From 67c20ae317bf8704ac13657bbfab9bee0bf7cf6a Mon Sep 17 00:00:00 2001 From: Benjamin Date: Sat, 13 Feb 2021 19:09:00 -0500 Subject: [PATCH 1/6] base ultralytics integration --- examples/ultralytics-sparseml/README.md | 100 ++++ examples/ultralytics-sparseml/main.py | 722 ++++++++++++++++++++++++ 2 files changed, 822 insertions(+) create mode 100644 examples/ultralytics-sparseml/README.md create mode 100644 examples/ultralytics-sparseml/main.py diff --git a/examples/ultralytics-sparseml/README.md b/examples/ultralytics-sparseml/README.md new file mode 100644 index 00000000000..67693654fec --- /dev/null +++ b/examples/ultralytics-sparseml/README.md @@ -0,0 +1,100 @@ + + +# SparseML-ultralytics/yolov5 integration +This directory provides a SparseML integrated training script for the popular +[ultralytics/yolov5](https://github.com/ultralytics/yolov5) +repository also known as [timm](https://pypi.org/project/timm/). + +Using this integration, you will be able to apply SparseML optimizations +to the powerful training flows provided in the yolov5 repository. + +Some of the tasks you can perform using this integration include, but are not limited to: +* model pruning +* quantization-aware-training +* sparse quantization-aware-training +* sparse transfer learning + +## Installation +To use both the script, clone both repositories, install their dependencies, +and copy the integrated training script into the yolov5 directory to run from. + +```bash +# clone +git clone https://github.com/ultralytics/yolov5.git +git clone https://github.com/neuralmagic/sparseml.git + +# copy script +cp sparseml/examples/ultralytics-sparseml/main.py yolov5 +cd yolov5 + +# install dependencies +pip install -r requirements.txt +pip install sparseml +``` + + +## Script +`examples/timm-sparseml/main.py` modifies +[`train.py`](https://github.com/ultralytics/yolov5/blob/master/train.py) +from yolov5 to include a `sparseml-recipe-path` argument +to run SparseML optimizations with. This can be a file path to a local +SparseML recipe or a SparseZoo model stub prefixed by `zoo:` such as +`zoo:cv/detection/yolo_v3-spp/pytorch/ultralytics/coco/pruned-aggressive`. + +Additionally, for sparse transfer learning, the flag `--sparse-transfer-learn` +was added. Running the script with this flag will add modifiers to the given +recipe that will keep the base sparsity constant during training, allowing +the model to learn the new dataset while keeping the same optimized structure. +If a SparseZoo recipe path is provided with sparse transfer learning enabled, +then the the model's specific "transfer" recipe will be loaded instead. + +To load the base weights for a SparseZoo recipe as the initial checkpoint, set +`--initial-checkpoint` to `zoo`. To use the weights of a SparseZoo model as the +initial checkpoint, pass that model's SparseZoo stub prefixed by `zoo:` to the +`--initial-checkpoint` argument. + +Running the script will +follow the normal yolov5 training flow with the given SparseML optimizations enabled. + +Some considerations: + +* `--sparseml-recipe-path` is a required parameter +* `--epochs` will now be overridden by the epochs set in the SparseML recipe +* if using learning rate schedulers both with the yolov5 script and your recipe, they +may conflict with each other causing unintended side effects, choose +hyperparameters accordingly. +* Modifiers will log their outputs to the console as well as to the tensorboard file +* After training is complete, the final model will be exported to ONNX using SparseML + +You can learn how to build or download a recipe using the +[SparseML](https://github.com/neuralmagic/sparseml) +or [SparseZoo](https://github.com/neuralmagic/sparsezoo) +documentation, or export one with [Sparsify](https://github.com/neuralmagic/sparsify). + +Documentation on the original script can be found +[here](https://github.com/ultralytics/yolov5). +The latest commit hash that `main.py` is based on is included in the docstring. + + +#### Example Command +Call the script from the `yolov5` directory, passing in the same arguments as +`train.py`, with the additional SparseML argument(s) included. +```bash +python main.py \ + --sparseml-recipe-path /PATH/TO/RECIPE/recipe.yaml \ + +``` diff --git a/examples/ultralytics-sparseml/main.py b/examples/ultralytics-sparseml/main.py new file mode 100644 index 00000000000..50cb3e7ca59 --- /dev/null +++ b/examples/ultralytics-sparseml/main.py @@ -0,0 +1,722 @@ +# neuralmagic: no copyright +# flake8: noqa +# fmt: off +# isort: skip_file + +""" +Integration between https://github.com/ultralytics/yolov5 and SparseML + +This script is adapted from https://github.com/ultralytics/yolov5/blob/master/train.py +to apply a SparseML recipe from the required `--sparseml-recipe-path` argument. +Integration lines are preceded by comment blocks. Run with `--help` for help printout, +more information can be found in the readme file. + +Latest yolov5 commit this script is based on: c9bda11 +""" +import argparse +import logging +import math +import os +import random +import time +from pathlib import Path +from threading import Thread + +import 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 tqdm + +import 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 +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 + +from sparseml.pytorch.optim import ScheduledModifierManager, ScheduledOptimizer +from sparseml.pytorch.utils import ModuleExporter, PythonLogger, TensorBoardLogger +from sparsezoo import Zoo + +logger = logging.getLogger(__name__) + + +def train(hyp, opt, device, tb_writer=None, wandb=None): + logger.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items())) + save_dir, epochs, batch_size, total_batch_size, weights, rank = \ + Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank + + # Directories + wdir = save_dir / 'weights' + wdir.mkdir(parents=True, exist_ok=True) # make dir + last = wdir / 'last.pt' + best = wdir / 'best.pt' + results_file = save_dir / 'results.txt' + + # Save run settings + with open(save_dir / 'hyp.yaml', 'w') as f: + yaml.dump(hyp, f, sort_keys=False) + with open(save_dir / 'opt.yaml', 'w') as f: + yaml.dump(vars(opt), f, sort_keys=False) + + # Configure + plots = not opt.evolve # create plots + cuda = device.type != 'cpu' + init_seeds(2 + rank) + with open(opt.data) as f: + data_dict = yaml.load(f, Loader=yaml.SafeLoader) # data dict + with torch_distributed_zero_first(rank): + check_dataset(data_dict) # check + train_path = data_dict['train'] + test_path = data_dict['val'] + nc = 1 if opt.single_cls else int(data_dict['nc']) # number of classes + names = ['item'] if opt.single_cls and len(data_dict['names']) != 1 else data_dict['names'] # class names + assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (len(names), nc, opt.data) # check + + # Model + pretrained = weights.endswith('.pt') + if pretrained: + with torch_distributed_zero_first(rank): + attempt_download(weights) # download if not found locally + ckpt = torch.load(weights, map_location=device) # load checkpoint + if hyp.get('anchors'): + ckpt['model'].yaml['anchors'] = round(hyp['anchors']) # force autoanchor + model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device) # create + exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [] # exclude keys + state_dict = ckpt['model'].float().state_dict() # to FP32 + state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude) # intersect + model.load_state_dict(state_dict, strict=False) # load + logger.info('Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights)) # report + else: + model = Model(opt.cfg, ch=3, nc=nc).to(device) # create + + # Freeze + freeze = [] # parameter names to freeze (full or partial) + for k, v in model.named_parameters(): + v.requires_grad = True # train all layers + if any(x in k for x in freeze): + print('freezing %s' % k) + v.requires_grad = False + + # Optimizer + nbs = 64 # nominal batch size + accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing + hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay + logger.info(f"Scaled weight_decay = {hyp['weight_decay']}") + + pg0, pg1, pg2 = [], [], [] # optimizer parameter groups + for k, v in model.named_modules(): + if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): + pg2.append(v.bias) # biases + if isinstance(v, nn.BatchNorm2d): + pg0.append(v.weight) # no decay + elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): + pg1.append(v.weight) # apply decay + + if opt.adam: + optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum + else: + optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) + + optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay + optimizer.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#OneCycleLR + if opt.linear_lr: + lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp['lrf']) + hyp['lrf'] # linear + else: + lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf'] + scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) + # plot_lr_scheduler(optimizer, scheduler, epochs) + + # Logging + if rank in [-1, 0] and wandb and wandb.run is None: + opt.hyp = hyp # add hyperparameters + wandb_run = wandb.init(config=opt, resume="allow", + project='YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem, + name=save_dir.stem, + id=ckpt.get('wandb_id') if 'ckpt' in locals() else None) + loggers = {'wandb': wandb} # loggers dict + + # Resume + start_epoch, best_fitness = 0, 0.0 + if pretrained: + # Optimizer + if ckpt['optimizer'] is not None: + optimizer.load_state_dict(ckpt['optimizer']) + best_fitness = ckpt['best_fitness'] + + # Results + if ckpt.get('training_results') is not None: + with open(results_file, 'w') as file: + file.write(ckpt['training_results']) # write results.txt + + # Epochs + 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 sizes + gs = int(model.stride.max()) # 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 mode + if cuda and rank == -1 and torch.cuda.device_count() > 1: + model = torch.nn.DataParallel(model) + + # SyncBatchNorm + if opt.sync_bn and cuda and rank != -1: + model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) + logger.info('Using SyncBatchNorm()') + + # EMA + ema = ModelEMA(model) if rank in [-1, 0] else None + + # DDP mode + if cuda and rank != -1: + model = DDP(model, device_ids=[opt.local_rank], output_device=opt.local_rank) + + # Trainloader + dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt, + hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, rank=rank, + world_size=opt.world_size, workers=opt.workers, + image_weights=opt.image_weights, quad=opt.quad, prefix=colorstr('train: ')) + mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class + nb = len(dataloader) # number of batches + assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (mlc, nc, opt.data, nc - 1) + + # Process 0 + if rank in [-1, 0]: + ema.updates = start_epoch * nb // accumulate # set EMA updates + testloader = create_dataloader(test_path, imgsz_test, batch_size * 2, gs, opt, # testloader + hyp=hyp, cache=opt.cache_images and not opt.notest, rect=True, rank=-1, + world_size=opt.world_size, workers=opt.workers, + pad=0.5, prefix=colorstr('val: '))[0] + + if not opt.resume: + labels = np.concatenate(dataset.labels, 0) + c = torch.tensor(labels[:, 0]) # classes + # cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency + # model._initialize_biases(cf.to(device)) + if plots: + plot_labels(labels, save_dir, loggers) + if tb_writer: + tb_writer.add_histogram('classes', c, 0) + + # Anchors + if not opt.noautoanchor: + check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) + + # Model parameters + hyp['box'] *= 3. / nl # scale to layers + hyp['cls'] *= nc / 80. * 3. / nl # scale to classes and layers + hyp['obj'] *= (imgsz / 640) ** 2 * 3. / nl # scale to image size and layers + model.nc = nc # attach number of classes to model + model.hyp = hyp # attach hyperparameters to model + model.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 weights + model.names = names + + #################################################################################### + # Start SparseML Integration + #################################################################################### + # determine recipe type to be used if loading from SparseZoo + if opt.sparseml_recipe_path.startswith("zoo:"): + zoo_recipe_type = "transfer" if opt.sparse_transfer_learn else "original" + else: + zoo_recipe_type = None + manager = ScheduledModifierManager.from_yaml( + opt.sparseml_recipe_path, zoo_recipe_type=zoo_recipe_type + ) + optimizer = ScheduledOptimizer( + optimizer, + model, + manager, + steps_per_epoch=len(dataloader), + loggers=[PythonLogger(), TensorBoardLogger(writer=tb_writer)] + ) + start_epoch = manager.min_epochs or start_epoch # override min_epochs + epochs = manager.max_epochs or epochs # override num_epochs + #################################################################################### + # End SparseML Integration + #################################################################################### + + # Start training + t0 = 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 training + maps = np.zeros(nc) # mAP per class + results = (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 move + scaler = amp.GradScaler(enabled=cuda) + compute_loss = ComputeLoss(model) # init loss class + logger.info(f'Image sizes {imgsz} train, {imgsz_test} test\n' + f'Using {dataloader.num_workers} dataloader workers\n' + f'Logging results to {save_dir}\n' + f'Starting training for {epochs} epochs...') + for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------ + model.train() + + # Update image weights (optional) + if opt.image_weights: + # Generate indices + if rank in [-1, 0]: + cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc # class weights + iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw) # image weights + dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n) # rand weighted idx + # Broadcast if DDP + if 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 borders + + mloss = torch.zeros(4, device=device) # mean losses + if rank != -1: + dataloader.sampler.set_epoch(epoch) + pbar = enumerate(dataloader) + logger.info(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'total', 'targets', 'img_size')) + if rank in [-1, 0]: + pbar = tqdm(pbar, total=nb) # progress bar + optimizer.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_blocking=True).float() / 255.0 # uint8 to float32, 0-255 to 0.0-1.0 + + # Warmup + if 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 lr0 + x['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-scale + if opt.multi_scale: + sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size + sf = sz / max(imgs.shape[2:]) # scale factor + if 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, size=ns, mode='bilinear', align_corners=False) + + # Forward + with amp.autocast(enabled=cuda): + pred = model(imgs) # forward + loss, loss_items = compute_loss(pred, targets.to(device)) # loss scaled by batch_size + if rank != -1: + loss *= opt.world_size # gradient averaged between devices in DDP mode + if opt.quad: + loss *= 4. + + # Backward + scaler.scale(loss).backward() + + # Optimize + if ni % accumulate == 0: + scaler.step(optimizer) # optimizer.step + scaler.update() + optimizer.zero_grad() + if ema: + ema.update(model) + + # Print + if rank in [-1, 0]: + mloss = (mloss * i + loss_items) / (i + 1) # update mean losses + mem = '%.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) + + # Plot + if plots and ni < 3: + f = save_dir / f'train_batch{ni}.jpg' # filename + Thread(target=plot_images, args=(imgs, targets, paths, f), daemon=True).start() + # if tb_writer: + # tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch) + # tb_writer.add_graph(model, imgs) # add model to tensorboard + elif plots and ni == 10 and wandb: + wandb.log({"Mosaics": [wandb.Image(str(x), caption=x.name) for x in save_dir.glob('train*.jpg') + if x.exists()]}, commit=False) + + # end batch ------------------------------------------------------------------------------------------------ + # end epoch ---------------------------------------------------------------------------------------------------- + + # Scheduler + lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard + scheduler.step() + + # DDP process 0 or single-GPU + if rank in [-1, 0]: + # mAP + if ema: + ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride', 'class_weights']) + final_epoch = epoch + 1 == epochs + if not opt.notest or final_epoch: # Calculate mAP + results, maps, times = test.test(opt.data, + batch_size=batch_size * 2, + imgsz=imgsz_test, + model=ema.ema, + single_cls=opt.single_cls, + dataloader=testloader, + save_dir=save_dir, + verbose=nc < 50 and final_epoch, + plots=plots and final_epoch, + log_imgs=opt.log_imgs if wandb else 0, + compute_loss=compute_loss) + + # Write + with open(results_file, 'a') as f: + f.write(s + '%10.4g' * 7 % results + '\n') # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls) + if len(opt.name) and opt.bucket: + os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name)) + + # Log + tags = ['train/box_loss', 'train/obj_loss', 'train/cls_loss', # train loss + 'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', + 'val/box_loss', 'val/obj_loss', 'val/cls_loss', # val loss + 'x/lr0', 'x/lr1', 'x/lr2'] # params + for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags): + if tb_writer: + tb_writer.add_scalar(tag, x, epoch) # tensorboard + if wandb: + wandb.log({tag: x}, step=epoch, commit=tag == tags[-1]) # W&B + + # Update best mAP + fi = fitness(np.array(results).reshape(1, -1)) # weighted combination of [P, R, mAP@.5, mAP@.5-.95] + if fi > best_fitness: + best_fitness = fi + + # Save model + save = (not opt.nosave) or (final_epoch and not opt.evolve) + if save: + with open(results_file, 'r') as f: # create checkpoint + ckpt = {'epoch': epoch, + 'best_fitness': best_fitness, + 'training_results': f.read(), + 'model': ema.ema, + 'optimizer': None if final_epoch else optimizer.state_dict(), + 'wandb_id': wandb_run.id if wandb else None} + + # Save last, best and delete + torch.save(ckpt, last) + if best_fitness == fi: + torch.save(ckpt, best) + del ckpt + # end epoch ---------------------------------------------------------------------------------------------------- + # end training + + if rank in [-1, 0]: + # Strip optimizers + final = best if best.exists() else last # final model + for f in [last, best]: + if f.exists(): + strip_optimizer(f) # strip optimizers + if opt.bucket: + os.system(f'gsutil cp {final} gs://{opt.bucket}/weights') # upload + + # Plots + if plots: + plot_results(save_dir=save_dir) # save as results.png + if wandb: + files = ['results.png', 'confusion_matrix.png', *[f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R')]] + wandb.log({"Results": [wandb.Image(str(save_dir / f), caption=f) for f in files + if (save_dir / f).exists()]}) + if opt.log_artifacts: + wandb.log_artifact(artifact_or_path=str(final), type='model', name=save_dir.stem) + + # Test best.pt + logger.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 COCO + for conf, iou, save_json in ([0.25, 0.45, False], [0.001, 0.65, True]): # speed, mAP tests + results, _, _ = test.test(opt.data, + batch_size=batch_size * 2, + imgsz=imgsz_test, + conf_thres=conf, + iou_thres=iou, + model=attempt_load(final, device).half(), + single_cls=opt.single_cls, + dataloader=testloader, + save_dir=save_dir, + save_json=save_json, + plots=False) + ################################################################################# + # Start SparseML ONNX Export + ################################################################################# + logger.info( + f"training complete, exporting ONNX to {save_dir}/model.onnx" + ) + exporter = ModuleExporter(model, save_dir) + exporter.export_onnx(torch.randn((1, 3, *imgsz))) + ################################################################################# + # End SparseML ONNX Export + ################################################################################# + + else: + dist.destroy_process_group() + + wandb.run.finish() if wandb and wandb.run else None + torch.cuda.empty_cache() + return results + + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + #################################################################################### + # Start SparseML arguments + #################################################################################### + parser.add_argument( + "--sparseml-recipe-path", + required=True, + type=str, + help="path to a SparseML recipe file or a SparseZoo model stub for a recipe to load. " + "SparseZoo stubs should be preceded by 'zoo:'. i.e. '/path/to/local/recipe.yaml', " + "'zoo:zoo/model/stub'" + ) + parser.add_argument( + "--sparse-transfer-learn", + action="store_true", + help="Enable sparse transfer learning modifiers to enforce the sparsity " + "if the recipe comes from a local file, modifiers will be added to the manager " + "to hold already sparse layers at the same sparsity level. If the recipe comes " + "from SparseZoo, the 'transfer' recipe for the model will be loaded instead", + ) + parser.add_argument('--weights', type=str, default='yolov5s.pt', help='initial weights path') + parser.add_argument( + "--weights", + type=str, + default="yolov5s.pt", + help="initial weights path. can be a local file path, can pass in 'zoo' if " + "using a SparseZoo recipe to load that recipes base weights, or pass in a " + "SparseZoo model stub, prefixed with 'zoo:' to load weights directly from " + "SparseZoo", + ) + #################################################################################### + # End SparseML arguments + #################################################################################### + parser.add_argument('--cfg', type=str, default='', help='model.yaml path') + parser.add_argument('--data', type=str, default='data/coco128.yaml', help='data.yaml path') + parser.add_argument('--hyp', type=str, default='data/hyp.scratch.yaml', help='hyperparameters path') + parser.add_argument('--epochs', type=int, default=300) + parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs') + parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='[train, test] image sizes') + parser.add_argument('--rect', action='store_true', help='rectangular training') + parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training') + parser.add_argument('--nosave', action='store_true', help='only save final checkpoint') + parser.add_argument('--notest', action='store_true', help='only test final epoch') + parser.add_argument('--noautoanchor', action='store_true', help='disable autoanchor check') + parser.add_argument('--evolve', action='store_true', help='evolve hyperparameters') + parser.add_argument('--bucket', type=str, default='', help='gsutil bucket') + parser.add_argument('--cache-images', action='store_true', help='cache images for faster training') + parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%') + parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class') + parser.add_argument('--adam', action='store_true', help='use torch.optim.Adam() optimizer') + parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode') + parser.add_argument('--local_rank', type=int, default=-1, help='DDP parameter, do not modify') + parser.add_argument('--log-imgs', type=int, default=16, help='number of images for W&B logging, max 100') + parser.add_argument('--log-artifacts', action='store_true', help='log artifacts, i.e. final trained model') + parser.add_argument('--workers', type=int, default=8, help='maximum number of dataloader workers') + parser.add_argument('--project', default='runs/train', help='save to project/name') + parser.add_argument('--name', default='exp', help='save to project/name') + parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') + parser.add_argument('--quad', action='store_true', help='quad dataloader') + parser.add_argument('--linear-lr', action='store_true', help='linear LR') + opt = parser.parse_args() + + # Set DDP variables + opt.world_size = int(os.environ['WORLD_SIZE']) if 'WORLD_SIZE' in os.environ else 1 + opt.global_rank = int(os.environ['RANK']) if 'RANK' in os.environ else -1 + set_logging(opt.global_rank) + if opt.global_rank in [-1, 0]: + check_git_status() + check_requirements() + + #################################################################################### + # Start - SparseML optional load weights from SparseZoo + #################################################################################### + if opt.weights == "zoo": + # Load checkpoint from base weights associated with given SparseZoo recipe + if opt.sparseml_recipe_path.startswith("zoo:"): + recipe_type = "transfer" if opt.sparse_transfer_learn else "original" + opt.weights = Zoo.download_recipe_base_framework_files( + opt.sparseml_recipe_path, + recipe_type=recipe_type, + extensions=[".pt", ".pth"] + )[0] + else: + raise ValueError( + "Attempting to load weights from SparseZoo recipe, but not given a " + "SparseZoo recipe stub. When --weights is set to 'zoo'. " + "sparseml-recipe-path must start with 'zoo:' and be a SparseZoo model " + f"stub. sparseml-recipe-path was set to {args.sparseml_recipe_path}" + ) + elif opt.weights.startswith("zoo:"): + # Load weights from a SparseZoo model stub + zoo_model = Zoo.load_model_from_stub(opt.weights) + args.initial_checkpoint = zoo_model.download_framework_files( + extensions=[".pt", ".pth"] + ) + #################################################################################### + # End - SparseML optional load weights from SparseZoo + #################################################################################### + + # Resume + if opt.resume: # resume an interrupted run + ckpt = opt.resume if isinstance(opt.resume, str) else get_latest_run() # specified or most recent path + assert os.path.isfile(ckpt), 'ERROR: --resume checkpoint does not exist' + apriori = opt.global_rank, opt.local_rank + with open(Path(ckpt).parent.parent / 'opt.yaml') as f: + opt = argparse.Namespace(**yaml.load(f, Loader=yaml.SafeLoader)) # replace + opt.cfg, opt.weights, opt.resume, opt.batch_size, opt.global_rank, opt.local_rank = '', ckpt, True, opt.total_batch_size, *apriori # reinstate + logger.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 files + assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified' + opt.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.name + opt.save_dir = increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok | opt.evolve) # increment run + + # DDP mode + opt.total_batch_size = opt.batch_size + device = select_device(opt.device, batch_size=opt.batch_size) + if opt.local_rank != -1: + assert torch.cuda.device_count() > opt.local_rank + torch.cuda.set_device(opt.local_rank) + device = torch.device('cuda', opt.local_rank) + dist.init_process_group(backend='nccl', init_method='env://') # distributed backend + assert opt.batch_size % opt.world_size == 0, '--batch-size must be multiple of CUDA device count' + opt.batch_size = opt.total_batch_size // opt.world_size + + # Hyperparameters + with open(opt.hyp) as f: + hyp = yaml.load(f, Loader=yaml.SafeLoader) # load hyps + + # Train + logger.info(opt) + try: + import wandb + except ImportError: + wandb = None + prefix = colorstr('wandb: ') + logger.info(f"{prefix}Install Weights & Biases for YOLOv5 logging with 'pip install wandb' (recommended)") + if not opt.evolve: + tb_writer = None # init loggers + if opt.global_rank in [-1, 0]: + logger.info(f'Start Tensorboard with "tensorboard --logdir {opt.project}", view at http://localhost:6006/') + tb_writer = SummaryWriter(opt.save_dir) # Tensorboard + train(hyp, opt, device, tb_writer, wandb) + + # 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 (SGD=1E-2, Adam=1E-3) + 'lrf': (1, 0.01, 1.0), # final OneCycleLR learning rate (lr0 * lrf) + 'momentum': (0.3, 0.6, 0.98), # SGD momentum/Adam beta1 + 'weight_decay': (1, 0.0, 0.001), # optimizer weight decay + 'warmup_epochs': (1, 0.0, 5.0), # warmup epochs (fractions ok) + 'warmup_momentum': (1, 0.0, 0.95), # warmup initial momentum + 'warmup_bias_lr': (1, 0.0, 0.2), # warmup initial bias lr + 'box': (1, 0.02, 0.2), # box loss gain + 'cls': (1, 0.2, 4.0), # cls loss gain + 'cls_pw': (1, 0.5, 2.0), # cls BCELoss positive_weight + 'obj': (1, 0.2, 4.0), # obj loss gain (scale with pixels) + 'obj_pw': (1, 0.5, 2.0), # obj BCELoss positive_weight + 'iou_t': (0, 0.1, 0.7), # IoU training threshold + 'anchor_t': (1, 2.0, 8.0), # anchor-multiple threshold + 'anchors': (2, 2.0, 10.0), # anchors per output grid (0 to ignore) + 'fl_gamma': (0, 0.0, 2.0), # focal loss gamma (efficientDet default gamma=1.5) + '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 0-0.001 + 'flipud': (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) + + assert opt.local_rank == -1, 'DDP mode not implemented for --evolve' + opt.notest, opt.nosave = True, True # only test/save final epoch + # ei = [isinstance(x, (int, float)) for x in hyp.values()] # evolvable indices + yaml_file = Path(opt.save_dir) / 'hyp_evolved.yaml' # save best result here + if opt.bucket: + os.system('gsutil cp gs://%s/evolve.txt .' % opt.bucket) # download evolve.txt if exists + + for _ in range(300): # generations to evolve + if Path('evolve.txt').exists(): # if evolve.txt exists: select best hyps and mutate + # Select parent(s) + parent = 'single' # parent selection method: 'single' or 'weighted' + x = np.loadtxt('evolve.txt', ndmin=2) + n = min(5, len(x)) # number of previous results to consider + x = x[np.argsort(-fitness(x))][:n] # top n mutations + w = fitness(x) - fitness(x).min() # weights + if parent == 'single' or len(x) == 1: + # x = x[random.randint(0, n - 1)] # random selection + x = x[random.choices(range(n), weights=w)[0]] # weighted selection + elif parent == 'weighted': + x = (x * w.reshape(n, 1)).sum(0) / w.sum() # weighted combination + + # Mutate + mp, s = 0.8, 0.2 # mutation probability, sigma + npr = np.random + npr.seed(int(time.time())) + g = np.array([x[0] for x in meta.values()]) # gains 0-1 + ng = 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 limits + for k, v in meta.items(): + hyp[k] = max(hyp[k], v[1]) # lower limit + hyp[k] = min(hyp[k], v[2]) # upper limit + hyp[k] = round(hyp[k], 5) # significant digits + + # Train mutation + results = train(hyp.copy(), opt, device, wandb=wandb) + + # Write mutation results + print_mutation(hyp.copy(), results, yaml_file, opt.bucket) + + # Plot results + plot_evolution(yaml_file) + print(f'Hyperparameter evolution complete. Best results saved as: {yaml_file}\n' + f'Command to train a new model with these hyperparameters: $ python train.py --hyp {yaml_file}') From da8821bf567e03aad05de810727ec03be5814e0c Mon Sep 17 00:00:00 2001 From: Benjamin Date: Mon, 22 Feb 2021 22:41:11 -0500 Subject: [PATCH 2/6] addressing review comments --- .../ultralytics}/README.md | 19 +++---- .../ultralytics}/main.py | 52 ++++++++----------- 2 files changed, 29 insertions(+), 42 deletions(-) rename {examples/ultralytics-sparseml => integrations/ultralytics}/README.md (80%) rename {examples/ultralytics-sparseml => integrations/ultralytics}/main.py (95%) diff --git a/examples/ultralytics-sparseml/README.md b/integrations/ultralytics/README.md similarity index 80% rename from examples/ultralytics-sparseml/README.md rename to integrations/ultralytics/README.md index 67693654fec..a100056147b 100644 --- a/examples/ultralytics-sparseml/README.md +++ b/integrations/ultralytics/README.md @@ -17,7 +17,7 @@ limitations under the License. # SparseML-ultralytics/yolov5 integration This directory provides a SparseML integrated training script for the popular [ultralytics/yolov5](https://github.com/ultralytics/yolov5) -repository also known as [timm](https://pypi.org/project/timm/). +repository. Using this integration, you will be able to apply SparseML optimizations to the powerful training flows provided in the yolov5 repository. @@ -38,7 +38,7 @@ git clone https://github.com/ultralytics/yolov5.git git clone https://github.com/neuralmagic/sparseml.git # copy script -cp sparseml/examples/ultralytics-sparseml/main.py yolov5 +cp sparseml/integrations/ultralytics/main.py yolov5 cd yolov5 # install dependencies @@ -48,20 +48,13 @@ pip install sparseml ## Script -`examples/timm-sparseml/main.py` modifies +`integrations/ultralytics/main.py` modifies [`train.py`](https://github.com/ultralytics/yolov5/blob/master/train.py) -from yolov5 to include a `sparseml-recipe-path` argument +from yolov5 to include a `sparseml-recipe` argument to run SparseML optimizations with. This can be a file path to a local SparseML recipe or a SparseZoo model stub prefixed by `zoo:` such as `zoo:cv/detection/yolo_v3-spp/pytorch/ultralytics/coco/pruned-aggressive`. -Additionally, for sparse transfer learning, the flag `--sparse-transfer-learn` -was added. Running the script with this flag will add modifiers to the given -recipe that will keep the base sparsity constant during training, allowing -the model to learn the new dataset while keeping the same optimized structure. -If a SparseZoo recipe path is provided with sparse transfer learning enabled, -then the the model's specific "transfer" recipe will be loaded instead. - To load the base weights for a SparseZoo recipe as the initial checkpoint, set `--initial-checkpoint` to `zoo`. To use the weights of a SparseZoo model as the initial checkpoint, pass that model's SparseZoo stub prefixed by `zoo:` to the @@ -72,7 +65,7 @@ follow the normal yolov5 training flow with the given SparseML optimizations ena Some considerations: -* `--sparseml-recipe-path` is a required parameter +* `--sparseml-recipe` is a required parameter * `--epochs` will now be overridden by the epochs set in the SparseML recipe * if using learning rate schedulers both with the yolov5 script and your recipe, they may conflict with each other causing unintended side effects, choose @@ -95,6 +88,6 @@ Call the script from the `yolov5` directory, passing in the same arguments as `train.py`, with the additional SparseML argument(s) included. ```bash python main.py \ - --sparseml-recipe-path /PATH/TO/RECIPE/recipe.yaml \ + --sparseml-recipe /PATH/TO/RECIPE/recipe.yaml \ ``` diff --git a/examples/ultralytics-sparseml/main.py b/integrations/ultralytics/main.py similarity index 95% rename from examples/ultralytics-sparseml/main.py rename to integrations/ultralytics/main.py index 50cb3e7ca59..31edbd0d2e1 100644 --- a/examples/ultralytics-sparseml/main.py +++ b/integrations/ultralytics/main.py @@ -7,11 +7,13 @@ Integration between https://github.com/ultralytics/yolov5 and SparseML This script is adapted from https://github.com/ultralytics/yolov5/blob/master/train.py -to apply a SparseML recipe from the required `--sparseml-recipe-path` argument. +to apply a SparseML recipe from the required `--sparseml-recipe` argument. Integration lines are preceded by comment blocks. Run with `--help` for help printout, more information can be found in the readme file. -Latest yolov5 commit this script is based on: c9bda11 +Latest yolov5 commit this script is based on: +https://github.com/ultralytics/yolov5/tree/c9bda112aebaa0be846864f9d224191d0e19d419 +commit hash: c9bda11 """ import argparse import logging @@ -244,14 +246,7 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): #################################################################################### # Start SparseML Integration #################################################################################### - # determine recipe type to be used if loading from SparseZoo - if opt.sparseml_recipe_path.startswith("zoo:"): - zoo_recipe_type = "transfer" if opt.sparse_transfer_learn else "original" - else: - zoo_recipe_type = None - manager = ScheduledModifierManager.from_yaml( - opt.sparseml_recipe_path, zoo_recipe_type=zoo_recipe_type - ) + manager = ScheduledModifierManager.from_yaml(opt.sparseml_recipe) optimizer = ScheduledOptimizer( optimizer, model, @@ -259,8 +254,15 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): steps_per_epoch=len(dataloader), loggers=[PythonLogger(), TensorBoardLogger(writer=tb_writer)] ) - start_epoch = manager.min_epochs or start_epoch # override min_epochs - epochs = manager.max_epochs or epochs # override num_epochs + # override lr scheduler if recipe makes any LR updates + if any("LearningRate" in str(modifier) for modifier in manager.modifiers): + logger.info("Disabling yolo LR scheduler, managing LR using SparseML recipe") + scheduler = None + if manager.max_epochs: + epochs = manager.max_epochs or epochs # override num_epochs + logger.info( + f"overriding number of epochs from SparseML manager to {manager.max_epochs}" + ) #################################################################################### # End SparseML Integration #################################################################################### @@ -271,7 +273,8 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training maps = np.zeros(nc) # mAP per class results = (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 move + if scheduler: + scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=cuda) compute_loss = ComputeLoss(model) # init loss class logger.info(f'Image sizes {imgsz} train, {imgsz_test} test\n' @@ -374,7 +377,8 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): # Scheduler lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard - scheduler.step() + if scheduler: + scheduler.step() # DDP process 0 or single-GPU if rank in [-1, 0]: @@ -496,21 +500,13 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): # Start SparseML arguments #################################################################################### parser.add_argument( - "--sparseml-recipe-path", + "--sparseml-recipe", required=True, type=str, help="path to a SparseML recipe file or a SparseZoo model stub for a recipe to load. " "SparseZoo stubs should be preceded by 'zoo:'. i.e. '/path/to/local/recipe.yaml', " "'zoo:zoo/model/stub'" ) - parser.add_argument( - "--sparse-transfer-learn", - action="store_true", - help="Enable sparse transfer learning modifiers to enforce the sparsity " - "if the recipe comes from a local file, modifiers will be added to the manager " - "to hold already sparse layers at the same sparsity level. If the recipe comes " - "from SparseZoo, the 'transfer' recipe for the model will be loaded instead", - ) parser.add_argument('--weights', type=str, default='yolov5s.pt', help='initial weights path') parser.add_argument( "--weights", @@ -568,19 +564,17 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): #################################################################################### if opt.weights == "zoo": # Load checkpoint from base weights associated with given SparseZoo recipe - if opt.sparseml_recipe_path.startswith("zoo:"): - recipe_type = "transfer" if opt.sparse_transfer_learn else "original" + if opt.sparseml_recipe.startswith("zoo:"): opt.weights = Zoo.download_recipe_base_framework_files( - opt.sparseml_recipe_path, - recipe_type=recipe_type, + opt.sparseml_recipe, extensions=[".pt", ".pth"] )[0] else: raise ValueError( "Attempting to load weights from SparseZoo recipe, but not given a " "SparseZoo recipe stub. When --weights is set to 'zoo'. " - "sparseml-recipe-path must start with 'zoo:' and be a SparseZoo model " - f"stub. sparseml-recipe-path was set to {args.sparseml_recipe_path}" + "sparseml-recipe must start with 'zoo:' and be a SparseZoo model " + f"stub. sparseml-recipe was set to {args.sparseml_recipe}" ) elif opt.weights.startswith("zoo:"): # Load weights from a SparseZoo model stub From ae7e602088c7ede981d66ce89fd8ca9e86f6598e Mon Sep 17 00:00:00 2001 From: Benjamin Date: Tue, 23 Feb 2021 09:34:30 -0500 Subject: [PATCH 3/6] renaming script to train.py --- integrations/ultralytics/README.md | 8 ++++---- integrations/ultralytics/{main.py => train.py} | 0 2 files changed, 4 insertions(+), 4 deletions(-) rename integrations/ultralytics/{main.py => train.py} (100%) diff --git a/integrations/ultralytics/README.md b/integrations/ultralytics/README.md index a100056147b..26ba6aebe1c 100644 --- a/integrations/ultralytics/README.md +++ b/integrations/ultralytics/README.md @@ -38,7 +38,7 @@ git clone https://github.com/ultralytics/yolov5.git git clone https://github.com/neuralmagic/sparseml.git # copy script -cp sparseml/integrations/ultralytics/main.py yolov5 +cp sparseml/integrations/ultralytics/train.py yolov5 cd yolov5 # install dependencies @@ -48,7 +48,7 @@ pip install sparseml ## Script -`integrations/ultralytics/main.py` modifies +`integrations/ultralytics/train.py` modifies [`train.py`](https://github.com/ultralytics/yolov5/blob/master/train.py) from yolov5 to include a `sparseml-recipe` argument to run SparseML optimizations with. This can be a file path to a local @@ -80,14 +80,14 @@ documentation, or export one with [Sparsify](https://github.com/neuralmagic/spar Documentation on the original script can be found [here](https://github.com/ultralytics/yolov5). -The latest commit hash that `main.py` is based on is included in the docstring. +The latest commit hash that `train.py` is based on is included in the docstring. #### Example Command Call the script from the `yolov5` directory, passing in the same arguments as `train.py`, with the additional SparseML argument(s) included. ```bash -python main.py \ +python train.py \ --sparseml-recipe /PATH/TO/RECIPE/recipe.yaml \ ``` diff --git a/integrations/ultralytics/main.py b/integrations/ultralytics/train.py similarity index 100% rename from integrations/ultralytics/main.py rename to integrations/ultralytics/train.py From 8ef1c013ef41129f5a1be2fae502e94dafa6daf6 Mon Sep 17 00:00:00 2001 From: Benjamin Fineran Date: Tue, 23 Feb 2021 15:44:33 -0500 Subject: [PATCH 4/6] delete repeated arg --- integrations/ultralytics/train.py | 1 - 1 file changed, 1 deletion(-) diff --git a/integrations/ultralytics/train.py b/integrations/ultralytics/train.py index 31edbd0d2e1..da62a23482e 100644 --- a/integrations/ultralytics/train.py +++ b/integrations/ultralytics/train.py @@ -507,7 +507,6 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): "SparseZoo stubs should be preceded by 'zoo:'. i.e. '/path/to/local/recipe.yaml', " "'zoo:zoo/model/stub'" ) - parser.add_argument('--weights', type=str, default='yolov5s.pt', help='initial weights path') parser.add_argument( "--weights", type=str, From 1f5b42701df516168a2562e6d62805411d6c921f Mon Sep 17 00:00:00 2001 From: Benjamin Date: Tue, 23 Feb 2021 17:41:51 -0500 Subject: [PATCH 5/6] disabling EMA by default, responding to review --- integrations/ultralytics/README.md | 16 +++++++++------- integrations/ultralytics/train.py | 21 +++++++++++++++++---- 2 files changed, 26 insertions(+), 11 deletions(-) diff --git a/integrations/ultralytics/README.md b/integrations/ultralytics/README.md index 26ba6aebe1c..11b420d2e4a 100644 --- a/integrations/ultralytics/README.md +++ b/integrations/ultralytics/README.md @@ -24,12 +24,12 @@ to the powerful training flows provided in the yolov5 repository. Some of the tasks you can perform using this integration include, but are not limited to: * model pruning -* quantization-aware-training -* sparse quantization-aware-training +* quantization-aware training +* sparse quantization-aware training * sparse transfer learning ## Installation -To use both the script, clone both repositories, install their dependencies, +To use the script, clone both repositories, install their dependencies, and copy the integrated training script into the yolov5 directory to run from. ```bash @@ -68,10 +68,12 @@ Some considerations: * `--sparseml-recipe` is a required parameter * `--epochs` will now be overridden by the epochs set in the SparseML recipe * if using learning rate schedulers both with the yolov5 script and your recipe, they -may conflict with each other causing unintended side effects, choose -hyperparameters accordingly. -* Modifiers will log their outputs to the console as well as to the tensorboard file +may conflict with each other causing unintended side effects, so choose +hyperparameters accordingly +* Modifiers will log their outputs to the console as well as to the TensorBoard file * After training is complete, the final model will be exported to ONNX using SparseML +* By default, EMA is disabled when using `train.py`. This is to allow for best compatibility +with pruning and quantization. To enable, set the `--use-ema` flag You can learn how to build or download a recipe using the [SparseML](https://github.com/neuralmagic/sparseml) @@ -89,5 +91,5 @@ Call the script from the `yolov5` directory, passing in the same arguments as ```bash python train.py \ --sparseml-recipe /PATH/TO/RECIPE/recipe.yaml \ - + ``` diff --git a/integrations/ultralytics/train.py b/integrations/ultralytics/train.py index da62a23482e..ae282a0c0d0 100644 --- a/integrations/ultralytics/train.py +++ b/integrations/ultralytics/train.py @@ -196,7 +196,13 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): logger.info('Using SyncBatchNorm()') # EMA - ema = ModelEMA(model) if rank in [-1, 0] else None + #################################################################################### + # Start SparseML Integration - optional EMA + #################################################################################### + ema = ModelEMA(model) if rank in [-1, 0] and opt.use_ema else None + #################################################################################### + # End SparseML Integration - optional EMA + #################################################################################### # DDP mode if cuda and rank != -1: @@ -213,7 +219,8 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): # Process 0 if rank in [-1, 0]: - ema.updates = start_epoch * nb // accumulate # set EMA updates + if ema: + ema.updates = start_epoch * nb // accumulate # set EMA updates testloader = create_dataloader(test_path, imgsz_test, batch_size * 2, gs, opt, # testloader hyp=hyp, cache=opt.cache_images and not opt.notest, rect=True, rank=-1, world_size=opt.world_size, workers=opt.workers, @@ -385,12 +392,13 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): # mAP if ema: ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride', 'class_weights']) + final_epoch = epoch + 1 == epochs if not opt.notest or final_epoch: # Calculate mAP results, maps, times = test.test(opt.data, batch_size=batch_size * 2, imgsz=imgsz_test, - model=ema.ema, + model=ema.ema if ema else model, single_cls=opt.single_cls, dataloader=testloader, save_dir=save_dir, @@ -428,7 +436,7 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): ckpt = {'epoch': epoch, 'best_fitness': best_fitness, 'training_results': f.read(), - 'model': ema.ema, + 'model': ema.ema if ema else model, 'optimizer': None if final_epoch else optimizer.state_dict(), 'wandb_id': wandb_run.id if wandb else None} @@ -516,6 +524,11 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): "SparseZoo model stub, prefixed with 'zoo:' to load weights directly from " "SparseZoo", ) + parser.add_argument( + "--use-ema", + action="store_true", + help="set flag to enable EMA updates. disabled by default in SparseML integration" + ) #################################################################################### # End SparseML arguments #################################################################################### From d8fd6d1efcc4348bc2c1877c1a0f29c72a34f440 Mon Sep 17 00:00:00 2001 From: Benjamin Date: Wed, 24 Feb 2021 12:34:59 -0500 Subject: [PATCH 6/6] disabling AMP by default for QAT compatibility --- integrations/ultralytics/README.md | 8 ++++++-- integrations/ultralytics/train.py | 10 ++++++++-- 2 files changed, 14 insertions(+), 4 deletions(-) diff --git a/integrations/ultralytics/README.md b/integrations/ultralytics/README.md index 11b420d2e4a..13bcddc9cc0 100644 --- a/integrations/ultralytics/README.md +++ b/integrations/ultralytics/README.md @@ -72,8 +72,12 @@ may conflict with each other causing unintended side effects, so choose hyperparameters accordingly * Modifiers will log their outputs to the console as well as to the TensorBoard file * After training is complete, the final model will be exported to ONNX using SparseML -* By default, EMA is disabled when using `train.py`. This is to allow for best compatibility -with pruning and quantization. To enable, set the `--use-ema` flag +* By default, EMA is disabled when using the integrated `train.py`. This is to allow +for best compatibility with pruning and quantization. To enable, set the `--use-ema` +flag +* By default, Automatic Mixed Precision (AMP) is disabled when using the integrated +`train.py`. This is because mixed precision is not supported for PyTorch +quantization-aware training. To enable, set the `--use-amp` flag You can learn how to build or download a recipe using the [SparseML](https://github.com/neuralmagic/sparseml) diff --git a/integrations/ultralytics/train.py b/integrations/ultralytics/train.py index ae282a0c0d0..e7ddbfdf51a 100644 --- a/integrations/ultralytics/train.py +++ b/integrations/ultralytics/train.py @@ -282,7 +282,7 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): results = (0, 0, 0, 0, 0, 0, 0) # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls) if scheduler: scheduler.last_epoch = start_epoch - 1 # do not move - scaler = amp.GradScaler(enabled=cuda) + scaler = amp.GradScaler(enabled=(cuda and opt.use_amp)) compute_loss = ComputeLoss(model) # init loss class logger.info(f'Image sizes {imgsz} train, {imgsz_test} test\n' f'Using {dataloader.num_workers} dataloader workers\n' @@ -341,7 +341,7 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) # Forward - with amp.autocast(enabled=cuda): + with amp.autocast(enabled=(cuda and opt.use_amp)): pred = model(imgs) # forward loss, loss_items = compute_loss(pred, targets.to(device)) # loss scaled by batch_size if rank != -1: @@ -529,6 +529,12 @@ def train(hyp, opt, device, tb_writer=None, wandb=None): action="store_true", help="set flag to enable EMA updates. disabled by default in SparseML integration" ) + parser.add_argument( + "--use-amp", + action="store_true", + help="set flag to enable Automatic Mixed Precision (AMP). disabled by default " + "in SparseML integration" + ) #################################################################################### # End SparseML arguments ####################################################################################