[fbsync] Additional SOTA ingredients on Classification Recipe (#4493)

Summary:
* Update EMA every X iters.

* Adding AdamW optimizer.

* Adjusting EMA decay scheme.

* Support custom weight decay for Normalization layers.

* Fix identation bug.

* Change EMA adjustment.

* Quality of life changes to faciliate testing

* ufmt format

* Fixing imports.

* Adding FixRes improvement.

* Support EMA in store_model_weights.

* Adding interpolation values.

* Change train_crop_size.

* Add interpolation option.

* Removing hardcoded interpolation and sizes from the scripts.

* Fixing linter.

* Incorporating feedback from code review.

Reviewed By: NicolasHug

Differential Revision: D31916313

fbshipit-source-id: 6136c02dd6d511d0f327b5a72c9056a134abc697

Author: datumbox

references/classification/README.md

@@ -31,6 +31,17 @@ Here `$MODEL` is one of `alexnet`, `vgg11`, `vgg13`, `vgg16` or `vgg19`. Note
 that `vgg11_bn`, `vgg13_bn`, `vgg16_bn`, and `vgg19_bn` include batch
 normalization and thus are trained with the default parameters.
 
+### Inception V3
+
+The weights of the Inception V3 model are ported from the original paper rather than trained from scratch.
+
+Since it expects tensors with a size of N x 3 x 299 x 299, to validate the model use the following command:
+
+```
+torchrun --nproc_per_node=8 train.py --model inception_v3
+      --val-resize-size 342 --val-crop-size 299 --train-crop-size 299 --test-only --pretrained
+```
+
 ### ResNext-50 32x4d
 ```
 torchrun --nproc_per_node=8 train.py\
@@ -79,6 +90,25 @@ The weights of the B0-B4 variants are ported from Ross Wightman's [timm repo](ht
 
 The weights of the B5-B7 variants are ported from Luke Melas' [EfficientNet-PyTorch repo](https://github.com/lukemelas/EfficientNet-PyTorch/blob/1039e009545d9329ea026c9f7541341439712b96/efficientnet_pytorch/utils.py#L562-L564).
 
+All models were trained using Bicubic interpolation and each have custom crop and resize sizes. To validate the models use the following commands:
+```
+torchrun --nproc_per_node=8 train.py --model efficientnet_b0 --interpolation bicubic\
+     --val-resize-size 256 --val-crop-size 224 --train-crop-size 224 --test-only --pretrained
+torchrun --nproc_per_node=8 train.py --model efficientnet_b1 --interpolation bicubic\
+      --val-resize-size 256 --val-crop-size 240 --train-crop-size 240 --test-only --pretrained
+torchrun --nproc_per_node=8 train.py --model efficientnet_b2 --interpolation bicubic\
+      --val-resize-size 288 --val-crop-size 288 --train-crop-size 288 --test-only --pretrained
+torchrun --nproc_per_node=8 train.py --model efficientnet_b3 --interpolation bicubic\
+      --val-resize-size 320 --val-crop-size 300 --train-crop-size 300 --test-only --pretrained
+torchrun --nproc_per_node=8 train.py --model efficientnet_b4 --interpolation bicubic\
+      --val-resize-size 384 --val-crop-size 380 --train-crop-size 380 --test-only --pretrained
+torchrun --nproc_per_node=8 train.py --model efficientnet_b5 --interpolation bicubic\
+      --val-resize-size 456 --val-crop-size 456 --train-crop-size 456 --test-only --pretrained
+torchrun --nproc_per_node=8 train.py --model efficientnet_b6 --interpolation bicubic\
+      --val-resize-size 528 --val-crop-size 528 --train-crop-size 528 --test-only --pretrained
+torchrun --nproc_per_node=8 train.py --model efficientnet_b7 --interpolation bicubic\
+      --val-resize-size 600 --val-crop-size 600 --train-crop-size 600 --test-only --pretrained
+```
 
 ### RegNet
 
@@ -181,3 +211,8 @@ For post training quant, device is set to CPU. For training, the device is set t
 ```
 python train_quantization.py --device='cpu' --test-only --backend='<backend>' --model='<model_name>'
 ```
+
+For inception_v3 you need to pass the following extra parameters:
+```
+--val-resize-size 342 --val-crop-size 299 --train-crop-size 299
+```

references/classification/presets.py

@@ -9,21 +9,22 @@ def __init__(
         crop_size,
         mean=(0.485, 0.456, 0.406),
         std=(0.229, 0.224, 0.225),
+        interpolation=InterpolationMode.BILINEAR,
         hflip_prob=0.5,
         auto_augment_policy=None,
         random_erase_prob=0.0,
     ):
-        trans = [transforms.RandomResizedCrop(crop_size)]
+        trans = [transforms.RandomResizedCrop(crop_size, interpolation=interpolation)]
         if hflip_prob > 0:
             trans.append(transforms.RandomHorizontalFlip(hflip_prob))
         if auto_augment_policy is not None:
             if auto_augment_policy == "ra":
-                trans.append(autoaugment.RandAugment())
+                trans.append(autoaugment.RandAugment(interpolation=interpolation))
             elif auto_augment_policy == "ta_wide":
-                trans.append(autoaugment.TrivialAugmentWide())
+                trans.append(autoaugment.TrivialAugmentWide(interpolation=interpolation))
             else:
                 aa_policy = autoaugment.AutoAugmentPolicy(auto_augment_policy)
-                trans.append(autoaugment.AutoAugment(policy=aa_policy))
+                trans.append(autoaugment.AutoAugment(policy=aa_policy, interpolation=interpolation))
         trans.extend(
             [
                 transforms.PILToTensor(),

references/classification/train.py

@@ -14,22 +14,20 @@
 from torchvision.transforms.functional import InterpolationMode
 
 
-def train_one_epoch(
-    model, criterion, optimizer, data_loader, device, epoch, print_freq, amp=False, model_ema=None, scaler=None
-):
+def train_one_epoch(model, criterion, optimizer, data_loader, device, epoch, args, model_ema=None, scaler=None):
     model.train()
     metric_logger = utils.MetricLogger(delimiter="  ")
     metric_logger.add_meter("lr", utils.SmoothedValue(window_size=1, fmt="{value}"))
     metric_logger.add_meter("img/s", utils.SmoothedValue(window_size=10, fmt="{value}"))
 
     header = "Epoch: [{}]".format(epoch)
-    for image, target in metric_logger.log_every(data_loader, print_freq, header):
+    for i, (image, target) in enumerate(metric_logger.log_every(data_loader, args.print_freq, header)):
         start_time = time.time()
         image, target = image.to(device), target.to(device)
         output = model(image)
 
         optimizer.zero_grad()
-        if amp:
+        if args.amp:
             with torch.cuda.amp.autocast():
                 loss = criterion(output, target)
             scaler.scale(loss).backward()
@@ -40,16 +38,19 @@ def train_one_epoch(
             loss.backward()
         optimizer.step()
 
+        if model_ema and i % args.model_ema_steps == 0:
+            model_ema.update_parameters(model)
+            if epoch < args.lr_warmup_epochs:
+                # Reset ema buffer to keep copying weights during warmup period
+                model_ema.n_averaged.fill_(0)
+
         acc1, acc5 = utils.accuracy(output, target, topk=(1, 5))
         batch_size = image.shape[0]
         metric_logger.update(loss=loss.item(), lr=optimizer.param_groups[0]["lr"])
         metric_logger.meters["acc1"].update(acc1.item(), n=batch_size)
         metric_logger.meters["acc5"].update(acc5.item(), n=batch_size)
         metric_logger.meters["img/s"].update(batch_size / (time.time() - start_time))
 
-    if model_ema:
-        model_ema.update_parameters(model)
-
 
 def evaluate(model, criterion, data_loader, device, print_freq=100, log_suffix=""):
     model.eval()
@@ -106,24 +107,8 @@ def _get_cache_path(filepath):
 def load_data(traindir, valdir, args):
     # Data loading code
     print("Loading data")
-    resize_size, crop_size = 256, 224
-    interpolation = InterpolationMode.BILINEAR
-    if args.model == "inception_v3":
-        resize_size, crop_size = 342, 299
-    elif args.model.startswith("efficientnet_"):
-        sizes = {
-            "b0": (256, 224),
-            "b1": (256, 240),
-            "b2": (288, 288),
-            "b3": (320, 300),
-            "b4": (384, 380),
-            "b5": (456, 456),
-            "b6": (528, 528),
-            "b7": (600, 600),
-        }
-        e_type = args.model.replace("efficientnet_", "")
-        resize_size, crop_size = sizes[e_type]
-        interpolation = InterpolationMode.BICUBIC
+    val_resize_size, val_crop_size, train_crop_size = args.val_resize_size, args.val_crop_size, args.train_crop_size
+    interpolation = InterpolationMode(args.interpolation)
 
     print("Loading training data")
     st = time.time()
@@ -138,7 +123,10 @@ def load_data(traindir, valdir, args):
         dataset = torchvision.datasets.ImageFolder(
             traindir,
             presets.ClassificationPresetTrain(
-                crop_size=crop_size, auto_augment_policy=auto_augment_policy, random_erase_prob=random_erase_prob
+                crop_size=train_crop_size,
+                interpolation=interpolation,
+                auto_augment_policy=auto_augment_policy,
+                random_erase_prob=random_erase_prob,
             ),
         )
         if args.cache_dataset:
@@ -156,7 +144,9 @@ def load_data(traindir, valdir, args):
     else:
         dataset_test = torchvision.datasets.ImageFolder(
             valdir,
-            presets.ClassificationPresetEval(crop_size=crop_size, resize_size=resize_size, interpolation=interpolation),
+            presets.ClassificationPresetEval(
+                crop_size=val_crop_size, resize_size=val_resize_size, interpolation=interpolation
+            ),
         )
         if args.cache_dataset:
             print("Saving dataset_test to {}".format(cache_path))
@@ -224,26 +214,30 @@ def main(args):
 
     criterion = nn.CrossEntropyLoss(label_smoothing=args.label_smoothing)
 
+    if args.norm_weight_decay is None:
+        parameters = model.parameters()
+    else:
+        param_groups = torchvision.ops._utils.split_normalization_params(model)
+        wd_groups = [args.norm_weight_decay, args.weight_decay]
+        parameters = [{"params": p, "weight_decay": w} for p, w in zip(param_groups, wd_groups) if p]
+
     opt_name = args.opt.lower()
     if opt_name.startswith("sgd"):
         optimizer = torch.optim.SGD(
-            model.parameters(),
+            parameters,
             lr=args.lr,
             momentum=args.momentum,
             weight_decay=args.weight_decay,
             nesterov="nesterov" in opt_name,
         )
     elif opt_name == "rmsprop":
         optimizer = torch.optim.RMSprop(
-            model.parameters(),
-            lr=args.lr,
-            momentum=args.momentum,
-            weight_decay=args.weight_decay,
-            eps=0.0316,
-            alpha=0.9,
+            parameters, lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay, eps=0.0316, alpha=0.9
         )
+    elif opt_name == "adamw":
+        optimizer = torch.optim.AdamW(parameters, lr=args.lr, weight_decay=args.weight_decay)
     else:
-        raise RuntimeError("Invalid optimizer {}. Only SGD and RMSprop are supported.".format(args.opt))
+        raise RuntimeError(f"Invalid optimizer {args.opt}. Only SGD, RMSprop and AdamW are supported.")
 
     scaler = torch.cuda.amp.GradScaler() if args.amp else None
 
@@ -288,13 +282,23 @@ def main(args):
 
     model_ema = None
     if args.model_ema:
-        model_ema = utils.ExponentialMovingAverage(model_without_ddp, device=device, decay=args.model_ema_decay)
+        # Decay adjustment that aims to keep the decay independent from other hyper-parameters originally proposed at:
+        # https://github.com/facebookresearch/pycls/blob/f8cd9627/pycls/core/net.py#L123
+        #
+        # total_ema_updates = (Dataset_size / n_GPUs) * epochs / (batch_size_per_gpu * EMA_steps)
+        # We consider constant = Dataset_size for a given dataset/setup and ommit it. Thus:
+        # adjust = 1 / total_ema_updates ~= n_GPUs * batch_size_per_gpu * EMA_steps / epochs
+        adjust = args.world_size * args.batch_size * args.model_ema_steps / args.epochs
+        alpha = 1.0 - args.model_ema_decay
+        alpha = min(1.0, alpha * adjust)
+        model_ema = utils.ExponentialMovingAverage(model_without_ddp, device=device, decay=1.0 - alpha)
 
     if args.resume:
         checkpoint = torch.load(args.resume, map_location="cpu")
         model_without_ddp.load_state_dict(checkpoint["model"])
-        optimizer.load_state_dict(checkpoint["optimizer"])
-        lr_scheduler.load_state_dict(checkpoint["lr_scheduler"])
+        if not args.test_only:
+            optimizer.load_state_dict(checkpoint["optimizer"])
+            lr_scheduler.load_state_dict(checkpoint["lr_scheduler"])
         args.start_epoch = checkpoint["epoch"] + 1
         if model_ema:
             model_ema.load_state_dict(checkpoint["model_ema"])
@@ -303,18 +307,18 @@ def main(args):
         # We disable the cudnn benchmarking because it can noticeably affect the accuracy
         torch.backends.cudnn.benchmark = False
         torch.backends.cudnn.deterministic = True
-
-        evaluate(model, criterion, data_loader_test, device=device)
+        if model_ema:
+            evaluate(model_ema, criterion, data_loader_test, device=device, log_suffix="EMA")
+        else:
+            evaluate(model, criterion, data_loader_test, device=device)
         return
 
     print("Start training")
     start_time = time.time()
     for epoch in range(args.start_epoch, args.epochs):
         if args.distributed:
             train_sampler.set_epoch(epoch)
-        train_one_epoch(
-            model, criterion, optimizer, data_loader, device, epoch, args.print_freq, args.amp, model_ema, scaler
-        )
+        train_one_epoch(model, criterion, optimizer, data_loader, device, epoch, args, model_ema, scaler)
         lr_scheduler.step()
         evaluate(model, criterion, data_loader_test, device=device)
         if model_ema:
@@ -362,6 +366,12 @@ def get_args_parser(add_help=True):
         help="weight decay (default: 1e-4)",
         dest="weight_decay",
     )
+    parser.add_argument(
+        "--norm-weight-decay",
+        default=None,
+        type=float,
+        help="weight decay for Normalization layers (default: None, same value as --wd)",
+    )
     parser.add_argument(
         "--label-smoothing", default=0.0, type=float, help="label smoothing (default: 0.0)", dest="label_smoothing"
     )
@@ -415,15 +425,33 @@ def get_args_parser(add_help=True):
     parser.add_argument(
         "--model-ema", action="store_true", help="enable tracking Exponential Moving Average of model parameters"
     )
+    parser.add_argument(
+        "--model-ema-steps",
+        type=int,
+        default=32,
+        help="the number of iterations that controls how often to update the EMA model (default: 32)",
+    )
     parser.add_argument(
         "--model-ema-decay",
         type=float,
-        default=0.9,
-        help="decay factor for Exponential Moving Average of model parameters(default: 0.9)",
+        default=0.99998,
+        help="decay factor for Exponential Moving Average of model parameters (default: 0.99998)",
     )
     parser.add_argument(
         "--use-deterministic-algorithms", action="store_true", help="Forces the use of deterministic algorithms only."
     )
+    parser.add_argument(
+        "--interpolation", default="bilinear", type=str, help="the interpolation method (default: bilinear)"
+    )
+    parser.add_argument(
+        "--val-resize-size", default=256, type=int, help="the resize size used for validation (default: 256)"
+    )
+    parser.add_argument(
+        "--val-crop-size", default=224, type=int, help="the central crop size used for validation (default: 224)"
+    )
+    parser.add_argument(
+        "--train-crop-size", default=224, type=int, help="the random crop size used for training (default: 224)"
+    )
 
     return parser
 

references/classification/train_quantization.py

@@ -236,6 +236,19 @@ def get_args_parser(add_help=True):
     parser.add_argument("--world-size", default=1, type=int, help="number of distributed processes")
     parser.add_argument("--dist-url", default="env://", help="url used to set up distributed training")
 
+    parser.add_argument(
+        "--interpolation", default="bilinear", type=str, help="the interpolation method (default: bilinear)"
+    )
+    parser.add_argument(
+        "--val-resize-size", default=256, type=int, help="the resize size used for validation (default: 256)"
+    )
+    parser.add_argument(
+        "--val-crop-size", default=224, type=int, help="the central crop size used for validation (default: 224)"
+    )
+    parser.add_argument(
+        "--train-crop-size", default=224, type=int, help="the random crop size used for training (default: 224)"
+    )
+
     return parser
 
 

references/classification/utils.py

@@ -380,6 +380,9 @@ def store_model_weights(model, checkpoint_path, checkpoint_key="model", strict=T
 
     # Load the weights to the model to validate that everything works
     # and remove unnecessary weights (such as auxiliaries, etc)
+    if checkpoint_key == "model_ema":
+        del checkpoint[checkpoint_key]["n_averaged"]
+        torch.nn.modules.utils.consume_prefix_in_state_dict_if_present(checkpoint[checkpoint_key], "module.")
     model.load_state_dict(checkpoint[checkpoint_key], strict=strict)
 
     tmp_path = os.path.join(output_dir, str(model.__hash__()))

test/test_ops.py

@@ -9,10 +9,10 @@
 import torch
 from common_utils import needs_cuda, cpu_and_gpu, assert_equal
 from PIL import Image
-from torch import Tensor
+from torch import nn, Tensor
 from torch.autograd import gradcheck
 from torch.nn.modules.utils import _pair
-from torchvision import ops
+from torchvision import models, ops
 
 
 class RoIOpTester(ABC):
@@ -1176,5 +1176,15 @@ def test_stochastic_depth(self, mode, p):
         assert p_value > 0.0001
 
 
+class TestUtils:
+    @pytest.mark.parametrize("norm_layer", [None, nn.BatchNorm2d, nn.LayerNorm])
+    def test_split_normalization_params(self, norm_layer):
+        model = models.mobilenet_v3_large(norm_layer=norm_layer)
+        params = ops._utils.split_normalization_params(model, None if norm_layer is None else [norm_layer])
+
+        assert len(params[0]) == 92
+        assert len(params[1]) == 82
+
+
 if __name__ == "__main__":
     pytest.main([__file__])