Source code for akida_models.transformers.model_vit

#!/usr/bin/env python
# ******************************************************************************
# Copyright 2022 Brainchip Holdings Ltd.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ******************************************************************************
"""
ViT model definition.
Inspired from https://github.com/faustomorales/vit-keras/blob/master/vit_keras/vit.py.
"""

__all__ = ["vit_imagenet", "vit_ti16", "bc_vit_ti16", "bc_vit_ti16_imagenet_pretrained", "vit_s16",
           "vit_s32", "vit_b16", "vit_b32", "vit_l16", "vit_l32", "CONFIG_TI", "CONFIG_S",
           "CONFIG_B", "CONFIG_L", "apply_embedding_weights"]

import keras
import warnings
import numpy as np
import scipy as sp
import typing_extensions as tx

from quantizeml.layers import AddPositionEmbs, ClassToken, ExtractToken
from quantizeml import load_model

from ..layer_blocks import transformer_block, norm_to_layer
from ..utils import fetch_file
from ..model_io import get_model_path


ConfigDict = tx.TypedDict(
    "ConfigDict",
    {
        "dropout": float,
        "mlp_dim": int,
        "num_heads": int,
        "num_blocks": int,
        "hidden_size": int,
    },
)

CONFIG_TI: ConfigDict = {
    "dropout": 0.0,
    "mlp_dim": 768,
    "num_heads": 3,
    "num_blocks": 12,
    "hidden_size": 192,
}

CONFIG_S: ConfigDict = {
    "dropout": 0.1,
    "mlp_dim": 1536,
    "num_heads": 6,
    "num_blocks": 12,
    "hidden_size": 384,
}

CONFIG_B: ConfigDict = {
    "dropout": 0.1,
    "mlp_dim": 3072,
    "num_heads": 12,
    "num_blocks": 12,
    "hidden_size": 768,
}

CONFIG_L: ConfigDict = {
    "dropout": 0.1,
    "mlp_dim": 4096,
    "num_heads": 16,
    "num_blocks": 24,
    "hidden_size": 1024,
}


def apply_embedding_weights(target_layer, source_weights, num_patches, num_tokens=1):
    """Apply embedding weights to a target layer.

    Args:
        target_layer (:obj:`keras.Layer`): The target layer to which weights will be applied.
        source_weights (list of :obj:`np.array`): The source weights.
        num_patches (int or tuple): Number of patches in xy spatial dimension, extracted from
            embedding layer.
        num_tokens (int, optional): Number of tokens. Defaults to 1.
    """
    assert isinstance(source_weights, list), "source_weights must be a list of numpy arrays"
    expected_shape = target_layer.weights[0].shape

    # number of patches constrains
    if isinstance(num_patches, int):
        num_patches = (num_patches, num_patches)
    assert len(num_patches) == 2, "num_patches must contain at most two values"
    assert all(x > 0 for x in num_patches), "every patch must be greater than 0"

    if expected_shape != source_weights[0].shape:
        token, grid = source_weights[0][0, :num_tokens], source_weights[0][0, num_tokens:]
        sin = int(np.sqrt(grid.shape[0]))
        sout_x = num_patches[0]
        sout_y = num_patches[1]
        warnings.warn(
            "Resizing position embeddings from " f"{sin}, {sin} to {sout_x}, {sout_y}",
            UserWarning,
        )
        zoom = (sout_y / sin, sout_x / sin, 1)
        grid = sp.ndimage.zoom(grid.reshape(sin, sin, -1), zoom, order=1).reshape(
            sout_x * sout_y, -1
        )
        new_weights = np.concatenate([token, grid], axis=0)[np.newaxis]
        source_weights = [new_weights] + source_weights[1:]
    target_layer.set_weights(source_weights)


[docs]def vit_imagenet(input_shape, patch_size, num_blocks, hidden_size, num_heads, name, mlp_dim, classes=1000, dropout=0.1, include_top=True, norm='LN', last_norm='LN', softmax='softmax', act="GeLU"): """Instantiates the ViT architecture. The Vision Transformer (ViT) is a model for image classification that employs a Transformer-like architecture over patches of the image. An image is split into fixed-size patches, each of them are then linearly embedded, position embeddings are added, and the resulting sequence of vectors are fed to a standard Transformer encoder. Please refer to https://arxiv.org/abs/2010.11929 for further details. Note: input preprocessing is included as part of the model (as a Rescaling layer). This model expects inputs to be float tensors of pixels with values in the [0, 255] range. Args: input_shape (tuple): image shape tuple patch_size (int): the size of each patch (must fit evenly in image size) num_blocks (int): the number of transformer blocks to use. hidden_size (int): the number of filters to use num_heads (int): the number of transformer heads name (str): the model name mlp_dim (int): the number of dimensions for the MLP output in the transformers. classes (int, optional): number of classes to classify images into, only to be specified if `include_top` is True. Defaults to 1000. dropout (float, optional): fraction of the units to drop for dense layers. Defaults to 0.1. include_top (bool, optional): whether to include the final classifier head. If False, the output will correspond to that of the transformer. Defaults to True. norm (str, optional): string that values in ['LN', 'GN1', 'BN', 'LMN'] and that allows to choose from LayerNormalization, GroupNormalization(groups=1, ...), BatchNormalization or LayerMadNormalization layers respectively in the model. Defaults to 'LN'. last_norm (str, optional): string that values in ['LN', 'BN'] and that allows to choose from LayerNormalization or BatchNormalization in the classifier network. Defaults to 'LN'. softmax (str, optional): string with values in ['softmax', 'softmax2'] that allows to choose between softmax and softmax2 in MHA. Defaults to 'softmax'. act (str, optional): string that values in ['GeLU', 'ReLUx', 'swish'] and that allows to choose from GeLU, ReLUx or swish activation in MLP block. Defaults to 'GeLU'. """ assert (input_shape[0] % patch_size == 0) and ( input_shape[1] % patch_size == 0), "image size must be a multiple of patch_size" if last_norm not in ('LN', 'BN'): raise NotImplementedError("last_norm should be in ['LN', 'BN']" f"but received {norm}.") # Normalize image adding rescaling layer x = keras.layers.Input(shape=input_shape, name="input") y = keras.layers.Rescaling(1 / 127.5, -1, name="Rescale")(x) # Build model y = keras.layers.Conv2D( filters=hidden_size, kernel_size=patch_size, strides=patch_size, padding="valid", name="Embedding", kernel_initializer=keras.initializers.TruncatedNormal(stddev=0.02), bias_initializer="zeros", )(y) y = keras.layers.Reshape((y.shape[1] * y.shape[2], hidden_size))(y) y = ClassToken(name="ClassToken")(y) y = AddPositionEmbs(name="Transformer/PosEmbed")(y) for n in range(num_blocks): y, _ = transformer_block( y, num_heads=num_heads, hidden_size=hidden_size, mlp_dim=mlp_dim, dropout=dropout, name=f"Transformer/EncoderBlock_{n}", norm=norm, softmax=softmax, mlp_act=act, ) # Include classification head if include_top: y = norm_to_layer(last_norm)( epsilon=1e-6, name="Transformer/EncoderNorm")(y) y = ExtractToken(token=0, name="ExtractToken")(y) y = keras.layers.Dense(classes, name="Head")(y) return keras.models.Model(inputs=x, outputs=y, name=name)
[docs]def vit_ti16(input_shape=(224, 224, 3), classes=1000, norm='LN', last_norm='LN', softmax='softmax', act='GeLU', include_top=True): """Instantiates the ViT-Tiny 16 architecture; that is a ViT architecture with 3 attention heads, 12 blocks and a patch size of 16. Args: input_shape (tuple, optional): input shape. Defaults to (224, 224, 3). classes (int, optional): number of classes. Defaults to 1000. norm (str, optional): string that values in ['LN', 'GN1', 'BN', 'LMN'] and that allows to choose from LayerNormalization, GroupNormalization(groups=1, ...), BatchNormalization or LayerMadNormalization layers respectively in the model. Defaults to 'LN'. last_norm (str, optional): string that values in ['LN', 'BN'] and that allows to choose from LayerNormalization or BatchNormalization in the classifier network. Defaults to 'LN'. softmax (str, optional): string with values in ['softmax', 'softmax2'] that allows to choose between softmax and softmax2 in attention block. Defaults to 'softmax'. act (str, optional): string that values in ['GeLU', 'ReLUx', 'swish'] and that allows to choose from GeLU, ReLUx or swish activation inside MLP. Defaults to 'GeLU'. include_top (bool, optional): whether to include the final classifier network. Defaults to True. Returns: keras.Model: the requested model """ return vit_imagenet( name="vit-tiny", patch_size=16, input_shape=input_shape, classes=classes, norm=norm, last_norm=last_norm, act=act, softmax=softmax, include_top=include_top, **CONFIG_TI )
[docs]def bc_vit_ti16(input_shape=(224, 224, 3), classes=1000, include_top=True, num_blocks=12): """Instantiates the ViT-Tiny 16 architecture adapted for implementation on hardware, that is: - LayerNormalization replaced by LayerMadNormalization, - GeLU replaced by ReLU8 activations, - Softmax replaced by shiftmax. Args: input_shape (tuple, optional): input shape. Defaults to (224, 224, 3). classes (int, optional): number of classes. Defaults to 1000. include_top (bool, optional): whether to include the final classifier network. Defaults to True. num_blocks (int, optional): the number of transformer blocks to use. Defaults to 12. Returns: keras.Model: the requested model """ config_ti = CONFIG_TI.copy() config_ti["num_blocks"] = num_blocks return vit_imagenet( name="vit-tiny", patch_size=16, input_shape=input_shape, classes=classes, norm="LMN", last_norm="BN", softmax="softmax2", act="ReLU8", include_top=include_top, **config_ti )
[docs]def bc_vit_ti16_imagenet_pretrained(quantized=True): """ Helper method to retrieve a ViT-Tiny 16 model adapted for implementation on hardware, that is: - LayerNormalization replaced by LayerMadNormalization, - GeLU replaced by ReLU8 activations, - Softmax replaced by shiftmax, and that was trained on ImageNet dataset. Args: quantized (bool, optional): a boolean indicating whether the model should be loaded quantized or not. Defaults to True. Returns: keras.Model: a Keras Model instance """ if quantized: model_name_v2 = 'bc_vit_ti16_224_i8_w8_a8.h5' file_hash_v2 = 'b147503b81991a5fbfabdd704fd17eb94fadb9faac1bb664bd195a16292d2e7f' else: model_name_v2 = 'bc_vit_ti16_224.h5' file_hash_v2 = '7ae299f698abc545e21338b129dfbc0ff97ee9b994319e53e8605d34e3ee1f82' model_path, model_name, file_hash = get_model_path("vit", model_name_v2=model_name_v2, file_hash_v2=file_hash_v2) model_path = fetch_file(model_path, fname=model_name, file_hash=file_hash, cache_subdir='models') return load_model(model_path)
[docs]def vit_s16(input_shape=(224, 224, 3), classes=1000, include_top=True): """Instantiates the ViT-Small 16 architecture; that is a ViT architecture with 6 attention heads, 12 blocks and a patch size of 16. Args: input_shape (tuple, optional): input shape. Defaults to (224, 224, 3). classes (int, optional): number of classes. Defaults to 1000. include_top (bool, optional): whether to include the final classifier network. Defaults to True. Returns: keras.Model: the requested model """ return vit_imagenet( name="vit-s16", patch_size=16, input_shape=input_shape, classes=classes, include_top=include_top, **CONFIG_S )
[docs]def vit_s32(input_shape=(224, 224, 3), classes=1000, include_top=True): """Instantiates the ViT-Small 32 architecture; that is a ViT architecture with 6 attention heads, 12 blocks and a patch size of 32. Args: input_shape (tuple, optional): input shape. Defaults to (224, 224, 3). classes (int, optional): number of classes. Defaults to 1000. include_top (bool, optional): whether to include the final classifier network. Defaults to True. Returns: keras.Model: the requested model """ return vit_imagenet( name="vit-s32", patch_size=32, input_shape=input_shape, classes=classes, include_top=include_top, **CONFIG_S )
[docs]def vit_b16(input_shape=(224, 224, 3), classes=1000, include_top=True): """Instantiates the ViT-B16 architecture; that is a ViT architecture with 12 attention heads, 12 blocks and a patch size of 16. Args: input_shape (tuple, optional): input shape. Defaults to (224, 224, 3). classes (int, optional): number of classes. Defaults to 1000. include_top (bool, optional): whether to include the final classifier network. Defaults to True. Returns: keras.Model: the requested model """ return vit_imagenet( name="vit-b16", patch_size=16, input_shape=input_shape, classes=classes, include_top=include_top, **CONFIG_B )
[docs]def vit_b32(input_shape=(224, 224, 3), classes=1000, include_top=True): """Instantiates the ViT-B32 architecture; that is a ViT architecture with 12 attention heads, 12 blocks and a patch size of 32. Args: input_shape (tuple, optional): input shape. Defaults to (224, 224, 3). classes (int, optional): number of classes. Defaults to 1000. include_top (bool, optional): whether to include the final classifier network. Defaults to True. Returns: keras.Model: the requested model """ return vit_imagenet( name="vit-b32", patch_size=32, input_shape=input_shape, classes=classes, include_top=include_top, **CONFIG_B )
[docs]def vit_l16(input_shape=(384, 384, 3), classes=1000, include_top=True): """Instantiates the ViT-L16 architecture; that is a ViT architecture with 16 attention heads, 24 blocks and a patch size of 16. Args: input_shape (tuple, optional): input shape. Defaults to (384, 384, 3). classes (int, optional): number of classes. Defaults to 1000. include_top (bool, optional): whether to include the final classifier network. Defaults to True. Returns: keras.Model: the requested model """ return vit_imagenet( name="vit-l16", patch_size=16, input_shape=input_shape, classes=classes, include_top=include_top, **CONFIG_L )
[docs]def vit_l32(input_shape=(384, 384, 3), classes=1000, include_top=True): """Instantiates the ViT-L32 architecture; that is a ViT architecture with 16 attention heads, 24 blocks and a patch size of 32. Args: input_shape (tuple, optional): input shape. Defaults to (384, 384, 3). classes (int, optional): number of classes. Defaults to 1000. include_top (bool, optional): whether to include the final classifier network. Defaults to True. Returns: keras.Model: the requested model """ return vit_imagenet( name="vit-l32", patch_size=32, input_shape=input_shape, classes=classes, include_top=include_top, **CONFIG_L )