无锡网站制作有哪些,中小型企业网站模板,电子政务网站建设法律法规,各学院二级网站建设通报YOLOV7剪枝流程
1、训练
1#xff09;划分数据集进行训练前的准备#xff0c;按正常的划分流程即可
2#xff09;修改train.py文件
第一次处在参数列表里添加剪枝的参数#xff0c;正常训练时设置为False#xff0c;剪枝后微调时设置为True
parser.add_argument(--pr…YOLOV7剪枝流程
1、训练
1划分数据集进行训练前的准备按正常的划分流程即可
2修改train.py文件
第一次处在参数列表里添加剪枝的参数正常训练时设置为False剪枝后微调时设置为True
parser.add_argument(--pruned, actionstore_true, defaultTrue, helppruned model fine train)第二处位置在
# Resume下面修改代码源代码为 # Resumestart_epoch, best_fitness 0, 0.0if pretrained:# Optimizerif ckpt[optimizer] is not None:optimizer.load_state_dict(ckpt[optimizer])best_fitness ckpt[best_fitness]# EMAif ema and ckpt.get(ema):ema.ema.load_state_dict(ckpt[ema].float().state_dict())ma.updates ckpt[updates]# Resultsif ckpt.get(training_results) is not None:results_file.write_text(ckpt[training_results]) # write results.txt修改为 # Resumestart_epoch, best_fitness 0, 0.0if pretrained:if not opt.pruned:# Optimizerif ckpt[optimizer] is not None:optimizer.load_state_dict(ckpt[optimizer])best_fitness ckpt[best_fitness]# EMAif ema and ckpt.get(ema):ema.ema.load_state_dict(ckpt[ema].float().state_dict())ema.updates ckpt[updates]# Resultsif ckpt.get(training_results) is not None:results_file.write_text(ckpt[training_results]) # write results.txt第三处位置在
# Epochs源代码为 # Epochsstart_epoch ckpt[epoch] 1if opt.resume:assert start_epoch 0, %s training to %g epochs is finished, nothing to resume. % (weights, epochs)if epochs start_epoch:logger.info(%s has been trained for %g epochs. Fine-tuning for %g additional epochs. %(weights, ckpt[epoch], epochs))epochs ckpt[epoch] # finetune additional epochsdel ckpt, state_dict修改为 # Epochsif not opt.pruned:start_epoch ckpt[epoch] 1if opt.resume:assert start_epoch 0, %s training to %g epochs is finished, nothing to resume. % (weights, epochs)elif opt.pruned:ckpt[epoch] 0start_epoch ckpt[epoch] 1if epochs start_epoch:logger.info(%s has been trained for %g epochs. Fine-tuning for %g additional epochs. %(weights, ckpt[epoch], epochs))epochs ckpt[epoch] # finetune additional epochsif not opt.pruned:del ckpt, state_dictelif opt.pruned:del ckpt第四处位置在
# Save model源代码为 # Save modelif (not opt.nosave) or (final_epoch and not opt.evolve): # if saveckpt {epoch: epoch,best_fitness: best_fitness,training_results: results_file.read_text(),model: deepcopy(model.module if is_parallel(model) else model).half(),ema: deepcopy(ema.ema).half(),updates: ema.updates,optimizer: optimizer.state_dict(),wandb_id: wandb_logger.wandb_run.id if wandb_logger.wandb else None}修改为 # Save modelif (not opt.nosave) or (final_epoch and not opt.evolve): # if saveif opt.pruned:ckpt {model: deepcopy(model.module if is_parallel(model) else model).half(),}elif not opt.pruned:ckpt {epoch: epoch,best_fitness: best_fitness,training_results: results_file.read_text(),model: deepcopy(model.module if is_parallel(model) else model).half(),ema: deepcopy(ema.ema).half(),updates: ema.updates,optimizer: optimizer.state_dict(),wandb_id: wandb_logger.wandb_run.id if wandb_logger.wandb else None}修改后的train.py整体代码如下
import argparse
import logging
import math
import os
import random
import time
from copy import deepcopy
from pathlib import Path
from threading import Threadimport numpy as np
import torch.distributed as dist
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torch.optim.lr_scheduler as lr_scheduler
import torch.utils.data
import yaml
from torch.cuda import amp
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdmimport test # import test.py to get mAP after each epoch
from models.experimental import attempt_load
from models.yolo import Model
from utils.autoanchor import check_anchors
from utils.datasets import create_dataloader
from utils.general import labels_to_class_weights, increment_path, labels_to_image_weights, init_seeds, \fitness, strip_optimizer, get_latest_run, check_dataset, check_file, check_git_status, check_img_size, \check_requirements, print_mutation, set_logging, one_cycle, colorstr
from utils.google_utils import attempt_download
from utils.loss import ComputeLoss, ComputeLossOTA
from utils.plots import plot_images, plot_labels, plot_results, plot_evolution
from utils.torch_utils import ModelEMA, select_device, intersect_dicts, torch_distributed_zero_first, is_parallel
from utils.wandb_logging.wandb_utils import WandbLogger, check_wandb_resumelogger logging.getLogger(__name__)def train(hyp, opt, device, tb_writerNone):logger.info(colorstr(hyperparameters: ) , .join(f{k}{v} for k, v in hyp.items()))save_dir, epochs, batch_size, total_batch_size, weights, rank, freeze \Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank, opt.freeze# Directorieswdir save_dir / weightswdir.mkdir(parentsTrue, exist_okTrue) # make dirlast wdir / last.ptbest wdir / best.ptresults_file save_dir / results.txt# Save run settingswith open(save_dir / hyp.yaml, w) as f:yaml.dump(hyp, f, sort_keysFalse)with open(save_dir / opt.yaml, w) as f:yaml.dump(vars(opt), f, sort_keysFalse)# Configureplots not opt.evolve # create plotscuda device.type ! cpuinit_seeds(2 rank)with open(opt.data) as f:data_dict yaml.load(f, Loaderyaml.SafeLoader) # data dictis_coco opt.data.endswith(coco.yaml)# Logging- Doing this before checking the dataset. Might update data_dictloggers {wandb: None} # loggers dictif rank in [-1, 0]:opt.hyp hyp # add hyperparametersrun_id torch.load(weights, map_locationdevice).get(wandb_id) if weights.endswith(.pt) and os.path.isfile(weights) else Nonewandb_logger WandbLogger(opt, Path(opt.save_dir).stem, run_id, data_dict)loggers[wandb] wandb_logger.wandbdata_dict wandb_logger.data_dictif wandb_logger.wandb:weights, epochs, hyp opt.weights, opt.epochs, opt.hyp # WandbLogger might update weights, epochs if resumingnc 1 if opt.single_cls else int(data_dict[nc]) # number of classesnames [item] if opt.single_cls and len(data_dict[names]) ! 1 else data_dict[names] # class namesassert len(names) nc, %g names found for nc%g dataset in %s % (len(names), nc, opt.data) # check# Modelpretrained weights.endswith(.pt)if pretrained:if opt.pruned:from models.yolo import attempt_load#model attempt_load(weights,map_locationdevice)ckpt torch.load(weights, map_locationdevice)model ckpt[model]else:with torch_distributed_zero_first(rank):attempt_download(weights) # download if not found locallyckpt torch.load(weights, map_locationdevice) # 加载模型# 模型的定义model Model(opt.cfg or ckpt[model].yaml, ch3, ncnc, anchorshyp.get(anchors)).to(device) # createexclude [anchor] if (opt.cfg or hyp.get(anchors))else [] # exclude keysstate_dict ckpt[model].float().state_dict() # to FP32 获得预权重的权值state_dict intersect_dicts(state_dict, model.state_dict(), excludeexclude) # intersect# 将权重加载到模型内model.load_state_dict(state_dict, strictFalse) # loadlogger.info(Transferred %g/%g items from %s % (len(state_dict), len(model.state_dict()), weights)) # report# with torch_distributed_zero_first(rank):# attempt_download(weights) # download if not found locally# ckpt torch.load(weights, map_locationdevice) # load checkpoint# model Model(opt.cfg or ckpt[model].yaml, ch3, ncnc, anchorshyp.get(anchors)).to(device) # create# exclude [anchor] if (opt.cfg or hyp.get(anchors)) and not opt.resume else [] # exclude keys# state_dict ckpt[model].float().state_dict() # to FP32# state_dict intersect_dicts(state_dict, model.state_dict(), excludeexclude) # intersect# model.load_state_dict(state_dict, strictFalse) # load# logger.info(Transferred %g/%g items from %s % (len(state_dict), len(model.state_dict()), weights)) # reportelse:model Model(opt.cfg, ch3, ncnc, anchorshyp.get(anchors)).to(device) # createwith torch_distributed_zero_first(rank):check_dataset(data_dict) # checktrain_path data_dict[train]test_path data_dict[val]# Freezefreeze [fmodel.{x}. for x in (freeze if len(freeze) 1 else range(freeze[0]))] # parameter names to freeze (full or partial)for k, v in model.named_parameters():v.requires_grad True # train all layersif any(x in k for x in freeze):print(freezing %s % k)v.requires_grad False# Optimizernbs 64 # nominal batch sizeaccumulate max(round(nbs / total_batch_size), 1) # accumulate loss before optimizinghyp[weight_decay] * total_batch_size * accumulate / nbs # scale weight_decaylogger.info(fScaled weight_decay {hyp[weight_decay]})pg0, pg1, pg2 [], [], [] # optimizer parameter groupsfor k, v in model.named_modules():if hasattr(v, bias) and isinstance(v.bias, nn.Parameter):pg2.append(v.bias) # biasesif isinstance(v, nn.BatchNorm2d):pg0.append(v.weight) # no decayelif hasattr(v, weight) and isinstance(v.weight, nn.Parameter):pg1.append(v.weight) # apply decayif hasattr(v, im):if hasattr(v.im, implicit): pg0.append(v.im.implicit)else:for iv in v.im:pg0.append(iv.implicit)if hasattr(v, imc):if hasattr(v.imc, implicit): pg0.append(v.imc.implicit)else:for iv in v.imc:pg0.append(iv.implicit)if hasattr(v, imb):if hasattr(v.imb, implicit): pg0.append(v.imb.implicit)else:for iv in v.imb:pg0.append(iv.implicit)if hasattr(v, imo):if hasattr(v.imo, implicit): pg0.append(v.imo.implicit)else:for iv in v.imo:pg0.append(iv.implicit)if hasattr(v, ia):if hasattr(v.ia, implicit): pg0.append(v.ia.implicit)else:for iv in v.ia:pg0.append(iv.implicit)if hasattr(v, attn):if hasattr(v.attn, logit_scale): pg0.append(v.attn.logit_scale)if hasattr(v.attn, q_bias): pg0.append(v.attn.q_bias)if hasattr(v.attn, v_bias): pg0.append(v.attn.v_bias)if hasattr(v.attn, relative_position_bias_table): pg0.append(v.attn.relative_position_bias_table)if hasattr(v, rbr_dense):if hasattr(v.rbr_dense, weight_rbr_origin): pg0.append(v.rbr_dense.weight_rbr_origin)if hasattr(v.rbr_dense, weight_rbr_avg_conv): pg0.append(v.rbr_dense.weight_rbr_avg_conv)if hasattr(v.rbr_dense, weight_rbr_pfir_conv): pg0.append(v.rbr_dense.weight_rbr_pfir_conv)if hasattr(v.rbr_dense, weight_rbr_1x1_kxk_idconv1): pg0.append(v.rbr_dense.weight_rbr_1x1_kxk_idconv1)if hasattr(v.rbr_dense, weight_rbr_1x1_kxk_conv2): pg0.append(v.rbr_dense.weight_rbr_1x1_kxk_conv2)if hasattr(v.rbr_dense, weight_rbr_gconv_dw): pg0.append(v.rbr_dense.weight_rbr_gconv_dw)if hasattr(v.rbr_dense, weight_rbr_gconv_pw): pg0.append(v.rbr_dense.weight_rbr_gconv_pw)if hasattr(v.rbr_dense, vector): pg0.append(v.rbr_dense.vector)if opt.adam:optimizer optim.Adam(pg0, lrhyp[lr0], betas(hyp[momentum], 0.999)) # adjust beta1 to momentumelse:optimizer optim.SGD(pg0, lrhyp[lr0], momentumhyp[momentum], nesterovTrue)optimizer.add_param_group({params: pg1, weight_decay: hyp[weight_decay]}) # add pg1 with weight_decayoptimizer.add_param_group({params: pg2}) # add pg2 (biases)logger.info(Optimizer groups: %g .bias, %g conv.weight, %g other % (len(pg2), len(pg1), len(pg0)))del pg0, pg1, pg2# Scheduler https://arxiv.org/pdf/1812.01187.pdf# https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLRif opt.linear_lr:lf lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp[lrf]) hyp[lrf] # linearelse:lf one_cycle(1, hyp[lrf], epochs) # cosine 1-hyp[lrf]scheduler lr_scheduler.LambdaLR(optimizer, lr_lambdalf)# plot_lr_scheduler(optimizer, scheduler, epochs)# EMAema ModelEMA(model) if rank in [-1, 0] else None# Resumestart_epoch, best_fitness 0, 0.0if pretrained:if not opt.pruned:# Optimizerif ckpt[optimizer] is not None:optimizer.load_state_dict(ckpt[optimizer])best_fitness ckpt[best_fitness]# EMAif ema and ckpt.get(ema):ema.ema.load_state_dict(ckpt[ema].float().state_dict())ema.updates ckpt[updates]# Resultsif ckpt.get(training_results) is not None:results_file.write_text(ckpt[training_results]) # write results.txt# Epochsif not opt.pruned:start_epoch ckpt[epoch] 1if opt.resume:assert start_epoch 0, %s training to %g epochs is finished, nothing to resume. % (weights, epochs)elif opt.pruned:ckpt[epoch] 0start_epoch ckpt[epoch] 1if epochs start_epoch:logger.info(%s has been trained for %g epochs. Fine-tuning for %g additional epochs. %(weights, ckpt[epoch], epochs))epochs ckpt[epoch] # finetune additional epochsif not opt.pruned:del ckpt, state_dictelif opt.pruned:del ckpt# start_epoch ckpt[epoch] 1# if opt.resume:# assert start_epoch 0, %s training to %g epochs is finished, nothing to resume. % (weights, epochs)# if epochs start_epoch:# logger.info(%s has been trained for %g epochs. Fine-tuning for %g additional epochs. %# (weights, ckpt[epoch], epochs))# epochs ckpt[epoch] # finetune additional epochs# del ckpt, state_dict# Image sizesgs max(int(model.stride.max()), 32) # grid size (max stride)nl model.model[-1].nl # number of detection layers (used for scaling hyp[obj])imgsz, imgsz_test [check_img_size(x, gs) for x in opt.img_size] # verify imgsz are gs-multiples# DP modeif cuda and rank -1 and torch.cuda.device_count() 1:model torch.nn.DataParallel(model)# SyncBatchNormif opt.sync_bn and cuda and rank ! -1:model torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)logger.info(Using SyncBatchNorm())# Trainloaderdataloader, dataset create_dataloader(train_path, imgsz, batch_size, gs, opt,hyphyp, augmentTrue, cacheopt.cache_images, rectopt.rect, rankrank,world_sizeopt.world_size, workersopt.workers,image_weightsopt.image_weights, quadopt.quad, prefixcolorstr(train: ))mlc np.concatenate(dataset.labels, 0)[:, 0].max() # max label classnb len(dataloader) # number of batchesassert mlc nc, Label class %g exceeds nc%g in %s. Possible class labels are 0-%g % (mlc, nc, opt.data, nc - 1)# Process 0if rank in [-1, 0]:testloader create_dataloader(test_path, imgsz_test, batch_size * 2, gs, opt, # testloaderhyphyp, cacheopt.cache_images and not opt.notest, rectTrue, rank-1,world_sizeopt.world_size, workersopt.workers,pad0.5, prefixcolorstr(val: ))[0]if not opt.resume:labels np.concatenate(dataset.labels, 0)c torch.tensor(labels[:, 0]) # classes# cf torch.bincount(c.long(), minlengthnc) 1. # frequency# model._initialize_biases(cf.to(device))if plots:#plot_labels(labels, names, save_dir, loggers)if tb_writer:tb_writer.add_histogram(classes, c, 0)# Anchorsif not opt.noautoanchor:check_anchors(dataset, modelmodel, thrhyp[anchor_t], imgszimgsz)model.half().float() # pre-reduce anchor precision# DDP modeif cuda and rank ! -1:model DDP(model, device_ids[opt.local_rank], output_deviceopt.local_rank,# nn.MultiheadAttention incompatibility with DDP https://github.com/pytorch/pytorch/issues/26698find_unused_parametersany(isinstance(layer, nn.MultiheadAttention) for layer in model.modules()))# Model parametershyp[box] * 3. / nl # scale to layershyp[cls] * nc / 80. * 3. / nl # scale to classes and layershyp[obj] * (imgsz / 640) ** 2 * 3. / nl # scale to image size and layershyp[label_smoothing] opt.label_smoothingmodel.nc nc # attach number of classes to modelmodel.hyp hyp # attach hyperparameters to modelmodel.gr 1.0 # iou loss ratio (obj_loss 1.0 or iou)model.class_weights labels_to_class_weights(dataset.labels, nc).to(device) * nc # attach class weightsmodel.names names# Start trainingt0 time.time()nw max(round(hyp[warmup_epochs] * nb), 1000) # number of warmup iterations, max(3 epochs, 1k iterations)# nw min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to 1/2 of trainingmaps np.zeros(nc) # mAP per classresults (0, 0, 0, 0, 0, 0, 0) # P, R, mAP.5, mAP.5-.95, val_loss(box, obj, cls)scheduler.last_epoch start_epoch - 1 # do not movescaler amp.GradScaler(enabledcuda)compute_loss_ota ComputeLossOTA(model) # init loss classcompute_loss ComputeLoss(model) # init loss classlogger.info(fImage sizes {imgsz} train, {imgsz_test} test\nfUsing {dataloader.num_workers} dataloader workers\nfLogging results to {save_dir}\nfStarting training for {epochs} epochs...)torch.save(model, wdir / init.pt)for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------model.train()# Update image weights (optional)if opt.image_weights:# Generate indicesif rank in [-1, 0]:cw model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc # class weightsiw labels_to_image_weights(dataset.labels, ncnc, class_weightscw) # image weightsdataset.indices random.choices(range(dataset.n), weightsiw, kdataset.n) # rand weighted idx# Broadcast if DDPif rank ! -1:indices (torch.tensor(dataset.indices) if rank 0 else torch.zeros(dataset.n)).int()dist.broadcast(indices, 0)if rank ! 0:dataset.indices indices.cpu().numpy()# Update mosaic border# b int(random.uniform(0.25 * imgsz, 0.75 * imgsz gs) // gs * gs)# dataset.mosaic_border [b - imgsz, -b] # height, width bordersmloss torch.zeros(4, devicedevice) # mean lossesif rank ! -1:dataloader.sampler.set_epoch(epoch)pbar enumerate(dataloader)logger.info((\n %10s * 8) % (Epoch, gpu_mem, box, obj, cls, total, labels, img_size))if rank in [-1, 0]:pbar tqdm(pbar, totalnb) # progress baroptimizer.zero_grad()for i, (imgs, targets, paths, _) in pbar: # batch -------------------------------------------------------------ni i nb * epoch # number integrated batches (since train start)imgs imgs.to(device, non_blockingTrue).float() / 255.0 # uint8 to float32, 0-255 to 0.0-1.0# Warmupif ni nw:xi [0, nw] # x interp# model.gr np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss 1.0 or iou)accumulate max(1, np.interp(ni, xi, [1, nbs / total_batch_size]).round())for j, x in enumerate(optimizer.param_groups):# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0x[lr] np.interp(ni, xi, [hyp[warmup_bias_lr] if j 2 else 0.0, x[initial_lr] * lf(epoch)])if momentum in x:x[momentum] np.interp(ni, xi, [hyp[warmup_momentum], hyp[momentum]])# Multi-scaleif opt.multi_scale:sz random.randrange(imgsz * 0.5, imgsz * 1.5 gs) // gs * gs # sizesf sz / max(imgs.shape[2:]) # scale factorif sf ! 1:ns [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]] # new shape (stretched to gs-multiple)imgs F.interpolate(imgs, sizens, modebilinear, align_cornersFalse)# Forwardwith amp.autocast(enabledcuda):pred model(imgs) # forwardif loss_ota not in hyp or hyp[loss_ota] 1:loss, loss_items compute_loss_ota(pred, targets.to(device), imgs) # loss scaled by batch_sizeelse:loss, loss_items compute_loss(pred, targets.to(device)) # loss scaled by batch_sizeif rank ! -1:loss * opt.world_size # gradient averaged between devices in DDP modeif opt.quad:loss * 4.# Backwardscaler.scale(loss).backward()# Optimizeif ni % accumulate 0:scaler.step(optimizer) # optimizer.stepscaler.update()optimizer.zero_grad()if ema:ema.update(model)# Printif rank in [-1, 0]:mloss (mloss * i loss_items) / (i 1) # update mean lossesmem %.3gG % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB)s (%10s * 2 %10.4g * 6) % (%g/%g % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1])pbar.set_description(s)# Plotif plots and ni 10:f save_dir / ftrain_batch{ni}.jpg # filenameThread(targetplot_images, args(imgs, targets, paths, f), daemonTrue).start()# if tb_writer:# tb_writer.add_image(f, result, dataformatsHWC, global_stepepoch)# tb_writer.add_graph(torch.jit.trace(model, imgs, strictFalse), []) # add model graphelif plots and ni 10 and wandb_logger.wandb:wandb_logger.log({Mosaics: [wandb_logger.wandb.Image(str(x), captionx.name) for x insave_dir.glob(train*.jpg) if x.exists()]})# end batch ------------------------------------------------------------------------------------------------# end epoch ----------------------------------------------------------------------------------------------------# Schedulerlr [x[lr] for x in optimizer.param_groups] # for tensorboardscheduler.step()# DDP process 0 or single-GPUif rank in [-1, 0]:# mAPema.update_attr(model, include[yaml, nc, hyp, gr, names, stride, class_weights])final_epoch epoch 1 epochsif not opt.notest or final_epoch: # Calculate mAPwandb_logger.current_epoch epoch 1results, maps, times test.test(data_dict,batch_sizebatch_size * 2,imgszimgsz_test,modelema.ema,single_clsopt.single_cls,dataloadertestloader,save_dirsave_dir,verbosenc 50 and final_epoch,plotsplots and final_epoch,wandb_loggerwandb_logger,compute_losscompute_loss,is_cocois_coco,v5_metricopt.v5_metric)# Writewith open(results_file, a) as f:f.write(s %10.4g * 7 % results \n) # append metrics, val_lossif len(opt.name) and opt.bucket:os.system(gsutil cp %s gs://%s/results/results%s.txt % (results_file, opt.bucket, opt.name))# Logtags [train/box_loss, train/obj_loss, train/cls_loss, # train lossmetrics/precision, metrics/recall, metrics/mAP_0.5, metrics/mAP_0.5:0.95,val/box_loss, val/obj_loss, val/cls_loss, # val lossx/lr0, x/lr1, x/lr2] # paramsfor x, tag in zip(list(mloss[:-1]) list(results) lr, tags):if tb_writer:tb_writer.add_scalar(tag, x, epoch) # tensorboardif wandb_logger.wandb:wandb_logger.log({tag: x}) # WB# Update best mAPfi fitness(np.array(results).reshape(1, -1)) # weighted combination of [P, R, mAP.5, mAP.5-.95]if fi best_fitness:best_fitness fiwandb_logger.end_epoch(best_resultbest_fitness fi)# Save modelif (not opt.nosave) or (final_epoch and not opt.evolve): # if saveif opt.pruned:ckpt {model: deepcopy(model.module if is_parallel(model) else model).half(),}elif not opt.pruned:ckpt {epoch: epoch,best_fitness: best_fitness,training_results: results_file.read_text(),model: deepcopy(model.module if is_parallel(model) else model).half(),ema: deepcopy(ema.ema).half(),updates: ema.updates,optimizer: optimizer.state_dict(),wandb_id: wandb_logger.wandb_run.id if wandb_logger.wandb else None}# Save last, best and deletetorch.save(ckpt, last)if best_fitness fi:torch.save(ckpt, best)if (best_fitness fi) and (epoch 200):torch.save(ckpt, wdir / best_{:03d}.pt.format(epoch))if epoch 0:torch.save(ckpt, wdir / epoch_{:03d}.pt.format(epoch))elif ((epoch1) % 25) 0:torch.save(ckpt, wdir / epoch_{:03d}.pt.format(epoch))elif epoch (epochs-5):torch.save(ckpt, wdir / epoch_{:03d}.pt.format(epoch))if wandb_logger.wandb:if ((epoch 1) % opt.save_period 0 and not final_epoch) and opt.save_period ! -1:wandb_logger.log_model(last.parent, opt, epoch, fi, best_modelbest_fitness fi)del ckpt# end epoch ----------------------------------------------------------------------------------------------------# end trainingif rank in [-1, 0]:# Plotsif plots:plot_results(save_dirsave_dir) # save as results.pngif wandb_logger.wandb:files [results.png, confusion_matrix.png, *[f{x}_curve.png for x in (F1, PR, P, R)]]wandb_logger.log({Results: [wandb_logger.wandb.Image(str(save_dir / f), captionf) for f in filesif (save_dir / f).exists()]})# Test best.ptlogger.info(%g epochs completed in %.3f hours.\n % (epoch - start_epoch 1, (time.time() - t0) / 3600))if opt.data.endswith(coco.yaml) and nc 80: # if COCOfor m in (last, best) if best.exists() else (last): # speed, mAP testsresults, _, _ test.test(opt.data,batch_sizebatch_size * 2,imgszimgsz_test,conf_thres0.001,iou_thres0.7,modelattempt_load(m, device).half(),single_clsopt.single_cls,dataloadertestloader,save_dirsave_dir,save_jsonTrue,plotsFalse,is_cocois_coco,v5_metricopt.v5_metric)# Strip optimizersfinal best if best.exists() else last # final modelfor f in last, best:if f.exists():strip_optimizer(f) # strip optimizersif opt.bucket:os.system(fgsutil cp {final} gs://{opt.bucket}/weights) # uploadif wandb_logger.wandb and not opt.evolve: # Log the stripped modelwandb_logger.wandb.log_artifact(str(final), typemodel,namerun_ wandb_logger.wandb_run.id _model,aliases[last, best, stripped])wandb_logger.finish_run()else:dist.destroy_process_group()torch.cuda.empty_cache()return resultsif __name__ __main__:parser argparse.ArgumentParser()parser.add_argument(--weights, typestr, defaultmy_dataset/layer_pruning.pt, helpinitial weights path)parser.add_argument(--cfg, typestr, defaultcfg/training/yolov7-tiny.yaml, helpmodel.yaml path)parser.add_argument(--data, typestr, defaultmy_dataset/coco.yaml, helpdata.yaml path)parser.add_argument(--hyp, typestr, defaultdata/hyp.scratch.p5.yaml, helphyperparameters path)parser.add_argument(--epochs, typeint, default600)parser.add_argument(--batch-size, typeint, default32, helptotal batch size for all GPUs)parser.add_argument(--img-size, nargs, typeint, default[640, 640], help[train, test] image sizes)parser.add_argument(--rect, actionstore_true, helprectangular training)parser.add_argument(--resume, nargs?, constTrue, defaultFalse, helpresume most recent training)parser.add_argument(--nosave, actionstore_true, helponly save final checkpoint)parser.add_argument(--notest, actionstore_true, helponly test final epoch)parser.add_argument(--noautoanchor, actionstore_true, helpdisable autoanchor check)parser.add_argument(--evolve, actionstore_true, helpevolve hyperparameters)parser.add_argument(--bucket, typestr, default, helpgsutil bucket)parser.add_argument(--cache-images, actionstore_true, helpcache images for faster training)parser.add_argument(--image-weights, actionstore_true, helpuse weighted image selection for training)parser.add_argument(--device, default0, helpcuda device, i.e. 0 or 0,1,2,3 or cpu)parser.add_argument(--multi-scale, actionstore_true, defaultTrue, helpvary img-size /- 50%%)parser.add_argument(--single-cls, actionstore_true, helptrain multi-class data as single-class)parser.add_argument(--adam, actionstore_true, helpuse torch.optim.Adam() optimizer)parser.add_argument(--sync-bn, actionstore_true, helpuse SyncBatchNorm, only available in DDP mode)parser.add_argument(--local_rank, typeint, default-1, helpDDP parameter, do not modify)parser.add_argument(--workers, typeint, default16, helpmaximum number of dataloader workers)parser.add_argument(--project, defaultruns/train, helpsave to project/name)parser.add_argument(--entity, defaultNone, helpWB entity)parser.add_argument(--name, defaultexp, helpsave to project/name)parser.add_argument(--exist-ok, actionstore_true, helpexisting project/name ok, do not increment)parser.add_argument(--quad, actionstore_true, helpquad dataloader)parser.add_argument(--linear-lr, actionstore_true, helplinear LR)parser.add_argument(--label-smoothing, typefloat, default0.0, helpLabel smoothing epsilon)parser.add_argument(--upload_dataset, actionstore_true, helpUpload dataset as WB artifact table)parser.add_argument(--bbox_interval, typeint, default-1, helpSet bounding-box image logging interval for WB)parser.add_argument(--save_period, typeint, default2, helpLog model after every save_period epoch)parser.add_argument(--artifact_alias, typestr, defaultlatest, helpversion of dataset artifact to be used)parser.add_argument(--freeze, nargs, typeint, default[0], helpFreeze layers: backbone of yolov750, first30 1 2)parser.add_argument(--v5-metric, actionstore_true, helpassume maximum recall as 1.0 in AP calculation)parser.add_argument(--pruned, actionstore_true, defaultTrue, helppruned model fine train)opt parser.parse_args()# Set DDP variablesopt.world_size int(os.environ[WORLD_SIZE]) if WORLD_SIZE in os.environ else 1opt.global_rank int(os.environ[RANK]) if RANK in os.environ else -1set_logging(opt.global_rank)#if opt.global_rank in [-1, 0]:# check_git_status()# check_requirements()# Resumewandb_run check_wandb_resume(opt)if opt.resume and not wandb_run: # resume an interrupted runckpt opt.resume if isinstance(opt.resume, str) else get_latest_run() # specified or most recent pathassert os.path.isfile(ckpt), ERROR: --resume checkpoint does not existapriori opt.global_rank, opt.local_rankwith open(Path(ckpt).parent.parent / opt.yaml) as f:opt argparse.Namespace(**yaml.load(f, Loaderyaml.SafeLoader)) # replaceopt.save_dir Path(ckpt).parent.parent # increment runopt.cfg, opt.weights, opt.resume, opt.batch_size, opt.global_rank, opt.local_rank , ckpt, True, opt.total_batch_size, *apriori # reinstatelogger.info(Resuming training from %s % ckpt)else:# opt.hyp opt.hyp or (hyp.finetune.yaml if opt.weights else hyp.scratch.yaml)opt.data, opt.cfg, opt.hyp check_file(opt.data), check_file(opt.cfg), check_file(opt.hyp) # check filesassert len(opt.cfg) or len(opt.weights), either --cfg or --weights must be specifiedopt.img_size.extend([opt.img_size[-1]] * (2 - len(opt.img_size))) # extend to 2 sizes (train, test)opt.name evolve if opt.evolve else opt.nameopt.save_dir increment_path(Path(opt.project) / opt.name, exist_okopt.exist_ok | opt.evolve) # increment run# DDP modeopt.total_batch_size opt.batch_sizedevice select_device(opt.device, batch_sizeopt.batch_size)if opt.local_rank ! -1:assert torch.cuda.device_count() opt.local_ranktorch.cuda.set_device(opt.local_rank)device torch.device(cuda, opt.local_rank)dist.init_process_group(backendnccl, init_methodenv://) # distributed backendassert opt.batch_size % opt.world_size 0, --batch-size must be multiple of CUDA device countopt.batch_size opt.total_batch_size // opt.world_size# Hyperparameterswith open(opt.hyp) as f:hyp yaml.load(f, Loaderyaml.SafeLoader) # load hyps# Trainlogger.info(opt)if not opt.evolve:tb_writer None # init loggersif opt.global_rank in [-1, 0]:prefix colorstr(tensorboard: )logger.info(f{prefix}Start with tensorboard --logdir {opt.project}, view at http://localhost:6006/)tb_writer SummaryWriter(opt.save_dir) # Tensorboardtrain(hyp, opt, device, tb_writer)# Evolve hyperparameters (optional)else:# Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)meta {lr0: (1, 1e-5, 1e-1), # initial learning rate (SGD1E-2, Adam1E-3)lrf: (1, 0.01, 1.0), # final OneCycleLR learning rate (lr0 * lrf)momentum: (0.3, 0.6, 0.98), # SGD momentum/Adam beta1weight_decay: (1, 0.0, 0.001), # optimizer weight decaywarmup_epochs: (1, 0.0, 5.0), # warmup epochs (fractions ok)warmup_momentum: (1, 0.0, 0.95), # warmup initial momentumwarmup_bias_lr: (1, 0.0, 0.2), # warmup initial bias lrbox: (1, 0.02, 0.2), # box loss gaincls: (1, 0.2, 4.0), # cls loss gaincls_pw: (1, 0.5, 2.0), # cls BCELoss positive_weightobj: (1, 0.2, 4.0), # obj loss gain (scale with pixels)obj_pw: (1, 0.5, 2.0), # obj BCELoss positive_weightiou_t: (0, 0.1, 0.7), # IoU training thresholdanchor_t: (1, 2.0, 8.0), # anchor-multiple thresholdanchors: (2, 2.0, 10.0), # anchors per output grid (0 to ignore)fl_gamma: (0, 0.0, 2.0), # focal loss gamma (efficientDet default gamma1.5)hsv_h: (1, 0.0, 0.1), # image HSV-Hue augmentation (fraction)hsv_s: (1, 0.0, 0.9), # image HSV-Saturation augmentation (fraction)hsv_v: (1, 0.0, 0.9), # image HSV-Value augmentation (fraction)degrees: (1, 0.0, 45.0), # image rotation (/- deg)translate: (1, 0.0, 0.9), # image translation (/- fraction)scale: (1, 0.0, 0.9), # image scale (/- gain)shear: (1, 0.0, 10.0), # image shear (/- deg)perspective: (0, 0.0, 0.001), # image perspective (/- fraction), range 0-0.001flipud: (1, 0.0, 1.0), # image flip up-down (probability)fliplr: (0, 0.0, 1.0), # image flip left-right (probability)mosaic: (1, 0.0, 1.0), # image mixup (probability)mixup: (1, 0.0, 1.0), # image mixup (probability)copy_paste: (1, 0.0, 1.0), # segment copy-paste (probability)paste_in: (1, 0.0, 1.0)} # segment copy-paste (probability)with open(opt.hyp, errorsignore) as f:hyp yaml.safe_load(f) # load hyps dictif anchors not in hyp: # anchors commented in hyp.yamlhyp[anchors] 3assert opt.local_rank -1, DDP mode not implemented for --evolveopt.notest, opt.nosave True, True # only test/save final epoch# ei [isinstance(x, (int, float)) for x in hyp.values()] # evolvable indicesyaml_file Path(opt.save_dir) / hyp_evolved.yaml # save best result hereif opt.bucket:os.system(gsutil cp gs://%s/evolve.txt . % opt.bucket) # download evolve.txt if existsfor _ in range(300): # generations to evolveif Path(evolve.txt).exists(): # if evolve.txt exists: select best hyps and mutate# Select parent(s)parent single # parent selection method: single or weightedx np.loadtxt(evolve.txt, ndmin2)n min(5, len(x)) # number of previous results to considerx x[np.argsort(-fitness(x))][:n] # top n mutationsw fitness(x) - fitness(x).min() # weightsif parent single or len(x) 1:# x x[random.randint(0, n - 1)] # random selectionx x[random.choices(range(n), weightsw)[0]] # weighted selectionelif parent weighted:x (x * w.reshape(n, 1)).sum(0) / w.sum() # weighted combination# Mutatemp, s 0.8, 0.2 # mutation probability, sigmanpr np.randomnpr.seed(int(time.time()))g np.array([x[0] for x in meta.values()]) # gains 0-1ng len(meta)v np.ones(ng)while all(v 1): # mutate until a change occurs (prevent duplicates)v (g * (npr.random(ng) mp) * npr.randn(ng) * npr.random() * s 1).clip(0.3, 3.0)for i, k in enumerate(hyp.keys()): # plt.hist(v.ravel(), 300)hyp[k] float(x[i 7] * v[i]) # mutate# Constrain to limitsfor k, v in meta.items():hyp[k] max(hyp[k], v[1]) # lower limithyp[k] min(hyp[k], v[2]) # upper limithyp[k] round(hyp[k], 5) # significant digits# Train mutationresults train(hyp.copy(), opt, device)# Write mutation resultsprint_mutation(hyp.copy(), results, yaml_file, opt.bucket)# Plot resultsplot_evolution(yaml_file)print(fHyperparameter evolution complete. Best results saved as: {yaml_file}\nfCommand to train a new model with these hyperparameters: $ python train.py --hyp {yaml_file})
2、剪枝
剪枝方法
YOLOv4剪枝【附代码】_strategy tp.strategy.l1strategy()-CSDN博客参考此篇博文进行的通道剪枝。Pruning Filters for Efficient ConvNets这篇论文的技术
添加prunmodel.py文件
下载torch_pruning模块时一定要使用0.2.5版本
pip install torch_pruning0.2.7
pip install loguru把训练好的模型路径放进去
layer_pruning(/home/jovyan/exp_3046/runs/train/best.pt) 定义剪枝后保存的模型路径
torch.save(model_, /home/jovyan/exp_3047/data/layer_pruning.pt)修改prunmodel.py文件中需要剪枝的权重路径。重点修改58~62行。这里是以修改model的前10层为例。head层不能剪我们选择backbone的层。 included_layers []for layer in model.model[:10]: # 获取backboneif type(layer) is Conv:included_layers.append(layer.conv)included_layers.append(layer.bn)下面代码是剪枝conv和BN层。【重点是tp.prune_conv】自己修改amout也就是剪枝率。 if isinstance(m, nn.Conv2d) and m in included_layers:# amount是剪枝率# 卷积剪枝pruning_plan DG.get_pruning_plan(m, tp.prune_conv, idxsstrategy(m.weight, amount0.8))logger.info(pruning_plan)# 执行剪枝pruning_plan.exec()if isinstance(m, nn.BatchNorm2d) and m in included_layers:# BN层剪枝pruning_plan DG.get_pruning_plan(m, tp.prune_batchnorm, idxsstrategy(m.weight, amount0.8))logger.info(pruning_plan)pruning_plan.exec()出现以下内容说明剪枝成功
2023-03-15 14:57:40.825 | INFO | __main__:layer_pruning:84 - Params: 37196556 36839795
2023-03-15 14:57:41.176 | INFO | __main__:layer_pruning:95 - 剪枝完成代码如下
import syssys.path.append(/home/jovyan/exp_3046)
# print(sys.path)import torch_pruning as tp
from loguru import logger
from models.common import *
from models.experimental import Ensemble
from utils.torch_utils import select_device
剪枝的时候根据模型结构去剪不要盲目的猜
剪枝完需要进行一个微调训练
# 加载模型
def attempt_load(weights, map_locationNone, inplaceTrue):from models.yolo import Detect, Modelmodel Ensemble()ckpt torch.load(weights, map_locationmap_location) # load weightsmodel.append(ckpt[ema if ckpt.get(ema) else model].float().eval()) # without layer fuse 权值的加载# Compatibility updatesfor m in model.modules(): # 取出每一层if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU, Detect, Model]:m.inplace inplace # pytorch 1.7.0 compatibilityif type(m) is Detect: # 判断是否为目标检测if not isinstance(m.anchor_grid, list): # new Detect Layer compatibilitydelattr(m, anchor_grid)setattr(m, anchor_grid, [torch.zeros(1)] * m.nl)elif type(m) is Conv: # 卷积层m._non_persistent_buffers_set set() # pytorch 1.6.0 compatibilityif len(model) 1:return model[-1], ckpt # return modelelse:print(fEnsemble created with {weights}\n)for k in [names]:setattr(model, k, getattr(model[-1], k))model.stride model[torch.argmax(torch.tensor([m.stride.max() for m in model])).int()].stride # max stridereturn model, ckpt # return ensemblelogger.catch
def layer_pruning(weights):logger.add(../logs/layer_pruning.log, rotation1 MB)device select_device(cpu)model, ckpt attempt_load(weights, map_locationdevice)for para in model.parameters():para.requires_grad True# 创建输入样例可在此修改输入大小x torch.zeros(1, 3, 640, 640)# -----------------对整个模型的剪枝--------------------strategy tp.strategy.L1Strategy() # L1策略DG tp.DependencyGraph() # 依赖图DG DG.build_dependency(model, example_inputsx)这里写要剪枝的层这里以backbone为例included_layers []for layer in model.model[:41]: # 获取backboneif type(layer) is Conv:included_layers.append(layer.conv)included_layers.append(layer.bn)logger.info(included_layers)# 获取未剪枝之前的参数量num_params_before_pruning tp.utils.count_params(model)# 模型遍历for m in model.modules():# 判断是否为卷积并且是否在需要剪枝的层里if isinstance(m, nn.Conv2d) and m in included_layers:# amount是剪枝率# 卷积剪枝# 层剪枝(需要筛选出不需要剪枝的层比如yolo需要把头部的预测部分取出来这个是不需要剪枝的) pruning_plan DG.get_pruning_plan(m, tp.prune_conv, idxsstrategy(m.weight, amount0.4))logger.info(pruning_plan)# 执行剪枝pruning_plan.exec()if isinstance(m, nn.BatchNorm2d) and m in included_layers:# BN层剪枝pruning_plan DG.get_pruning_plan(m, tp.prune_batchnorm, idxsstrategy(m.weight, amount0.3))logger.info(pruning_plan)pruning_plan.exec()# 获得剪枝以后的参数量num_params_after_pruning tp.utils.count_params(model)# 输出一下剪枝前后的参数量logger.info( Params: %s %s\n % (num_params_before_pruning, num_params_after_pruning))# 剪枝完以后模型的保存(不要用torch.save(model.state_dict(),...))model_ {model: model.half(),# optimizer: ckpt[optimizer],# training_results: ckpt[training_results],epoch: ckpt[epoch]}torch.save(model_, /home/jovyan/exp_3046/my_dataset/layer_pruning.pt)del model_, ckptlogger.info(剪枝完成\n)layer_pruning(/home/jovyan/exp_3046/runs/train/best.pt) 3、微调
运行train.py
权重为剪枝之后的layer_pruning.pt
parser.add_argument(--pruned, actionstore_true, defaultTrue, helppruned model fine train)4、效果
使用初始训练好的权重进行预测用时 249.217s
使用剪枝之后的模型进行预测用时172.418s p.prune_batchnorm, idxsstrategy(m.weight, amount0.3))logger.info(pruning_plan)pruning_plan.exec()# 获得剪枝以后的参数量num_params_after_pruning tp.utils.count_params(model)# 输出一下剪枝前后的参数量logger.info( Params: %s %s\n % (num_params_before_pruning, num_params_after_pruning))# 剪枝完以后模型的保存(不要用torch.save(model.state_dict(),...))model_ {model: model.half(),# optimizer: ckpt[optimizer],# training_results: ckpt[training_results],epoch: ckpt[epoch]}torch.save(model_, /home/jovyan/exp_3046/my_dataset/layer_pruning.pt)del model_, ckptlogger.info(剪枝完成\n)layer_pruning(/home/jovyan/exp_3046/runs/train/best.pt) 3、微调
运行train.py
权重为剪枝之后的layer_pruning.pt
parser.add_argument(--pruned, actionstore_true, defaultTrue, helppruned model fine train)4、效果
使用初始训练好的权重进行预测用时 249.217s 使用剪枝之后的模型进行预测用时172.418s
