Mel_Vision_Transformer_ComP.../models/transformer_model_vertical.py
2021-04-02 08:45:11 +02:00

113 lines
4.7 KiB
Python

import inspect
from argparse import Namespace
import warnings
import torch
from einops import repeat
from torch import nn
from ml_lib.metrics.multi_class_classification import MultiClassScores
from ml_lib.modules.blocks import TransformerModule
from ml_lib.modules.util import (LightningBaseModule, AutoPadToShape, F_x, SlidingWindow)
from util.module_mixins import CombinedModelMixins
MIN_NUM_PATCHES = 16
class VerticalVisualTransformer(CombinedModelMixins, LightningBaseModule):
def __init__(self, in_shape, n_classes, weight_init, activation,
embedding_size, heads, attn_depth, patch_size, use_residual, variable_length,
use_bias, use_norm, dropout, lat_dim, loss, scheduler, mlp_dim, head_dim,
lr, weight_decay, sto_weight_avg, lr_scheduler_parameter, opt_reset_interval,
return_logits=False):
# 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(VerticalVisualTransformer, self).__init__(params)
self.in_shape = in_shape
self.n_classes = n_classes
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
self.height = height
self.channels = channels
self.new_width = ((width - self.params.patch_size)//1) + 1
num_patches = self.new_width - (self.params.patch_size // 2)
patch_dim = channels * self.params.patch_size * self.height
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'
# Utility Modules
self.autopad = AutoPadToShape((self.height, self.new_width))
self.dropout = nn.Dropout(self.params.dropout)
self.slider = SlidingWindow((channels, *self.autopad.target_shape), (self.height, self.params.patch_size),
keepdim=False)
# Modules with Parameters
self.transformer = TransformerModule(in_shape=self.embed_dim, mlp_dim=self.params.mlp_dim,
head_dim=self.params.head_dim,
heads=self.params.heads, depth=self.params.attn_depth,
dropout=self.params.dropout, use_norm=self.params.use_norm,
activation=self.params.activation, use_residual=self.params.use_residual
)
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.to_cls_token = nn.Identity()
logits = self.params.n_classes if self.params.n_classes > 2 else 1
outbound_activation = nn.Softmax if logits > 1 else nn.Sigmoid
self.mlp_head = nn.Sequential(
nn.LayerNorm(self.embed_dim),
nn.Linear(self.embed_dim, self.params.lat_dim),
self.params.activation(),
nn.Dropout(self.params.dropout),
nn.Linear(self.params.lat_dim, logits),
outbound_activation()
)
def forward(self, x, mask=None, return_attn_weights=False):
"""
:param x: the sequence to the encoder (required).
:param mask: the mask for the src sequence (optional).
:param return_attn_weights: wether to return the attn weights (optional)
:return:
"""
tensor = self.autopad(x)
tensor = self.slider(tensor)
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)
if return_attn_weights:
tensor, attn_weights = self.transformer(tensor, mask, return_attn_weights)
else:
attn_weights = None
tensor = self.transformer(tensor, mask)
tensor = self.to_cls_token(tensor[:, 0])
tensor = self.mlp_head(tensor)
return Namespace(main_out=tensor, attn_weights=attn_weights)