139 lines
6.6 KiB
Python
139 lines
6.6 KiB
Python
import inspect
|
|
from argparse import Namespace
|
|
|
|
import warnings
|
|
|
|
import torch
|
|
from torch import nn
|
|
|
|
from einops import rearrange, repeat
|
|
|
|
from ml_lib.metrics.multi_class_classification import MultiClassScores
|
|
from ml_lib.modules.util import (LightningBaseModule, AutoPadToShape, F_x)
|
|
from util.module_mixins import CombinedModelMixins
|
|
|
|
MIN_NUM_PATCHES = 16
|
|
|
|
try:
|
|
from performer_pytorch import Performer
|
|
|
|
class VisualPerformer(CombinedModelMixins,
|
|
LightningBaseModule
|
|
):
|
|
|
|
def __init__(self, in_shape, n_classes, weight_init, activation,
|
|
embedding_size, heads, attn_depth, patch_size, use_residual,
|
|
use_bias, use_norm, dropout, lat_dim, loss, scheduler,
|
|
lr, weight_decay, sto_weight_avg, lr_warm_restart_epochs, opt_reset_interval):
|
|
|
|
# TODO: Move this to parent class, or make it much easieer to access... But How...
|
|
a = dict(locals())
|
|
params = {arg: a[arg] for arg in inspect.signature(self.__init__).parameters.keys() if arg != 'self'}
|
|
super(VisualPerformer, self).__init__(params)
|
|
|
|
self.in_shape = in_shape
|
|
assert len(self.in_shape) == 3, 'There need to be three Dimensions'
|
|
channels, height, width = self.in_shape
|
|
|
|
# Model Paramters
|
|
# =============================================================================
|
|
# Additional parameters
|
|
self.embed_dim = self.params.embedding_size
|
|
|
|
# Automatic Image Shaping
|
|
self.patch_size = self.params.patch_size
|
|
image_size = (max(height, width) // self.patch_size) * self.patch_size
|
|
self.image_size = image_size + self.patch_size if image_size < max(height, width) else image_size
|
|
|
|
# This should be obsolete
|
|
assert self.image_size % self.patch_size == 0, 'image dimensions must be divisible by the patch size'
|
|
|
|
num_patches = (self.image_size // self.patch_size) ** 2
|
|
patch_dim = channels * self.patch_size ** 2
|
|
assert num_patches >= MIN_NUM_PATCHES, f'your number of patches ({num_patches}) is way too small for ' + \
|
|
f'attention. Try decreasing your patch size'
|
|
|
|
# Correct the Embedding Dim
|
|
if not self.embed_dim % self.params.heads == 0:
|
|
self.embed_dim = (self.embed_dim // self.params.heads) * self.params.heads
|
|
message = ('Embedding Dimension was fixed to be devideable by the number' +
|
|
f' of attention heads, is now: {self.embed_dim}')
|
|
for func in print, warnings.warn:
|
|
func(message)
|
|
|
|
# Utility Modules
|
|
self.autopad = AutoPadToShape((self.image_size, self.image_size))
|
|
|
|
# Modules with Parameters
|
|
self.performer = Performer(
|
|
dim=self.embed_dim, # dimension
|
|
depth=self.params.attn_depth, # layers
|
|
heads=self.params.heads, # heads
|
|
causal=True, # auto-regressive or not
|
|
nb_features=None, # 256, # number of random features, if not set, will default to
|
|
# (d * log(d)), where d is the dimension of each head
|
|
feature_redraw_interval=1000, # how frequently to redraw the projection matrix,
|
|
# the more frequent, the slower the training
|
|
generalized_attention=False, # defaults to softmax approximation,
|
|
# but can be set to True for generalized attention
|
|
kernel_fn=self.params.activation(), # the kernel function to be used,
|
|
# if generalized attention is turned on, defaults to Relu
|
|
reversible=True, # reversible layers, from Reformer paper
|
|
ff_chunks=10, # chunk feedforward layer, from Reformer paper
|
|
use_scalenorm=False, # use scale norm, from 'Transformers without Tears' paper
|
|
use_rezero=False, # use rezero, from 'Rezero is all you need' paper
|
|
ff_glu=True, # use GLU variant for feedforward
|
|
ff_dropout=self.params.dropout, # feedforward dropout
|
|
attn_dropout=self.params.dropout, # post-attn dropout
|
|
local_attn_heads=self.params.heads // 2, # 4 heads are local attention, 4 others are global performers
|
|
local_window_size=(patch_dim // self.params.heads) * 2 # window size of local attention
|
|
)
|
|
|
|
self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, self.embed_dim))
|
|
self.patch_to_embedding = nn.Linear(patch_dim, self.embed_dim) if self.params.embedding_size \
|
|
else F_x(self.embed_dim)
|
|
self.cls_token = nn.Parameter(torch.randn(1, 1, self.embed_dim))
|
|
self.dropout = nn.Dropout(self.params.dropout)
|
|
|
|
self.to_cls_token = nn.Identity()
|
|
|
|
self.mlp_head = nn.Sequential(
|
|
nn.LayerNorm(self.embed_dim),
|
|
nn.Linear(self.embed_dim, self.params.lat_dim),
|
|
nn.GELU(),
|
|
nn.Dropout(self.params.dropout),
|
|
nn.Linear(self.params.lat_dim, n_classes),
|
|
nn.Softmax()
|
|
)
|
|
|
|
def forward(self, x):
|
|
"""
|
|
:param x: the sequence to the encoder (required).
|
|
:return:
|
|
"""
|
|
tensor = self.autopad(x)
|
|
p = self.params.patch_size
|
|
tensor = rearrange(tensor, 'b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1=p, p2=p)
|
|
|
|
tensor = self.patch_to_embedding(tensor)
|
|
b, n, _ = tensor.shape
|
|
|
|
cls_tokens = repeat(self.cls_token, '() n d -> b n d', b=b)
|
|
|
|
tensor = torch.cat((cls_tokens, tensor), dim=1)
|
|
tensor += self.pos_embedding[:, :(n + 1)]
|
|
tensor = self.dropout(tensor)
|
|
|
|
tensor = self.performer(tensor)
|
|
|
|
tensor = self.to_cls_token(tensor[:, 0])
|
|
tensor = self.mlp_head(tensor)
|
|
return Namespace(main_out=tensor)
|
|
|
|
def additional_scores(self, outputs):
|
|
return MultiClassScores(self)(outputs)
|
|
|
|
except ImportError: # pragma: do not provide model class
|
|
print('You want to use `performer_pytorch` plugins which are not installed yet,' # pragma: no-cover
|
|
' install it with `pip install performer_pytorch`.')
|