From 2af8354a07224c2e414b7d0f916e1c6ea4a3931e Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Matt=C3=A9o=20Delabre?= <spam@delab.re>
Date: Fri, 13 Dec 2019 12:11:09 -0500
Subject: [PATCH] Add initial implementation

---
 .gitignore          |  1 +
 ConstellationNet.py | 78 +++++++++++++++++++++++++++++++++++++++++++++
 train.py            | 47 +++++++++++++++++++++++++++
 3 files changed, 126 insertions(+)
 create mode 100644 .gitignore
 create mode 100644 ConstellationNet.py
 create mode 100644 train.py

diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..bee8a64
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1 @@
+__pycache__
diff --git a/ConstellationNet.py b/ConstellationNet.py
new file mode 100644
index 0000000..60f5db3
--- /dev/null
+++ b/ConstellationNet.py
@@ -0,0 +1,78 @@
+import torch.nn as nn
+
+
+class ConstellationNet(nn.Module):
+    def __init__(
+        self,
+        order=2,
+        encoder_layers_sizes=(),
+        decoder_layers_sizes=()
+    ):
+        """
+        Create an encoder-decoder network to automatically shape a
+        constellation of symbols for efficient communication over an optical
+        fiber channel.
+
+        :param order: Order of the constellation, i.e. the number of messages
+        that are to be transmitted or equivalently the number of symbols whose
+        placements in the constellation have to be learned.
+        :param encoder_layers_sizes: Shape of the encoder’s hidden layers. The
+        size of this sequence is the number of hidden layers, with each element
+        being a number which specifies the number of neurons in its channel.
+        :param decoder_layers_sizes: Shape of the decoder’s hidden layers. Uses
+        the same convention as `encoder_layers_sizes` above.
+        """
+        super().__init__()
+
+        # Build the encoder network taking a one-hot encoded message as input
+        # and outputting an I/Q vector. The network additionally uses hidden
+        # layers as specified in `encoder_layers_sizes`
+        prev_layer_size = order
+        encoder_layers = []
+
+        for layer_size in encoder_layers_sizes:
+            encoder_layers.append(nn.Linear(prev_layer_size, layer_size))
+            encoder_layers.append(nn.ReLU())
+            prev_layer_size = layer_size
+
+        encoder_layers += [
+            nn.Linear(prev_layer_size, 2),
+            nn.ReLU(),
+            # TODO: Normalization step
+        ]
+
+        self.encoder = nn.Sequential(*encoder_layers)
+
+        # TODO: Add real channel model
+        self.channel = nn.Identity()
+
+        # Build the decoder network taking the noisy I/Q vector received from
+        # the channel as input and outputting a probability vector for each
+        # original message. The network additionally uses hidden layers as
+        # specified in `decoder_layers_sizes`
+        prev_layer_size = 2
+        decoder_layers = []
+
+        for layer_size in decoder_layers_sizes:
+            decoder_layers.append(nn.Linear(prev_layer_size, layer_size))
+            decoder_layers.append(nn.ReLU())
+            prev_layer_size = layer_size
+
+        # Softmax is not used at the end of the network because the
+        # CrossEntropyLoss criterion is used for training, which includes
+        # LogSoftmax
+        decoder_layers.append(nn.Linear(prev_layer_size, order),)
+
+        self.decoder = nn.Sequential(*decoder_layers)
+
+    def forward(self, x):
+        """
+        Perform encoding and decoding of an input vector and compute its
+        reconstructed vector.
+
+        :param x: Original one-hot encoded data.
+        :return: Reconstructed vector.
+        """
+        symbol = self.encoder(x)
+        noisy_symbol = self.channel(symbol)
+        return self.decoder(noisy_symbol)
diff --git a/train.py b/train.py
new file mode 100644
index 0000000..ad571eb
--- /dev/null
+++ b/train.py
@@ -0,0 +1,47 @@
+from ConstellationNet import ConstellationNet
+import torch
+
+# Number of symbols to learn
+order = 4
+
+# Number of training examples in an epoch
+epoch_size = 10000
+
+# Number of epochs
+num_epochs = 25000
+
+# Number of epochs to skip between every loss report
+loss_report_epoch_skip = 200
+
+model = ConstellationNet(order=order)
+criterion = torch.nn.CrossEntropyLoss()
+optimizer = torch.optim.Adam(model.parameters())
+
+running_loss = 0
+
+for epoch in range(num_epochs):
+    classes_dataset = torch.randint(0, order, (epoch_size,))
+    onehot_dataset = torch.nn.functional.one_hot(
+        classes_dataset,
+        num_classes=order
+    ).float()
+
+    optimizer.zero_grad()
+    predictions = model(onehot_dataset)
+    loss = criterion(predictions, classes_dataset)
+    loss.backward()
+    optimizer.step()
+
+    # Report loss
+    running_loss += loss.item()
+
+    if epoch % loss_report_epoch_skip == loss_report_epoch_skip - 1:
+        print('Epoch {}/{}'.format(epoch + 1, num_epochs))
+        print('Loss is {}'.format(running_loss))
+        running_loss = 0
+
+# Test the model with class 1
+print(model(torch.nn.functional.one_hot(torch.tensor(0), num_classes=order).float()))
+
+# Test the model with class 2
+print(model(torch.nn.functional.one_hot(torch.tensor(1), num_classes=order).float()))