Neural network PyTorch迭代修剪
根据教程，我一直在尝试使用“`torch.nn.utils.prune”实现修剪。出于实验目的，我使用LeNet-300-100稠密神经网络和可以访问的MNIST数据集 然后，我使用以下代码片段训练模型：Neural network PyTorch迭代修剪,neural-network,pytorch,Neural Network,Pytorch,根据教程，我一直在尝试使用“`torch.nn.utils.prune”实现修剪。出于实验目的，我使用LeNet-300-100稠密神经网络和可以访问的MNIST数据集 然后，我使用以下代码片段训练模型： # Training loop- for epoch in range(num_epochs): running_loss = 0.0 running_corrects = 0.0 if loc_patience >= patience:
# Training loop-
for epoch in range(num_epochs):
running_loss = 0.0
running_corrects = 0.0
if loc_patience >= patience:
print("\n'EarlyStopping' called!\n")
break
running_loss, running_corrects = train_model(best_model, train_loader)
epoch_loss = running_loss / len(train_dataset)
epoch_acc = running_corrects.double() / len(train_dataset)
# epoch_acc = 100 * running_corrects / len(trainset)
# print(f"\nepoch: {epoch + 1} training loss = {epoch_loss:.4f}, training accuracy = {epoch_acc * 100:.2f}%\n")
running_loss_val, correct, total = test_model(best_model, test_loader)
epoch_val_loss = running_loss_val / len(test_dataset)
val_acc = 100 * (correct / total)
# print(f"\nepoch: {epoch + 1} training loss = {epoch_loss:.4f}, training accuracy = {epoch_acc * 100:.2f}%, val_loss = {epoch_val_loss:.4f} & val_accuracy = {val_acc:.2f}%\n")
print(f"\nepoch: {epoch + 1} training loss = {epoch_loss:.4f}, training accuracy = {epoch_acc * 100:.2f}%, val_loss = {epoch_val_loss:.4f} & val_accuracy = {val_acc:.2f}%")
curr_params = count_params(best_model)
print(f"Number of parameters = {curr_params}\n")
percentage_pruned = ((orig_params - curr_params.numpy()) / orig_params * 100).numpy()
# Code for manual Early Stopping:
# if np.abs(epoch_val_loss < best_val_loss) >= minimum_delta:
if (epoch_val_loss < best_val_loss) and np.abs(epoch_val_loss - best_val_loss) >= minimum_delta:
# print(f"epoch_val_loss = {epoch_val_loss:.4f}, best_val_loss = {best_val_loss:.4f}")
# update 'best_val_loss' variable to lowest loss encountered so far-
best_val_loss = epoch_val_loss
# reset 'loc_patience' variable-
loc_patience = 0
print(f"\nSaving model with lowest val_loss = {epoch_val_loss:.4f}")
# Save trained model with validation accuracy-
# torch.save(model.state_dict, f"LeNet-300-100_Trained_{val_acc}.pth")
torch.save(best_model.state_dict(), f"LeNet-300-100_{percentage_pruned:.2f}.pth")
else: # there is no improvement in monitored metric 'val_loss'
loc_patience += 1 # number of epochs without any improvement
#训练循环-
对于范围内的历元（num_历元）：
运行损耗=0.0
运行_校正=0.0
如果loc_耐心>=耐心：
打印（“\n'earlysting'调用！\n”）
打破
运行损耗，运行修正=列车模型（最佳模型，列车装载机）
历元损耗=运行损耗/长度（列车数据集）
epoch\u acc=运行\u纠正.double（）/len（列车数据集）
#epoch\u acc=100*运行\u校正/长度（列车组）
#打印（f“\nepoch:{epoch+1}训练损失={epoch\u损失：.4f}，训练精度={epoch\u acc*100:.2f}%\n”）
运行损耗值，正确，总计=测试模型（最佳模型，测试加载程序）
历元值损失=运行损失值/长度（测试数据集）
val_acc=100*（正确/总计）
#打印（f“\nepoch:{epoch+1}训练损失={epoch_损失：.4f}，训练精度={epoch_acc*100:.2f}%，val_损失={epoch_val_损失：.4f}和val_精度={val_acc:.2f}%\n”）
打印（f“\nepoch:{epoch+1}训练损失={epoch_损失：.4f}，训练精度={epoch_acc*100:.2f}%，val_损失={epoch_val_损失：.4f}和val_精度={val_acc:.2f}%”）
当前参数=计数参数（最佳模型）
打印（f“参数数={curr\u params}\n”）
已修剪百分比=（（原始参数-curr\u params.numpy（））/orig\u params*100）.numpy（）
#手动提前停止代码：
#如果np.abs（历元价值损失<最佳价值损失）>=最小价值增量：
如果（历元价值损失<最佳价值损失）和np.abs（历元价值损失-最佳价值损失）>=最小增量：
#打印（f“epoch_val_loss={epoch_val_loss:.4f}，best_val_loss={best_val_loss:.4f}”）
#将“best_val_loss”变量更新为迄今为止遇到的最低损失-
最佳价值损失=时代价值损失
#重置“loc_耐心”变量-
loc_耐心=0
打印（f“\n保存具有最低val_损失的模型={epoch_val_损失：.4f}”）
#保存经过训练的模型，确保验证的准确性-
#火炬保存（模型状态，f“LeNet-300-100”{val_acc}.pth”）
保存（最佳模型状态dict（），f“LeNet-300-100{percentage\u pruned:.2f}.pth”）
否则：#监控指标“val_损失”没有改善
loc_patience+=1#没有任何改进的历代次数
# Training loop-
for epoch in range(num_epochs):
running_loss = 0.0
running_corrects = 0.0
if loc_patience >= patience:
print("\n'EarlyStopping' called!\n")
break
running_loss, running_corrects = train_model(best_model, train_loader)
epoch_loss = running_loss / len(train_dataset)
epoch_acc = running_corrects.double() / len(train_dataset)
# epoch_acc = 100 * running_corrects / len(trainset)
# print(f"\nepoch: {epoch + 1} training loss = {epoch_loss:.4f}, training accuracy = {epoch_acc * 100:.2f}%\n")
running_loss_val, correct, total = test_model(best_model, test_loader)
epoch_val_loss = running_loss_val / len(test_dataset)
val_acc = 100 * (correct / total)
# print(f"\nepoch: {epoch + 1} training loss = {epoch_loss:.4f}, training accuracy = {epoch_acc * 100:.2f}%, val_loss = {epoch_val_loss:.4f} & val_accuracy = {val_acc:.2f}%\n")
print(f"\nepoch: {epoch + 1} training loss = {epoch_loss:.4f}, training accuracy = {epoch_acc * 100:.2f}%, val_loss = {epoch_val_loss:.4f} & val_accuracy = {val_acc:.2f}%")
curr_params = count_params(best_model)
print(f"Number of parameters = {curr_params}\n")
percentage_pruned = ((orig_params - curr_params.numpy()) / orig_params * 100).numpy()
# Code for manual Early Stopping:
# if np.abs(epoch_val_loss < best_val_loss) >= minimum_delta:
if (epoch_val_loss < best_val_loss) and np.abs(epoch_val_loss - best_val_loss) >= minimum_delta:
# print(f"epoch_val_loss = {epoch_val_loss:.4f}, best_val_loss = {best_val_loss:.4f}")
# update 'best_val_loss' variable to lowest loss encountered so far-
best_val_loss = epoch_val_loss
# reset 'loc_patience' variable-
loc_patience = 0
print(f"\nSaving model with lowest val_loss = {epoch_val_loss:.4f}")
# Save trained model with validation accuracy-
# torch.save(model.state_dict, f"LeNet-300-100_Trained_{val_acc}.pth")
torch.save(best_model.state_dict(), f"LeNet-300-100_{percentage_pruned:.2f}.pth")
else: # there is no improvement in monitored metric 'val_loss'
loc_patience += 1 # number of epochs without any improvement