Save trained models and plot encoding

constellation/__init__.py

@@ -0,0 +1,2 @@
+from constellation.ConstellationNet import ConstellationNet
+import constellation.util

constellation/util.py

@@ -0,0 +1,36 @@
+import torch
+
+
+def get_random_messages(count, order):
+    """
+    Generate a list of messages.
+
+    :param count: Number of messages to generate.
+    :param order: Number of possible messages.
+    :return: One-dimensional vector with each entry being the index of the
+    generated message which is between 0, inclusive, and `order`, exclusive.
+
+    >>> get_random_messages(5)
+    torch.tensor([0, 2, 0, 3, 4])
+    """
+    return torch.randint(0, order, (count,))
+
+
+def messages_to_onehot(messages, order):
+    """
+    Convert messages represented as indexes to one-hot encoding.
+
+    :param messages: List of messages to convert.
+    :param order: Number of possible messages.
+    :return: One-hot encoded messages.
+
+    >>> messages_to_onehot(torch.tensor([0, 2, 0, 3, 4]))
+    torch.tensor([
+        [1., 0., 0., 0., 0.],
+        [0., 0., 1., 0., 0.],
+        [1., 0., 0., 0., 0.],
+        [0., 0., 0., 1., 0.],
+        [0., 0., 0., 0., 1.],
+    ])
+    """
+    return torch.nn.functional.one_hot(messages, num_classes=order).float()

output/.gitignore

@@ -0,0 +1,5 @@
+# Ignore all files in this directory…
+*
+
+# …except for this one.
+!.gitignore

plot.py

@@ -0,0 +1,26 @@
+import constellation
+from constellation import util
+import torch
+from matplotlib import pyplot
+
+# Number learned symbols
+order = 4
+
+# File in which the trained model is saved
+input_file = 'output/constellation-net.tc'
+
+model = constellation.ConstellationNet(order=order)
+model.load_state_dict(torch.load(input_file))
+
+# Compute encoded vectors
+with torch.no_grad():
+    encoded_vectors = model.encoder(
+        util.messages_to_onehot(
+            torch.arange(0, order),
+            order
+        )
+    ).tolist()
+
+fig, axis = pyplot.subplots()
+axis.scatter(*zip(*encoded_vectors))
+pyplot.show()

train.py

@@ -1,4 +1,5 @@
-from ConstellationNet import ConstellationNet
+import constellation
+from constellation import util
 import torch
 
 # Number of symbols to learn
@@ -8,23 +9,25 @@ order = 4
 epoch_size = 10000
 
 # Number of epochs
-num_epochs = 25000
+num_epochs = 20000
 
 # Number of epochs to skip between every loss report
-loss_report_epoch_skip = 200
+loss_report_epoch_skip = 500
 
-model = ConstellationNet(order=order)
+# File in which the trained model is saved
+output_file = 'output/constellation-net.tc'
+
+print('Starting training with {} epochs\n'.format(num_epochs))
+
+model = constellation.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,))
+    classes_dataset = util.get_random_messages(epoch_size, order)
-    onehot_dataset = torch.nn.functional.one_hot(
+    onehot_dataset = util.messages_to_onehot(classes_dataset, order)
-        classes_dataset,
-        num_classes=order
-    ).float()
 
     optimizer.zero_grad()
     predictions = model(onehot_dataset)
@@ -40,8 +43,29 @@ for epoch in range(num_epochs):
         print('Loss is {}'.format(running_loss))
         running_loss = 0
 
-# Test the model with class 1
+print('\nFinished training\n')
-print(model(torch.nn.functional.one_hot(torch.tensor(0), num_classes=order).float()))
 
-# Test the model with class 2
+# Print some examples of reconstruction
-print(model(torch.nn.functional.one_hot(torch.tensor(1), num_classes=order).float()))
+with torch.no_grad():
+    num_examples = 5
+
+    classes_example = util.get_random_messages(num_examples, order)
+    onehot_example = util.messages_to_onehot(classes_example, order)
+    raw_output = model(onehot_example)
+    raw_output.required_grad = False
+    reconstructed_example = torch.nn.functional.softmax(raw_output, dim=1)
+
+    print('Reconstruction examples:')
+    print('Input vector\t\t\tOutput vector after softmax')
+
+    for example_index in range(num_examples):
+        print('{}\t\t{}'.format(
+            onehot_example[example_index].tolist(),
+            '[{}]'.format(', '.join(
+                '{:.5f}'.format(x)
+                for x in reconstructed_example[example_index].tolist()
+            ))
+        ))
+
+print('\nSaving model as {}'.format(output_file))
+torch.save(model.state_dict(), output_file)