def plot_model_exploration(model: ModelWrapper,
img: np.array,
min_z: float = -6,
max_z: float = 6,
num_steps: int = 10):
img_height, img_width, channels = model.output_shape
steps = np.linspace(-15, 15, num_steps)
canvas = np.zeros(
(model.latent_dim * img_height, num_steps * img_width, channels))
fig, ax = plt.subplots()
z = model.get_latent(img)
for j in range(model.latent_dim):
z_temp = np.copy(z)
for i, step in enumerate(steps):
z_temp[j] = step
reconstruction = model.get_reconstruction(z_temp)
canvas[j * img_width:(j + 1) * img_width,
i * img_height:(i + 1) * img_height] = reconstruction
ax.imshow(canvas)
start_range_y = img_height // 2
end_range_y = model.latent_dim * img_height + start_range_y
pixel_range_y = np.arange(start_range_y, end_range_y, img_height)
ax.set_yticks(pixel_range_y)
ax.set_yticklabels(list(range(model.latent_dim)), fontsize=8)
ax.set_xticks([])
ax.set_xticklabels([])
return fig, ax
def prepare_model(learning_rate, momentum, checkpoint_file):
"""Prepare a ResNet-34 model with CrossEntropyLoss and SGD.
Args:
learning_rate (float): The learning rate for SGD.
momentum (float): The momentum for SGD.
checkpoint_file (str or None): If not `None`, the path of the
checkpoint file to load.
Returns:
model.Model: The prepared model object.
"""
# Load model.
resnet = torchvision.models.resnet34()
resnet.conv1 = torch.nn.Conv2d(1,
64,
kernel_size=3,
stride=1,
padding=1,
bias=False)
resnet.avgpool = torch.nn.AvgPool2d(2)
# Prepare loss function and optimizer.
loss_function = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(resnet.parameters(),
lr=learning_rate,
momentum=momentum)
# Wrap model object and load checkpoint file if provided.
model = ModelWrapper(resnet, loss_function, optimizer)
if checkpoint_file:
model.load(checkpoint_file)
return model
def __init__(self, data_frame, label_name, feature_names,\
categorical_feature_names, k=5):
ModelWrapper.__init__(self, data_frame, \
label_name, feature_names, categorical_feature_names)
self.k = k
self._kfold = None
self._generate_kfold()
def train(args):
loss_fn = nn.CrossEntropyLoss(ignore_index=0)
adaptive = ModelWrapper(args, loss_fn, get_loader)
optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad,
adaptive.model.parameters()),
lr=args.learning_rate,
betas=(args.alpha, args.beta),
weight_decay=args.l2_rate)
if adaptive.train(optimizer, args):
return 0
def main(args):
loss_fn = nn.CrossEntropyLoss(ignore_index=0)
adaptive = ModelWrapper(args, loss_fn)
with open(args.word2idx_path, 'r') as fr:
word2idx = json.loads(fr.read())
with open(args.sememe2idx_path, 'r') as fr:
sememe2idx = json.loads(fr.read())
results = ResDataset(args.gen_file_path, word2idx, sememe2idx)
res_loader = data.DataLoader(dataset=results, batch_size=1, shuffle=False)
scores = adaptive.score(res_loader)
with codecs.open(args.output_path, 'w', 'utf-8') as fw:
fw.write('\n'.join(scores))
return 0
def __init__(self, environment: Env,
memory: AbstractMemory,
policy: AbstractPolicy,
model: K.Model,
optimizer: K.optimizers.Optimizer,
logger: Logger):
self.environment = environment
self.memory = memory
self.policy = policy
self.model = ModelWrapper(model, optimizer)
self.logger = logger
self.history = []
logger.create_settings_model_file(model)
logger.create_settings_agent_file(self)
def upload():
print 'uploading...'
fin = request.files['file']
data = fin.read()
user_id = request.form['openid']
user_dir = get_user_dir(user_id)
if not os.path.isdir(user_dir):
os.mkdir(user_dir) # TODO: delete zombie cache
img = cv2.imdecode(np.fromstring(data, dtype=np.uint8), cv2.IMREAD_COLOR)
predicted = ModelWrapper.predict(img)
assets = json.load(open(ASSETS_DIR + ASSETS_JSON_FILE))
resp_dict = assets[predicted.decode('utf-8')]
resp_dict.update({'predicted': predicted})
resp = flask.jsonify(resp_dict)
print predicted
print resp
with open(user_dir + CACHE_IMAGE_FILE, 'wb') as fout:
fout.write(data)
json_data = {
'predicted': predicted,
'description': resp_dict['description']
}
json.dump(json_data, open(user_dir + CACHE_JSON_FILE, 'w'))
return resp
def predict(batch: Batch):
# Remove empty texts
batch.texts = list(filter(lambda x: len(x) > 0, batch.texts))
model = ModelWrapper()
res = model(batch)
return {"documents": res}
def get(self, id):
m = ModelWrapper("pickle_model.pkl")
customer = data[data.cuid == id]
if len(customer) == 0:
print("Customer not found")
return {"conv": 0, "revenue": 0}
print("Customer found")
res = m.predict(customer)
print("res[0]", res[0], "res[1]", res[1])
mes = metrics(customer)
return {
"conv": res[0][0],
"revenue": str(res[1].values[0]),
"message": mes
}
def __init__(self):
""" Khởi động ứng dụng """
super(UiOutputDialog, self).__init__()
self.model_wrapper = ModelWrapper(model_path)
self.capture = None
self.timer = QTimer(self)
self.timer.timeout.connect(self.update_frame)
loadUi("window.ui", self)
self.holistic = mp.solutions.holistic.Holistic()
self.face = mp.solutions.face_mesh.FaceMesh(max_num_faces=100)
self.hand = mp.solutions.hands.Hands(max_num_hands=100)
now = QDate.currentDate()
current_date = now.toString('ddd dd MMMM yyyy')
current_time = datetime.datetime.now().strftime("%I:%M %p")
self.Date_Label.setText(current_date)
self.Time_Label.setText(current_time)
self.image = None
self.settings = UISettings()
self.Settings_Button.clicked.connect(
lambda: self.handle_setting_button())
self.warnings = []
self.prev_waring_code = -1
self.prev_waring_code_1 = -1
self.prev_waring_code_2 = -1
self.warning_history = UIWarnings()
self.Warning_History.clicked.connect(
lambda: self.handle_waring_history_button())
self.net = cv2.dnn.readNet("yolov3-tiny.weights", "yolov3-tiny.cfg")
with open("yolov3.txt", 'r') as f:
self.classes = [line.strip() for line in f.readlines()]
self.colors = np.random.uniform(0, 255, size=(len(self.classes), 3))
def test_can_fit_model(self):
""" This test check ability of fitting model in PER to random vector. """
state_shape = (4, )
action_space = 2
model = PrioritizedExperienceReplayTests._create_model(
state_shape, action_space)
PER = PrioritizedExperienceReplay(maxlen=1,
model=model,
key_scaling=10,
gamma=1)
model_wrapper = ModelWrapper(
model=model, optimizer=K.optimizers.Adam(learning_rate=0.01))
model_wrapper.compile()
sample = Sample(action=np.random.randint(0, action_space),
state=np.random.rand(state_shape[0]),
reward=10,
next_state=None)
PER.add(samples=[sample])
history_of_loss = []
fit_vector = np.zeros((action_space, ))
fit_vector[sample.action] = sample.reward
for _ in range(100):
model_wrapper.fit(sample.state, fit_vector)
history_of_loss.append(PER._loss_calculate(sample=sample))
for idx, loss in enumerate(history_of_loss[:-1]):
self.assertGreater(loss, history_of_loss[idx + 1])
def __init__(self,
environment: Env,
memory: AbstractMemory,
policy: AbstractPolicy,
model: K.Model,
logger: Logger,
gamma: float,
optimizer: K.optimizers.Optimizer,
n_step: int = 1):
self.model = ModelWrapper(model, optimizer)
#self.model.compile()
self.current_model = None
self.gamma = gamma
self.n_step = n_step
super(DQN, self).__init__(environment=environment,
memory=memory,
policy=policy,
model=model,
optimizer=optimizer,
logger=logger)
RNN(HIDDEN_SIZE),
layers.RepeatVector(3),
RNN(128, return_sequences=True),
layers.TimeDistributed(layers.Dense(len(CHARS), activation='softmax'))
])
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
model.summary()
train_generator = encode_generator(training_generator, BATCH_SIZE)
hist = model.fit_generator(train_generator,
steps_per_epoch=STEPS_PER_EPOCH,
epochs=EPOCHS,
verbose=1,
use_multiprocessing=True,
workers=-2,
callbacks=callbacks,
validation_data=train_generator, validation_steps=30)
score = model.evaluate_generator(encode_generator(
test_generator, BATCH_SIZE), steps=STEPS_PER_EPOCH)
print(score)
config = build_config(MODEL_NAME, LEARNING_RATE, BATCH_SIZE,
EPOCHS, STEPS_PER_EPOCH, score[0], score[1])
wrapper = ModelWrapper(model, config=config)
wrapper.save_model()
shuffle=True,
collate_fn=data.image_label_list_of_masks_collate_function)
validation_dataset_fid = DataLoader(
data.Places365(path_to_index_file=args.path_to_places365,
index_file_name='val.txt',
max_length=6000,
validation=True),
batch_size=args.batch_size,
num_workers=args.batch_size,
shuffle=False,
collate_fn=data.image_label_list_of_masks_collate_function)
validation_dataset = data.Places365(
path_to_index_file=args.path_to_places365, index_file_name='val.txt')
# Init model wrapper
model_wrapper = ModelWrapper(
generator=generator,
discriminator=discriminator,
vgg16=vgg16,
training_dataset=training_dataset,
validation_dataset=validation_dataset,
validation_dataset_fid=validation_dataset_fid,
generator_optimizer=generator_optimizer,
discriminator_optimizer=discriminator_optimizer)
# Perform training
if args.train:
model_wrapper.train(epochs=args.epochs, device=args.device)
# Perform testing
if args.test:
print('FID=', model_wrapper.validate(device=args.device))
model_wrapper.inference(device=args.device)
class UiOutputDialog(QDialog):
def __init__(self):
""" Khởi động ứng dụng """
super(UiOutputDialog, self).__init__()
self.model_wrapper = ModelWrapper(model_path)
self.capture = None
self.timer = QTimer(self)
self.timer.timeout.connect(self.update_frame)
loadUi("window.ui", self)
self.holistic = mp.solutions.holistic.Holistic()
self.face = mp.solutions.face_mesh.FaceMesh(max_num_faces=100)
self.hand = mp.solutions.hands.Hands(max_num_hands=100)
now = QDate.currentDate()
current_date = now.toString('ddd dd MMMM yyyy')
current_time = datetime.datetime.now().strftime("%I:%M %p")
self.Date_Label.setText(current_date)
self.Time_Label.setText(current_time)
self.image = None
self.settings = UISettings()
self.Settings_Button.clicked.connect(
lambda: self.handle_setting_button())
self.warnings = []
self.prev_waring_code = -1
self.prev_waring_code_1 = -1
self.prev_waring_code_2 = -1
self.warning_history = UIWarnings()
self.Warning_History.clicked.connect(
lambda: self.handle_waring_history_button())
self.net = cv2.dnn.readNet("yolov3-tiny.weights", "yolov3-tiny.cfg")
with open("yolov3.txt", 'r') as f:
self.classes = [line.strip() for line in f.readlines()]
self.colors = np.random.uniform(0, 255, size=(len(self.classes), 3))
def handle_waring_history_button(self):
self.pause_video()
self.warning_history.handle_show(self.warnings)
def handle_setting_button(self):
self.pause_video()
self.settings.show()
def init_video(self, camera):
self.capture = camera
def start_video(self):
self.timer.start(1)
def pause_video(self):
self.timer.stop()
def process_frame(self, img):
skeletons = self.model_wrapper.process_image(img)
skeleton_drawer = vis.SkeletonDrawer(img, draw_config)
for skeleton in skeletons:
skeleton.draw_skeleton(skeleton_drawer.joint_draw,
skeleton_drawer.kpt_draw)
return img
def face_rec(self, image):
"""
Nhận diện dáng người và tay từ ảnh đã cho
:param image: ảnh từ camera
:return: image: ảnh đuọc xử lý sau khi nhận diện
"""
width = image.shape[1]
height = image.shape[0]
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
raw_img = image
image = self.process_frame(image)
# Bắt đầu nhận diện
faces = self.face.process(raw_img)
hands = self.hand.process(raw_img)
image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
if not faces.multi_face_landmarks:
self.Warnings_List.clear()
return image
count = 0
for face_landmarks in faces.multi_face_landmarks:
if self.settings.print_results:
draw_landmarks(image, face_landmarks,
mp.solutions.face_mesh.FACEMESH_CONTOURS)
"""
nose_tip 1
chin = 199
left_eye_outer = 249
right_eye_outer = 7
mouth_left = 291
mouth_right = 61
"""
nose_tip = normalize_to_pixel_coordinates(
face_landmarks.landmark[1].x, face_landmarks.landmark[1].y,
width, height)
chin = normalize_to_pixel_coordinates(
face_landmarks.landmark[199].x, face_landmarks.landmark[199].y,
width, height)
left_eye_outer = normalize_to_pixel_coordinates(
face_landmarks.landmark[249].x, face_landmarks.landmark[249].y,
width, height)
right_eye_outer = normalize_to_pixel_coordinates(
face_landmarks.landmark[7].x, face_landmarks.landmark[7].y,
width, height)
mouth_left = normalize_to_pixel_coordinates(
face_landmarks.landmark[291].x, face_landmarks.landmark[291].y,
width, height)
mouth_right = normalize_to_pixel_coordinates(
face_landmarks.landmark[61].x, face_landmarks.landmark[61].y,
width, height)
if chin is None or left_eye_outer is None or right_eye_outer is None or mouth_left is None or mouth_right is None:
return image
image_points = np.array([
nose_tip, chin, left_eye_outer, right_eye_outer, mouth_left,
mouth_right
],
dtype="double")
focal_length = width
camera_matrix = np.array(
[[focal_length, 0, width / 2], [0, focal_length, height / 2],
[0, 0, 1]],
dtype="double")
dist_coeffs = np.zeros((4, 1))
(success, rotation_vector,
translation_vector) = cv2.solvePnP(model_points,
image_points,
camera_matrix,
dist_coeffs,
flags=cv2.SOLVEPNP_ITERATIVE)
if self.Warnings_List.topLevelItemCount() <= count:
item = QTreeWidgetItem(["Thí sinh #" + str(count + 1)])
item.addChildren([QTreeWidgetItem(), QTreeWidgetItem()])
self.Warnings_List.addTopLevelItem(item)
self.Warnings_List.topLevelItem(count).setExpanded(True)
rotation_mat, _ = cv2.Rodrigues(rotation_vector)
pose_mat = cv2.hconcat((rotation_mat, translation_vector))
_, _, _, _, _, _, euler_angles = cv2.decomposeProjectionMatrix(
pose_mat)
angle = round(abs(euler_angles[1][0]), 2)
self.Warnings_List.topLevelItem(count).child(0).setText(
0, "Đầu di chuyển ngang " + str(angle) + " độ so với camera")
color = int(angle / self.settings.head_angle_limit * 255)
if color > 255:
if self.prev_waring_code_1 != 3:
self.warnings.append(
QDateTime.currentDateTime().toString() + " - " +
"Thí sinh #" + str(count + 1) +
" - : Đầu di chuyển ngang " + str(angle) +
" độ so với camera")
self.prev_waring_code_1 = 3
color = 255
self.Warnings_List.topLevelItem(count).child(0).setBackground(
0, QColor(color, 255 - color, 0))
angle = round(abs(euler_angles[2][0]), 2)
self.Warnings_List.topLevelItem(count).child(1).setText(
0, "Đầu di chuyển dọc " + str(angle) + " độ so với camera")
color = int(angle / self.settings.head_angle_limit * 255)
if color > 255:
if self.prev_waring_code_2 != 5:
self.warnings.append(
QDateTime.currentDateTime().toString() + " - " +
"Thí sinh #" + str(count + 1) +
" - : Đầu di chuyển dọc " + str(angle) +
" độ so với camera")
self.prev_waring_code_2 = 5
color = 255
self.Warnings_List.topLevelItem(count).child(1).setBackground(
0, QColor(color, 255 - color, 0))
if self.settings.print_results:
(nose_end_point2D,
jacobian) = cv2.projectPoints(np.array([
(0.0, 0.0, 500.0)
]), rotation_vector, translation_vector, camera_matrix,
dist_coeffs)
'''
for p in image_points:
cv2.circle(image, (int(p[0]), int(p[1])), 2, (0, 0, 255), -1)
'''
cv2.line(image,
(int(image_points[0][0]), int(image_points[0][1])),
(int(nose_end_point2D[0][0][0]),
int(nose_end_point2D[0][0][1])), (255, 0, 0), 2)
count += 1
while self.Warnings_List.topLevelItemCount() > len(
faces.multi_face_landmarks):
self.Warnings_List.takeTopLevelItem(len(
faces.multi_face_landmarks))
if not hands.multi_hand_landmarks:
for i in range(len(faces.multi_face_landmarks)):
if self.Warnings_List.topLevelItem(i).childCount() == 2:
self.Warnings_List.topLevelItem(i).addChild(
QTreeWidgetItem())
self.Warnings_List.topLevelItem(i).child(2).setText(
0, "Không tìm thấy tay!")
self.Warnings_List.topLevelItem(i).child(2).setBackground(
0, QColor("red"))
if self.prev_waring_code != 0:
self.warnings.append(
QDateTime.currentDateTime().toString() + " - " +
"Thí sinh #" + str(i + 1) + " - : Không tìm thấy tay!")
self.prev_waring_code = 0
return image
face_list = {}
for hand_landmarks in hands.multi_hand_landmarks:
if self.settings.print_results:
draw_landmarks(image, hand_landmarks,
mp.solutions.hands.HAND_CONNECTIONS)
hand_x = hand_landmarks.landmark[0].x
hand_y = hand_landmarks.landmark[0].y
best_face = -1
best_face_value = 1
count = 0
for face_landmarks in faces.multi_face_landmarks:
if count in face_list and face_list[count] == 2:
count += 1
continue
face_x = face_landmarks.landmark[0].x
face_y = face_landmarks.landmark[0].y
dist = math.sqrt(
pow(abs(face_x - hand_x), 2) +
pow(abs(face_y - hand_y), 2))
if dist < best_face_value:
best_face_value = dist
best_face = count
count += 1
if best_face != -1:
if best_face in face_list:
face_list[best_face] = 2
else:
face_list[best_face] = 1
for i in range(len(faces.multi_face_landmarks)):
if self.Warnings_List.topLevelItem(i).childCount() == 2:
self.Warnings_List.topLevelItem(i).addChild(QTreeWidgetItem())
if i in face_list:
if face_list[i] == 1:
self.Warnings_List.topLevelItem(i).child(2).setText(
0, "Chỉ phát hiện được một tay!")
self.Warnings_List.topLevelItem(i).child(2).setBackground(
0, QColor("yellow"))
if self.prev_waring_code != 1:
self.warnings.append(
QDateTime.currentDateTime().toString() + " - " +
"Thí sinh #" + str(count + 1) +
" - Chỉ phát hiện được một tay!")
self.prev_waring_code = 1
else:
if self.Warnings_List.topLevelItem(i).childCount() == 3:
self.Warnings_List.topLevelItem(i).takeChild(2)
else:
self.Warnings_List.topLevelItem(i).child(2).setText(
0, "Không tìm thấy tay!")
self.Warnings_List.topLevelItem(i).child(2).setBackground(
0, QColor("red"))
if self.prev_waring_code != 0:
self.warnings.append(
QDateTime.currentDateTime().toString() + " - " +
"Thí sinh #" + str(count + 1) +
" - : Không tìm thấy tay!")
self.prev_waring_code = 0
"""
blob = cv2.dnn.blobFromImage(raw_img, 0.00392, (416, 416), (0, 0, 0), True, crop=False)
self.net.setInput(blob)
outs = self.net.forward(get_output_layers(self.net))
class_ids = []
confidences = []
boxes = []
for out in outs:
for detection in out:
scores = detection[5:]
class_id = np.argmax(scores)
confidence = scores[class_id]
if confidence > 0.5:
center_x = int(detection[0] * width)
center_y = int(detection[1] * height)
w = int(detection[2] * width)
h = int(detection[3] * height)
x = center_x - w / 2
y = center_y - h / 2
class_ids.append(class_id)
confidences.append(float(confidence))
boxes.append([x, y, w, h])
indices = cv2.dnn.NMSBoxes(boxes, confidences, 0.5, 0.4)
for i in indices:
i = i[0]
box = boxes[i]
x = box[0]
y = box[1]
w = box[2]
h = box[3]
color = self.colors[class_ids[i]]
cv2.rectangle(image, (x, y), (x + w, y + h), color, 2)
cv2.putText(image, str(self.classes[class_ids[i]]), (x - 10, y - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, color,
2)
"""
return image
def update_frame(self):
"""
Cập nhật hình ảnh lên ứng dụng
"""
ret, self.image = self.capture.read()
self.display_image(self.image, 1)
def display_image(self, image, windowed=1):
"""
Phát hình ảnh đã được xử lý lên ứng dụng
:param image: ảnh từ camera
:param windowed: số window đang hiện
"""
image = cv2.resize(image, (640, 480))
image = cv2.flip(image, 1)
try:
image = self.face_rec(image)
except Exception as e:
print(e)
qformat = QImage.Format_Indexed8
if len(image.shape) == 3:
if image.shape[2] == 4:
qformat = QImage.Format_RGBA8888
else:
qformat = QImage.Format_RGB888
out_image = QImage(image, image.shape[1], image.shape[0],
image.strides[0], qformat)
out_image = out_image.rgbSwapped()
if windowed == 1:
self.imgLabel.setPixmap(QPixmap.fromImage(out_image))
self.imgLabel.setScaledContents(True)
from tensorflow.keras import Model
import os
from utils import display_activation
from model_wrapper import ModelWrapper
from data_generator import test_generator
def reshape_data(a):
x, _, _ = a
x = x.reshape(-1, 28, 84, 1)
return x
def data_flatten(generator, batch_size=32):
return map(reshape_data, generator(batch_size=batch_size))
X = next(data_flatten(test_generator, 512))
for model_dir in os.listdir('classification_model_results'):
try:
model = ModelWrapper.load_model(
'classification_model_results/{}'.format(model_dir))
layer_outputs = [layer.output for layer in model.layers]
activation_model = Model(inputs=model.input, outputs=layer_outputs)
activations = activation_model.predict(X, verbose=1)
display_activation(activations,
'classification_model_results/{}'.format(model_dir))
except:
print(model_dir)
def main():
# Grab the dataset from scikit-learn
data = datasets.load_iris()
X = data['data']
y = data['target']
target_names = data['target_names']
feature_names = [
f.replace(' (cm)', '').replace(' ', '_') for f in data.feature_names
]
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.3,
random_state=42)
# Build and train the model
model = RandomForestClassifier(random_state=101)
model.fit(X_train, y_train)
print("Score on the training set is: {:2}".format(
model.score(X_train, y_train)))
print("Score on the test set is: {:.2}".format(model.score(X_test,
y_test)))
# CHANGES HERE >>>>>>>>>>
X_mean = X_train.mean(axis=0).round(1)
feature_defaults = dict(zip(feature_names, X_mean.tolist()))
wrapped = ModelWrapper(model_name='iris-rf',
model_version=MODEL_VERSION,
model_object=model,
class_labels=target_names.tolist(),
feature_defaults=feature_defaults)
# Save the model
model_filename = 'iris-rf-v{}.pkl'.format(MODEL_VERSION)
print("Saving model to {}...".format(model_filename))
joblib.dump(wrapped, model_filename)
# ***** Generate test data *****
print('Generating test data...')
all_probs = model.predict_proba(X_test)
all_test_cases = prep_test_cases(X_test, all_probs, feature_names,
target_names)
test_data_fname = 'testdata_iris_v{}.json'.format(MODEL_VERSION)
with open(test_data_fname, 'w') as fout:
json.dump(all_test_cases, fout)
# ***** Generate test data (Missing) *****
print('Generating test data with missing values...')
missing_grps = [(0, ), (1, ), (2, ), (0, 1), (0, 2), (1, 2), (0, 1, 2)]
X_mean = X_train.mean(axis=0).round(1)
all_features = []
all_probs = []
for missing_cols in missing_grps:
X_missing = X_test.copy().astype('object')
X_scored = X_test.copy()
for col in missing_cols:
X_missing[:, col] = None
X_scored[:, col] = X_mean[col]
# Use the imputed one to find expected probabilities
all_probs.extend(model.predict_proba(X_scored))
all_features.extend(X_missing)
all_test_cases_missing = prep_test_cases(all_features, all_probs,
feature_names, target_names)
test_data_fname = 'testdata_iris_missing_v{}.json'.format(MODEL_VERSION)
with open(test_data_fname, 'w') as fout:
json.dump(all_test_cases_missing, fout)
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# @Time : 17/11/14 PM2:21
# @Author : shaoguang.csg
# @File : main.py
from parse_conf import DataConf, ModelConf
from model_wrapper import ModelWrapper
from utils.logger import logger
from time import time
start = time()
data_conf = DataConf()
model_conf = ModelConf()
model = ModelWrapper(data_conf=data_conf, model_conf=model_conf)
model.train()
logger.info(model.get_weight())
model.evaluate()
result = model.predict()
end = time()
logger.info('time: {}'.format(end - start))
# 2 core 1 threads 116
# 1 core 228
model_list = os.listdir(model_dir)
print('\nFind model:')
for num, model in enumerate(model_list):
print(num, model)
model_name = model_list[int(input('\nChoose a model (enter a number): '))]
print('Load Model {}'.format(model_name))
model_path = os.path.join(model_dir, model_name)
learning_rate = float(model_name.split('-')[2])
img_size = int(model_name.split('-')[3])
model = ModelWrapper(learning_rate, img_size).model
model.load(os.path.join(model_path, model_name))
print('Model loaded!')
test_set_processed_data_path = os.path.join(processed_data_dir,
'test_{}.npy'.format(img_size))
if os.path.exists(test_set_processed_data_path):
test_set_data = np.load(test_set_processed_data_path)
print('Data loaded!')
else:
test_set_data = process_data.process_test_set_data(
img_size, test_set_raw_data_dir, test_set_processed_data_path)
print('Data processed!')
fig = plt.figure()
class AbstractAgent:
def __init__(self, environment: Env,
memory: AbstractMemory,
policy: AbstractPolicy,
model: K.Model,
optimizer: K.optimizers.Optimizer,
logger: Logger):
self.environment = environment
self.memory = memory
self.policy = policy
self.model = ModelWrapper(model, optimizer)
self.logger = logger
self.history = []
logger.create_settings_model_file(model)
logger.create_settings_agent_file(self)
def _explore_env(self, batch_size: int, number_of_game: int = 10) -> Tuple[float, List[Sample]]:
""" Return tuple of mean gain from all games and list of samples. """
data = []
gains = []
state = self.environment.reset()
previous_sample = None
current_gain = 0
n_sample = 0
n_game = 0
while n_game <= number_of_game or n_sample <= batch_size:
q_values = self.model.predict(state)
action = self.policy(q_values)
next_state, reward, done, _ = self.environment.step(action)
current_gain += reward
if previous_sample is None:
previous_sample = Sample(state, action, next_state, reward)
else:
current_sample = Sample(state, action, next_state, reward)
previous_sample.next_sample = current_sample
data.append(previous_sample)
previous_sample = current_sample
if done:
data.append(current_sample)
gains.append(current_gain)
current_gain = 0
previous_sample = None
state = self.environment.reset()
n_game += 1
else:
state = next_state
n_sample += 1
self.environment.close()
return np.mean(gains), rand.sample(data, batch_size)
def learn(self, epochs: int,
batch_size_in_step: int,
min_n_game_in_exploration: int,
batch_size_in_exploration: int,
change_model_delay: int):
raise NotImplementedError
def __str__(self):
raise NotImplementedError
shots = args.shots
token_limit = args.token_limit
model_type = args.model
model_size = args.model_size
data_size = args.data_size
lr_base = args.lr
eval_data_size = args.eval_data_size
eval_batch_size = args.eval_batch_size
train_batch_size = args.train_batch_size
local_rank = args.local_rank
fp16 = args.fp16
device, n_gpu = setup_device(local_rank)
if model_type == "bert":
lm = ModelWrapper('bert',
f'bert-{model_size}-uncased',
token_limit=token_limit,
device=device)
elif model_type == "roberta":
lm = ModelWrapper('roberta',
f'roberta-{model_size}',
token_limit=token_limit,
device=device)
elif model_type == "longformer":
lm = ModelWrapper('longformer',
f'allenai/longformer-{model_size}-4096',
token_limit=token_limit,
device=device)
else:
raise KeyError(f"model type {model_type} not supported")
if do_mlm:
classes = [['yes', 'right'], ['maybe'], ['wrong', 'no']]
class DQN(AbstractAgent):
def __init__(self,
environment: Env,
memory: AbstractMemory,
policy: AbstractPolicy,
model: K.Model,
logger: Logger,
gamma: float,
optimizer: K.optimizers.Optimizer,
n_step: int = 1):
self.model = ModelWrapper(model, optimizer)
#self.model.compile()
self.current_model = None
self.gamma = gamma
self.n_step = n_step
super(DQN, self).__init__(environment=environment,
memory=memory,
policy=policy,
model=model,
optimizer=optimizer,
logger=logger)
def _bellman_equation(self, batch: List[Sample]) -> np.ndarray:
state = np.array([sample.state for sample in batch])
q_values = self.current_model.predict(state)
for idx in range(q_values.shape[0]):
q_values[idx][batch[idx].action] = batch[idx].reward
if not batch[idx].is_done():
best_action_for_q_next = np.argmax(
self.model.predict(batch[idx].next_state))
q_next = self.model.predict(
batch[idx].next_state)[0][best_action_for_q_next]
q_values[idx][batch[idx].action] += self.gamma * q_next
return q_values
def learn(self, epochs: int, batch_size_in_step: int,
min_n_game_in_exploration: int, batch_size_in_exploration: int,
change_model_delay: int):
self.model.compile()
self.current_model = self.model.clone()
self.current_model.compile()
eval_score, starting_experience = self._explore_env(self.memory.maxlen)
self.memory.add(starting_experience)
for epoch in tqdm(range(epochs), desc='Learning in progress: '):
if epoch % change_model_delay == 0:
self.model = self.current_model.clone()
self.model.compile()
if type(self.memory) == PrioritizedExperienceReplay:
self.memory.update_model(self.model)
eval_score, batch = self._explore_env(
batch_size_in_exploration, min_n_game_in_exploration)
self.memory.add(batch)
batch = self.memory.sample(batch_size_in_step)
q_values = self._bellman_equation(batch)
state = np.array([sample.state for sample in batch])
loss = self.current_model.fit(state, q_values)
self.policy.update()
self.logger.add_event({
'loss_value': loss,
'mean_gain': eval_score,
'epoch': epoch
})
def __str__(self):
return "Agent: " + self.__class__.__name__ + "\n\n" + \
"Discount value: " + str(self.gamma) + "\n"\
"N-step: " + str(self.n_step) + "\n\n"\
"Environment:\n" + str(self.environment) + "\n\n" + \
"Memory:\n" + str(self.memory) + "\n" + \
"Policy:\n" + str(self.policy)
shuffle=True)
validation_dataset = DataLoader(
CellInstanceSegmentation(path=os.path.join(args.path_to_data, "val"),
augmentation_p=0.0, two_classes=not args.three_classes),
collate_fn=collate_function_cell_instance_segmentation, batch_size=1, num_workers=1, shuffle=False)
test_dataset = DataLoader(
CellInstanceSegmentation(path=os.path.join(args.path_to_data, "test"),
augmentation_p=0.0, two_classes=not args.three_classes),
collate_fn=collate_function_cell_instance_segmentation, batch_size=1, num_workers=1, shuffle=False)
# Model wrapper
model_wrapper = ModelWrapper(detr=detr,
detr_optimizer=detr_optimizer,
detr_segmentation_optimizer=detr_segmentation_optimizer,
training_dataset=training_dataset,
validation_dataset=validation_dataset,
test_dataset=test_dataset,
loss_function=InstanceSegmentationLoss(
segmentation_loss=SegmentationLoss(),
ohem=args.ohem,
ohem_faction=args.ohem_fraction),
device=device)
# Perform training
if args.train:
model_wrapper.train(epochs=args.epochs,
optimize_only_segmentation_head_after_epoch=args.only_train_segmentation_head_after_epoch)
# Perform validation
if args.val:
model_wrapper.validate(number_of_plots=30)
# Perform testing
if args.test:
model_wrapper.test()
#
#
flag_load_data = True
# data
if flag_load_data:
dataset.load_preprocessed_data()
else:
dataset.pretrained_emb_file = None
dataset.emb_dim = 200
dataset.max_seq_len = 200
dataset.prepare_preprocessed_data(load_vocab=False)
#
data_train, data_test = dataset.split_train_and_test()
#
#
config = ModelSettings()
config.vocab = dataset.vocab
config.model_tag = model_tag
config.model_graph = build_graph
config.is_train = True
config.check_settings()
#
model = ModelWrapper(config)
model.prepare_for_train_and_valid()
#
model.train_and_valid(data_train, data_test)
#
if not os.path.exists(processed_data_dir):
os.makedirs(processed_data_dir)
train_set_processed_data_path = os.path.join(processed_data_dir,
'train_{}.npy'.format(img_size))
if os.path.exists(train_set_processed_data_path):
train_set_data = np.load(train_set_processed_data_path)
print('Data loaded!')
else:
train_set_data = process_data.process_train_set_data(
img_size, train_set_raw_data_dir, train_set_processed_data_path)
print('Data processed!')
model_wrapper = ModelWrapper(learning_rate,
img_size,
tensorboard_dir=tensorboard_dir)
model = model_wrapper.model
model_path = os.path.join(model_dir, model_wrapper.name)
if os.path.exists(model_path):
model.load(os.path.join(model_path, model_wrapper.name))
print('Model loaded!')
train_set_data, validation_set_data = train_set_data[:-500], train_set_data[
-500:]
train_x = np.array([i[0]
for i in train_set_data]).reshape(-1, img_size, img_size,
1)
train_y = [i[1] for i in train_set_data]
# # data dataset = Dataset() dataset.load_vocab_tokens_and_emb() # # config = ModelSettings() config.vocab = dataset.vocab config.model_tag = model_tag config.model_graph = build_graph config.is_train = False config.check_settings() # model = ModelWrapper(config) model.prepare_for_prediction() # text_raw = ["这本书不错"] """ work_book = xlrd.open_workbook(file_raw) data_sheet = work_book.sheets()[0] text_raw = data_sheet.col_values(0) """ # preds_list = [] logits_list = [] # for item in text_raw:
def generate(args):
# Load word2idx
adaptive = ModelWrapper(args, data_loader=get_loader)
if adaptive.generate(args):
return 0