SalazarPevelll
fe
8fcf809
raw
history blame
13.3 kB
import os, sys
# sys.path.append("..")
import torch
import numpy as np
from umap.umap_ import find_ab_params
import pickle
import gc
import base64
from .backend_adapter import TimeVisBackend, ActiveLearningTimeVisBackend, AnormalyTimeVisBackend
timevis_path = "../../DLVisDebugger"
sys.path.append(timevis_path)
from singleVis.SingleVisualizationModel import VisModel
from singleVis.losses import SingleVisLoss, UmapLoss, ReconstructionLoss, SmoothnessLoss, HybridLoss
from singleVis.trainer import SingleVisTrainer, HybridVisTrainer
from singleVis.data import NormalDataProvider, ActiveLearningDataProvider, DenseActiveLearningDataProvider
from singleVis.eval.evaluator import Evaluator
from singleVis.visualizer import visualizer, DenseALvisualizer
from singleVis.projector import Projector, ALProjector, DenseALProjector
from singleVis.segmenter import Segmenter
def initialize_backend(CONTENT_PATH, dense_al=False):
import config
# load hyperparameters
CLASSES = config["CLASSES"]
DATASET = config["DATASET"]
GPU_ID = config["GPU"]
DEVICE = torch.device("cuda:{}".format(GPU_ID) if torch.cuda.is_available() else "cpu")
################################################# VISUALIZATION PARAMETERS ########################################
PREPROCESS = config["VISUALIZATION"]["PREPROCESS"]
B_N_EPOCHS = config["VISUALIZATION"]["BOUNDARY"]["B_N_EPOCHS"]
L_BOUND = config["VISUALIZATION"]["BOUNDARY"]["L_BOUND"]
LAMBDA = config["VISUALIZATION"]["LAMBDA"]
# HIDDEN_LAYER = config["VISUALIZATION"]["HIDDEN_LAYER"]
ENCODER_DIMS = config["VISUALIZATION"]["ENCODER_DIMS"]
DECODER_DIMS = config["VISUALIZATION"]["DECODER_DIMS"]
N_NEIGHBORS = config["VISUALIZATION"]["N_NEIGHBORS"]
MAX_EPOCH = config["VISUALIZATION"]["MAX_EPOCH"]
S_N_EPOCHS = config["VISUALIZATION"]["S_N_EPOCHS"]
PATIENT = config["VISUALIZATION"]["PATIENT"]
VIS_MODEL_NAME = config["VISUALIZATION"]["VIS_MODEL_NAME"]
RESOLUTION = config["VISUALIZATION"]["RESOLUTION"]
EVALUATION_NAME = config["VISUALIZATION"]["EVALUATION_NAME"]
NET = config["TRAINING"]["NET"]
SETTING = config["SETTING"] # active learning
if SETTING == "normal" or SETTING == "abnormal":
EPOCH_START = config["EPOCH_START"]
EPOCH_END = config["EPOCH_END"]
EPOCH_PERIOD = config["EPOCH_PERIOD"]
INIT_NUM = config["VISUALIZATION"]["INIT_NUM"]
ALPHA = config["VISUALIZATION"]["ALPHA"]
BETA = config["VISUALIZATION"]["BETA"]
MAX_HAUSDORFF = config["VISUALIZATION"]["MAX_HAUSDORFF"]
T_N_EPOCHS = config["VISUALIZATION"]["T_N_EPOCHS"]
elif SETTING == "active learning":
BASE_ITERATION = config["BASE_ITERATION"]
TOTAL_EPOCH = config["TRAINING"]["total_epoch"]
else:
raise NotImplementedError
import Model.model as subject_model
net = eval("subject_model.{}()".format(NET))
# ########################################################################################################################
# TRAINING SETTING #
# ########################################################################################################################
# model = SingleVisualizationModel(input_dims=512, output_dims=2, units=256, hidden_layer=HIDDEN_LAYER)
model = VisModel(ENCODER_DIMS, DECODER_DIMS)
if SETTING == "normal" or SETTING == "abnormal":
data_provider = NormalDataProvider(CONTENT_PATH, net, EPOCH_START, EPOCH_END, EPOCH_PERIOD, device=DEVICE, classes=CLASSES,epoch_name="Epoch", verbose=1)
segmenter = Segmenter(data_provider=data_provider, threshold=78.5, range_s=EPOCH_START, range_e=EPOCH_END, range_p=EPOCH_PERIOD)
SEGMENTS = segmenter.segment()
# SEGMENTS = config["VISUALIZATION"]["SEGMENTS"]
projector = Projector(vis_model=model, content_path=CONTENT_PATH, segments=SEGMENTS, device=DEVICE)
elif SETTING == "active learning":
DENSE_VIS_MODEL_NAME = config["VISUALIZATION"]["DENSE_VIS_MODEL_NAME"]
if dense_al:
data_provider = DenseActiveLearningDataProvider(CONTENT_PATH, net, BASE_ITERATION, epoch_num=TOTAL_EPOCH, split=-1, device=DEVICE, classes=CLASSES,verbose=1)
projector = DenseALProjector(vis_model=model, content_path=CONTENT_PATH, vis_model_name=DENSE_VIS_MODEL_NAME, device=DEVICE)
else:
data_provider = ActiveLearningDataProvider(CONTENT_PATH, net, BASE_ITERATION, split=-1, device=DEVICE, classes=CLASSES, verbose=1)
projector = ALProjector(vis_model=model, content_path=CONTENT_PATH, vis_model_name=VIS_MODEL_NAME, device=DEVICE)
# ########################################################################################################################
# # TRAIN #
# ########################################################################################################################
if SETTING == "active learning":
negative_sample_rate = 5
min_dist = .1
_a, _b = find_ab_params(1.0, min_dist)
umap_loss_fn = UmapLoss(negative_sample_rate, DEVICE, _a, _b, repulsion_strength=1.0)
recon_loss_fn = ReconstructionLoss(beta=1.0)
if dense_al:
smooth_loss_fn = SmoothnessLoss(margin=1.)
S_LAMBDA = config["VISUALIZATION"]["S_LAMBDA"]
criterion = HybridLoss(umap_loss_fn, recon_loss_fn, smooth_loss_fn, lambd1=LAMBDA, lambd2=S_LAMBDA)
optimizer = torch.optim.Adam(model.parameters(), lr=.01, weight_decay=1e-5)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=4, gamma=.1)
trainer = HybridVisTrainer(model, criterion, optimizer, lr_scheduler,edge_loader=None, DEVICE=DEVICE)
else:
criterion = SingleVisLoss(umap_loss_fn, recon_loss_fn, lambd=LAMBDA)
optimizer = torch.optim.Adam(model.parameters(), lr=.01, weight_decay=1e-5)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=4, gamma=.1)
trainer = SingleVisTrainer(model, criterion, optimizer, lr_scheduler,edge_loader=None, DEVICE=DEVICE)
# ########################################################################################################################
# # EVALUATION #
# ########################################################################################################################
if dense_al:
vis = DenseALvisualizer(data_provider, projector, RESOLUTION)
else:
vis = visualizer(data_provider, projector, RESOLUTION)
evaluator = Evaluator(data_provider, projector)
if SETTING == "normal":
timevis = TimeVisBackend(data_provider, projector, vis, evaluator, **config)
elif SETTING == "abnormal":
timevis = AnormalyTimeVisBackend(data_provider, projector, vis, evaluator, period=100, **config)
elif SETTING == "active learning":
timevis = ActiveLearningTimeVisBackend(data_provider, projector, trainer, vis, evaluator, dense_al, **config)
del config
gc.collect()
return timevis
def update_epoch_projection(timevis, EPOCH, predicates):
train_data = timevis.data_provider.train_representation(EPOCH)
test_data = timevis.data_provider.test_representation(EPOCH)
all_data = np.concatenate((train_data, test_data), axis=0)
fname = "Epoch" if timevis.data_provider.mode == "normal" or timevis.data_provider.mode == "abnormal" else "Iteration"
embedding_path = os.path.join(timevis.data_provider.model_path, "{}_{}".format(fname, EPOCH), "embedding.npy")
if os.path.exists(embedding_path):
embedding_2d = np.load(embedding_path)
else:
embedding_2d = timevis.projector.batch_project(EPOCH, all_data)
np.save(embedding_path, embedding_2d)
train_labels = timevis.data_provider.train_labels(EPOCH)
test_labels = timevis.data_provider.test_labels(EPOCH)
labels = np.concatenate((train_labels, test_labels), axis=0).tolist()
training_data_number = timevis.hyperparameters["TRAINING"]["train_num"]
testing_data_number = timevis.hyperparameters["TRAINING"]["test_num"]
training_data_index = list(range(training_data_number))
testing_data_index = list(range(training_data_number, training_data_number + testing_data_number))
# return the image of background
# read cache if exists
fname = "Epoch" if timevis.data_provider.mode == "normal" or timevis.data_provider.mode == "abnormal" else "Iteration"
bgimg_path = os.path.join(timevis.data_provider.model_path, "{}_{}".format(fname, EPOCH), "bgimg.png")
grid_path = os.path.join(timevis.data_provider.model_path, "{}_{}".format(fname, EPOCH), "grid.pkl")
if os.path.exists(bgimg_path) and os.path.exists(grid_path):
with open(os.path.join(grid_path), "rb") as f:
grid = pickle.load(f)
with open(bgimg_path, 'rb') as img_f:
img_stream = img_f.read()
b_fig = base64.b64encode(img_stream).decode()
else:
x_min, y_min, x_max, y_max, b_fig = timevis.vis.get_background(EPOCH, timevis.hyperparameters["VISUALIZATION"]["RESOLUTION"])
grid = [x_min, y_min, x_max, y_max]
# formating
grid = [float(i) for i in grid]
b_fig = str(b_fig, encoding='utf-8')
# save results, grid and decision_view
save_path = timevis.data_provider.model_path
iteration_name = "Epoch" if timevis.data_provider.mode == "normal" or timevis.data_provider.mode == "abnormal" else "Iteration"
save_path = os.path.join(save_path, "{}_{}".format(iteration_name, EPOCH))
with open(os.path.join(save_path, "grid.pkl"), "wb") as f:
pickle.dump(grid, f)
np.save(os.path.join(save_path, "embedding.npy"), embedding_2d)
color = timevis.vis.get_standard_classes_color() * 255
color = color.astype(int).tolist()
# TODO fix its structure
file_name = timevis.hyperparameters["VISUALIZATION"]["EVALUATION_NAME"]
evaluation = timevis.evaluator.get_eval(file_name=file_name)
eval_new = dict()
# eval_new["nn_train_15"] = evaluation["15"]['nn_train'][str(EPOCH)]
# eval_new['nn_test_15'] = evaluation["15"]['nn_test'][str(EPOCH)]
# eval_new['bound_train_15'] = evaluation["15"]['b_train'][str(EPOCH)]
# eval_new['bound_test_15'] = evaluation["15"]['b_test'][str(EPOCH)]
# eval_new['ppr_train'] = evaluation["ppr_train"][str(EPOCH)]
# eval_new['ppr_test'] = evaluation["ppr_test"][str(EPOCH)]
# eval_new["nn_train_15"] = 1
# eval_new['nn_test_15'] = 1
# eval_new['bound_train_15'] = 1
# eval_new['bound_test_15'] = 1
# eval_new['ppr_train'] = 1
# eval_new['ppr_test'] = 1
eval_new["train_acc"] = evaluation["train_acc"][str(EPOCH)]
eval_new["test_acc"] = evaluation["test_acc"][str(EPOCH)]
label_color_list = []
label_list = []
label_name_dict = dict()
for i, label in enumerate(timevis.hyperparameters["CLASSES"]):
label_name_dict[i] = label
for label in labels:
label_color_list.append(color[int(label)])
label_list.append(timevis.hyperparameters["CLASSES"][int(label)])
prediction_list = []
prediction = timevis.data_provider.get_pred(EPOCH, all_data).argmax(1)
for i in range(len(prediction)):
prediction_list.append(timevis.hyperparameters["CLASSES"][prediction[i]])
if timevis.hyperparameters["SETTING"] == "normal" or timevis.hyperparameters["SETTING"] == "abnormal":
max_iter = (timevis.hyperparameters["EPOCH_END"] - timevis.hyperparameters["EPOCH_START"]) // timevis.hyperparameters["EPOCH_PERIOD"] + 1
elif timevis.hyperparameters["SETTING"] == "active learning":
# TODO fix this, could be larger than EPOCH
max_iter = timevis.get_max_iter()
# max_iter = max(timevis.hyperparameters["BASE_ITERATION"], EPOCH)
# current_index = timevis.get_epoch_index(EPOCH)
# selected_points = np.arange(training_data_number + testing_data_number)[current_index]
selected_points = np.arange(training_data_number + testing_data_number)
for key in predicates.keys():
if key == "label":
tmp = np.array(timevis.filter_label(predicates[key]))
elif key == "type":
tmp = np.array(timevis.filter_type(predicates[key], int(EPOCH)))
else:
tmp = np.arange(training_data_number + testing_data_number)
selected_points = np.intersect1d(selected_points, tmp)
properties = np.concatenate((np.zeros(training_data_number, dtype=np.int16), 2*np.ones(testing_data_number, dtype=np.int16)), axis=0)
lb = timevis.get_epoch_index(EPOCH)
ulb = np.setdiff1d(training_data_index, lb)
properties[ulb] = 1
return embedding_2d.tolist(), grid, b_fig, label_name_dict, label_color_list, label_list, max_iter, training_data_index, testing_data_index, eval_new, prediction_list, selected_points, properties