深度学习和目标检测系列教程 11-300：小麦数据集训练Faster-RCNN模型

Posted 2021-08-06 刘润森！

@Author：Runsen

上次训练的Faster-RCNN的数据格式是xml和jpg图片提供，在很多Object-Detection中，数据有的是csv格式，

• 数据集来源：https://www.kaggle.com/c/global-wheat-detection

width和heigth是图片的长和宽，bbox是框的位置。

我们需要在小麦植物的室外图像中检测小麦头，分类的类别只有一个。

我们来看一个牛逼人的做法：https://www.kaggle.com/pestipeti/pytorch-starter-fasterrcnn-train

这次使用torch训练Faster-RCNN和之前的没有什么不一样。

import pandas as pd
import numpy as np
import cv2
import os
import re

from PIL import Image

import albumentations as A
from albumentations.pytorch.transforms import ToTensorV2

import torch
import torchvision

from torchvision.models.detection.faster_rcnn import FastRCNNPredictor
from torchvision.models.detection import FasterRCNN
from torchvision.models.detection.rpn import AnchorGenerator

from torch.utils.data import DataLoader, Dataset
from torch.utils.data.sampler import SequentialSampler

from matplotlib import pyplot as plt

DIR_INPUT = '/kaggle/input/global-wheat-detection'
DIR_TRAIN = f'{DIR_INPUT}/train'
DIR_TEST = f'{DIR_INPUT}/test'

train_df = pd.read_csv(f'{DIR_INPUT}/train.csv')

train_df['x'] = -1
train_df['y'] = -1
train_df['w'] = -1
train_df['h'] = -1

def expand_bbox(x):
    r = np.array(re.findall("([0-9]+[.]?[0-9]*)", x))
    if len(r) == 0:
        r = [-1, -1, -1, -1]
    return r
# 读取'x', 'y', 'w', 'h'
train_df[['x', 'y', 'w', 'h']] = np.stack(train_df['bbox'].apply(lambda x: expand_bbox(x)))
train_df.drop(columns=['bbox'], inplace=True)
train_df['x'] = train_df['x'].astype(np.float)
train_df['y'] = train_df['y'].astype(np.float)
train_df['w'] = train_df['w'].astype(np.float)
train_df['h'] = train_df['h'].astype(np.float)

image_ids = train_df['image_id'].unique()
valid_ids = image_ids[-665:]
train_ids = image_ids[:-665]

# Albumentations
def get_train_transform():
    return A.Compose([
        A.Flip(0.5),
        ToTensorV2(p=1.0)
    ], bbox_params={'format': 'pascal_voc', 'label_fields': ['labels']})

def get_valid_transform():
    return A.Compose([
        ToTensorV2(p=1.0)
    ], bbox_params={'format': 'pascal_voc', 'label_fields': ['labels']})

# load a model; pre-trained on COCO
model = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True)

num_classes = 2  # 1 class (wheat) + background

# get number of input features for the classifier
in_features = model.roi_heads.box_predictor.cls_score.in_features

# replace the pre-trained head with a new one
model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes)

不同的是定义了Averager类，这一个类来保存对应的loss。

device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
images, targets, image_ids = next(iter(train_data_loader))
images = list(image.to(device) for image in images)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
boxes = targets[2]['boxes'].cpu().numpy().astype(np.int32)
sample = images[2].permute(1,2,0).cpu().numpy()

fig, ax = plt.subplots(1, 1, figsize=(16, 8))

for box in boxes:
    cv2.rectangle(sample,
                  (box[0], box[1]),
                  (box[2], box[3]),
                  (220, 0, 0), 3)
    
ax.set_axis_off()
ax.imshow(sample)

model.to(device)
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params, lr=0.005, momentum=0.9, weight_decay=0.0005)
# lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=3, gamma=0.1)
lr_scheduler = None

num_epochs = 2

loss_hist = Averager()
itr = 1

for epoch in range(num_epochs):
    loss_hist.reset()
    
    for images, targets, image_ids in train_data_loader:
        
        images = list(image.to(device) for image in images)
        targets = [{k: v.to(device) for k, v in t.items()} for t in targets]

        loss_dict = model(images, targets)

        losses = sum(loss for loss in loss_dict.values())
        loss_value = losses.item()

        loss_hist.send(loss_value)

        optimizer.zero_grad()
        losses.backward()
        optimizer.step()

        if itr % 50 == 0:
            print(f"Iteration #{itr} loss: {loss_value}")

        itr += 1
    
    # update the learning rate
    if lr_scheduler is not None:
        lr_scheduler.step()

    print(f"Epoch #{epoch} loss: {loss_hist.value}")   

images, targets, image_ids = next(iter(valid_data_loader))
images = list(img.to(device) for img in images)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
boxes = targets[1]['boxes'].cpu().numpy().astype(np.int32)
sample = images[1].permute(1,2,0).cpu().numpy()
model.eval()
cpu_device = torch.device("cpu")

outputs = model(images)
outputs = [{k: v.to(cpu_device) for k, v in t.items()} for t in outputs]
fig, ax = plt.subplots(1, 1, figsize=(16, 8))

for box in boxes:
    cv2.rectangle(sample,
                  (box[0], box[1]),
                  (box[2], box[3]),
                  (220, 0, 0), 3)
    
ax.set_axis_off()
ax.imshow(sample)
torch.save(model.state_dict(), 'fasterrcnn_resnet50_fpn.pth')

这个代码真的值得学习和模仿：
https://www.kaggle.com/pestipeti/pytorch-starter-fasterrcnn-train