timm
/

Image Classification
timm
PyTorch
Safetensors
rwightman HF staff commited on
Commit
d1104e0
1 Parent(s): f342a2f
Files changed (4) hide show
  1. README.md +171 -0
  2. config.json +40 -0
  3. model.safetensors +3 -0
  4. pytorch_model.bin +3 -0
README.md ADDED
@@ -0,0 +1,171 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ tags:
3
+ - image-classification
4
+ - timm
5
+ library_name: timm
6
+ license: apache-2.0
7
+ datasets:
8
+ - imagenet-1k
9
+ ---
10
+ # Model card for lambda_resnet50ts.a1h_in1k
11
+
12
+ A LambdaNet image classification model (based on ResNet architecture). Trained on ImageNet-1k in `timm` by Ross Wightman.
13
+
14
+ NOTE: this model did not adhere to any specific paper configuration, it was tuned for reasonable training times and reduced frequency of self-attention blocks.
15
+
16
+ Recipe details:
17
+ * Based on [ResNet Strikes Back](https://arxiv.org/abs/2110.00476) `A1` recipe
18
+ * LAMB optimizer
19
+ * Stronger dropout, stochastic depth, and RandAugment than paper `A1` recipe
20
+ * Cosine LR schedule with warmup
21
+
22
+ This model architecture is implemented using `timm`'s flexible [BYOBNet (Bring-Your-Own-Blocks Network)](https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/byobnet.py).
23
+
24
+ BYOB (with BYOANet attention specific blocks) allows configuration of:
25
+ * block / stage layout
26
+ * block-type interleaving
27
+ * stem layout
28
+ * output stride (dilation)
29
+ * activation and norm layers
30
+ * channel and spatial / self-attention layers
31
+
32
+ ...and also includes `timm` features common to many other architectures, including:
33
+ * stochastic depth
34
+ * gradient checkpointing
35
+ * layer-wise LR decay
36
+ * per-stage feature extraction
37
+
38
+
39
+ ## Model Details
40
+ - **Model Type:** Image classification / feature backbone
41
+ - **Model Stats:**
42
+ - Params (M): 21.5
43
+ - GMACs: 5.1
44
+ - Activations (M): 17.5
45
+ - Image size: 256 x 256
46
+ - **Papers:**
47
+ - LambdaNetworks: Modeling Long-Range Interactions Without Attention: https://arxiv.org/abs/2102.08602
48
+ - ResNet strikes back: An improved training procedure in timm: https://arxiv.org/abs/2110.00476
49
+ - **Dataset:** ImageNet-1k
50
+
51
+ ## Model Usage
52
+ ### Image Classification
53
+ ```python
54
+ from urllib.request import urlopen
55
+ from PIL import Image
56
+ import timm
57
+
58
+ img = Image.open(urlopen(
59
+ 'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
60
+ ))
61
+
62
+ model = timm.create_model('lambda_resnet50ts.a1h_in1k', pretrained=True)
63
+ model = model.eval()
64
+
65
+ # get model specific transforms (normalization, resize)
66
+ data_config = timm.data.resolve_model_data_config(model)
67
+ transforms = timm.data.create_transform(**data_config, is_training=False)
68
+
69
+ output = model(transforms(img).unsqueeze(0)) # unsqueeze single image into batch of 1
70
+
71
+ top5_probabilities, top5_class_indices = torch.topk(output.softmax(dim=1) * 100, k=5)
72
+ ```
73
+
74
+ ### Feature Map Extraction
75
+ ```python
76
+ from urllib.request import urlopen
77
+ from PIL import Image
78
+ import timm
79
+
80
+ img = Image.open(urlopen(
81
+ 'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
82
+ ))
83
+
84
+ model = timm.create_model(
85
+ 'lambda_resnet50ts.a1h_in1k',
86
+ pretrained=True,
87
+ features_only=True,
88
+ )
89
+ model = model.eval()
90
+
91
+ # get model specific transforms (normalization, resize)
92
+ data_config = timm.data.resolve_model_data_config(model)
93
+ transforms = timm.data.create_transform(**data_config, is_training=False)
94
+
95
+ output = model(transforms(img).unsqueeze(0)) # unsqueeze single image into batch of 1
96
+
97
+ for o in output:
98
+ # print shape of each feature map in output
99
+ # e.g.:
100
+ # torch.Size([1, 64, 128, 128])
101
+ # torch.Size([1, 256, 64, 64])
102
+ # torch.Size([1, 512, 32, 32])
103
+ # torch.Size([1, 1024, 16, 16])
104
+ # torch.Size([1, 2048, 8, 8])
105
+
106
+ print(o.shape)
107
+ ```
108
+
109
+ ### Image Embeddings
110
+ ```python
111
+ from urllib.request import urlopen
112
+ from PIL import Image
113
+ import timm
114
+
115
+ img = Image.open(urlopen(
116
+ 'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
117
+ ))
118
+
119
+ model = timm.create_model(
120
+ 'lambda_resnet50ts.a1h_in1k',
121
+ pretrained=True,
122
+ num_classes=0, # remove classifier nn.Linear
123
+ )
124
+ model = model.eval()
125
+
126
+ # get model specific transforms (normalization, resize)
127
+ data_config = timm.data.resolve_model_data_config(model)
128
+ transforms = timm.data.create_transform(**data_config, is_training=False)
129
+
130
+ output = model(transforms(img).unsqueeze(0)) # output is (batch_size, num_features) shaped tensor
131
+
132
+ # or equivalently (without needing to set num_classes=0)
133
+
134
+ output = model.forward_features(transforms(img).unsqueeze(0))
135
+ # output is unpooled, a (1, 2048, 8, 8) shaped tensor
136
+
137
+ output = model.forward_head(output, pre_logits=True)
138
+ # output is a (1, num_features) shaped tensor
139
+ ```
140
+
141
+ ## Model Comparison
142
+ Explore the dataset and runtime metrics of this model in timm [model results](https://github.com/huggingface/pytorch-image-models/tree/main/results).
143
+
144
+ ## Citation
145
+ ```bibtex
146
+ @misc{rw2019timm,
147
+ author = {Ross Wightman},
148
+ title = {PyTorch Image Models},
149
+ year = {2019},
150
+ publisher = {GitHub},
151
+ journal = {GitHub repository},
152
+ doi = {10.5281/zenodo.4414861},
153
+ howpublished = {\url{https://github.com/huggingface/pytorch-image-models}}
154
+ }
155
+ ```
156
+ ```bibtex
157
+ @article{Bello2021LambdaNetworksML,
158
+ title={LambdaNetworks: Modeling Long-Range Interactions Without Attention},
159
+ author={Irwan Bello},
160
+ journal={ArXiv},
161
+ year={2021},
162
+ volume={abs/2102.08602}
163
+ }
164
+ ```
165
+ ```bibtex
166
+ @inproceedings{wightman2021resnet,
167
+ title={ResNet strikes back: An improved training procedure in timm},
168
+ author={Wightman, Ross and Touvron, Hugo and Jegou, Herve},
169
+ booktitle={NeurIPS 2021 Workshop on ImageNet: Past, Present, and Future}
170
+ }
171
+ ```
config.json ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture": "lambda_resnet50ts",
3
+ "num_classes": 1000,
4
+ "num_features": 2048,
5
+ "pretrained_cfg": {
6
+ "tag": "a1h_in1k",
7
+ "custom_load": false,
8
+ "input_size": [
9
+ 3,
10
+ 256,
11
+ 256
12
+ ],
13
+ "min_input_size": [
14
+ 3,
15
+ 128,
16
+ 128
17
+ ],
18
+ "fixed_input_size": false,
19
+ "interpolation": "bicubic",
20
+ "crop_pct": 0.95,
21
+ "crop_mode": "center",
22
+ "mean": [
23
+ 0.485,
24
+ 0.456,
25
+ 0.406
26
+ ],
27
+ "std": [
28
+ 0.229,
29
+ 0.224,
30
+ 0.225
31
+ ],
32
+ "num_classes": 1000,
33
+ "pool_size": [
34
+ 8,
35
+ 8
36
+ ],
37
+ "first_conv": "stem.conv1.conv",
38
+ "classifier": "head.fc"
39
+ }
40
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bcb80eaa77910c72c9c0ff9e83f99d550270fc23ab0ccac69d01e35cb269cdf8
3
+ size 86405518
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aa21c2821d9d6fa9c5af9629e78b54dab1b47a4f61497ed728fc7759e4a0fbc1
3
+ size 86506541