[Paddle學習筆記][11][基於YOLOv3的昆蟲檢測-訓練模型]
阿新 • • 發佈：2020-09-17
說明：
本例程使用YOLOv3進行昆蟲檢測。例程分為資料處理、模型設計、損失函式、訓練模型、模型預測和測試模型六個部分。本篇為第四部分，使用Momentum優化演算法訓練YOLOv3網路權重。資料集訓練一輪就使用驗證集計算驗證損失，如果當前為最好驗證損失，則儲存網路權重。
實驗程式碼：
訓練模型：
from source.train import train # 訓練模型 # train(epoch_numb=1, # learning_rate=[0.001, 0.0005, 0.00025, 0.00002, 0.00001], # lr_iterations=[4250, 8500, 12750, 17000],# model_path='./output/darknet53-yolov3', # train_path='./dataset/train/', # valid_path='./dataset/val/') # train(epoch_numb=100, # learning_rate=[0.001, 0.0005, 0.00025, 0.00002, 0.00001], # lr_iterations=[8500, 17000, 25500, 34000], # model_path='./output/darknet53-yolov3', # train_path='./dataset/train/',# valid_path='./dataset/val/') train(epoch_numb=150, learning_rate=[0.001, 0.0005, 0.00025, 0.00002, 0.00001], lr_iterations=[12750, 25500, 38250, 51000], model_path='./output/darknet53-yolov3', train_path='./dataset/train/', valid_path='./dataset/val/')
結果：
valid - iter: 63601, epoch: 149, loss: 95.885
best - epoch: 74, loss:92.725, time: 43219s
訓練結果
loss_cls + loss_obj + loss_loc + ignore + wtloc: base lr:0.001, epoch:50, best loss: 100.772
loss_cls + loss_obj + loss_loc + ignore + wtloc: base lr:0.001, epoch:100, best loss: 93.092
loss_cls + loss_obj + loss_loc + ignore + wtloc: base lr:0.001, epoch:150, best loss: 92.725
train.py檔案
import time import numpy as np import paddle.fluid as fluid from paddle.fluid.dygraph.base import to_variable from paddle.utils.plot import Ploter from source.data import multip_thread_reader from source.model import YOLOv3 from source.loss import get_sum_loss num_classes = 7 # 類別數量 anchor_size = [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326] # 錨框大小 anchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]] # 錨框掩碼 ignore_threshold = 0.7 # 忽略閾值 downsample_ratio = 32 # 下采樣率 def train(epoch_numb=50, learning_rate=[0.001, 0.0005, 0.00025, 0.00002, 0.00001], lr_iterations=[4250, 8500, 12750, 17000], model_path='./output/darknet53-yolov3', train_path='./dataset/train/', valid_path='./dataset/val/'): """ 功能: 訓練模型 輸入: epoch_numb - 迭代週期 learning_rate - 學習率 lr_iterations - 學習率迭代數 model_path - 模型儲存路徑 train_path - 訓練資料路徑 valid_path - 驗證資料路徑 輸出: """ with fluid.dygraph.guard(): # 動態圖域 # 準備資料 train_reader = multip_thread_reader(train_path, 4, 'train') # 讀取訓練資料 valid_reader = multip_thread_reader(valid_path, 2, 'valid') # 讀取驗證資料 # 宣告模型 model = YOLOv3(num_classes=num_classes, anchor_mask=anchor_mask) # 優化演算法 optimizer = fluid.optimizer.Momentum( learning_rate=fluid.layers.piecewise_decay(boundaries=lr_iterations, values=learning_rate), momentum=0.9, parameter_list=model.parameters()) # 開始訓練 train_prompt = "Train loss" # 訓練標籤 valid_prompt = "Valid loss" # 驗證標籤 ploter = Ploter(train_prompt, valid_prompt) # 損失影象 iterator = 1 # 迭代次數 iter_time = time.time() # 迭代時間 best_epoch = 0 # 最好週期 best_loss = 100000.0 # 最好損失 for epoch_id in range(epoch_numb): # 訓練模型 model.train() # 設定訓練 for batch_id, train_data in enumerate(train_reader()): # 讀取資料 image, gtbox, gtcls, image_size = train_data # 讀取一條資料 image = to_variable(image) # 轉換資料格式 # 前向傳播 infer = model(image) # 計算損失 loss = get_sum_loss( infer, gtbox, gtcls, num_classes, anchor_size, anchor_mask, ignore_threshold, downsample_ratio) # 反向傳播 loss.backward() # 反向傳播 optimizer.minimize(loss) # 更新權重 model.clear_gradients() # 清除梯度 # 顯示損失 if iterator % 10 == 0: # 顯示損失次數 ploter.append(train_prompt, iterator, loss.numpy()) # 新增損失數值 ploter.plot() # 顯示損失影象 print("train - iter: {:5d}, epoch: {:3d}, loss: {[0]:.3f}, best loss:{:.3f}".format( iterator, epoch_id, loss.numpy(), best_loss)) iterator += 1 # 增加損失次數 # 驗證模型 loss_list = [] # 損失列表 model.eval() # 設定驗證 for batch_id, valid_data in enumerate(valid_reader()): # 讀取資料 image, gtbox, gtcls, image_size = valid_data # 讀取一條資料 image = to_variable(image) # 轉換資料格式 # 前向傳播 infer = model(image) # 計算損失 loss = get_sum_loss( infer, gtbox, gtcls, num_classes, anchor_size, anchor_mask, ignore_threshold, downsample_ratio) # 記錄損失 loss_list.append(loss.numpy()) # 顯示損失 mean_loss = np.mean(loss_list) # 計算驗證損失 ploter.append(valid_prompt, iterator, mean_loss) # 新增損失數值 ploter.plot() # 顯示損失影象 print("valid - iter: {:5d}, epoch: {:3d}, loss: {:.3f}".format(iterator, epoch_id, mean_loss)) # 儲存模型 if mean_loss < best_loss: fluid.save_dygraph(model.state_dict(), model_path) # 儲存模型 best_loss = mean_loss # 更新損失 best_epoch = epoch_id # 更新迭代 # 顯示時間 iter_time = time.time() - iter_time # 總的時間 print("best - epoch:{:4d}, loss:{:.3f}, time: {:.0f}s".format(best_epoch ,best_loss, iter_time))
參考資料：
https://blog.csdn.net/litt1e/article/details/88814417
https://blog.csdn.net/litt1e/article/details/88852745
https://blog.csdn.net/litt1e/article/details/88907542
https://aistudio.baidu.com/aistudio/projectdetail/742781
https://aistudio.baidu.com/aistudio/projectdetail/672017
https://aistudio.baidu.com/aistudio/projectdetail/868589
https://aistudio.baidu.com/aistudio/projectdetail/122277