Ross Wightman commited on
Commit
10d04ca
1 Parent(s): 42234a7

Initial commit w/ README, config, weights

Browse files
Files changed (3) hide show
  1. README.md +103 -0
  2. config.json +32 -0
  3. pytorch_model.bin +3 -0
README.md ADDED
@@ -0,0 +1,103 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ tags:
3
+ - image-classification
4
+ - timm
5
+ - normalization-free
6
+ - efficient-channel-attention
7
+ license: apache-2.0
8
+ datasets:
9
+ - imagenet
10
+ inference: false
11
+ ---
12
+
13
+ # ECA-NFNet-L0
14
+
15
+ Pretrained model on [ImageNet](http://www.image-net.org/), this is a variant of the [NFNet (Normalization Free)](https://arxiv.org/abs/2102.06171) model family.
16
+
17
+ ## Model description
18
+
19
+ This model variant was slimmed down from the original F0 variant in the paper for improved runtime characteristics (throughput, memory use) in PyTorch, on a GPU accelerator. It utilizes [Efficient Channel Attention (ECA)](https://arxiv.org/abs/1910.03151) instead of Squeeze-Excitation. It also features SiLU activations instead of the usual GELU.
20
+
21
+ Like other models in the NF family, this model contains no normalization layers (batch, group, etc). The models make use of [Weight Standardized](https://arxiv.org/abs/1903.10520) convolutions with additional scaling values in lieu of normalization layers.
22
+
23
+ ## Intended uses & limitations
24
+ You can use the raw model to classify images along the 1,000 ImageNet labels, but you can also change its head
25
+ to fine-tune it on a downstream task (another classification task with different labels, image segmentation or
26
+ object detection, to name a few).
27
+
28
+ ### How to use
29
+ You can use this model with the usual factory method in `timm`:
30
+ ```python
31
+ import PIL
32
+ import timm
33
+ import torch
34
+
35
+ model = timm.create_model("timm/eca_nfnet_l0")
36
+
37
+ config = model.default_cfg
38
+ img_size = config["test_input_size"][-1] if "test_input_size" in config else config["input_size"][-1]
39
+ transform = timm.data.transforms_factory.transforms_imagenet_eval(
40
+ img_size=img_size,
41
+ interpolation=config["interpolation"],
42
+ mean=config["mean"],
43
+ std=config["std"],
44
+ crop_pct=config["crop_pct"],
45
+ )
46
+
47
+ img = PIL.Image.open(path_to_an_image)
48
+ img = img.convert("RGB")
49
+ input_tensor = transform(cat_img)
50
+ input_tensor = input_tensor.unsqueeze(0)
51
+ # ^ batch size = 1
52
+ with torch.no_grad():
53
+ output = model(input_tensor)
54
+ probs = output.squeeze(0).softmax(dim=0)
55
+ ```
56
+
57
+ ### Limitations and bias
58
+ The training images in the dataset are usually photos clearly representing one of the 1,000 labels. The model will
59
+ probably not generalize well on drawings or images containing multiple objects with different labels.
60
+ The training images in the dataset come mostly from the US (45.4%) and Great Britain (7.6%). As such the model or
61
+ models created by fine-tuning this model will work better on images picturing scenes from these countries (see
62
+ [this paper](https://arxiv.org/abs/1906.02659) for examples).
63
+ More generally, [recent research](https://arxiv.org/abs/2010.15052) has shown that even models trained in an
64
+ unsupervised fashion on ImageNet (i.e. without using the labels) will pick up racial and gender bias represented in
65
+ the training images.
66
+
67
+ ## Training data
68
+ This model was pretrained on [ImageNet](http://www.image-net.org/), a dataset consisting of 14 millions of
69
+ hand-annotated images with 1,000 categories.
70
+
71
+ ## Training procedure
72
+ For stability during training it is highly recommended to train all NFNet variants with gradient clipping enabled. This model was trained with an Adaptive Gradient Clipping (AGC) factor of 0.015 as described in [the paper](https://arxiv.org/abs/2102.06171). Similar to the paper, a cosine learning rate decay was employed using SGD w/ nesterov. Moderate to heavy augmentation ([RandAugment](https://arxiv.org/abs/1909.13719)) and regularization (dropout, stochastic depth) is recommended for training.
73
+
74
+ ### Preprocessing
75
+ The images are resized using bicubic interpolation to 288x288 and normalized with the usual ImageNet statistics.
76
+
77
+ ## Evaluation results
78
+ This model has a top1-accuracy of 82.6% and a top-5 accuracy of 96.5% on the ImageNet evaluation set.
79
+
80
+ ### BibTeX entry and citation info
81
+
82
+ NFNet model architecture:
83
+ ```bibtex
84
+ @article{brock2021high,
85
+ author={Andrew Brock and Soham De and Samuel L. Smith and Karen Simonyan},
86
+ title={High-Performance Large-Scale Image Recognition Without Normalization},
87
+ journal={arXiv preprint arXiv:2102.06171},
88
+ year={2021}
89
+ }
90
+ ```
91
+
92
+ L0 model variant & pretraining:
93
+ ```bibtex
94
+ @misc{rw2019timm,
95
+ author = {Ross Wightman},
96
+ title = {PyTorch Image Models},
97
+ year = {2019},
98
+ publisher = {GitHub},
99
+ journal = {GitHub repository},
100
+ doi = {10.5281/zenodo.4414861},
101
+ howpublished = {\url{https://github.com/rwightman/pytorch-image-models}}
102
+ }
103
+ ```
config.json ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "num_classes": 1000,
3
+ "input_size": [
4
+ 3,
5
+ 224,
6
+ 224
7
+ ],
8
+ "pool_size": [
9
+ 7,
10
+ 7
11
+ ],
12
+ "crop_pct": 1.0,
13
+ "interpolation": "bicubic",
14
+ "mean": [
15
+ 0.485,
16
+ 0.456,
17
+ 0.406
18
+ ],
19
+ "std": [
20
+ 0.229,
21
+ 0.224,
22
+ 0.225
23
+ ],
24
+ "first_conv": "stem.conv1",
25
+ "classifier": "head.fc",
26
+ "test_input_size": [
27
+ 3,
28
+ 288,
29
+ 288
30
+ ],
31
+ "architecture": "eca_nfnet_l0"
32
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:57290e127ff515f3cfafce1f69af2aa10cb7ecc8db7eafa0ecac3a906c198072
3
+ size 96611441