vanillanet

Keras VanillaNet

---

Summary

• Keras implementation of Github huawei-noah/VanillaNet. Paper PDF 2305.12972 VanillaNet: the Power of Minimalism in Deep Learning.
• Model weights ported from official publication.

---

Models

Model	Params	FLOPs	Input	Top1 Acc	Download
VanillaNet5	22.33M	8.46G	224	72.49	vanillanet_5_imagenet.h5
- deploy=True	15.52M	5.17G	224	72.49	vanillanet_5_deploy_imagenet.h5
VanillaNet6	56.12M	10.11G	224	76.36	vanillanet_6_imagenet.h5
- deploy=True	32.51M	6.00G	224	76.36	vanillanet_6_deploy_imagenet.h5
VanillaNet7	56.67M	11.84G	224	77.98	vanillanet_7_imagenet.h5
- deploy=True	32.80M	6.90G	224	77.98	vanillanet_7_deploy_imagenet.h5
VanillaNet8	65.18M	13.50G	224	79.13	vanillanet_8_imagenet.h5
- deploy=True	37.10M	7.75G	224	79.13	vanillanet_8_deploy_imagenet.h5
VanillaNet9	73.68M	15.17G	224	79.87	vanillanet_9_imagenet.h5
- deploy=True	41.40M	8.59G	224	79.87	vanillanet_9_deploy_imagenet.h5
VanillaNet10	82.19M	16.83G	224	80.57	vanillanet_10_imagenet.h5
- deploy=True	45.69M	9.43G	224	80.57	vanillanet_10_deploy_imagenet.h5
VanillaNet11	90.69M	18.49G	224	81.08
- deploy=True	50.00M	10.27G	224	81.08
VanillaNet12	99.20M	20.16G	224	81.55
- deploy=True	54.29M	11.11G	224	81.55
VanillaNet13	107.7M	21.82G	224	82.05
- deploy=True	58.59M	11.96G	224	82.05

Usage

from keras_cv_attention_models import vanillanet, test_images

# Will download and load pretrained imagenet weights.
model = vanillanet.VanillaNet5(pretrained="imagenet")
# >>>> Load pretrained from: ~/.keras/models/vanillanet_5_imagenet.h5

# Run prediction
preds = model(model.preprocess_input(test_images.cat()))
print(model.decode_predictions(preds))
# [('n02124075', 'Egyptian_cat', 0.9908214), ('n02123045', 'tabby', 0.008346258), ...]

Set deploy=True for a fused model, and keep output same. Evaluation only, not good for training.

from keras_cv_attention_models import vanillanet, test_images
model = vanillanet.VanillaNet5(deploy=True, pretrained="imagenet")
# >>>> Load pretrained from: ~/.keras/models/vanillanet_5_deploy_imagenet.h5

preds = model(model.preprocess_input(test_images.cat()))
print(model.decode_predictions(preds))
# [('n02124075', 'Egyptian_cat', 0.9908214), ('n02123045', 'tabby', 0.008346282), ...]

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

from keras_cv_attention_models import vanillanet
model = vanillanet.VanillaNet6(input_shape=(None, None, 3), num_classes=0)
# >>>> Load pretrained from: ~/.keras/models/vanillanet_6_imagenet.h5
print(model.output_shape)
# (None, None, None, 4096)

print(model(np.ones([1, 223, 123, 3])).shape)
# (1, 6, 3, 4096)
print(model(np.ones([1, 32, 526, 3])).shape)
# (1, 1, 16, 4096)

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

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

from keras_cv_attention_models import vanillanet, test_images
model = vanillanet.VanillaNet6(deploy=True)
# >>>> Using PyTorch backend
# >>>> Aligned input_shape: [3, 224, 224]
# >>>> Load pretrained from: ~/.keras/models/vanillanet_6_deploy_imagenet.h5

preds = model(model.preprocess_input(test_images.cat()))
print(model.decode_predictions(preds))
# [('n02124075', 'Egyptian_cat', 0.9729379), ('n02123045', 'tabby', 0.008538764), ...]

Train and Deploy

• Currently only works for Tensorflow
• set_leaky_relu_alpha is introduced in the paper, used for changing activation alpha while training. Say epochs=300, decay_epochs=100, then alpha = (current_epoch / decay_epochs) if current_epoch < decay_epochs else 1.
• switch_to_deploy is used for converting a custom pretrained model to deploy version. The process is fuse conv bn -> remove leaky_relu with alpha=1 -> fuse 2 sequential conv.
• Training process could be:

  from keras_cv_attention_models import vanillanet
  
  """ Create a `deploy=False` model for custom training """
  model = vanillanet.VanillaNet5(deploy=False, pretrained="imagenet")
  print(f"{model.count_params() = }")
  # model.count_params() = 22369712
  # ...
  
  """ Run some training step with changing leaky_relu alpha """
  decay_epochs = 100
  current_epoch = 10
  model.set_leaky_relu_alpha(alpha=(current_epoch / decay_epochs) if current_epoch < decay_epochs else 1)
  
  # Or set in actual callbacks
  # on_epoch_end = lambda epoch, logs: model.set_leaky_relu_alpha(alpha=(epoch / decay_epochs) if epoch < decay_epochs else 1)
  # act_learn = keras.callbacks.LambdaCallback(on_epoch_end=on_epoch_end)
  # model.fit(..., callbacks=[..., act_learn])
  # ...
  
  """ Fuse to deploy model """
  model.set_leaky_relu_alpha(alpha=1)  # This should already set when training reach `decay_epochs`
  tt = model.switch_to_deploy()
  deploy_save_name = model.name + "_deploy.h5"
  tt.save(deploy_save_name)
  
  """ Further evaluation with deploy=True """
  bb = vanillanet.VanillaNet5(deploy=True, pretrained=deploy_save_name)
  print(f"{bb.count_params() = }")
  # bb.count_params() = 15523304
  print(f"{np.allclose(model(tf.ones([1, 224, 224, 3])), bb(tf.ones([1, 224, 224, 3])), atol=1e-5) = }")
  # np.allclose(model(tf.ones([1, 224, 224, 3])), bb(tf.ones([1, 224, 224, 3])), atol=1e-5) = True

Verification with PyTorch version

""" PyTorch vanillanet_6 """
sys.path.append('../VanillaNet/')
sys.path.append('../pytorch-image-models/')  # Needs timm
import torch
from models import vanillanet as torch_vanillanet

torch_model = torch_vanillanet.vanillanet_6()
_ = torch_model.eval()
ss = torch.load('vanillanet_6.pth'.format(id), map_location='cpu')
torch_model.load_state_dict(ss['model_ema'])

""" Keras VanillaNet6 """
from keras_cv_attention_models import vanillanet
mm = vanillanet.VanillaNet6(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