-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathdvi_main.py
202 lines (170 loc) · 9.92 KB
/
dvi_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
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
########################################################################################################################
# IMPORT #
########################################################################################################################
import torch
import sys
import os
import json
import time
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, TemporalLoss, DVILoss, SingleVisLoss, DummyTemporalLoss
from singleVis.edge_dataset import DVIDataHandler
from singleVis.trainer import DVITrainer, DVIALMODITrainer, OriginDVITrainer
from singleVis.data import NormalDataProvider
from singleVis.spatial_edge_constructor import SingleEpochSpatialEdgeConstructor, OriginSingleEpochSpatialEdgeConstructor, PredDistSingleEpochSpatialEdgeConstructor
from singleVis.projector import DVIProjector
from singleVis.eval.evaluator import Evaluator
from singleVis.utils import find_neighbor_preserving_rate
########################################################################################################################
# DVI PARAMETERS #
########################################################################################################################
"""This serve as an example of DeepVisualInsight implementation in pytorch."""
VIS_METHOD = "DVI" # DeepVisualInsight
########################################################################################################################
# LOAD PARAMETERS #
########################################################################################################################
parser = argparse.ArgumentParser(description='Process hyperparameters...')
parser.add_argument('--content_path', type=str, default='/home/yiming/ContrastDebugger/EXP/cifar10')
args = parser.parse_args()
CONTENT_PATH = args.content_path
sys.path.append(CONTENT_PATH)
with open(os.path.join(CONTENT_PATH, "config_dvi_modi.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"]
# 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"]
LAMBDA1 = VISUALIZATION_PARAMETER["LAMBDA1"]
LAMBDA2 = VISUALIZATION_PARAMETER["LAMBDA2"]
B_N_EPOCHS = VISUALIZATION_PARAMETER["BOUNDARY"]["B_N_EPOCHS"]
L_BOUND = VISUALIZATION_PARAMETER["BOUNDARY"]["L_BOUND"]
ENCODER_DIMS = VISUALIZATION_PARAMETER["ENCODER_DIMS"]
DECODER_DIMS = VISUALIZATION_PARAMETER["DECODER_DIMS"]
S_N_EPOCHS = VISUALIZATION_PARAMETER["S_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"]
# 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 #
########################################################################################################################
# Define data_provider
data_provider = NormalDataProvider(CONTENT_PATH, net, EPOCH_START, EPOCH_END, EPOCH_PERIOD, device=DEVICE, classes=CLASSES, epoch_name='Epoch', verbose=1)
if PREPROCESS:
data_provider._meta_data()
if B_N_EPOCHS >0:
data_provider._estimate_boundary(LEN//10, l_bound=L_BOUND)
# Define visualization models
model = VisModel(ENCODER_DIMS, DECODER_DIMS)
# Define Losses
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)
single_loss_fn = SingleVisLoss(umap_loss_fn, recon_loss_fn, lambd=LAMBDA1)
# Define Projector
projector = DVIProjector(vis_model=model, content_path=CONTENT_PATH, vis_model_name=VIS_MODEL_NAME, device=DEVICE)
start_flag = 1
prev_model = VisModel(ENCODER_DIMS, DECODER_DIMS)
for iteration in range(EPOCH_START, EPOCH_END+EPOCH_PERIOD, EPOCH_PERIOD):
# Define DVI Loss
if start_flag:
temporal_loss_fn = DummyTemporalLoss(DEVICE)
criterion = DVILoss(umap_loss_fn, recon_loss_fn, temporal_loss_fn, lambd1=LAMBDA1, lambd2=0.0, device=DEVICE)
start_flag = 0
else:
# TODO AL mode, redefine train_representation
prev_data = data_provider.train_representation(iteration-EPOCH_PERIOD)
curr_data = data_provider.train_representation(iteration)
npr = find_neighbor_preserving_rate(prev_data, curr_data, N_NEIGHBORS)
temporal_loss_fn = TemporalLoss(w_prev, DEVICE)
criterion = DVILoss(umap_loss_fn, recon_loss_fn, temporal_loss_fn, lambd1=LAMBDA1, lambd2=torch.from_numpy(LAMBDA2*npr), device=DEVICE)
# Define training parameters
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)
# Define Edge dataset
t0 = time.time()
spatial_cons = OriginSingleEpochSpatialEdgeConstructor(data_provider, iteration, S_N_EPOCHS, B_N_EPOCHS, N_NEIGHBORS)
edge_to, edge_from, probs, feature_vectors, attention = spatial_cons.construct()
t1 = time.time()
probs = probs / (probs.max()+1e-3)
eliminate_zeros = probs>5e-2 #1e-3
edge_to = edge_to[eliminate_zeros]
edge_from = edge_from[eliminate_zeros]
probs = probs[eliminate_zeros]
dataset = DVIDataHandler(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=2000, sampler=sampler, num_workers=8, prefetch_factor=10)
########################################################################################################################
# TRAIN #
########################################################################################################################
trainer = OriginDVITrainer(model, criterion, optimizer, lr_scheduler, edge_loader=edge_loader, DEVICE=DEVICE)
# trainer = DVIALMODITrainer(model, criterion, optimizer, lr_scheduler, edge_loader=edge_loader, DEVICE=DEVICE)
t2=time.time()
trainer.train(PATIENT, MAX_EPOCH)
t3 = time.time()
# save result
save_dir = data_provider.model_path
trainer.record_time(save_dir, "time_{}".format(VIS_MODEL_NAME), "complex_construction", str(iteration), t1-t0)
trainer.record_time(save_dir, "time_{}".format(VIS_MODEL_NAME), "training", str(iteration), t3-t2)
save_dir = os.path.join(data_provider.model_path, "Epoch_{}".format(iteration))
trainer.save(save_dir=save_dir, file_name="{}".format(VIS_MODEL_NAME))
print("Finish epoch {}...".format(iteration))
prev_model.load_state_dict(model.state_dict())
for param in prev_model.parameters():
param.requires_grad = False
w_prev = dict(prev_model.named_parameters())
########################################################################################################################
# VISUALIZATION #
########################################################################################################################
from singleVis.visualizer import visualizer
vis = visualizer(data_provider, projector, 200, "tab10")
save_dir = os.path.join(data_provider.content_path, "imgptDVI_")
if not os.path.exists(save_dir):
os.mkdir(save_dir)
for i in range(EPOCH_START, EPOCH_END+1, EPOCH_PERIOD):
vis.savefig(i, path=os.path.join(save_dir, "{}_{}_{}.png".format(VIS_MODEL_NAME, i, VIS_METHOD)))
########################################################################################################################
# EVALUATION #
########################################################################################################################
# eval_epochs = range(EPOCH_START, EPOCH_END+1, EPOCH_PERIOD)
# EVAL_EPOCH_DICT = {
# "mnist":[1,10,15],
# "fmnist":[1,25,50],
# "cifar10":[1,100,199]
# }
# eval_epochs = EVAL_EPOCH_DICT[DATASET]
# 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))