from_dict transform, CompositeLoss, DANN

ovotools/__init__.py

 from .params import AttrDict
+import .params
+import .pytorch
+import .ignite
\ No newline at end of file

ovotools/ignite/__init__.py

+from .transforms import from_dict
\ No newline at end of file

ovotools/ignite/transforms.py

+def from_dict(key):
+    '''
+    Creates transform function extracting value from dict if input is dict
+    :param key: string
+    :return: functions: x -> x[key] if x is dict, x otherwise
+    '''
+
+    def call(data):
+        if isinstance(data, dict):
+            return data[key]
+        elif isinstance(data, tuple):
+            return tuple(call(di) for di in data)
+        else:
+            return data
+
+    return call
\ No newline at end of file

ovotools/pytorch/__init__.py

 from .data import CachedDataSet, BatchThreadingDataLoader, ThreadingDataLoader
 
-from .losses import MeanLoss
+from .losses import CompositeLoss, MeanLoss
+
+from .modules import ReverseLayerF, DANN_module, Dann_Head, DannEncDecNet
 
 from .utils import set_reproducibility
 from .utils import reproducibility_worker_init_fn

ovotools/pytorch/losses/__init__.py

+from .composite_loss import CompositeLoss
 from .mean_loss import MeanLoss

ovotools/pytorch/losses/composite_loss.py

+import torch
+
+class CompositeLoss(torch.nn.modules.loss._Loss):
+    def __init__(self, loss_funcs):
+        super(CompositeLoss, self).__init__()
+        self.loss_funcs = loss_funcs
+        self.loss_vals = [None] * (len(self.loss_funcs) + 1)
+
+    def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor) -> torch.Tensor:
+        self.loss_vals = [loss_fn(y_pred, y_true) for (loss_fn, _,) in self.loss_funcs]
+        res = sum([w * self.loss_vals[i] for i, (_, w,) in enumerate(self.loss_funcs)])
+        self.loss_vals.append(res)
+        return res
+
+    def __len__(self):
+        return len(self.loss_funcs)
+
+    def get_val(self, index):
+        def call(*kargs, **kwargs):
+            return self.loss_vals[index]
+        return call
+
+
+

ovotools/pytorch/modules/__init__.py

+from .dann import ReverseLayerF, DANN_module, Dann_Head, DannEncDecNet
\ No newline at end of file

ovotools/pytorch/modules/dann.py

+'''
+DANN module. See  https://arxiv.org/abs/1505.07818, https://arxiv.org/abs/1409.7495
+'''
+
+import torch
+import torch.nn as nn
+from torch.autograd import Function
+
+class ReverseLayerF(Function):
+
+    @staticmethod
+    def forward(ctx, x, alpha):
+        ctx.alpha = alpha
+        return x.view_as(x)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        output = grad_output.neg() * ctx.alpha
+        return output, None
+
+
+class DANN_module(nn.Module):
+    def __init__(self, gamma = 10., lambda_max = 1., **kwargs):
+        super(DANN_module, self).__init__()
+        self.gamma = gamma
+        self.lambda_max = lambda_max
+        self.progress = nn.Parameter(torch.Tensor([0,]), requires_grad = False) # must be updated from outside
+
+    def set_progress(self, value):
+        self.progress.data = torch.Tensor([value,]).to(self.progress.device)
+
+    def forward(self, input_data):
+        lambda_p = self.lambda_max * (2. / (1. + torch.exp(-self.gamma * self.progress.data)) - 1)
+        reverse_feature = ReverseLayerF.apply(input_data, lambda_p)
+        return reverse_feature
+
+
+class Dann_Head(nn.Module):
+    def __init__(self, input_dims, num_classes, **kwargs):
+        super(Dann_Head, self).__init__()
+        self.input_dims = input_dims
+        self.pooling_depth = 100
+        self.dann_module = DANN_module(**kwargs)
+        self.pooling_modules = [nn.Sequential(
+                                    nn.Conv2d(dim, self.pooling_depth, 3),
+                                    nn.AdaptiveMaxPool2d(1)
+                                )
+                                for dim in input_dims]
+        for i, m in enumerate(self.pooling_modules):
+            self.add_module("pooling_module"+str(i), m)
+        self.domain_classifier = nn.Sequential()
+        self.domain_classifier.add_module('dann_fc1', nn.Linear(self.pooling_depth*len(input_dims), 100))
+        self.domain_classifier.add_module('dann_bn1', nn.BatchNorm1d(100)) # - dows not work for batch = 1
+        #self.domain_classifier.add_module('dann_bn1', nn.GroupNorm(5, 100)) # - dows not work for batch = 1
+        self.domain_classifier.add_module('dann_relu1', nn.ReLU(True))
+        self.domain_classifier.add_module('dann_fc2', nn.Linear(100, num_classes))
+        self.domain_classifier.add_module('dann_softmax', nn.LogSoftmax(dim=1))
+        self.loss = nn.NLLLoss()
+
+    def set_progress(self, value):
+            self.dann_module.set_progress(value)
+
+    def forward(self, inputs, y_true: torch.Tensor):
+        features = []
+        for i, x in enumerate(inputs[:len(self.input_dims)]):
+            assert x.shape[1] == self.input_dims[i]
+            x = self.dann_module(x)
+            x = self.pooling_modules[i](x)
+            features.append(x)
+        feature = torch.cat([f.flatten(start_dim=1) for f in features], dim=1)
+        domain_output = self.domain_classifier(feature)
+        loss = self.loss(domain_output, y_true.long())
+        return loss
+
+
+class DannEncDecNet(nn.Module):
+    def __init__(self, base_net, input_dim, num_classes, **enc_dec_params):
+        super(DannEncDecNet, self).__init__()
+        self.net = base_net
+        self.dann_head = Dann_Head(input_dim, num_classes, **enc_dec_params)
+        self.bottleneck_data = None
+        self.dann_loss = None
+
+    def set_progress(self, value):
+        self.dann_head.set_progress(value)
+
+    def forward(self, x):
+        """Sequentially pass `x` trough model`s `encoder` and `decoder` (return logits!)"""
+        x = self.net.encoder(x)
+        self.bottleneck_data = x
+        x = self.net.decoder(x)
+        return x
+
+    def calc_dann_loss(self, y_pred, y_true):
+        self.dann_loss = self.dann_head(self.bottleneck_data, y_true)
+        return self.dann_loss
+