diff --git a/README.md b/README.md
index 2eb9db8..90d83ce 100644
--- a/README.md
+++ b/README.md
@@ -1,19 +1,17 @@
# SwiftFormer
### **SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications**
-[Abdelrahman Shaker](https://scholar.google.com/citations?hl=en&user=eEz4Wu4AAAAJ),
-[Muhammad Maaz](https://scholar.google.com/citations?user=vTy9Te8AAAAJ&hl=en&authuser=1&oi=sra),
-[Hanoona Rasheed](https://scholar.google.com/citations?user=yhDdEuEAAAAJ&hl=en&authuser=1&oi=sra),
-[Salman Khan](https://salman-h-khan.github.io),
-[Ming-Hsuan Yang](https://scholar.google.com/citations?user=p9-ohHsAAAAJ&hl=en),
-and [Fahad Shahbaz Khan](https://scholar.google.es/citations?user=zvaeYnUAAAAJ&hl=en)
+
+[Abdelrahman Shaker](https://scholar.google.com/citations?hl=en&user=eEz4Wu4AAAAJ)*1, [Muhammad Maaz](https://scholar.google.com/citations?user=vTy9Te8AAAAJ&hl=en&authuser=1&oi=sra)1, [Hanoona Rasheed](https://scholar.google.com/citations?user=yhDdEuEAAAAJ&hl=en&authuser=1&oi=sra)1, [Salman Khan](https://salman-h-khan.github.io/)1, [Ming-Hsuan Yang](https://scholar.google.com/citations?user=p9-ohHsAAAAJ&hl=en)2,3 and [Fahad Shahbaz Khan](https://scholar.google.es/citations?user=zvaeYnUAAAAJ&hl=en)1,4
+Mohamed Bin Zayed University of Artificial Intelligence1, University of California Merced2, Google Research3, Linkoping University4
[](https://arxiv.org/abs/2303.15446)
## :rocket: News
+* **(Jul 14, 2023):** SwiftFormer has been accepted at ICCV 2023. :fire::fire:
* **(Mar 27, 2023):** Classification training and evaluation codes along with pre-trained models are released.
@@ -47,10 +45,10 @@ Self-attention has become a defacto choice for capturing global context in vario
| Model | Top-1 accuracy | #params | GMACs | Latency | Ckpt | CoreML|
|:---------------|:----:|:---:|:--:|:--:|:--:|:--:|
-| SwiftFormer-XS | 75.7% | 3.5M | 0.4G | 0.7ms | [XS](https://drive.google.com/file/d/15Ils-U96pQePXQXx2MpmaI-yAceFAr2x/view?usp=sharing) | [XS](https://drive.google.com/file/d/1tZVxtbtAZoLLoDc5qqoUGulilksomLeK/view?usp=sharing) |
-| SwiftFormer-S | 78.5% | 6.1M | 1.0G | 0.8ms | [S](https://drive.google.com/file/d/1_0eWwgsejtS0bWGBQS3gwAtYjXdPRGlu/view?usp=sharing) | [S](https://drive.google.com/file/d/13EOCZmtvbMR2V6UjezSZnbBz2_-59Fva/view?usp=sharing) |
-| SwiftFormer-L1 | 80.9% | 12.1M | 1.6G | 1.1ms | [L1](https://drive.google.com/file/d/1jlwrwWQ0SQzDRc5adtWIwIut5d1g9EsM/view?usp=sharing) | [L1](https://drive.google.com/file/d/1c3VUsi4q7QQ2ykXVS2d4iCRL478fWF3e/view?usp=sharing) |
-| SwiftFormer-L3 | 83.0% | 28.5M | 4.0G | 1.9ms | [L3](https://drive.google.com/file/d/1ypBcjx04ShmPYRhhjBRubiVjbExUgSa7/view?usp=sharing) | [L3](https://drive.google.com/file/d/1svahgIjh7da781jHOHjX58mtzCzYXSsJ/view?usp=sharing) |
+| SwiftFormer-XS | 75.7% | 3.5M | 0.6G | 0.7ms | [XS](https://drive.google.com/file/d/12RchxzyiJrtZS-2Bur9k4wcRQMItA43S/view?usp=sharing) | [XS](https://drive.google.com/file/d/1bkAP_BD6CdDqlbQsStZhLa0ST2NZTIvH/view?usp=sharing) |
+| SwiftFormer-S | 78.5% | 6.1M | 1.0G | 0.8ms | [S](https://drive.google.com/file/d/1awpcXAaHH38WaHrOmUM8updxQazUZ3Nb/view?usp=sharing) | [S](https://drive.google.com/file/d/1qNAhecWIeQ1YJotWhbnLTCR5Uv1zBaf1/view?usp=sharing) |
+| SwiftFormer-L1 | 80.9% | 12.1M | 1.6G | 1.1ms | [L1](https://drive.google.com/file/d/1SDzauVmpR5uExkOv3ajxdwFnP-Buj9Uo/view?usp=sharing) | [L1](https://drive.google.com/file/d/1CowZE7-lbxz93uwXqefe-HxGOHUdvX_a/view?usp=sharing) |
+| SwiftFormer-L3 | 83.0% | 28.5M | 4.0G | 1.9ms | [L3](https://drive.google.com/file/d/1DAxMe6FlnZBBIpR-HYIDfFLWJzIgiF0Y/view?usp=sharing) | [L3](https://drive.google.com/file/d/1SO3bRWd9oWJemy-gpYUcwP-B4bJ-dsdg/view?usp=sharing) |
## Detection and Segmentation Qualitative Results
@@ -77,6 +75,7 @@ conda activate swiftformer
pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 --extra-index-url https://download.pytorch.org/whl/cu113
pip install timm
+pip install coremltools==5.2.0
```
### Data preparation
@@ -98,7 +97,7 @@ To train SwiftFormer models on an 8-GPU machine:
sh dist_train.sh /path/to/imagenet 8
```
-Note: specify which model command you want to run in the script. To reproduce the results of the paper, use 16-GPU machine with batch-size of 128 or 8-GPU machine with batch size of 256. Auto Augmentation, CutMix, MixUp are disabled for SwiftFormer-XS only.
+Note: specify which model command you want to run in the script. To reproduce the results of the paper, use 16-GPU machine with batch-size of 128 or 8-GPU machine with batch size of 256. Auto Augmentation, CutMix, MixUp are disabled for SwiftFormer-XS, and CutMix, MixUp are disabled for SwiftFormer-S.
### Multi-node training
diff --git a/dist_train.sh b/dist_train.sh
index 0f81d00..2578b11 100644
--- a/dist_train.sh
+++ b/dist_train.sh
@@ -4,18 +4,18 @@
IMAGENET_PATH=$1
nGPUs=$2
-## SwiftFormer-XS
+## SwiftFormer-XS training
python -m torch.distributed.launch --nproc_per_node=$nGPUs --use_env main.py --model SwiftFormer_XS --aa="" --mixup 0 --cutmix 0 --data-path "$IMAGENET_PATH" \
--output_dir SwiftFormer_XS_results
-## SwiftFormer-S
-python -m torch.distributed.launch --nproc_per_node=$nGPUs --use_env main.py --model SwiftFormer_S --data-path "$IMAGENET_PATH" \
+## SwiftFormer-S training
+python -m torch.distributed.launch --nproc_per_node=$nGPUs --use_env main.py --model SwiftFormer_S --mixup 0 --cutmix 0 --data-path "$IMAGENET_PATH" \
--output_dir SwiftFormer_S_results
-## SwiftFormer-L1
+## SwiftFormer-L1 training
python -m torch.distributed.launch --nproc_per_node=$nGPUs --use_env main.py --model SwiftFormer_L1 --data-path "$IMAGENET_PATH" \
--output_dir SwiftFormer_L1_results
-## SwiftFormer-L3
+## SwiftFormer-L3 training
python -m torch.distributed.launch --nproc_per_node=$nGPUs --use_env main.py --model SwiftFormer_L3 --data-path "$IMAGENET_PATH" \
--output_dir SwiftFormer_L3_results
diff --git a/models/swiftformer.py b/models/swiftformer.py
index 7243fae..1b74936 100644
--- a/models/swiftformer.py
+++ b/models/swiftformer.py
@@ -25,9 +25,6 @@ SwiftFormer_depth = {
'l3': [4, 4, 12, 6],
}
-CoreMLConversion = False
-
-
def stem(in_chs, out_chs):
"""
Stem Layer that is implemented by two layers of conv.
@@ -144,8 +141,8 @@ class Mlp(nn.Module):
class EfficientAdditiveAttnetion(nn.Module):
"""
Efficient Additive Attention module for SwiftFormer.
- Input: tensor in shape [B, C, H, W]
- Output: tensor in shape [B, C, H, W]
+ Input: tensor in shape [B, N, D]
+ Output: tensor in shape [B, N, D]
"""
def __init__(self, in_dims=512, token_dim=256, num_heads=2):
@@ -163,26 +160,23 @@ class EfficientAdditiveAttnetion(nn.Module):
query = self.to_query(x)
key = self.to_key(x)
- if not CoreMLConversion:
- # torch.nn.functional.normalize is not supported by the ANE of iPhone devices.
- # Using this layer improves the accuracy by ~0.1-0.2%
- query = torch.nn.functional.normalize(query, dim=-1)
- key = torch.nn.functional.normalize(key, dim=-1)
+ query = torch.nn.functional.normalize(query, dim=-1) #BxNxD
+ key = torch.nn.functional.normalize(key, dim=-1) #BxNxD
- query_weight = query @ self.w_g
- A = query_weight * self.scale_factor
+ query_weight = query @ self.w_g # BxNx1 (BxNxD @ Dx1)
+ A = query_weight * self.scale_factor # BxNx1
- A = A.softmax(dim=-1)
+ A = torch.nn.functional.normalize(A, dim=1) # BxNx1
- G = torch.sum(A * query, dim=1)
+ G = torch.sum(A * query, dim=1) # BxD
G = einops.repeat(
G, "b d -> b repeat d", repeat=key.shape[1]
- )
+ ) # BxNxD
- out = self.Proj(G * key) + query
+ out = self.Proj(G * key) + query #BxNxD
- out = self.final(out)
+ out = self.final(out) # BxNxD
return out
@@ -505,3 +499,4 @@ def SwiftFormer_L3(pretrained=False, **kwargs):
**kwargs)
model.default_cfg = _cfg(crop_pct=0.9)
return model
+
diff --git a/slurm_train.sh b/slurm_train.sh
index 17c4650..85f485a 100644
--- a/slurm_train.sh
+++ b/slurm_train.sh
@@ -15,9 +15,8 @@ srun python main.py --model "$MODEL" \
--data-path "$IMAGENET_PATH" \
--batch-size 128 \
--epochs 300 \
---aa="" --mixup 0 --cutmix 0
-## Note: Disable aa, mixup, and cutmix for SwiftFormer-XS only
+## Note: Disable aa, mixup, and cutmix for SwiftFormer-XS, and disable mixup, and cutmix for SwiftFormer-S.
## By default, this script requests total 16 GPUs on 4 nodes. The batch size per gpu is set to 128,
-## tha sums to 128*16=2048 in total.
\ No newline at end of file
+## tha sums to 128*16=2048 in total.