-
Notifications
You must be signed in to change notification settings - Fork 11
/
train_transitionNet.py
194 lines (172 loc) · 7.73 KB
/
train_transitionNet.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
import copy
import os
import re
from argparse import ArgumentParser
import pytorch_lightning as pl
import torch
from pytorch_lightning import Trainer
from pytorch_lightning import loggers
from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint
from pytorch_lightning.profiler import SimpleProfiler
from pytorch_lightning.utilities.seed import seed_everything
from src.Datasets.BaseLoader import WindowBasedLoader
from src.Datasets.Style100Processor import StyleLoader, Swap100StyJoints
from src.utils import BVH_mod as BVH
from src.utils.motion_process import subsample
def setup_seed(seed:int):
seed_everything(seed,True)
def test_model():
dict = {}
#dict['fast_dev_run'] = 1 # only run 1 train, val, test batch and program ends
dict['limit_train_batches'] = 1.
dict['limit_val_batches'] = 1.
return dict
def detect_nan_par():
'''track_grad_norm": 'inf'''
return { "detect_anomaly":True}
def select_gpu_par():
return {"accelerator":'gpu', "auto_select_gpus":True, "devices":-1}
def create_common_states(prefix:str):
log_name = prefix+'/'
'''test upload'''
parser = ArgumentParser()
parser.add_argument("--dev_run", action="store_true")
parser.add_argument("--version", type=str, default="-1")
parser.add_argument("--epoch",type=str,default="last")
parser.add_argument("--resume", action="store_true")
parser.add_argument("--test",action="store_true")
parser.add_argument("--moe_model",type=str,default="./results/StyleVAE2_style100/myResults/55/m_save_model_332")
parser.add_argument("--pretrained",action="store_true")
parser.add_argument("--predict_phase",action="store_true")
args = parser.parse_args()
ckpt_path_prefix = "results/"
if (args.version != "-1"):
version = args.version
else:
version = None
'''Create Loggers tensorboard'''
if args.dev_run:
log_name += "dev_run"
else:
log_name += "myResults"
tb_logger = pl.loggers.TensorBoardLogger(save_dir="tensorboard_logs/", name=log_name, version=None)
#load_ckpt_path = os.path.join(ckpt_path_prefix, prefix+'/myResults', str(version))
load_ckpt_path = os.path.join(ckpt_path_prefix, prefix+'/myResults', str(version))
save_ckpt_path = os.path.join(ckpt_path_prefix, log_name, str(tb_logger.version))
if (args.resume == True):
check_file = load_ckpt_path+"/"
if (args.epoch == "last"):
check_file += "last.ckpt"
else:
dirs = os.listdir(check_file)
for dir in dirs:
st = "epoch=" + args.epoch + "-step=\d+.ckpt"
out = re.findall(st, dir)
if (len(out) > 0):
check_file += out[0]
print(check_file)
break
resume_from_checkpoint = check_file # results/version/last.ckpt"
else:
resume_from_checkpoint = None
checkpoint_callback = [ModelCheckpoint(dirpath=save_ckpt_path + "/", save_top_k=-1, save_last=False, every_n_epochs=2,save_weights_only=True),
ModelCheckpoint(dirpath=save_ckpt_path + "/", save_top_k=1, monitor="val_loss", save_last=True, every_n_epochs=1,save_weights_only=True),
# EMA(0.99)
]
'''Train'''
checkpoint_callback[0].CHECKPOINT_NAME_LAST = "last"
profiler = SimpleProfiler()#PyTorchProfiler(filename="profiler")
trainer_dict = {
"callbacks":checkpoint_callback,
"profiler":profiler,
"logger":tb_logger
}
return args,trainer_dict,load_ckpt_path
def read_style_bvh(style,content,clip=None):
swap_joints = Swap100StyJoints()
anim = BVH.read_bvh(os.path.join("MotionData/100STYLE/",style,style+"_"+content+".bvh"),remove_joints=swap_joints)
if (clip != None):
anim.quats = anim.quats[clip[0]:clip[1], ...]
anim.hip_pos = anim.hip_pos[clip[0]:clip[1], ...]
anim = subsample(anim,ratio=2)
return anim
def training_style100_phase():
from src.Datasets.StyleVAE_DataModule import StyleVAE_DataModule
from src.Net.TransitionPhaseNet import TransitionNet_phase,Application_phase
prefix = "Transitionv2"
data_set = "style100"
prefix += "_" + data_set
args, trainer_dict, ckpt_path = create_common_states(prefix)
moe_net = torch.load(args.moe_model)
if(args.pretrained==True):
from src.utils.locate_model import locate_model
pretrained_file = locate_model(ckpt_path+"/",args.epoch)
pre_trained = torch.load(pretrained_file)
else:
pre_trained = None
loader = WindowBasedLoader(61, 21, 1)
dt = 1. / 30.
phase_dim = 10
phase_file = "+phase_gv10"
style_file_name = phase_file + WindowBasedLoader(120,0,1).get_postfix_str()
if (args.test == False):
'''Create the model'''
style_loader = StyleLoader()
data_module = StyleVAE_DataModule(style_loader, phase_file + loader.get_postfix_str(),style_file_name, dt=dt,
batch_size=32,mirror=0.0) # when apply phase, should avoid mirror
stat = style_loader.load_part_to_binary("motion_statistics")
mode = "pretrain"
model = TransitionNet_phase(moe_net, data_module.skeleton, pose_channels=9,stat=stat ,phase_dim=phase_dim,
dt=dt,mode=mode,pretrained_model=pre_trained,predict_phase=args.predict_phase)
if (args.dev_run):
trainer = Trainer(**trainer_dict, **test_model(),
**select_gpu_par(), precision=32,reload_dataloaders_every_n_epochs=1,
log_every_n_steps=5, flush_logs_every_n_steps=10,
weights_summary='full')
else:
trainer = Trainer(**trainer_dict, max_epochs=10000,reload_dataloaders_every_n_epochs=1,gradient_clip_val=1.0,
**select_gpu_par(), log_every_n_steps=50,check_val_every_n_epoch=2,
flush_logs_every_n_steps=100)
trainer.fit(model, datamodule=data_module)
else:
style_loader = StyleLoader()
data_module = StyleVAE_DataModule(style_loader, phase_file + loader.get_postfix_str(),style_file_name, dt=dt,batch_size=32,mirror=0.0)
data_module.setup()
check_file = ckpt_path + "/"
if (args.epoch == "last"):
check_file += "last.ckpt"
print(check_file)
else:
dirs = os.listdir(check_file)
for dir in dirs:
st = "epoch=" + args.epoch + "-step=\d+.ckpt"
out = re.findall(st, dir)
if (len(out) > 0):
check_file += out[0]
print(check_file)
break
model = TransitionNet_phase.load_from_checkpoint(check_file, moe_decoder=moe_net, pose_channels=9,phase_dim=phase_dim,
dt=dt,mode='fine_tune',strict=False)
model = model.cuda()
data_module.mirror = 0
app = Application_phase(model, data_module)
model.eval()
app = app.float()
key = "HighKnees"
sty_key = "HighKnees"
cid = 61
sid = 4
src_motion = app.data_module.test_set.dataset[key][cid]
target_motion = app.data_module.test_set_sty.dataset[sty_key][sid]
app.setSource(src_motion)
app.setTarget(target_motion)
source = BVH.read_bvh("source.bvh")
output = copy.deepcopy(source)
output.hip_pos, output.quats = app.forward(t=2., x=0.)
BVH.save_bvh("test_net.bvh", output)
output.hip_pos, output.quats = app.get_source()
BVH.save_bvh("source.bvh", output)
torch.save(model, ckpt_path + "/m_save_model_" + str(args.epoch))
if __name__ == '__main__':
setup_seed(3407)
training_style100_phase()