inceptionnext

Keras InceptionNeXt

---

Summary

• Keras implementation of Github sail-sg/inceptionnext. Paper PDF 2303.16900 InceptionNeXt: When Inception Meets ConvNeXt.
• Model weights ported from official publication.

---

Models

Model	Params	FLOPs	Input	Top1 Acc	Download
InceptionNeXtTiny	28.05M	4.21G	224	82.3	inceptionnext_tiny_imagenet.h5
InceptionNeXtSmall	49.37M	8.39G	224	83.5	inceptionnext_small_imagenet.h5
InceptionNeXtBase	86.67M	14.88G	224	84.0	inceptionnext_base_224_imagenet.h5
	86.67M	43.73G	384	85.2	inceptionnext_base_384_imagenet.h5

Usage

from keras_cv_attention_models import inceptionnext

# Will download and load pretrained imagenet weights.
model = inceptionnext.InceptionNeXtTiny(pretrained="imagenet")

# Run prediction
from skimage.data import chelsea # Chelsea the cat
preds = model(model.preprocess_input(chelsea()))
print(model.decode_predictions(preds))
# [('n02124075', 'Egyptian_cat', 0.8221698), ('n02123159', 'tiger_cat', 0.019049658), ...]

Use dynamic input resolution by set input_shape=(None, None, 3).

from keras_cv_attention_models import inceptionnext
model = inceptionnext.InceptionNeXtTiny(input_shape=(None, None, 3), num_classes=0)
# >>>> Load pretrained from: ~/.keras/models/inceptionnext_tiny_imagenet.h5
print(model.output_shape)
# (None, None, None, 768)

print(model(np.ones([1, 223, 123, 3])).shape)
# (1, 6, 3, 768)
print(model(np.ones([1, 32, 526, 3])).shape)
# (1, 1, 16, 768)

Using PyTorch backend by set KECAM_BACKEND='torch' environment variable.

os.environ['KECAM_BACKEND'] = 'torch'

from keras_cv_attention_models import inceptionnext
model = inceptionnext.InceptionNeXtTiny(input_shape=(None, None, 3), num_classes=0)
# >>>> Using PyTorch backend
# >>>> Aligned input_shape: [3, None, None]
# >>>> Load pretrained from: ~/.keras/models/inceptionnext_tiny_imagenet.h5
print(model.output_shape)
# (None, 768, None, None)

import torch
print(model(torch.ones([1, 3, 223, 123])).shape)
# (1, 768, 6, 3 )
print(model(torch.ones([1, 3, 32, 526])).shape)
# (1, 768, 1, 16)

Verification with PyTorch version

""" PyTorch inceptionnext_tiny """
sys.path.append('../inceptionnext/')
sys.path.append('../pytorch-image-models/')  # Needs timm
import torch
from models import inceptionnext as inceptionnext_torch

torch_model = inceptionnext_torch.inceptionnext_tiny(pretrained=True)
_ = torch_model.eval()

""" Keras InceptionNeXtTiny """
from keras_cv_attention_models import inceptionnext
mm = inceptionnext.InceptionNeXtTiny(pretrained="imagenet", classifier_activation=None)

""" Verification """
inputs = np.random.uniform(size=(1, *mm.input_shape[1:3], 3)).astype("float32")
torch_out = torch_model(torch.from_numpy(inputs).permute(0, 3, 1, 2)).detach().numpy()
keras_out = mm(inputs).numpy()
print(f"{np.allclose(torch_out, keras_out, atol=5e-5) = }")
# np.allclose(torch_out, keras_out, atol=5e-5) = True