))
episode += 1
steps = 0
score = 0
dead = False
life = 5
self.init_state()
if dead:
dead = False
self.init_state()
def data_generator(df, batch_size=128):
"""
yields a pair (X, Y) where X and Y are both numpy arrays of length `batch_size`
"""
n_rows = df.shape[0]
while True:
# Shuffle the data frame rows after every complete cycle through the data
df = df.sample(frac=1).reset_index(drop=True)
for index in range(0, n_rows, batch_size):
df_batch = df[index:index + batch_size]
# Ignoring the last batch which is smaller than the requested batch size
if (df_batch.shape[0] == batch_size):
X_batch = np.array([
pre_process(get_image(row))
for i, row in df_batch.iterrows()
])
y_batch = np.array(
[row['angle'] for i, row in df_batch.iterrows()])
yield X_batch, y_batch
def show_bad_cases(test_fname, param):
# load test data
X_test, y_test = utils.load_data(test_fname)
X_test_normed = np.array(
[utils.pre_process(X_test[i]) for i in range(len(X_test))],
dtype=np.float32)
n_data, n_rows, n_cols, n_channels = X_test.shape
param._n_rows = n_rows
param._n_cols = n_cols
param._n_channels = n_channels
# load model
n_classes = int(np.max(y_test) + 1)
tf.reset_default_graph()
net = network.TrafficSignNet(n_classes, param)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
saver.restore(sess, param._model_fname)
# test
preds_test = sess.run(net._preds, {
net._X: X_test_normed,
net._is_training: False
})
test_accuracy = utils.classification_accuracy(y_test, preds_test)
print('test accuracy: ', test_accuracy)
sess.close()
X_test_normed = None
# show test images that are not correctly classified
err_indices = np.where(preds_test != y_test)[0]
utils.show_images(X_test,
y_test,
err_indices,
n_cols=5,
num_images=200,
preds=preds_test)
def generate(dataset, weights, json_path):
raw_text = open(dataset).read()
pre_text = utils.pre_process(raw_text)
char_map = utils.map_chars_to_int(pre_text)
int_map = utils.map_int_to_char(pre_text)
params = {
"seq_length":80,
"n_chars" : len(pre_text),
"n_vocab" : len(char_map)
}
testX, _ = utils.prepare_dset(pre_text, char_map, params)
params["n_patterns"] = len(testX)
with open(json_path, 'r') as json_file:
json_model = json_file.read()
model = model_from_json(json_model)
model.load_weights(weights)
model.compile(loss="categorical_crossentropy", optimizer="adam")
start = np.random.randint(0 , len(testX) - 1)
sentence = testX[start]
output = []
for i in range(1500):
x = np.reshape(sentence, (1, len(sentence), 1))
x = x / float(params["n_vocab"])
prediction = model.predict(x, verbose=0)
index = np.argmax(prediction)
result = int_map[index]
output.append(result)
sentence.append(index)
sentence = sentence[1:len(sentence)]
print("".join(output))
def test_pre_process(self):
# 3 1d_bins: (0, 85, 170, 255)
n_1d_bins = 3
resolution = 2
image = np.array(
[[[50, 50, 50], [0, 25, 100]],
[[0, 50, 100], [100, 150, 200]]], dtype=np.uint8
)
# note the (width, height, 1) size instead of (width, height)
expected_luminance = np.array([[[-75], [-94]],
[[-75], [26]]], dtype=int)
expected_ab_bins = np.array([[[4], [3]],
[[4], [4]]], dtype=int)
luminance, ab_bins = pre_process(image, resolution, n_1d_bins)
np.testing.assert_equal(luminance, expected_luminance)
np.testing.assert_equal(ab_bins, expected_ab_bins)
self.assertEqual(luminance.dtype, expected_luminance.dtype)
self.assertEqual(ab_bins.dtype, expected_ab_bins.dtype)
def forward(self, x, ret_edge=False):
if ret_edge:
edge_list = []
x, params = utils.pre_process(x, self.tkargs[1], self.eval())
z = torch.cat([self.PPCONV[i](self.INCONV[i](x)) for i in range(3)],
dim=1)
hiz = self.HPF(z)
for i in range(self.iters):
print(f"i = {i}")
z0 = (1 - self.alpha[i]) * z + self.beta[i] * hiz
z = self.LPF[i](z0, ret_edge=ret_edge)
if ret_edge:
z, edge = z
edge_list.append(edge)
z = z0 + z
z = (1 - self.alpha[-1]) * z + self.beta[-1] * hiz
edge = self.topK(z)
z = self.GCout(z, edge)
x = utils.post_process(x + z, params)
if ret_edge:
edge_list.append(edge)
return x, edge_list
return x
def main(img_path, side_num):
img = read_img(img_path)
opened = pre_process(img)
# find poly or circle
if side_num == 0:
approx_contours = find_circle(img, opened)
else:
approx_contours = find_poly(img, opened, side_num=side_num) # temp
# find momentums
momentum_img = get_blank_img(img.shape)
cv2.drawContours(momentum_img, approx_contours, -1, (0, 0, 0), 2)
for contour in approx_contours:
center_x, center_y = get_momentum(contour)
print("x,y: {},{}".format(center_x, center_y))
cv2.circle(momentum_img, (center_x, center_y), 7, 128, -1) # 绘制中心点
# show result
show_img(momentum_img, 'momentums')
cv2.imwrite("./image/momentum.png", momentum_img)
cv2.waitKey(0)
cv2.destroyAllWindows()
def generator(X_train, y_train, batch_size, training=False):
'''
Data generator used for training the model.
- Reades images from path
- Preprocess images
- Augments training data only.
'''
number_samples = len(y_train)
while 1:
shuffle(X_train, y_train)
for offset in range(0, number_samples, batch_size):
X, y = X_train[offset:offset + batch_size], y_train[offset:offset +
batch_size]
X = [utils.open_image(x) for x in X]
X = [utils.pre_process(x) for x in X]
if training:
for i, (image, label) in enumerate(zip(X, y)):
X[i] = utils.augment(image)
X = np.array([keras_image.img_to_array(x) for x in X])
yield shuffle(X, y)
def main():
# get playlist id
ids = utils.get_user_playlists(sp, login).id
# get dataframe with all tracks from playlist
playlist_tracks_df = pd.DataFrame()
for i in ids:
temp = utils.get_playlist_tracks(sp, i)
playlist_tracks_df = pd.concat([playlist_tracks_df, temp])
playlist_tracks_df = playlist_tracks_df.reset_index(drop=True)
# k-means clustering flow
preprocessor = utils.pre_process()
playlist_df_clusters = utils.train_model(preprocessor, playlist_tracks_df,
7)
clusters = utils.cluster_averages(playlist_df_clusters).sort_values(
'danceability')
clusters['duration_seconds'] = clusters.duration_ms / 1000 / 60
print(clusters)
return
# -*- coding: utf -8 -*-
import tensorflow as tf
from word2vec import comment_embedding, WordModel
import gensim.models.keyedvectors as word2vec
import numpy as np
from utils import pre_process
from underthesea import word_tokenize
text = ["San pham dep, minh rat thich chat vai nay!!!! 👍👍👍", "Làm ăn như ccccc 😡😡😡", "<3", "💥💥💥Giao hàng nhanh.👏 💥💥💥", "🔥 <3", "Shop chán vkl", "Nchung là ổn, giao hàng nhanh kb cho ⭐⭐⭐⭐⭐"]
text = pre_process(text)
print(text)
for i in range(len(text)):
text[i] = word_tokenize(text[i])
print(text)
def start_tracking(self):
time_list = []
# get the image of the first frame... (read as gray scale image...)
init_img = cv2.imread(self.frame_lists[0])
init_frame = cv2.cvtColor(init_img, cv2.COLOR_BGR2GRAY)
init_frame = init_frame.astype(np.float32)
# get the init ground truth.. [x, y, width, height]
init_gt = cv2.selectROI('demo', init_img, False, False)
init_gt = np.array(init_gt).astype(np.int64)
# start to draw the gaussian response...
response_map = self._get_gauss_response(init_frame, init_gt)
# start to create the training set ...
# get the goal..
g = response_map[init_gt[1]:init_gt[1]+init_gt[3], init_gt[0]:init_gt[0]+init_gt[2]]
fi = init_frame[init_gt[1]:init_gt[1]+init_gt[3], init_gt[0]:init_gt[0]+init_gt[2]]
G = np.fft.fft2(g)
# start to do the pre-training...
Ai, Bi = self._pre_training(fi, G)
# start the tracking...
i=0
for idx in range(len(self.frame_lists)):
start = time.time()
current_frame = cv2.imread(self.frame_lists[idx])
frame_gray = cv2.cvtColor(current_frame, cv2.COLOR_BGR2GRAY)
frame_gray = frame_gray.astype(np.float32)
if idx == 0:
Ai = self.args.lr * Ai
Bi = self.args.lr * Bi
pos = init_gt.copy()
clip_pos = np.array([pos[0], pos[1], pos[0]+pos[2], pos[1]+pos[3]]).astype(np.int64)
else:
Hi = Ai / Bi
fi = frame_gray[clip_pos[1]:clip_pos[3], clip_pos[0]:clip_pos[2]]
fi = pre_process(cv2.resize(fi, (init_gt[2], init_gt[3])))
Gi = Hi * np.fft.fft2(fi)
gi = linear_mapping(np.fft.ifft2(Gi))
# find the max pos...
max_value = np.max(gi)
max_pos = np.where(gi == max_value)
dy = int(np.mean(max_pos[0]) - gi.shape[0] / 2)
dx = int(np.mean(max_pos[1]) - gi.shape[1] / 2)
# update the position...
pos[0] = pos[0] + dx
pos[1] = pos[1] + dy
# trying to get the clipped position [xmin, ymin, xmax, ymax]
clip_pos[0] = np.clip(pos[0], 0, current_frame.shape[1])
clip_pos[1] = np.clip(pos[1], 0, current_frame.shape[0])
clip_pos[2] = np.clip(pos[0]+pos[2], 0, current_frame.shape[1])
clip_pos[3] = np.clip(pos[1]+pos[3], 0, current_frame.shape[0])
clip_pos = clip_pos.astype(np.int64)
# get the current fi..
fi = frame_gray[clip_pos[1]:clip_pos[3], clip_pos[0]:clip_pos[2]]
fi = pre_process(cv2.resize(fi, (init_gt[2], init_gt[3])))
# online update...
Ai = self.args.lr * (G * np.conjugate(np.fft.fft2(fi))) + (1 - self.args.lr) * Ai
Bi = self.args.lr * (np.fft.fft2(fi) * np.conjugate(np.fft.fft2(fi))) + (1 - self.args.lr) * Bi
# visualize the tracking process...
cv2.rectangle(current_frame, (pos[0], pos[1]), (pos[0]+pos[2], pos[1]+pos[3]), (255, 0, 0), 2)
#out.write(current_frame)
cv2.imshow('demo11', current_frame)
cv2.imwrite('goog/'+str(i)+'.jpg', current_frame)
i += 1
cv2.waitKey(10)
# if record... save the frames..
# if self.args.record:
# frame_path = 'record_frames/' + self.img_path.split('/')[1] + '/'
# if not os.path.exists(frame_path):
# os.mkdir(frame_path)
# cv2.imwrite(frame_path + str(idx).zfill(5) + '.png', current_frame)
# #out.write(current_frame)
end = time.time()
time_list.append(end-start)
out.release()
print('视频写入成功！')
return time_list
selected_VGG_layer_weights, selected_VGG_layers)
encoded = keras.models.Model(
inputs=inputs,
outputs=model.get_layer('encoder').get_layer('encoded').output)
return model, encoder, decoder, lossModel, encoded
if __name__ == '__main__':
nc = Dataset('../Data/tohoku_2020.nc', 'r')
Z = np.flip(np.array(nc.variables['elevation']), axis=0)
x = 96
y = 96
X_train, X_val, X_test = pre_process(Z, x, y)
#build model
inputs = Input(shape=(y, x, 1), name='encoder_inputs')
# type:
#'AE'=autoencoder
#'VAE'=variational autoencoder
#'DFC_VAE'=deep feature consistent variational autoencoder
model, encoder, decoder, lossModel, encoded = build_model(type='DFC_VAE')
#hyper parameters
epochs = 50
batch_size = 128
learning_rate = 1e-6
def train(train_queue, model, cnn_optimizer, grad_scalar, global_step,
warmup_iters, writer, logging):
alpha_i = utils.kl_balancer_coeff(num_scales=model.num_latent_scales,
groups_per_scale=model.groups_per_scale,
fun='square')
nelbo = utils.AvgrageMeter()
model.train()
for step, x in enumerate(train_queue):
x = x[0] if len(x) > 1 else x
x = x.half().cuda()
# change bit length
x = utils.pre_process(x, args.num_x_bits)
# warm-up lr
if global_step < warmup_iters:
lr = args.learning_rate * float(global_step) / warmup_iters
for param_group in cnn_optimizer.param_groups:
param_group['lr'] = lr
# sync parameters, it may not be necessary
if step % 100 == 0:
utils.average_params(model.parameters(), args.distributed)
cnn_optimizer.zero_grad()
with autocast():
logits, log_q, log_p, kl_all, kl_diag = model(x)
output = model.decoder_output(logits)
kl_coeff = utils.kl_coeff(
global_step, args.kl_anneal_portion * args.num_total_iter,
args.kl_const_portion * args.num_total_iter,
args.kl_const_coeff)
recon_loss = utils.reconstruction_loss(output,
x,
crop=model.crop_output)
balanced_kl, kl_coeffs, kl_vals = utils.kl_balancer(
kl_all, kl_coeff, kl_balance=True, alpha_i=alpha_i)
nelbo_batch = recon_loss + balanced_kl
loss = torch.mean(nelbo_batch)
norm_loss = model.spectral_norm_parallel()
bn_loss = model.batchnorm_loss()
# get spectral regularization coefficient (lambda)
if args.weight_decay_norm_anneal:
assert args.weight_decay_norm_init > 0 and args.weight_decay_norm > 0, 'init and final wdn should be positive.'
wdn_coeff = (1. - kl_coeff) * np.log(
args.weight_decay_norm_init) + kl_coeff * np.log(
args.weight_decay_norm)
wdn_coeff = np.exp(wdn_coeff)
else:
wdn_coeff = args.weight_decay_norm
loss += norm_loss * wdn_coeff + bn_loss * wdn_coeff
grad_scalar.scale(loss).backward()
utils.average_gradients(model.parameters(), args.distributed)
grad_scalar.step(cnn_optimizer)
grad_scalar.update()
nelbo.update(loss.data, 1)
if (global_step + 1) % 100 == 0:
if (global_step + 1) % 1000 == 0: # reduced frequency
n = int(np.floor(np.sqrt(x.size(0))))
x_img = x[:n * n]
output_img = output.mean if isinstance(
output, torch.distributions.bernoulli.Bernoulli
) else output.sample()
output_img = output_img[:n * n]
x_tiled = utils.tile_image(x_img, n)
output_tiled = utils.tile_image(output_img, n)
in_out_tiled = torch.cat((x_tiled, output_tiled), dim=2)
writer.add_image('reconstruction', in_out_tiled, global_step)
# norm
writer.add_scalar('train/norm_loss', norm_loss, global_step)
writer.add_scalar('train/bn_loss', bn_loss, global_step)
writer.add_scalar('train/norm_coeff', wdn_coeff, global_step)
utils.average_tensor(nelbo.avg, args.distributed)
logging.info('train %d %f', global_step, nelbo.avg)
writer.add_scalar('train/nelbo_avg', nelbo.avg, global_step)
writer.add_scalar(
'train/lr',
cnn_optimizer.state_dict()['param_groups'][0]['lr'],
global_step)
writer.add_scalar('train/nelbo_iter', loss, global_step)
writer.add_scalar('train/kl_iter', torch.mean(sum(kl_all)),
global_step)
writer.add_scalar(
'train/recon_iter',
torch.mean(
utils.reconstruction_loss(output,
x,
crop=model.crop_output)),
global_step)
writer.add_scalar('kl_coeff/coeff', kl_coeff, global_step)
total_active = 0
for i, kl_diag_i in enumerate(kl_diag):
utils.average_tensor(kl_diag_i, args.distributed)
num_active = torch.sum(kl_diag_i > 0.1).detach()
total_active += num_active
# kl_ceoff
writer.add_scalar('kl/active_%d' % i, num_active, global_step)
writer.add_scalar('kl_coeff/layer_%d' % i, kl_coeffs[i],
global_step)
writer.add_scalar('kl_vals/layer_%d' % i, kl_vals[i],
global_step)
writer.add_scalar('kl/total_active', total_active, global_step)
global_step += 1
utils.average_tensor(nelbo.avg, args.distributed)
return nelbo.avg, global_step
shape=[
None,
],
name='image_lables_tensor')
images_path_tensor_val = tf.placeholder(tf.string,
shape=[
None,
],
name='images_path_tensor_val')
# # A hack to add validation accuracy in tensorboard
val_accuracy = tf.placeholder(tf.double, shape=[], name='val_accuracy')
# Training
print('[INFO]: getting training model')
net = Network(FLAGS)
images_tensor = pre_process(images_path_tensor, FLAGS)
_print_shape = tf.Print(images_tensor, [tf.shape(images_tensor)],
message="[INFO] current train batch shape: ",
first_n=1)
with tf.control_dependencies([_print_shape]):
train = net(images_tensor, images_label_tensor)
# Validation
val_image_tensor = pre_process(images_path_tensor_val, FLAGS, mode='val')
_print_val_shape = tf.Print(val_image_tensor, [tf.shape(val_image_tensor)],
message="[INFO] current val batch shape: ",
first_n=1)
with tf.control_dependencies([_print_val_shape]):
val_forward_pass = net.forward_pass(val_image_tensor)
# Add summaries
def start_tracking(self):
# get the image of the first frame... (read as gray scale image...)
init_img = cv2.imread(self.frame_lists[0])
init_frame = cv2.cvtColor(init_img, cv2.COLOR_BGR2GRAY)
init_frame = init_frame.astype(np.float32)
# get the init ground truth.. [x, y, width, height]
init_gt = cv2.selectROI('demo', init_img, False, False)
init_gt = np.array(init_gt).astype(np.int64)
# start to draw the gaussian response...
response_map = self._get_gauss_response(init_frame, init_gt)
# start to create the training set ...
# get the goal..
g = response_map[init_gt[1]:init_gt[1]+init_gt[3], init_gt[0]:init_gt[0]+init_gt[2]]
fi = init_frame[init_gt[1]:init_gt[1]+init_gt[3], init_gt[0]:init_gt[0]+init_gt[2]]
G = np.fft.fft2(g)
# start to do the pre-training...
Ai, Bi = self._pre_training(fi, G)
# start the tracking...
for idx in range(len(self.frame_lists)):
current_frame = cv2.imread(self.frame_lists[idx])
frame_gray = cv2.cvtColor(current_frame, cv2.COLOR_BGR2GRAY)
frame_gray = frame_gray.astype(np.float32)
if idx == 0:
Ai = self.args.lr * Ai
Bi = self.args.lr * Bi
pos = init_gt.copy()
clip_pos = np.array([pos[0], pos[1], pos[0]+pos[2], pos[1]+pos[3]]).astype(np.int64)
else:
Hi = Ai / Bi
fi = frame_gray[clip_pos[1]:clip_pos[3], clip_pos[0]:clip_pos[2]]
fi = pre_process(cv2.resize(fi, (init_gt[2], init_gt[3])))
Gi = Hi * np.fft.fft2(fi)
gi = linear_mapping(np.fft.ifft2(Gi))
# find the max pos...
max_value = np.max(gi)
max_pos = np.where(gi == max_value)
dy = int(np.mean(max_pos[0]) - gi.shape[0] / 2)
dx = int(np.mean(max_pos[1]) - gi.shape[1] / 2)
# update the position...
pos[0] = pos[0] + dx
pos[1] = pos[1] + dy
# trying to get the clipped position [xmin, ymin, xmax, ymax]
clip_pos[0] = np.clip(pos[0], 0, current_frame.shape[1])
clip_pos[1] = np.clip(pos[1], 0, current_frame.shape[0])
clip_pos[2] = np.clip(pos[0]+pos[2], 0, current_frame.shape[1])
clip_pos[3] = np.clip(pos[1]+pos[3], 0, current_frame.shape[0])
clip_pos = clip_pos.astype(np.int64)
# get the current fi..
fi = frame_gray[clip_pos[1]:clip_pos[3], clip_pos[0]:clip_pos[2]]
fi = pre_process(cv2.resize(fi, (init_gt[2], init_gt[3])))
# online update...
Ai = self.args.lr * (G * np.conjugate(np.fft.fft2(fi))) + (1 - self.args.lr) * Ai
Bi = self.args.lr * (np.fft.fft2(fi) * np.conjugate(np.fft.fft2(fi))) + (1 - self.args.lr) * Bi
# visualize the tracking process...
cv2.rectangle(current_frame, (pos[0], pos[1]), (pos[0]+pos[2], pos[1]+pos[3]), (255, 0, 0), 2)
cv2.imshow('demo', current_frame)
cv2.waitKey(100)
# if record... save the frames..
if self.args.record:
frame_path = 'record_frames/' + self.img_path.split('/')[1] + '/'
if not os.path.exists(frame_path):
os.mkdir(frame_path)
cv2.imwrite(frame_path + str(idx).zfill(5) + '.png', current_frame)
