swin ported. ./run.sh to reproduce

.gitignore

@@ -6,4 +6,27 @@ checkpoints/
 .idea/
 wandb/
 *.pth
-debug*
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST

datasets_builders.py

@@ -8,7 +8,6 @@
 
 from timm.data.constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
 from timm.data import create_transform
-from datasets import load_dataset
 from debug import DEBUG
 
 class INatDataset(ImageFolder):
@@ -65,6 +64,7 @@ def build_dataset(is_train, data_path, args):
         dataset = datasets.ImageFolder(root, transform=transform)
         nb_classes = 1000
     elif args.data_set == 'HUGGINGFACE':
+        from datasets import load_dataset
         def huggingface_transform(examples):
             examples["image"] = [transform(x.convert(mode="RGB")) for x in examples["image"]]
             return examples

debug.py

@@ -0,0 +1 @@
+DEBUG=False

main.py

@@ -228,6 +228,9 @@ def get_args_parser():
                         help='use scale-aware embeds')
     parser.add_argument('--grid-to-random-ratio', default=0.7, type=float, help='hybrid sampler grid to random ratio')
 
+    parser.add_argument('--model-type', default='deit', type=str, choices=["deit", "swin"])
+
+
     return parser
 
 
@@ -356,19 +359,35 @@ def log_to_wandb(log_dict, step):
     mixup_fn = None
 
     print(f"Creating model: {args.model}")
-    model = create_model(
-        args.model,
-        pretrained=False,
-        num_classes=args.nb_classes,
-        drop_rate=args.drop,
-        drop_path_rate=args.drop_path,
-        drop_block_rate=None,
-        img_size=args.input_size,
-
-        Patch_layer=PatchEmbedHybrid,
-        use_learned_pos_embed = args.use_learned_pos_embed,
-        quantize_pos_embed = args.quantize_pos_embed
-    )
+    if args.model_type == "deit":
+        model = create_model(
+            args.model,
+            pretrained=False,
+            num_classes=args.nb_classes,
+            drop_rate=args.drop,
+            drop_path_rate=args.drop_path,
+            drop_block_rate=None,
+            img_size=args.input_size,
+
+            Patch_layer=PatchEmbedHybrid,
+            use_learned_pos_embed = args.use_learned_pos_embed,
+            quantize_pos_embed = args.quantize_pos_embed
+        )
+    elif args.model_type == "swin":
+        from swin.config import get_config
+        from swin.models import build_model
+        class _Args:
+            pass
+        _args = _Args()
+        cfg_path = "swin/swin_base_patch4_window7_224.yaml"
+        setattr(_args, "cfg", cfg_path)
+        setattr(_args, "opts", [])
+        setattr(_args, "local_rank", 0)
+        setattr(_args, "data_path", args.data_path)
+        config = get_config(_args)
+        model = build_model(config)
+    else:
+        assert False
 
     if args.finetune:
         if args.finetune.startswith('https'):

run.sh

@@ -0,0 +1,54 @@
+#!/bin/bash
+
+model_name=STD_SWINB
+model_path=models/swin_base_patch4_window7_224.pth
+input_size=224
+patch_size=16
+patch_shake=0
+patch_dropout=0
+patch_select=0
+patch_select_mode=drop
+patch_zoom=0
+
+
+#model_name=$1
+#model_path=$2
+#input_size=$3
+#patch_size=$4
+#patch_shake=$5
+#patch_dropout=$6
+#patch_select=$7
+#patch_select_mode=$8
+#patch_zoom=$9
+
+run_name="${model_name}_GRID${input_size}x${patch_size}_SHAKE${patch_shake}_DROP${patch_dropout}_SELECT${patch_select}_MODE${patch_select_mode}_ZOOM${patch_zoom}"
+echo $run_name
+if [[ $model_name == *"STD"* ]]; then
+    quantize_pos_embed=1
+    use_learned_pos_embed=1
+else
+    quantize_pos_embed=0
+    use_learned_pos_embed=0
+fi
+
+python main.py \
+    --model-type swin \
+    --no-log-to-wandb \
+    --hard_dropout 0 \
+    --merge_patches 1 \
+    --model _ \
+    --data-path ~/datasets/imagenet/ \
+    --batch 16 --input-size $input_size --eval-crop-ratio 1.0 --seed 0 \
+    --grid-patch-size $patch_size \
+    --num_workers 10 \
+    --eval --resume=$model_path \
+    --patch_shake=$patch_shake \
+    --patch_dropout=$patch_dropout \
+    --patch_select=$patch_select \
+    --patch_select_mode=$patch_select_mode \
+    --quantize_pos_embed=$quantize_pos_embed \
+    --use_learned_pos_embed=$use_learned_pos_embed \
+    --patch_zoom=$patch_zoom \
+    --modify_patches
+
+    #--log-to-wandb --wandb_run_name=$run_name --wandb_model_name=$model_name \

scripts/run_all.sh

@@ -1,18 +1,20 @@
 #!/bin/bash
 
-source scripts/sweep_grid_scale.sh
+set -e
 
-#source scripts/sweep_zoom_exp.sh
-#source scripts/sweep_dropout_exp.sh
-#source scripts/sweep_shake_exp.sh
+#source scripts/sweep_grid_scale.sh
+
+source scripts/sweep_zoom_exp.sh
+source scripts/sweep_dropout_exp.sh
+source scripts/sweep_shake_exp.sh
 
 source scripts/sweep_zoom_shake_exp.sh
 source scripts/sweep_zoom_dropout_exp.sh 
-#source scripts/sweep_dropout_shake_exp.sh
+source scripts/sweep_dropout_shake_exp.sh
 #source random
 
 #source edge counting
 #source scripts/sweep_central_sampling_exp.sh
 
 #source 30/70, 70/30 vits
-#source scripts/sweep_dropout_vs_rescale_exp.sh
+source scripts/sweep_dropout_vs_rescale_exp.sh

scripts/single_grid_mod_exp.sh

@@ -10,7 +10,27 @@ patch_select=$7
 patch_select_mode=$8
 patch_zoom=$9
 
-run_name="${model_name}_GRID${input_size}x${patch_size}_SHAKE${patch_shake}_DROP${patch_dropout}_SELECT${patch_select}_MODE${patch_select_mode}_ZOOM${patch_zoom}"
+
+# change if needed
+hard_dropout=0
+merge_patches=1
+wandb="--log-to-wandb"
+data_set="IMNET"
+data_path="~/datasets/imagenet"
+# change if needed
+
+if [[ $model_name == *"PVT"* ]]; then
+    model_type="pvt"
+    model="pvt_v2_b5"
+elif [[ $model_name == *"SWIN"* ]]; then
+    model_type="swin"
+    model="swinb_p4_w7"
+else
+    model_type="deit"
+    model="deit_base_patch16_LS"
+fi
+
+run_name="${model_type}_${model_name}_HARD_DROP${hard_dropout}_MERGE_PATCHES_${merge_patches}_GRID${input_size}x${patch_size}_SHAKE${patch_shake}_DROP${patch_dropout}_SELECT${patch_select}_MODE${patch_select_mode}_ZOOM${patch_zoom}"
 echo $run_name
 if [[ $model_name == *"STD"* ]]; then
     quantize_pos_embed=1
@@ -20,14 +40,19 @@ else
     use_learned_pos_embed=0
 fi
 
+
 python main.py \
-    --model deit_base_patch16_LS \
-    --data-set HUGGINGFACE \
-    --data-path /home/jan.olszewski/datasets/imagenet-1k/default/1.0.0/a1e9bfc56c3a7350165007d1176b15e9128fcaf9ab972147840529aed3ae52bc \
+    --model-type $model_type \
+    --hard_dropout $hard_dropout \
+    --merge_patches $merge_patches \
+    --model $model \
+    --data-set $data_set \
+    --data-path $data_path \
     --batch 16 --input-size $input_size --eval-crop-ratio 1.0 --seed 0 \
     --grid-patch-size $patch_size \
     --num_workers 10 \
-    --log-to-wandb --wandb_run_name=$run_name --wandb_model_name=$model_name \
+    $wandb \
+    --wandb_run_name=$run_name --wandb_model_name=$model_name \
     --eval --resume=$model_path \
     --patch_shake=$patch_shake \
     --patch_dropout=$patch_dropout \
@@ -36,4 +61,4 @@ python main.py \
     --quantize_pos_embed=$quantize_pos_embed \
     --use_learned_pos_embed=$use_learned_pos_embed \
     --patch_zoom=$patch_zoom \
-    --modify_patches
+    --modify_patches

scripts/sweep_dropout_exp.sh

@@ -1,22 +1,9 @@
 #!/bin/bash
 
+set -e
 
-model_name=ELASTIC_VITB
-model_path=models/base_448.pth
-input_size=448
-patch_size=32
-patch_shake=0
-patch_dropout=0
-patch_select=0
-patch_select_mode=drop
-patch_zoom=0
-
-for ((patch_dropout=0; patch_dropout<=160; patch_dropout+=20)); do
-    source scripts/single_grid_mod_exp.sh $model_name $model_path $input_size $patch_size $patch_shake $patch_dropout $patch_select $patch_select_mode $patch_zoom
-done
-
-model_name=STD_VITB
-model_path=https://dl.fbaipublicfiles.com/deit/deit_3_base_224_1k.pth
+model_name=STD_SWINB
+model_path=models/swin_base_patch4_window7_224.pth
 input_size=224
 patch_size=16
 patch_shake=0

scripts/sweep_dropout_shake_exp.sh

@@ -1,24 +1,9 @@
 #!/bin/bash
 
+set -e
 
-model_name=ELASTIC_VITB
-model_path=models/base_448.pth
-input_size=448
-patch_size=32
-patch_shake=0
-patch_dropout=0
-patch_select=0
-patch_select_mode=drop
-patch_zoom=0
-
-for ((i=0; i<=8; i+=1)); do
-    patch_shake=$((2*i))
-    patch_dropout=$((20*i))
-    source scripts/single_grid_mod_exp.sh $model_name $model_path $input_size $patch_size $patch_shake $patch_dropout $patch_select $patch_select_mode $patch_zoom
-done
-
-model_name=STD_VITB
-model_path=https://dl.fbaipublicfiles.com/deit/deit_3_base_224_1k.pth
+model_name=STD_SWINB
+model_path=models/swin_base_patch4_window7_224.pth
 input_size=224
 patch_size=16
 patch_shake=0
@@ -31,4 +16,4 @@ for ((i=0; i<=8; i+=1)); do
     patch_shake=$i
     patch_dropout=$((20*i))
     source scripts/single_grid_mod_exp.sh $model_name $model_path $input_size $patch_size $patch_shake $patch_dropout $patch_select $patch_select_mode $patch_zoom
-done
+done

scripts/sweep_dropout_vs_rescale_exp.sh

@@ -1,39 +1,13 @@
 #!/bin/bash
 
-
-model_name=ELASTIC_VITB
-model_path=models/base_448.pth
-input_size=448
-patch_size=32
-patch_shake=0
-patch_dropout=0
-patch_select=0
-patch_select_mode=drop
-patch_selects=(30 40 49)
-patch_zoom=0
-
-for patch_select in ${patch_selects[@]}; do
-    source scripts/single_grid_mod_exp.sh $model_name $model_path $input_size $patch_size $patch_shake $patch_dropout $patch_select $patch_select_mode $patch_zoom
-done
+set -e
 
 
 patch_selects=(0 10 20 30 40 49)
-model_name=ELASTIC_VITB
-model_path=models/base_448.pth
-input_size=448
-patch_size=32
-patch_shake=0
-patch_dropout=0
-patch_select=0
-patch_select_mode=resize
-patch_zoom=0
 
-for patch_select in ${patch_selects[@]}; do
-    source scripts/single_grid_mod_exp.sh $model_name $model_path $input_size $patch_size $patch_shake $patch_dropout $patch_select $patch_select_mode $patch_zoom
-done
 
-model_name=STD_VITB
-model_path=https://dl.fbaipublicfiles.com/deit/deit_3_base_224_1k.pth
+model_name=STD_SWINB
+model_path=models/swin_base_patch4_window7_224.pth
 input_size=224
 patch_size=16
 patch_shake=0
@@ -47,8 +21,8 @@ for patch_select in ${patch_selects[@]}; do
 done
 
 
-model_name=STD_VITB
-model_path=https://dl.fbaipublicfiles.com/deit/deit_3_base_224_1k.pth
+model_name=STD_SWINB
+model_path=models/swin_base_patch4_window7_224.pth
 input_size=224
 patch_size=16
 patch_shake=0
@@ -59,4 +33,4 @@ patch_zoom=0
 
 for patch_select in ${patch_selects[@]}; do
     source scripts/single_grid_mod_exp.sh $model_name $model_path $input_size $patch_size $patch_shake $patch_dropout $patch_select $patch_select_mode $patch_zoom
-done
+done