-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathtimevis_main.py
170 lines (144 loc) · 8.67 KB
/
timevis_main.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
import torch
import sys
import os
import time
import json
import numpy as np
import argparse
from torch.utils.data import DataLoader
from torch.utils.data import WeightedRandomSampler
from umap.umap_ import find_ab_params
from singleVis.custom_weighted_random_sampler import CustomWeightedRandomSampler
from singleVis.SingleVisualizationModel import VisModel
from singleVis.losses import UmapLoss, ReconstructionLoss, SingleVisLoss
from singleVis.edge_dataset import DataHandler
from singleVis.trainer import SingleVisTrainer
from singleVis.data import NormalDataProvider
from singleVis.spatial_edge_constructor import kcSpatialEdgeConstructor
from singleVis.temporal_edge_constructor import GlobalTemporalEdgeConstructor
from singleVis.projector import TimeVisProjector
from singleVis.eval.evaluator import Evaluator
########################################################################################################################
# VISUALIZATION SETTING #
########################################################################################################################
VIS_METHOD= "TimeVis"
########################################################################################################################
# LOAD PARAMETERS #
########################################################################################################################
parser = argparse.ArgumentParser(description='Process hyperparameters...')
parser.add_argument('--content_path', '-c', type=str)
args = parser.parse_args()
CONTENT_PATH = args.content_path
sys.path.append(CONTENT_PATH)
with open(os.path.join(CONTENT_PATH, "config.json"), "r") as f:
config = json.load(f)
config = config[VIS_METHOD]
# record output information
# now = time.strftime("%Y-%m-%d-%H_%M_%S", time.localtime(time.time()))
# sys.stdout = open(os.path.join(CONTENT_PATH, now+".txt"), "w")
SETTING = config["SETTING"]
CLASSES = config["CLASSES"]
DATASET = config["DATASET"]
PREPROCESS = config["VISUALIZATION"]["PREPROCESS"]
GPU_ID = config["GPU"]
EPOCH_START = config["EPOCH_START"]
EPOCH_END = config["EPOCH_END"]
EPOCH_PERIOD = config["EPOCH_PERIOD"]
EPOCH_NAME = config["EPOCH_NAME"]
# Training parameter (subject model)
TRAINING_PARAMETER = config["TRAINING"]
NET = TRAINING_PARAMETER["NET"]
LEN = TRAINING_PARAMETER["train_num"]
# Training parameter (visualization model)
VISUALIZATION_PARAMETER = config["VISUALIZATION"]
LAMBDA = VISUALIZATION_PARAMETER["LAMBDA"]
B_N_EPOCHS = VISUALIZATION_PARAMETER["BOUNDARY"]["B_N_EPOCHS"]
L_BOUND = VISUALIZATION_PARAMETER["BOUNDARY"]["L_BOUND"]
INIT_NUM = VISUALIZATION_PARAMETER["INIT_NUM"]
ALPHA = VISUALIZATION_PARAMETER["ALPHA"]
BETA = VISUALIZATION_PARAMETER["BETA"]
# MAX_HAUSDORFF = VISUALIZATION_PARAMETER["MAX_HAUSDORFF"]
ENCODER_DIMS = VISUALIZATION_PARAMETER["ENCODER_DIMS"]
DECODER_DIMS = VISUALIZATION_PARAMETER["DECODER_DIMS"]
S_N_EPOCHS = VISUALIZATION_PARAMETER["S_N_EPOCHS"]
T_N_EPOCHS = VISUALIZATION_PARAMETER["T_N_EPOCHS"]
N_NEIGHBORS = VISUALIZATION_PARAMETER["N_NEIGHBORS"]
PATIENT = VISUALIZATION_PARAMETER["PATIENT"]
MAX_EPOCH = VISUALIZATION_PARAMETER["MAX_EPOCH"]
VIS_MODEL_NAME = VISUALIZATION_PARAMETER["VIS_MODEL_NAME"]
EVALUATION_NAME = VISUALIZATION_PARAMETER["EVALUATION_NAME"]
SEGMENTS = [(EPOCH_START, EPOCH_END)]
# define hyperparameters
DEVICE = torch.device("cuda:{}".format(GPU_ID) if torch.cuda.is_available() else "cpu")
import Model.model as subject_model
net = eval("subject_model.{}()".format(NET))
########################################################################################################################
# TRAINING SETTING #
########################################################################################################################
data_provider = NormalDataProvider(CONTENT_PATH, net, EPOCH_START, EPOCH_END, EPOCH_PERIOD, device=DEVICE, classes=CLASSES, epoch_name=EPOCH_NAME, verbose=1)
if PREPROCESS:
data_provider._meta_data()
if B_N_EPOCHS >0:
data_provider._estimate_boundary(LEN//10, l_bound=L_BOUND)
model = VisModel(ENCODER_DIMS, DECODER_DIMS)
projector = TimeVisProjector(vis_model=model, content_path=CONTENT_PATH, vis_model_name=VIS_MODEL_NAME, device=DEVICE)
########################################################################################################################
# EDGE DATASET #
########################################################################################################################
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)
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)
t0 = time.time()
spatial_cons = kcSpatialEdgeConstructor(data_provider=data_provider, init_num=INIT_NUM, s_n_epochs=S_N_EPOCHS, b_n_epochs=B_N_EPOCHS, n_neighbors=N_NEIGHBORS, MAX_HAUSDORFF=None, ALPHA=ALPHA, BETA=BETA)
s_edge_to, s_edge_from, s_probs, feature_vectors, time_step_nums, time_step_idxs_list, knn_indices, sigmas, rhos, attention = spatial_cons.construct()
temporal_cons = GlobalTemporalEdgeConstructor(X=feature_vectors, time_step_nums=time_step_nums, sigmas=sigmas, rhos=rhos, n_neighbors=N_NEIGHBORS, n_epochs=T_N_EPOCHS)
t_edge_to, t_edge_from, t_probs = temporal_cons.construct()
t1 = time.time()
edge_to = np.concatenate((s_edge_to, t_edge_to),axis=0)
edge_from = np.concatenate((s_edge_from, t_edge_from), axis=0)
probs = np.concatenate((s_probs, t_probs), axis=0)
probs = probs / (probs.max()+1e-3)
eliminate_zeros = probs>1e-3
edge_to = edge_to[eliminate_zeros]
edge_from = edge_from[eliminate_zeros]
probs = probs[eliminate_zeros]
dataset = DataHandler(edge_to, edge_from, feature_vectors, attention)
n_samples = int(np.sum(S_N_EPOCHS * probs) // 1)
# chose sampler based on the number of dataset
if len(edge_to) > pow(2,24):
sampler = CustomWeightedRandomSampler(probs, n_samples, replacement=True)
else:
sampler = WeightedRandomSampler(probs, n_samples, replacement=True)
edge_loader = DataLoader(dataset, batch_size=1000, sampler=sampler, num_workers=4, prefetch_factor=10)
########################################################################################################################
# TRAIN #
########################################################################################################################
trainer = SingleVisTrainer(model, criterion, optimizer, lr_scheduler, edge_loader=edge_loader, DEVICE=DEVICE)
t2=time.time()
trainer.train(PATIENT, MAX_EPOCH)
t3 = time.time()
save_dir = data_provider.model_path
trainer.record_time(save_dir, "time_{}.json".format(VIS_MODEL_NAME), "complex_construction", t1-t0)
trainer.record_time(save_dir, "time_{}.json".format(VIS_MODEL_NAME), "training", t3-t2)
trainer.save(save_dir=save_dir, file_name="{}".format(VIS_MODEL_NAME))
########################################################################################################################
# VISUALIZATION #
########################################################################################################################
from singleVis.visualizer import visualizer
vis = visualizer(data_provider, projector, 200)
save_dir = os.path.join(data_provider.content_path, "img")
os.makedirs(save_dir, exist_ok=True)
for i in range(EPOCH_START, EPOCH_END+1, EPOCH_PERIOD*4):
vis.save_default_fig(i, path=os.path.join(save_dir, "{}_{}_{}.png".format(DATASET, i, VIS_METHOD)))
########################################################################################################################
# EVALUATION #
########################################################################################################################
# eval_epochs = range(EPOCH_START, EPOCH_END, EPOCH_PERIOD)
# evaluator = Evaluator(data_provider, projector)
# for eval_epoch in eval_epochs:
# evaluator.save_epoch_eval(eval_epoch, 15, temporal_k=5, file_name="{}".format(EVALUATION_NAME))