def create_segmenter(net, pretrained, num_classes):
"""Create Encoder; for now only ResNet [50,101,152]"""
from models.resnet import rf_lw50, rf_lw101, rf_lw152
if str(net) == '50':
return rf_lw50(num_classes, imagenet=pretrained)
elif str(net) == '101':
return rf_lw101(num_classes, imagenet=pretrained)
elif str(net) == '152':
return rf_lw152(num_classes, imagenet=pretrained)
else:
raise ValueError("{} is not supported".format(str(net)))
def get_segmenter(
enc_backbone,
enc_pretrained,
num_classes,
):
"""Create Encoder-Decoder; for now only ResNet [50,101,152] Encoders are supported"""
if enc_backbone == "50":
return rf_lw50(num_classes, imagenet=enc_pretrained)
elif enc_backbone == "101":
return rf_lw101(num_classes, imagenet=enc_pretrained)
elif enc_backbone == "152":
return rf_lw152(num_classes, imagenet=enc_pretrained)
else:
raise ValueError("{} is not supported".format(str(enc_backbone)))
def create_segmenter(net, pretrained, num_classes):
"""Create Encoder; for now only ResNet [50,101,152]"""
import sys
sys.path.append("../")
from models.resnet import rf_lw50, rf_lw101, rf_lw152
init_model = '../models/resnet/50_person.ckpt'
if str(net) == '50':
return rf_lw50(num_classes, model_path=init_model, imagenet=pretrained)
elif str(net) == '101':
return rf_lw101(num_classes,
model_path=init_model,
imagenet=pretrained)
elif str(net) == '152':
return rf_lw152(num_classes, imagenet=pretrained)
else:
raise ValueError("{} is not supported".format(str(net)))
def get_model(model_name, classes, pre_train=False, mode='train'):
if model_name == 'ESpnet_2_8_decoder':
from models import Espnet
if pre_train:
pre_train_path = os.path.join(pre_train)
model = Espnet.ESPNet(classes, 2, 8, pre_train_path, mode=mode)
else:
model = Espnet.ESPNet(classes, 2, 8, mode=mode)
elif model_name == 'ESpnet_2_8':
from models import Espnet
model = Espnet.ESPNet_Encoder(classes, 2, 8)
elif model_name == 'EDAnet':
from models import EDANet
model = EDANet.EDANet(classes)
elif model_name == 'ERFnet':
from models import ERFnet
model = ERFnet.Net(classes)
elif model_name == 'Enet':
from models import Enet
model = Enet.ENet(classes)
elif model_name == 'IRRnet_2_8':
from models import Irregularity_conv
model = Irregularity_conv.ESPNet(classes, 2, 8, mode=mode)
elif model_name == 'MOBILE_V2':
from models import mobilenet
model = mobilenet.mbv2(classes)
elif model_name == 'RF_LW_resnet_50':
from models import resnet
model = resnet.rf_lw50(classes)
elif model_name == 'RF_LW_resnet_101':
from models import resnet
model = resnet.rf_lw101(classes)
elif model_name == 'RF_LW_resnet_152':
from models import resnet
model = resnet.rf_lw152(classes)
elif model_name == 'Bisenet':
from models import BiSeNet
model = BiSeNet.BiSeNet(out_class=classes)
elif model_name == 'Basenet':
from models import Basenet
model = Basenet.Basenet(classes)
else:
raise NotImplementedError
return model
def create_multiNet(net,
num_classes,
num_depths=10,
pretrained=None,
task_type=2):
from models.multitask import multiNet
from models.resnet import rf_lw50, rf_lw101, rf_lw152
if str(net) == 'multi':
return multiNet(num_classes, num_depths, task_type)
elif str(net) == '50':
return rf_lw50(num_classes, imagenet=pretrained)
elif str(net) == '101':
return rf_lw101(num_classes, imagenet=pretrained)
elif str(net) == '152':
return rf_lw152(num_classes, imagenet=pretrained)
else:
raise ValueError("{} is not supported".format(str(net)))
import torch.nn as nn
from skimage import io, color
from scipy.special import softmax
from PIL import Image
cmap = np.load('../../utils/cmap.npy')
has_cuda = torch.cuda.is_available()
img_dir = '../../../test/rgb/0/'
# img_dir = "../imgs/ho3d_test/"
imgs = glob.glob('{}*[0-9].png'.format(img_dir))
n_classes = 3
# net = rf_lw50(n_classes, pretrained=True).eval()
# iter = 0
net = nn.DataParallel(rf_lw50(num_classes=3, pretrained=True))
weights_path = "../../checkpoints/checkpoint_9070.pth.tar"
# weights_path = "../../checkpoints.iter1/checkpoint_9254.pth.tar"
net.load_state_dict(torch.load(weights_path)["model"])
if has_cuda:
net = net.cuda()
# cv2.namedWindow("preview")
# vc = cv2.VideoCapture(0)
#
# if vc.isOpened(): # try to get the first frame
# rval, frame = vc.read()
# else:
# rval = False
#
import sys
sys.path.append('../../')
import collections
import cv2
from models.resnet import rf_lw50
import cv2
import numpy as np
import torch
has_cuda = torch.cuda.is_available()
n_classes = 11
net = rf_lw50(n_classes, imagenet=False, pretrained=False)
cpkt = torch.load("../face/checkpoint.pth.tar")['segmenter']
weights = collections.OrderedDict()
for key in cpkt:
print(key.split('.', 1))
weights[key.split('.', 1)[1]] = cpkt[key]
net.load_state_dict(weights)
net = net.cuda()
net.eval()
dummy_input1 = torch.randn(1, 3, 512, 512)
dummy_input1 = dummy_input1.cuda()
torch.onnx.export(net, dummy_input1, "refinenet.onnx", verbose=True)
