-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathpltrain.py
254 lines (229 loc) · 9.73 KB
/
pltrain.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
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
import torch.optim as optim
import numpy as np
import time,os,random
import pytorch_lightning as pl
from pytorch_lightning.loggers import WandbLogger
import torchvision.transforms as T
import functools as ft
import warnings
warnings.filterwarnings('ignore')
from TAE import TAE,Sampler,TAETrainer,TAETunner
from TAE3D import VoxTAE,VoxSampler,VoxTAETrainer
from datasets.babyarc import ConceptCompositionDataset,NoiseBabyARC,collate_fn_lwp,ParseBabyARC,ConceptDataset
from datasets.omniglot import OmniGlot,OmniGlotDataset,collate_fn_ogp,OmniGlotParse
from datasets.shapenetvox import ShapeNetVoxDataset
""" 1. Setting """
dataset='omniglot' # lineworld omniglot shapenet lineworld_parse omniglot_parse
project='TAE_'+dataset
PATH = os.path.join('.','ckpts',dataset)
if not os.path.exists(PATH): os.makedirs(PATH)
ind=1 # Update this for each machine!
name=str(ind)
PATH = os.path.join(PATH,name)
accelerator="cuda"
devices="auto" # [1,2,3,4,5,6] #"auto"
strategy="auto"
log_every_n_steps=10
draw_freq=1
noise=False
use_small=True
mapper_small=True
use_aug=True
PE_mode='rand' # rand fix none
opt=ft.partial(optim.Adam)
sched=ft.partial(optim.lr_scheduler.StepLR,step_size=30,gamma=0.95)
# sched = ft.partial(optim.lr_scheduler.CosineAnnealingWarmRestarts, T_0=30, T_mult=1, eta_min=1e-5)
batch_size=96
eval_size=96
epochs=1000
c_m=1
mask_p=0#0.8
mask_ds=2
Niter=10 # kmeans N iters
use_zca=False # use zca in kmeans
dropout=0.1
use_out_res=False
loss_option='mse_10.0' # focal_1.0-dice_1.0-bce_1.0-smoothl1_1.0-mse_1.0
lambda_r=10 # for tune, p2 ratio
use_attn=False
use_self_attn=True
n_heads=8
d_head=8
context_dim=64
share_mapper=False
use_ldm=False
ldm_out=8
ldm_ds=[1,2,1,1]
lr=1e-3
sampler_dim=32
embed_dim=256
mapper_mid=256 # clustering cnn feature dim
method='EM'
blur_sigma=0 # sigma for gaussian blur of masks
quota=32 # quota for each player
N=6
NP1=6
NP2=6
NH1=4
NH2=4
NK=2
K=3
t=100
sigma=2.5
lambda_P=2e-2
focal_alpha=0.95 # alpha for focal loss
alpha_overlap=0.005 # overlap
alpha_l2=1e-6 # l2 to control m scale
alpha_resources=0.1 # resources limit
jump_alpha=25 #25
jump_shift=0.1 # threshhold for jumping to 1
jump_smooth=False
gamma_cluster=1e-2 # clustering
memlen=3000
threshold=(0.01,0.2)
cluster_start_epoch=0
beta=0 # sde overall
use_rl=False
use_ae=False
rl_start_epoch=0
critic_mid=128
ppo_gamma=0.95
ppo_lamda=0.95
ppo_epsilon=0.5
ppo_entropy_coef=0.2
ppo_K_epochs=3
ppo_use_grad_clip=True
ppo_use_lr_decay=False
ppo_reward_norm=True
ppo_human_tune=False
ppo_inc_reward=False
ppo_sparse_reward=False
ppo_prefer_last=False
ppo_max_train_steps=1e6
ppo_update_every=3
ppo_pc_rand=10
reward_option='loss_1.0-prefer_0.0'
prefer_option='ct_1.0-cp_1.0-sm_1.0-z_0.0:spl'
lr_a=1e-3
lr_c=1e-3
if use_ldm: name+='-ldm'
if use_rl:
name+='-rl'
if not use_ae: name+='N'
if use_attn: name+='-a'
if ppo_human_tune: name+='-H'
""" 2. Dataset """
num_workers=16
if dataset=='lineworld':
c_in=3
datasize=(16,16)
lwname='LineWorld40k'
trainset=NoiseBabyARC(lwname+'_train',datasize,noise=noise)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size,
shuffle=True, num_workers=num_workers)
devset=NoiseBabyARC(lwname+'_dev',datasize,noise=noise)
devloader = torch.utils.data.DataLoader(devset, batch_size=eval_size,
shuffle=True, num_workers=num_workers)
testset=NoiseBabyARC(lwname+'_test',datasize,noise=noise)
testloader = torch.utils.data.DataLoader(testset, batch_size=eval_size,
shuffle=True, num_workers=num_workers)
elif dataset=='omniglot':
c_in=1
size=28 #16 32
og_dir='./datasets/files'
transforms = torch.nn.Sequential(
T.RandomHorizontalFlip(p=0.5),
T.RandomVerticalFlip(p=0.5),
T.RandomRotation(degrees=(0, 180))
) if use_aug else None
trainset=OmniGlotDataset(og_dir,split='train',transform=transforms,size=size)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size,
shuffle=True, num_workers=num_workers)
testset=OmniGlotDataset(og_dir,split='test',size=size)
testloader = torch.utils.data.DataLoader(testset, batch_size=eval_size,
shuffle=True, num_workers=num_workers)
devset=OmniGlotDataset(og_dir,split='dev',size=size)
devloader = torch.utils.data.DataLoader(devset, batch_size=eval_size,
shuffle=True, num_workers=num_workers)
elif dataset=='shapenet':
c_in=1
snv16=np.load('./datasets/files/shapenetvox16.npy')
seq=list(range(len(snv16)))
random.shuffle(seq)
snv16=snv16[seq]
trainset=ShapeNetVoxDataset(snv16)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size,
shuffle=True, num_workers=num_workers)
testset=ShapeNetVoxDataset(snv16,split='test')
testloader = torch.utils.data.DataLoader(testset, batch_size=eval_size,
shuffle=True, num_workers=num_workers)
devset=ShapeNetVoxDataset(snv16,split='dev')
devloader = torch.utils.data.DataLoader(devset, batch_size=eval_size,
shuffle=True, num_workers=num_workers)
elif dataset=='lineworld_parse':
c_in=3
NP2=5 # Need to replace predicate head when finetuning!
datasize=(16,16)
lwname='LineWorldParse'
trainset=ParseBabyARC(lwname+'_train')
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size,
shuffle=True, num_workers=num_workers,collate_fn=collate_fn_lwp)
devset=ParseBabyARC(lwname+'_dev')
devloader = torch.utils.data.DataLoader(devset, batch_size=eval_size,
shuffle=True, num_workers=num_workers,collate_fn=collate_fn_lwp)
testset=ParseBabyARC(lwname+'_test')
testloader = torch.utils.data.DataLoader(testset, batch_size=eval_size,
shuffle=True, num_workers=num_workers,collate_fn=collate_fn_lwp)
elif dataset=='omniglot_parse':
c_in=1
size=28 #16 32
og_dir='./datasets/files'
transforms = torch.nn.Sequential(
T.RandomHorizontalFlip(p=0.5),
T.RandomVerticalFlip(p=0.5),
T.RandomRotation(degrees=(0, 180))
) if use_aug else None
trainset=OmniGlotParse(og_dir,split='train',transform=transforms,size=size)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size,
shuffle=True, num_workers=num_workers,collate_fn=collate_fn_ogp)
testset=OmniGlotParse(og_dir,split='test',size=size)
testloader = torch.utils.data.DataLoader(testset, batch_size=eval_size,
shuffle=True, num_workers=num_workers,collate_fn=collate_fn_ogp)
devset=OmniGlotParse(og_dir,split='dev',size=size)
devloader = torch.utils.data.DataLoader(devset, batch_size=eval_size,
shuffle=True, num_workers=num_workers,collate_fn=collate_fn_ogp)
""" 3. Training """
Sampler_=VoxSampler if dataset=='shapenet' else Sampler
TAE_=VoxTAE if dataset=='shapenet' else TAE
TAETrainer_=VoxTAETrainer if dataset=='shapenet' else TAETrainer
if 'parse' in dataset.split('_'):
TAETrainer_=TAETunner
loss_option='mse_1.0'
use_rl=False
if use_rl and not use_ae: rl=1e-10
sampler=Sampler_(c_in=c_in,c_m=c_m,dim=sampler_dim,method=method,embed_dim=embed_dim,sigma=sigma,use_small=use_small,K=K,t=t,ldm_ds=ldm_ds,
lambda_P=lambda_P,NP1=NP1,NP2=NP2,NH1=NH1,NH2=NH2,NK=NK,memlen=memlen,mapper_mid=mapper_mid,dropout=dropout,use_rl=use_rl,
threshold=threshold,use_out_res=use_out_res,Niter=Niter,use_zca=use_zca,use_self_attn=use_self_attn,use_ldm=use_ldm,
n_heads=n_heads, d_head=d_head,context_dim=context_dim, share_mapper=share_mapper,critic_mid=critic_mid,ldm_out=ldm_out,mapper_small=mapper_small)
tae=TAE_(sampler,N=N,embed_dim=embed_dim,c_m=c_m,quota=quota,focal_alpha=focal_alpha,alpha_overlap=alpha_overlap,jump_alpha=jump_alpha,jump_smooth=jump_smooth,
alpha_l2=alpha_l2,alpha_resources=alpha_resources,cluster_start_epoch=cluster_start_epoch,PE_mode=PE_mode,jump_shift=jump_shift,
beta=beta,gamma_cluster=gamma_cluster,blur_sigma=blur_sigma,mask_ds=mask_ds,mask_p=mask_p,loss_option=loss_option,lambda_r=lambda_r,
use_rl=use_rl,ppo_gamma=ppo_gamma,ppo_lamda=ppo_lamda,ppo_K_epochs=ppo_K_epochs,ppo_use_grad_clip=ppo_use_grad_clip,lr_a=lr_a,lr_c=lr_c,
ppo_use_lr_decay=ppo_use_lr_decay,ppo_max_train_steps=ppo_max_train_steps,ppo_update_every=ppo_update_every,reward_option=reward_option,
ppo_epsilon=ppo_epsilon,ppo_entropy_coef=ppo_entropy_coef,rl_start_epoch=rl_start_epoch,use_ldm=use_ldm,ppo_reward_norm=ppo_reward_norm,
ppo_inc_reward=ppo_inc_reward,prefer_option=prefer_option,ppo_human_tune=ppo_human_tune,ppo_pc_rand=ppo_pc_rand,
ppo_sparse_reward=ppo_sparse_reward,ppo_prefer_last=ppo_prefer_last)
model=TAETrainer_(tae,opt=opt,sched=sched,lr=lr,dataset=dataset,draw_freq=draw_freq,use_rl=use_rl)
if use_rl: strategy=pl.strategies.DDPStrategy(find_unused_parameters=True, static_graph=True)
wandb_logger = WandbLogger(name=name,project=project,save_dir=PATH)
lr_monitor = pl.callbacks.LearningRateMonitor(logging_interval='step')
early_stopping = pl.callbacks.EarlyStopping('val_loss', mode="min", patience=100)
checkpoint_callback = pl.callbacks.ModelCheckpoint(monitor='val_loss')
trainer = pl.Trainer(max_epochs=epochs,default_root_dir=PATH,log_every_n_steps=log_every_n_steps,
logger=wandb_logger,accelerator=accelerator, strategy=strategy, devices=devices,
callbacks=[lr_monitor,early_stopping,checkpoint_callback])
trainer.fit(model, trainloader, devloader)