121 lines
5.4 KiB
Python
121 lines
5.4 KiB
Python
from argparse import Namespace
|
|
|
|
import torch
|
|
from torch import nn
|
|
from torch.nn import ModuleList
|
|
|
|
from ml_lib.modules.blocks import ConvModule, LinearModule
|
|
from ml_lib.modules.utils import (LightningBaseModule, Flatten, HorizontalSplitter)
|
|
from util.module_mixins import (BaseOptimizerMixin, BaseTrainMixin, BaseValMixin, BinaryMaskDatasetMixin,
|
|
BaseDataloadersMixin)
|
|
|
|
|
|
class BandwiseConvMultiheadClassifier(BinaryMaskDatasetMixin,
|
|
BaseDataloadersMixin,
|
|
BaseTrainMixin,
|
|
BaseValMixin,
|
|
BaseOptimizerMixin,
|
|
LightningBaseModule
|
|
):
|
|
|
|
def training_step(self, batch_xy, batch_nb, *args, **kwargs):
|
|
batch_x, batch_y = batch_xy
|
|
y = self(batch_x)
|
|
y, bands_y = y.main_out, y.bands
|
|
bands_y_losses = [self.bce_loss(band_y, batch_y) for band_y in bands_y]
|
|
return_dict = {f'band_{band_idx}_loss': band_y for band_idx, band_y in enumerate(bands_y_losses)}
|
|
|
|
last_bce_loss = self.bce_loss(y, batch_y)
|
|
return_dict.update(last_bce_loss=last_bce_loss)
|
|
|
|
bands_y_losses.append(last_bce_loss)
|
|
combined_loss = torch.stack(bands_y_losses).mean()
|
|
|
|
return_dict.update(loss=combined_loss)
|
|
return return_dict
|
|
|
|
def validation_step(self, batch_xy, batch_idx, *args, **kwargs):
|
|
batch_x, batch_y = batch_xy
|
|
y = self(batch_x)
|
|
y, bands_y = y.main_out, y.bands
|
|
bands_y_losses = [self.bce_loss(band_y, batch_y) for band_y in bands_y]
|
|
return_dict = {f'band_{band_idx}_val_loss': band_y for band_idx, band_y in enumerate(bands_y_losses)}
|
|
|
|
last_bce_loss = self.bce_loss(y, batch_y)
|
|
return_dict.update(last_val_bce_loss=last_bce_loss)
|
|
|
|
bands_y_losses.append(last_bce_loss)
|
|
combined_loss = torch.stack(bands_y_losses).mean()
|
|
|
|
return_dict.update(val_bce_loss=combined_loss,
|
|
batch_idx=batch_idx, y=y, batch_y=batch_y
|
|
)
|
|
return return_dict
|
|
|
|
def __init__(self, hparams):
|
|
super(BandwiseConvMultiheadClassifier, self).__init__(hparams)
|
|
|
|
# Dataset
|
|
# =============================================================================
|
|
self.dataset = self.build_dataset()
|
|
|
|
# Model Paramters
|
|
# =============================================================================
|
|
# Additional parameters
|
|
self.in_shape = self.dataset.train_dataset.sample_shape
|
|
self.conv_filters = self.params.filters
|
|
self.n_band_sections = 4
|
|
k = 3 # Base Kernel Value
|
|
|
|
# Modules
|
|
# =============================================================================
|
|
self.split = HorizontalSplitter(self.in_shape, self.n_band_sections)
|
|
|
|
self.band_list = ModuleList()
|
|
for band in range(self.n_band_sections):
|
|
last_shape = self.split.shape
|
|
conv_list = ModuleList()
|
|
for filters in self.conv_filters:
|
|
conv_list.append(ConvModule(last_shape, filters, (k, k), conv_stride=(2, 2), conv_padding=2,
|
|
**self.params.module_kwargs))
|
|
last_shape = conv_list[-1].shape
|
|
# self.conv_list.append(ConvModule(last_shape, 1, 1, conv_stride=1, **self.params.module_kwargs))
|
|
# last_shape = self.conv_list[-1].shape
|
|
self.band_list.append(conv_list)
|
|
|
|
self.bandwise_deep_list_1 = ModuleList([
|
|
LinearModule(self.band_list[0][-1].shape, self.params.lat_dim, **self.params.module_kwargs)
|
|
for _ in range(self.n_band_sections)])
|
|
self.bandwise_deep_list_2 = ModuleList([
|
|
LinearModule(self.params.lat_dim, self.params.lat_dim * 2, **self.params.module_kwargs)
|
|
for _ in range(self.n_band_sections)])
|
|
self.bandwise_latent_list = ModuleList([
|
|
LinearModule(self.params.lat_dim * 2, self.params.lat_dim, **self.params.module_kwargs)
|
|
for _ in range(self.n_band_sections)])
|
|
self.bandwise_classifier_list = ModuleList([
|
|
LinearModule(self.params.lat_dim, 1, bias=self.params.bias, activation=nn.Sigmoid)
|
|
for _ in range(self.n_band_sections)])
|
|
|
|
self.full_1 = LinearModule(self.n_band_sections, self.params.lat_dim, **self.params.module_kwargs)
|
|
self.full_2 = LinearModule(self.full_1.shape, self.params.lat_dim * 2, **self.params.module_kwargs)
|
|
self.full_3 = LinearModule(self.full_2.shape, self.params.lat_dim, **self.params.module_kwargs)
|
|
self.full_out = LinearModule(self.full_3.shape, 1, bias=self.params.bias, activation=nn.Sigmoid)
|
|
|
|
def forward(self, batch, **kwargs):
|
|
tensors = self.split(batch)
|
|
for idx, (tensor, convs) in enumerate(zip(tensors, self.band_list)):
|
|
for conv in convs:
|
|
tensor = conv(tensor)
|
|
|
|
tensor = self.bandwise_deep_list_1[idx](tensor)
|
|
tensor = self.bandwise_deep_list_2[idx](tensor)
|
|
tensor = self.bandwise_latent_list[idx](tensor)
|
|
tensors[idx] = self.bandwise_classifier_list[idx](tensor)
|
|
|
|
tensor = torch.cat(tensors, dim=1)
|
|
tensor = self.full_1(tensor)
|
|
tensor = self.full_2(tensor)
|
|
tensor = self.full_3(tensor)
|
|
tensor = self.full_out(tensor)
|
|
return Namespace(main_out=tensor, bands=tensors)
|