208 lines
8.0 KiB
Python
208 lines
8.0 KiB
Python
import numpy as np
|
|
import tensorflow as tf
|
|
from tensorflow.keras.layers import (Conv2D, Dense, DepthwiseConv2D,
|
|
GlobalAveragePooling2D, Layer,
|
|
LayerNormalization)
|
|
|
|
|
|
def drop_path(inputs, drop_prob, is_training):
|
|
# borrowed from https://github.com/rishigami/Swin-Transformer-TF/blob/main/swintransformer/model.py
|
|
if (not is_training) or (drop_prob == 0.):
|
|
return inputs
|
|
|
|
# Compute keep_prob
|
|
keep_prob = 1.0 - drop_prob
|
|
|
|
# Compute drop_connect tensor
|
|
random_tensor = keep_prob
|
|
shape = (tf.shape(inputs)[0],) + (1,) * \
|
|
(len(tf.shape(inputs)) - 1)
|
|
random_tensor += tf.random.uniform(shape, dtype=inputs.dtype)
|
|
binary_tensor = tf.floor(random_tensor)
|
|
output = tf.math.divide(inputs, keep_prob) * binary_tensor
|
|
return output
|
|
|
|
|
|
class DropPath(tf.keras.layers.Layer):
|
|
# borrowed from https://github.com/rishigami/Swin-Transformer-TF/blob/main/swintransformer/model.py
|
|
def __init__(self, drop_prob=None):
|
|
super().__init__()
|
|
self.drop_prob = drop_prob
|
|
|
|
def call(self, x, training=None):
|
|
return drop_path(x, self.drop_prob, training)
|
|
|
|
|
|
class Block(Layer):
|
|
r""" ConvNeXt Block. There are two equivalent implementations:
|
|
(1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W)
|
|
(2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back
|
|
We use (2) as we find it slightly faster in PyTorch
|
|
|
|
Args:
|
|
dim (int): Number of input channels.
|
|
drop_path (float): Stochastic depth rate. Default: 0.0
|
|
layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
|
|
"""
|
|
|
|
def __init__(self, dim, drop_path=0., layer_scale_init_value=1e-6, prefix=''):
|
|
super().__init__()
|
|
self.dwconv = DepthwiseConv2D(
|
|
kernel_size=7, padding='same') # depthwise conv
|
|
self.norm = LayerNormalization(epsilon=1e-6)
|
|
# pointwise/1x1 convs, implemented with linear layers
|
|
self.pwconv1 = Dense(4 * dim)
|
|
self.act = tf.keras.activations.gelu
|
|
self.pwconv2 = Dense(dim)
|
|
self.drop_path = DropPath(drop_path)
|
|
self.dim = dim
|
|
self.layer_scale_init_value = layer_scale_init_value
|
|
self.prefix = prefix
|
|
|
|
def build(self, input_shape):
|
|
self.gamma = tf.Variable(
|
|
initial_value=self.layer_scale_init_value * tf.ones((self.dim)),
|
|
trainable=True,
|
|
name=f'{self.prefix}/gamma')
|
|
self.built = True
|
|
|
|
def call(self, x):
|
|
input = x
|
|
x = self.dwconv(x)
|
|
# x = x.permute(0, 2, 3, 1) # (N, C, H, W) -> (N, H, W, C)
|
|
x = self.norm(x)
|
|
x = self.pwconv1(x)
|
|
x = self.act(x)
|
|
x = self.pwconv2(x)
|
|
if self.gamma is not None:
|
|
x = self.gamma * x
|
|
# x = x.permute(0, 3, 1, 2) # (N, H, W, C) -> (N, C, H, W)
|
|
|
|
x = input + self.drop_path(x)
|
|
return x
|
|
|
|
|
|
class ConvNeXt(tf.keras.Model):
|
|
r""" ConvNeXt
|
|
A Tensorflow keras impl of : `A ConvNet for the 2020s` -
|
|
https://arxiv.org/pdf/2201.03545.pdf
|
|
|
|
Args:
|
|
num_classes (int): Number of classes for classification head. Default: 1000
|
|
depths (tuple(int)): Number of blocks at each stage. Default: [3, 3, 9, 3]
|
|
dims (int): Feature dimension at each stage. Default: [96, 192, 384, 768]
|
|
include_top (bool): whether to add head or just use it as feature extractor. Default: True
|
|
drop_path_rate (float): Stochastic depth rate. Default: 0.
|
|
layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
|
|
head_init_scale (float): Init scaling value for classifier weights and biases. Default: 1.
|
|
"""
|
|
|
|
def __init__(self, num_classes=1000,
|
|
depths=[3, 3, 9, 3], dims=[96, 192, 384, 768], include_top=True,
|
|
drop_path_rate=0., layer_scale_init_value=1e-6, head_init_scale=1.,
|
|
):
|
|
super().__init__()
|
|
self.include_top = include_top
|
|
self.downsample_layers = [] # stem and 3 intermediate downsampling conv layers
|
|
stem = tf.keras.Sequential([
|
|
Conv2D(dims[0], kernel_size=4, strides=4, padding='same'),
|
|
LayerNormalization(epsilon=1e-6)]
|
|
)
|
|
self.downsample_layers.append(stem)
|
|
for i in range(3):
|
|
downsample_layer = tf.keras.Sequential([
|
|
LayerNormalization(epsilon=1e-6),
|
|
Conv2D(dims[i+1], kernel_size=2, strides=2, padding='same')]
|
|
)
|
|
self.downsample_layers.append(downsample_layer)
|
|
|
|
self.stages = [] # 4 feature resolution stages, each consisting of multiple residual blocks
|
|
dp_rates = [x for x in np.linspace(0, drop_path_rate, sum(depths))]
|
|
cur = 0
|
|
for i in range(4):
|
|
stage = tf.keras.Sequential(
|
|
[Block(dim=dims[i], drop_path=dp_rates[cur + j],
|
|
layer_scale_init_value=layer_scale_init_value, prefix=f'block{i}') for j in range(depths[i])]
|
|
)
|
|
self.stages.append(stage)
|
|
cur += depths[i]
|
|
|
|
if self.include_top:
|
|
self.avg = GlobalAveragePooling2D()
|
|
self.norm = LayerNormalization(epsilon=1e-6) # final norm layer
|
|
self.head = Dense(num_classes)
|
|
else:
|
|
self.avg = None
|
|
self.norm = None
|
|
self.head = None
|
|
|
|
def forward_features(self, x):
|
|
for i in range(4):
|
|
x = self.downsample_layers[i](x)
|
|
x = self.stages[i](x)
|
|
return x
|
|
|
|
def call(self, x):
|
|
x = self.forward_features(x)
|
|
if self.include_top:
|
|
x = self.avg(x)
|
|
x = self.norm(x)
|
|
x = self.head(x)
|
|
return x
|
|
|
|
|
|
model_urls = {
|
|
"convnext_tiny_224": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_tiny_1k_224_ema.h5",
|
|
"convnext_small_224": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_small_1k_224_ema.h5",
|
|
"convnext_base_224": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_base_22k_1k_224.h5",
|
|
"convnext_base_384": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_base_22k_1k_384.h5",
|
|
"convnext_large_224": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_large_22k_1k_224.h5",
|
|
"convnext_large_384": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_large_22k_1k_384.h5",
|
|
"convnext_xlarge_224": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_xlarge_22k_1k_224_ema.h5",
|
|
"convnext_xlarge_384": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_xlarge_22k_1k_384_ema.h5",
|
|
}
|
|
|
|
|
|
model_configs = dict(
|
|
convnext_tiny=dict(
|
|
depths=[3, 3, 9, 3],
|
|
dims=[8, 16, 32, 512]
|
|
),
|
|
convnext_small=dict(
|
|
depths=[3, 3, 27, 3],
|
|
dims=[96, 192, 384, 768]
|
|
),
|
|
convnext_base=dict(
|
|
depths=[3, 3, 27, 3],
|
|
dims=[128, 256, 512, 1024]
|
|
),
|
|
convnext_large=dict(
|
|
depths=[3, 3, 27, 3],
|
|
dims=[192, 384, 768, 1536]
|
|
),
|
|
convnext_xlarge=dict(
|
|
depths=[3, 3, 27, 3],
|
|
dims=[256, 512, 1024, 2048]
|
|
),
|
|
)
|
|
|
|
|
|
def create_model(model_name='convnext_tiny_1k', input_shape=(224, 224), num_classes=1000, include_top=True, pretrained=True, use_tpu=False, **kwargs):
|
|
cfg = model_configs['_'.join(model_name.split('_')[:2])]
|
|
net = ConvNeXt(num_classes, cfg['depths'],
|
|
cfg['dims'], include_top, **kwargs)
|
|
net(tf.keras.Input(shape=(*input_shape, 3)))
|
|
if pretrained is True:
|
|
url = model_urls[model_name]
|
|
pretrained_ckpt = tf.keras.utils.get_file(
|
|
f'{model_name}.h5', url, untar=False)
|
|
if use_tpu:
|
|
load_locally = tf.saved_model.LoadOptions(
|
|
experimental_io_device='/job:localhost')
|
|
net.load_weights(
|
|
pretrained_ckpt, options=load_locally, skip_mismatch=True)
|
|
else:
|
|
net.load_weights(pretrained_ckpt, skip_mismatch=True, by_name=True)
|
|
|
|
return net
|