iformer

Keras InceptionTransformer

---

Summary

• Keras implementation of Github sail-sg/iFormer. Paper PDF 2205.12956 Inception Transformer.
• Model weights ported from official publication.

---

Models

Model	Params	FLOPs	Input	Top1 Acc	Download
IFormerSmall	19.9M	4.88G	224	83.4	iformer_small_224_imagenet.h5
	20.9M	16.29G	384	84.6	iformer_small_384_imagenet.h5
IFormerBase	47.9M	9.44G	224	84.6	iformer_base_224_imagenet.h5
	48.9M	30.86G	384	85.7	iformer_base_384_imagenet.h5
IFormerLarge	86.6M	14.12G	224	84.6	iformer_large_224_imagenet.h5
	87.7M	45.74G	384	85.8	iformer_large_384_imagenet.h5

Usage

from keras_cv_attention_models import iformer

# Will download and load pretrained imagenet weights.
mm = iformer.IFormerSmall(pretrained="imagenet")

# Run prediction
import tensorflow as tf
from tensorflow import keras
from skimage.data import chelsea
imm = keras.applications.imagenet_utils.preprocess_input(chelsea(), mode='torch') # Chelsea the cat
pred = mm(tf.expand_dims(tf.image.resize(imm, mm.input_shape[1:3]), 0)).numpy()
print(keras.applications.imagenet_utils.decode_predictions(pred)[0])
# [('n02124075', 'Egyptian_cat', 0.7471715), ('n02123159', 'tiger_cat', 0.035306472), ...]

Change input resolution.

from keras_cv_attention_models import iformer
mm = iformer.IFormerSmall(input_shape=(512, 393, 3), pretrained="imagenet")
# >>>> Load pretrained from: ~/.keras/models/iformer_small_384_imagenet.h5
# >>>> Reload mismatched weights: 384 -> (512, 393)
# >>>> Reload layer: stack1_positional_embedding
# ...

# Run prediction
from skimage.data import chelsea
preds = mm(mm.preprocess_input(chelsea()))
print(mm.decode_predictions(preds))
# [('n02124075', 'Egyptian_cat', 0.72780704), ('n02123159', 'tiger_cat', 0.11522171), ...]

Verification with PyTorch version

""" PyTorch iformer_small """
sys.path.append('../iFormer/')
sys.path.append('../pytorch-image-models/')  # Needs timm
import torch
from models import inception_transformer

torch_model = inception_transformer.iformer_small(pretrained=True)
_ = torch_model.eval()

""" Keras IFormerSmall """
from keras_cv_attention_models import iformer
mm = iformer.IFormerSmall(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=1e-5) = }")
# np.allclose(torch_out, keras_out, atol=1e-5) = True