Views
No views yet
RmsNorm on top of the encoder patch tokens, ready for fine-tuning a linear head. The architecture is a custom ViT with 2D RoPE, gated MLP (GELU-tanh), RMS normalization on Q/K/V, and a 4-norm sandwich block layout.transformers Gemma4VisionModel implementation bit-for-bit on matching inputs.[0, 1]-range pixel tensors — the model maps them internally to [-1, 1] via 2 * (x - 0.5). The pretrained_cfg therefore declares mean=(0, 0, 0) / std=(1, 1, 1) to disable normalization in timm's standard transform pipeline and avoid double-normalization.timm.data.naflex_loader / --naflex-loader in train.py). Both raw (B, C, H, W) images and pre-patchified (B, N, ...) tensors (plus patch_coord / patch_valid) are accepted — so the same model can be trained or fine-tuned with dynamic batch sizing that preserves native aspect ratios at flexible resolutions, or run at a fixed square input size.1from urllib.request import urlopen
2from PIL import Image
3import timm
4
5img = Image.open(urlopen(
6 'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
7))
8
9model = timm.create_model('gemma4_vit_167m.gemma4_e4b_it', pretrained=True)
10model = model.eval()
11
12# get model specific transforms (normalization, resize)
13data_config = timm.data.resolve_model_data_config(model)
14transforms = timm.data.create_transform(**data_config, is_training=False)
15
16output = model(transforms(img).unsqueeze(0)) # unsqueeze single image into batch of 1
17
18top5_probabilities, top5_class_indices = torch.topk(output.softmax(dim=1) * 100, k=5)1from urllib.request import urlopen
2from PIL import Image
3import timm
4
5img = Image.open(urlopen(
6 'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
7))
8
9model = timm.create_model(
10 'gemma4_vit_167m.gemma4_e4b_it',
11 pretrained=True,
12 features_only=True,
13)
14model = model.eval()
15
16# get model specific transforms (normalization, resize)
17data_config = timm.data.resolve_model_data_config(model)
18transforms = timm.data.create_transform(**data_config, is_training=False)
19
20output = model(transforms(img).unsqueeze(0)) # unsqueeze single image into batch of 1
21
22for o in output:
23 # print shape of each feature map in output
24 # e.g.:
25 # torch.Size([1, 768, 48, 48])
26 # torch.Size([1, 768, 48, 48])
27 # torch.Size([1, 768, 48, 48])
28
29 print(o.shape)1from urllib.request import urlopen
2from PIL import Image
3import timm
4
5img = Image.open(urlopen(
6 'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
7))
8
9model = timm.create_model(
10 'gemma4_vit_167m.gemma4_e4b_it',
11 pretrained=True,
12 num_classes=0, # remove classifier nn.Linear
13)
14model = model.eval()
15
16# get model specific transforms (normalization, resize)
17data_config = timm.data.resolve_model_data_config(model)
18transforms = timm.data.create_transform(**data_config, is_training=False)
19
20output = model(transforms(img).unsqueeze(0)) # output is (batch_size, num_features) shaped tensor
21
22# or equivalently (without needing to set num_classes=0)
23
24output = model.forward_features(transforms(img).unsqueeze(0))
25# output is unpooled, a (1, 2304, 768) shaped tensor
26
27output = model.forward_head(output, pre_logits=True)
28# output is a (1, num_features) shaped tensor1@misc{gemma4_2025,
2 title={Gemma 4},
3 author={{Gemma Team, Google DeepMind}},
4 year={2025},
5 howpublished={\url{https://ai.google.dev/gemma/docs/core/model_card_4}}
6}1@misc{rw2019timm,
2 author = {Ross Wightman},
3 title = {PyTorch Image Models},
4 year = {2019},
5 publisher = {GitHub},
6 journal = {GitHub repository},
7 doi = {10.5281/zenodo.4414861},
8 howpublished = {\url{https://github.com/huggingface/pytorch-image-models}}
9}