Merge pull request #95 from eth-cscs/release-0.6.3

Release 0.6.3

Makefile

@@ -28,6 +28,7 @@ unittest:
 	python3 -m unittest discover -s tests -v -p "*_tests.py" || (echo "Error in standard unit tests."; exit 1)
 	@# remove temporary files created during testing
 	@if test -f net.pth; then rm net.pth; fi
+	@if test -f net_with_discard_wrapper.pth; then rm net_with_discard_wrapper.pth; fi
 	@if test -f scaler.pkl; then rm scaler.pkl; fi
 	@if test -f tmp.jnl; then rm tmp.jnl; fi
 	@if test -f journal_tests_testfile.pkl; then rm journal_tests_testfile.pkl; fi
@@ -40,6 +41,10 @@ unittest_mpi:
 exampletest: $(MAKEDIRS)
 	@echo "Testing standard examples.."
 	python3 -m unittest -v tests/test_examples.py || (echo "Error in example tests."; exit 1)
+	@if test -f scaler.pkl; then rm scaler.pkl; fi
+	@if test -f seminn_net.pth; then rm seminn_net.pth; fi
+	@if test -f triplet_net.pth; then rm triplet_net.pth; fi
+	@if test -f tmp.jnl; then rm tmp.jnl; fi
 
 exampletest_mpi:
 	@echo "Testing MPI backend examples.."

README.md

@@ -44,6 +44,7 @@ Additional **features** are:
 * several methods for summary selection:
   * [Semi-automatic summary selection (with Neural networks)](http://proceedings.mlr.press/v97/wiqvist19a/wiqvist19a.pdf)
   * [summary selection using distance learning (with Neural networks)](https://link.springer.com/article/10.1007/s13571-019-00208-8)
+  * [Sufficient statistics of exponential family approximating the likelihood (with Neural networks)](https://arxiv.org/abs/2012.10903)
 * [Random Forest Model Selection Scheme](https://academic.oup.com/bioinformatics/article/32/6/859/1744513)
 
 
@@ -130,9 +131,9 @@ Publications in which ABCpy was applied:
 
 * R. Dutta, K. Zouaoui-Boudjeltia, C. Kotsalos, A. Rousseau, D. Ribeiro de Sousa, J. M. Desmet, A. Van Meerhaeghe, A. Mira, and B. Chopard. "Interpretable pathological test for Cardio-vascular disease: Approximate Bayesian computation with distance learning.", 2020, arXiv:2010.06465.
 
-* R. Dutta, S. Gomes, D. Kalise, L. Pacchiardi. "Using mobility data in the design of optimal lockdown strategies for the COVID-19 pandemic in England.", 2020, arXiv:2006.16059.
+* R. Dutta, S. Gomes, D. Kalise, L. Pacchiardi. "Using mobility data in the design of optimal lockdown strategies for the COVID-19 pandemic in England.", 2021,  PLOS Computational Biology, 17(8), e1009236.
 
-* L. Pacchiardi, P. Künzli, M. Schöngens, B. Chopard, R. Dutta, "Distance-Learning for Approximate Bayesian Computation to Model a Volcanic Eruption", 2020, Sankhya B, 1-30.
+* L. Pacchiardi, P. Künzli, M. Schöngens, B. Chopard, R. Dutta, "Distance-Learning for Approximate Bayesian Computation to Model a Volcanic Eruption", 2021, Sankhya B, 83(1), 288-317.
 
 * R. Dutta, J. P.  Onnela, A. Mira, "Bayesian Inference of Spreading Processes on Networks", 2018, Proceedings of Royal Society A, 474(2215), 20180129.
 

VERSION

@@ -1 +1 @@
-0.6.2
+0.6.3

abcpy/NN_utilities/algorithms.py

@@ -81,13 +81,13 @@ def contrastive_training(samples, similarity_set, embedding_net, cuda, batch_siz
         scheduler = scheduler(optimizer, **scheduler_kwargs)
 
     # now train:
-    fit(pairs_train_loader, model_contrastive, loss_fn, optimizer, scheduler, n_epochs, cuda,
+    train_losses, test_losses = fit(pairs_train_loader, model_contrastive, loss_fn, optimizer, scheduler, n_epochs, cuda,
         val_loader=pairs_train_loader_val,
         early_stopping=early_stopping, start_epoch_early_stopping=start_epoch_early_stopping,
         epochs_early_stopping_interval=epochs_early_stopping_interval, start_epoch_training=start_epoch_training,
         use_tqdm=use_tqdm)
 
-    return embedding_net
+    return embedding_net, train_losses, test_losses
 
 
 def triplet_training(samples, similarity_set, embedding_net, cuda, batch_size=16, n_epochs=400,
@@ -155,12 +155,12 @@ def triplet_training(samples, similarity_set, embedding_net, cuda, batch_size=16
         scheduler = scheduler(optimizer, **scheduler_kwargs)
 
     # now train:
-    fit(triplets_train_loader, model_triplet, loss_fn, optimizer, scheduler, n_epochs, cuda,
+    train_losses, test_losses = fit(triplets_train_loader, model_triplet, loss_fn, optimizer, scheduler, n_epochs, cuda,
         val_loader=triplets_train_loader_val,
         early_stopping=early_stopping, start_epoch_early_stopping=start_epoch_early_stopping,
         epochs_early_stopping_interval=epochs_early_stopping_interval, start_epoch_training=start_epoch_training, use_tqdm=use_tqdm)
 
-    return embedding_net
+    return embedding_net, train_losses, test_losses
 
 
 def FP_nn_training(samples, target, embedding_net, cuda, batch_size=1, n_epochs=50, samples_val=None, target_val=None,
@@ -222,9 +222,9 @@ def FP_nn_training(samples, target, embedding_net, cuda, batch_size=1, n_epochs=
         scheduler = scheduler(optimizer, **scheduler_kwargs)
 
     # now train:
-    fit(data_loader_FP_nn, embedding_net, loss_fn, optimizer, scheduler, n_epochs, cuda,
+    train_losses, test_losses = fit(data_loader_FP_nn, embedding_net, loss_fn, optimizer, scheduler, n_epochs, cuda,
         val_loader=data_loader_FP_nn_val,
         early_stopping=early_stopping, start_epoch_early_stopping=start_epoch_early_stopping,
         epochs_early_stopping_interval=epochs_early_stopping_interval, start_epoch_training=start_epoch_training, use_tqdm=use_tqdm)
 
-    return embedding_net
+    return embedding_net, train_losses, test_losses

abcpy/NN_utilities/losses.py

@@ -1,3 +1,4 @@
+import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
@@ -37,3 +38,24 @@ def forward(self, anchor, positive, negative, size_average=True):
         distance_negative = (anchor - negative).pow(2).sum(1)  # .pow(.5)
         losses = F.relu(distance_positive - distance_negative + self.margin)
         return losses.mean() if size_average else losses.sum()
+
+
+def Fisher_divergence_loss(first_der_t, second_der_t, eta, lam=0):
+    """lam is the regularization parameter of the Kingma & LeCun (2010) regularization"""
+    inner_prod_second_der_eta = torch.bmm(second_der_t, eta.unsqueeze(-1))  # this is used twice
+
+    if lam == 0:
+        return sum(
+            (0.5 * torch.bmm(first_der_t, eta.unsqueeze(-1)) ** 2 + inner_prod_second_der_eta).view(-1))
+    else:
+        return sum(
+            (0.5 * torch.bmm(first_der_t, eta.unsqueeze(-1)) ** 2 +
+             inner_prod_second_der_eta + lam * inner_prod_second_der_eta ** 2).view(-1))
+
+
+def Fisher_divergence_loss_with_c_x(first_der_t, second_der_t, eta, lam=0):
+    # this enables to use the term c(x) in the approximating family, ie a term that depends only on x and not on theta.
+    new_eta = torch.cat((eta, torch.ones(eta.shape[0], 1).to(eta)),
+                        dim=1)  # the one tensor need to be on same device as eta.
+    # then call the other loss function with this new_eta:
+    return Fisher_divergence_loss(first_der_t, second_der_t, new_eta, lam=lam)