update readme and modify some hyper parameter

pprp · Jun 18, 2021 · 2730a1f · 2730a1f
1 parent 569d5e7
commit 2730a1f
Show file tree

Hide file tree

Showing 6 changed files with 43 additions and 18 deletions.
diff --git a/NAS/cifar100/README.md b/NAS/cifar100/README.md
@@ -14,6 +14,7 @@
 - [ ] experimental results on cifar100 and cifar10
 - [ ] distributed data parrallel to support
 - [ ] support transformer models
+- [ ] support differenet datasets such as tiny imagenet
 
 ## Get Started
 
@@ -26,8 +27,28 @@
 Training Details:
 
 - total epoch: 300 
-- lr scheduler: 
-- weight decay: 
+- lr scheduler: CosineAnnealingLR
+- weight decay: 5e-4
+
+### CIFAR10
+
+| Network        | Params(M) | Train loss | Train top1 | Val loss | Val top1 | Hyper              | GPU(M) |
+| -------------- | --------- | ---------- | ---------- | -------- | -------- | ------------------ | ------ |
+| densenet_cifar | 4.4       | 0.00156    | 99.99%     | 0.24     | 94.83%   | 0.1/256/w/o cutout | 7303   |
+| dla            | 63        | 0.00164    | 99.99%     | 0.20     | 95.57%   | 0.1/256/w/o cutout | 5555   |
+|                |           |            |            |          |          |                    |        |
+|                |           |            |            |          |          |                    |        |
+|                |           |            |            |          |          |                    |        |
+|                |           |            |            |          |          |                    |        |
+|                |           |            |            |          |          |                    |        |
+|                |           |            |            |          |          |                    |        |
+|                |           |            |            |          |          |                    |        |
+
+
+
+
+
+### CIFAR100
 
 | dataset  | network            | params | top1 err | top5 err | epoch(lr = 0.1) | epoch(lr = 0.02) | epoch(lr = 0.004) | epoch(lr = 0.0008) | total epoch |
 | -------- | ------------------ | ------ | -------- | -------- | --------------- | ---------------- | ----------------- | ------------------ | ----------- |

diff --git a/NAS/cifar100/configs/meta.yml b/NAS/cifar100/configs/meta.yml
@@ -27,11 +27,11 @@ save_dir: record
 dataset: cifar100
 # support cifar100 or cifar10
 
-epochs: 200
+epochs: 300
 batch_size: 2048
-learning_rate: 0.1 
+learning_rate: 0.1
 num_workers: 3
-model_type: sample
+model_type: dla
 finetune: false 
 distill: false
 
@@ -41,7 +41,6 @@ seed: 0
 
 report_freq: 2
 gpu: 0
-classes: 100
 
 # 算法细节
 sandwich_N: 3

diff --git a/NAS/cifar100/datasets/dataset.py b/NAS/cifar100/datasets/dataset.py
@@ -176,7 +176,7 @@ def get_train_loader(batch_size, num_workers, clss='cifar10', cutout=0):
     # 3. get dataloader
 
     train_loader = torch.utils.data.DataLoader(
-        train_dataset, num_workers=num_workers, pin_memory=True, batch_size=batch_size, drop_last=True)
+        train_dataset, num_workers=num_workers, pin_memory=True, batch_size=batch_size, drop_last=True, shuffle=True)
 
     return train_loader
 

diff --git a/NAS/cifar100/datasets/transforms.py b/NAS/cifar100/datasets/transforms.py
@@ -33,6 +33,8 @@ def __call__(self, img):
         return img
 
 
+
+
 class DatasetTransforms:
     def __init__(self, clss, cutout=0):
         if clss == 'cifar10':
@@ -55,8 +57,8 @@ def _get_default_transforms(self):
         default_configure = T.Compose([
             T.RandomCrop(32, 4),
             T.RandomHorizontalFlip(),
-            T.RandomResizedCrop((32, 32)),  # for cifar10 or cifar100
-            T.RandomRotation(15)
+            # T.RandomResizedCrop((32, 32)),  # for cifar10 or cifar100
+            # T.RandomRotation(15)
         ])
         return default_configure
 

diff --git a/NAS/cifar100/script.sh → NAS/cifar100/scripts/train_script.sh b/NAS/cifar100/script.sh → NAS/cifar100/scripts/train_script.sh
diff --git a/NAS/cifar100/train.py b/NAS/cifar100/train.py
@@ -25,17 +25,17 @@
 
 print = functools.partial(print, flush=True)
 
-parser = argparse.ArgumentParser("ResNet20-cifar100")
+parser = argparse.ArgumentParser("cifar")
 
 parser.add_argument('--local_rank', type=int, default=0,
                     help='local rank for distributed training')
-parser.add_argument('--batch_size', type=int, default=128,
+parser.add_argument('--batch_size', type=int, default=256,
                     help='batch size')  # 8192
 parser.add_argument('--learning_rate', type=float,
                     default=0.1, help='init learning rate')  # 0.8
 parser.add_argument('--num_workers', type=int,
                     default=3, help='num of workers')
-parser.add_argument('--model-type', type=str, default="densenet",
+parser.add_argument('--model-type', type=str, default="resnet50",
                     help="type of model(sample masked dynamic independent slimmable original)")
 
 parser.add_argument('--finetune', action='store_true',
@@ -51,9 +51,9 @@
                     default=5e-4, help='weight decay')
 
 parser.add_argument('--report_freq', type=float,
-                    default=2, help='report frequency')
+                    default=5, help='report frequency')
 parser.add_argument('--gpu', type=int, default=0, help='gpu device id')
-parser.add_argument('--epochs', type=int, default=200,
+parser.add_argument('--epochs', type=int, default=300,
                     help='num of training epochs')
 
 parser.add_argument('--classes', type=int, default=100,
@@ -66,7 +66,7 @@
 parser.add_argument('--config', help="configuration file",
                     type=str, default="configs/meta.yml")
 parser.add_argument('--save_dir', type=str,
-                    help="save exp floder name", default="densenet")
+                    help="save exp floder name", default="resnet50")
 args = parser.parse_args()
 
 # process argparse & yaml
@@ -155,7 +155,7 @@ def main():
                                 weight_decay=args.weight_decay)
 
     scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
-        optimizer, 200, eta_min=0.0005)
+        optimizer, 300, eta_min=0.0005)
 
     # Prepare data
     train_loader = get_train_loader(
@@ -316,8 +316,11 @@ def train(train_dataloader, val_dataloader, optimizer, scheduler, model, archloa
             writer.add_scalar("Train/acc1", top1_.avg, step +
                               len(train_dataloader) * epoch * args.batch_size)
 
-        logging.info('{} |=> Train loss = {} Train acc = {}'.format(
-            losses_.avg, top1_.avg))
+        now = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
+
+        if step % 10 == 0:
+            logging.info('{} |=> Train loss = {} Train acc = {}'.format(now,
+                                                                        losses_.avg, top1_.avg))
 
 
 def infer(train_loader, val_loader, model, criterion,  archloader, args, epoch):