from squid.data import RandomCropPhoto2PhotoData
from squid.model import SuperviseModel
import torch
import torch.nn as nn
from squid.loss import VGGLoss
from squid.net import AOD_Deep1_Residual_Net
target_net = AOD_Deep1_Residual_Net()
target_net = nn.DataParallel(target_net).cuda()
model = SuperviseModel({
'net': target_net,
'optimizer': torch.optim.Adam([{'name':'net_params', 'params':target_net.parameters(), 'base_lr':1e-5}], betas=(0.9, 0.999), weight_decay=0.0005),
'lr_step_ratio': 0.5,
'lr_step_size': 8,
'supervise':{
'out': {'MSE_loss': {'obj': nn.MSELoss(size_average=True), 'factor':1.0, 'weight':1.0}},
},
'metrics': {}
})
train_dataset = Photo2PhotoData({
'data_root': DATASET_DIR,
'desc_file_path': os.path.join(DATASET_TXT_DIR, DATASET_ID, 'train.txt'),
})
valid_dataset = Photo2PhotoData({
'data_root': DATASET_DIR,
'desc_file_path': os.path.join(DATASET_TXT_DIR, DATASET_ID, 'val.txt'),
})
model = SuperviseModel({
'net':
target_net,
'optimizer':
torch.optim.Adam([{
'name': 'net_params',
'params': target_net.parameters(),
'base_lr': 1e-4
}],
betas=(0.9, 0.999),
weight_decay=0.0005),
'lr_step_ratio':
0.5,
'lr_step_size':
2,
'supervise': {
'out': {
'MSE_loss': {
'obj': nn.MSELoss(size_average=True),
'factor': 1.0,
'weight': 1.0
},
'VGG_loss': {
'obj':
VGGLoss1(vgg19_feature_model_path=
'/root/group-competition/pretrain_model/vgg_tf.pth',
out_feature_level=18),
'factor':
1e-7,
'weight':
1.0
}
},
},
'metrics': {}
})
model = SuperviseModel({
'net':
target_net,
# 'optimizer': torch.optim.Adam([{'name':'net_params', 'params':target_net.parameters(), 'base_lr':1e-4,'warm_epoch':None,'total_epoch':epochs}], betas=(0.9, 0.999), weight_decay=0.0005),
'optimizer':
torch.optim.SGD([{
'name': 'net_params',
'params': target_net.parameters(),
'base_lr': 1e-8,
'warm_epoch': None,
'total_epoch': epochs
}],
lr=1e-8,
momentum=0.9,
weight_decay=0.0005),
# 'lr_step_ratio': 10,
# 'lr_step_size': 1,
'supervise': {
'out': {
'L2_loss': {
'obj': nn.MSELoss(size_average=True),
'factor': 1.0,
'weight': 1.0
}
}
},
'metrics': {
'out': {
'psnr': {
'obj': PSNR()
}
}
},
'not_show_gradient':
True
})
target_net = DarkNet()
model = SuperviseModel({
'net':
target_net,
'optimizer':
torch.optim.Adam([{
'name': 'net_params',
'params': target_net.parameters(),
'base_lr': 1e-3,
'warm_epoch': 2,
'total_epoch': epochs
}],
betas=(0.9, 0.999),
weight_decay=0.0005),
'not_show_gradient':
True,
'supervise': {
'out': {
'L1_loss': {
'obj': nn.L1Loss(size_average=True),
'factor': 1.0,
'weight': 1.0
}
},
},
'metrics': {}
})
# =================== dataset ==========================================================================================
from data import DarkPipeline
from squid.model import SuperviseModel
from squid.metric import IouScore
from squid.metric import AccScore
from squid.data import Photo2MaskData
target_net = ICNet(nclass=15)
model = SuperviseModel({
'net': target_net,
'optimizer': torch.optim.Adam([{'name':'net_params', 'params':target_net.parameters(), 'base_lr':1e-3}], betas=(0.9, 0.999), weight_decay=0.0005),
'lr_step_ratio': 0.5,
'lr_step_size': 500,
'supervise':{
'out1': {'cross_entrypy': {'obj': nn.CrossEntropyLoss(size_average=True), 'factor':1.0, 'weight': 1.0}},
'out2': {'cross_entrypy': {'obj': nn.CrossEntropyLoss(size_average=True), 'factor':1.0, 'weight': 1.0}},
'out3': {'cross_entrypy': {'obj': nn.CrossEntropyLoss(size_average=True), 'factor':1.0, 'weight': 1.0}},
'out4': {'cross_entrypy': {'obj': nn.CrossEntropyLoss(size_average=True), 'factor':1.0, 'weight': 1.0}},
},
'metrics':{
'mask_out': {'iou': {'obj': IouScore(nclass=15)},
'acc': {'obj': AccScore()}
},
},
})
# =================== dataset =====================================================================
train_dataset = Photo2MaskData({
'desc_file_path':'/3T/xcb/pytorch-srgan/examples/face_parse/txt/train_for_dev.txt',
})
valid_dataset = Photo2MaskData({
