Я использую код FasterRCNN и экспериментировал с различными магистралями (resnet18 ... 101). Я создал класс FasterRCNN с функцией инициализации. Во время обучения на моем linux сервере я сталкиваюсь с ошибкой «Пустой параметр» (можно увидеть на прикрепленном изображении 1) 
Пока тот же код выполняется без каких-либо ошибка на моей локальной машине. Я использую дистрибутив Anaconda 4.0.1 и Python 3.7 с pytorch 1.2. Ниже вы можете сослаться на мой код (модель и поезд) -
class Faster_RCNN(nn.Module):
def __init__(self, num_classes=2, backbone = 'mobilenet_v2', test= False):
super(Faster_RCNN, self).__init__()
self.classes = num_classes # replace the classifier with a new one, that has num_classes 2 for our use.
# 1 class (bar) + background
self.test = test
self.backbone = backbone
print('....Initializing Model....\n')
print(f'....This model uses {self.backbone} as backbone....\n')
if self.backbone == 'mobilenet_v2':
backbone = torchvision.models.mobilenet_v2(pretrained=True).features
backbone.out_channels = 1280
anchor_generator = AnchorGenerator(sizes=((32, 64, 128),),
aspect_ratios=((1.0),))
roi_pooler = torchvision.ops.MultiScaleRoIAlign(featmap_names=[0],
output_size=7,
sampling_ratio=2)
self.model = FasterRCNN(backbone,
num_classes=2,
rpn_anchor_generator=anchor_generator,
box_roi_pool=roi_pooler)
if self.backbone == 'resnet18':
bb = torchvision.models.resnet18(pretrained=True)
backbone = nn.Sequential(*list(bb.children())[:-2])
backbone.out_channels = 512
anchor_generator = AnchorGenerator(sizes=((32, 64, 128),),
aspect_ratios=((1.0),))
roi_pooler = torchvision.ops.MultiScaleRoIAlign(featmap_names=[0],
output_size=7,
sampling_ratio=2)
self.model = FasterRCNN(backbone,
num_classes=2,
rpn_anchor_generator=anchor_generator,
box_roi_pool=roi_pooler)
if self.backbone == 'resnet34':
bb = torchvision.models.resnet34(pretrained=True)
backbone = nn.Sequential(*list(bb.children())[:-2])
backbone.out_channels = 512
anchor_generator = AnchorGenerator(sizes=((32, 64, 128),),
aspect_ratios=((1.0),))
roi_pooler = torchvision.ops.MultiScaleRoIAlign(featmap_names=[0],
output_size=7,
sampling_ratio=2)
self.model = FasterRCNN(backbone,
num_classes=2,
rpn_anchor_generator=anchor_generator,
box_roi_pool=roi_pooler)
if self.backbone == 'resnet50':
bb = torchvision.models.resnet50(pretrained=True)
backbone = nn.Sequential(*list(bb.children())[:-2])
backbone.out_channels = 2048
anchor_generator = AnchorGenerator(sizes=((32, 64, 128),),
aspect_ratios=((1.0),))
roi_pooler = torchvision.ops.MultiScaleRoIAlign(featmap_names=[0],
output_size=7,
sampling_ratio=2)
self.model = FasterRCNN(backbone,
num_classes=2,
rpn_anchor_generator=anchor_generator,
box_roi_pool=roi_pooler)
if self.backbone == 'resnet101':
bb = torchvision.models.resnet50(pretrained=True)
backbone = nn.Sequential(*list(bb.children())[:-2])
backbone.out_channels = 2048
anchor_generator = AnchorGenerator(sizes=((32, 64, 128),),
aspect_ratios=((1.0),))
roi_pooler = torchvision.ops.MultiScaleRoIAlign(featmap_names=[0],
output_size=7,
sampling_ratio=2)
self.model = FasterRCNN(backbone,
num_classes=2,
rpn_anchor_generator=anchor_generator,
box_roi_pool=roi_pooler)
def forward(self, x,y):
if self.test:
pred = self.model(x)
else:
pred = self.model(x,y)
return pred
Вот код поезда:
import dataloader
import model as md
import numpy as np
#import matplotlib.pyplot as plt
import torch
from torch.optim import Adam
from torch.utils.tensorboard import SummaryWriter
import argparse
writer = SummaryWriter()
bkbone = ['resnet18','resenet34','resenet50','resnet101']
for bbone in bkbone:
ap = argparse.ArgumentParser()
ap.add_argument("-b", "--backbone", required = False, default = bbone, help = "resenet18, resnet34, resnet50, resenet101 and mobilenet_v2 can be given")
args = vars(ap.parse_args())
epochs = 1000
batch_size = 10
lr = 0.005
path = 'D:\\beantech_Data\\objtect_detection'
# path = '/media/TBDataStudent/pankaj/beantech/object_detection'
data= dataloader.Bar(root=path, batch_size=batch_size)
model = md.Faster_RCNN(backbone=args['backbone']).cuda()
model.train()
optimizer = Adam(model.parameters(), lr = lr, weight_decay=0.0005)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,step_size=3,gamma=0.1)
### Train ###
for epoch in range(epochs):
loss_classifier=[]
loss_box_reg =[]
loss_objectness = []
loss_rpn_box_reg = []
Tloss = []
for _, (img, label) in enumerate(data.train_loader):
images = [image.cuda() for image in img ]
for i in range(len(label)):
label[i]['boxes']=label[i]['boxes'].cuda()
label[i]['labels']=label[i]['labels'].cuda()
optimizer.zero_grad()
output = model(images, label)
l1, l2, l3, l4 = output
loss_classifier.append(output[l1].item())
loss_box_reg.append(output[l2].item())
loss_objectness.append(output[l3].item())
loss_rpn_box_reg.append(output[l4].item())
loss = sum(list(output.values()))
loss.backward()
optimizer.step()
Tloss.append(loss.item())
writer.add_scalar(l1, np.mean(loss_classifier), epoch)
writer.add_scalar(l2, np.mean(loss_box_reg), epoch)
writer.add_scalar(l3, np.mean(loss_objectness), epoch)
writer.add_scalar(l4, np.mean(loss_rpn_box_reg), epoch)
writer.add_scalar('Total Loss', np.mean(Tloss), epoch)
print(f'\n\n --{l1}: {np.mean(loss_classifier)}\n --{l2}: {np.mean(loss_box_reg)}\n --{l3}: {np.mean(loss_objectness)}\n --{l4}: {np.mean(loss_rpn_box_reg)}')
print(f'Total loss of epoch {epoch}is: {np.mean(Tloss)}')
writer.close()
torch.save(model.state_dict(), f'FasterRCNN_{args["backbone"]}'+'.pt')