DECO / models /deco.py
ac5113's picture
added files
99a05f0
raw
history blame
3.19 kB
from models.components import Encoder, Cross_Att, Decoder, Classifier
import torch.nn as nn
import torch
class DECO(nn.Module):
def __init__(self, encoder, context, device):
super(DECO, self).__init__()
self.encoder_type = encoder
self.context = context
self.encoder_sem = Encoder(encoder=encoder).to(device)
self.encoder_part = Encoder(encoder=encoder).to(device)
if self.encoder_type == 'hrnet':
if self.context:
self.decoder_sem = Decoder(480, 133, encoder=encoder).to(device)
self.decoder_part = Decoder(480, 26, encoder=encoder).to(device)
self.sem_pool = nn.AdaptiveAvgPool2d((1))
self.part_pool = nn.AdaptiveAvgPool2d((1))
self.cross_att = Cross_Att(480, 480).to(device)
self.classif = Classifier(480).to(device)
elif self.encoder_type == 'swin':
self.correction_conv = nn.Conv1d(768, 1024, 1).to(device)
if self.context:
self.decoder_sem = Decoder(1, 133, encoder=encoder).to(device)
self.decoder_part = Decoder(1, 26, encoder=encoder).to(device)
self.cross_att = Cross_Att(1024, 1024).to(device)
self.classif = Classifier(1024).to(device)
else:
NotImplementedError('Encoder type not implemented')
self.device = device
def forward(self, img):
if self.encoder_type == 'hrnet':
sem_enc_out = self.encoder_sem(img)
part_enc_out = self.encoder_part(img)
if self.context:
sem_mask_pred = self.decoder_sem(sem_enc_out)
part_mask_pred = self.decoder_part(part_enc_out)
sem_enc_out = self.sem_pool(sem_enc_out)
sem_enc_out = sem_enc_out.squeeze(2)
sem_enc_out = sem_enc_out.squeeze(2)
sem_enc_out = sem_enc_out.unsqueeze(1)
part_enc_out = self.part_pool(part_enc_out)
part_enc_out = part_enc_out.squeeze(2)
part_enc_out = part_enc_out.squeeze(2)
part_enc_out = part_enc_out.unsqueeze(1)
att = self.cross_att(sem_enc_out, part_enc_out)
cont = self.classif(att)
else:
sem_enc_out = self.encoder_sem(img)
part_enc_out = self.encoder_part(img)
sem_seg = torch.reshape(sem_enc_out, (-1, 768, 1))
part_seg = torch.reshape(part_enc_out, (-1, 768, 1))
sem_seg = self.correction_conv(sem_seg)
part_seg = self.correction_conv(part_seg)
sem_seg = torch.reshape(sem_seg, (-1, 1, 32, 32))
part_seg = torch.reshape(part_seg, (-1, 1, 32, 32))
if self.context:
sem_mask_pred = self.decoder_sem(sem_seg)
part_mask_pred = self.decoder_part(part_seg)
sem_enc_out = torch.reshape(sem_seg, (-1, 1, 1024))
part_enc_out = torch.reshape(part_seg, (-1, 1, 1024))
att = self.cross_att(sem_enc_out, part_enc_out)
cont = self.classif(att)
if self.context: return cont, sem_mask_pred, part_mask_pred
return cont