Support for all VSA models in the embeddings module (#111)

* Update empty, identity and random embeddings

* Level embeddings

* [github-action] formatting fixes

* Implement VSA model choice for Thermometer and Circular

* [github-action] formatting fixes

* Update examples to latest version of torchmetrics

* [github-action] formatting fixes

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

Author: mikeheddes

docs/embeddings.rst

@@ -9,6 +9,7 @@ torchhd.embeddings
     :toctree: generated/
     :template: class.rst
 
+    Empty
     Identity
     Random
     Level

examples/emg_hand_gestures.py

@@ -66,7 +66,8 @@ def experiment(subjects=[0]):
     train_ld = data.DataLoader(train_ds, batch_size=BATCH_SIZE, shuffle=True)
     test_ld = data.DataLoader(test_ds, batch_size=BATCH_SIZE, shuffle=False)
 
-    model = Model(len(ds.classes), ds[0][0].size(-2), ds[0][0].size(-1))
+    num_classes = len(ds.classes)
+    model = Model(num_classes, ds[0][0].size(-2), ds[0][0].size(-1))
     model = model.to(device)
 
     with torch.no_grad():
@@ -79,7 +80,7 @@ def experiment(subjects=[0]):
 
         model.classify.weight[:] = F.normalize(model.classify.weight)
 
-    accuracy = torchmetrics.Accuracy()
+    accuracy = torchmetrics.Accuracy("multiclass", num_classes=num_classes)
 
     with torch.no_grad():
         for samples, labels in tqdm(test_ld, desc="Testing"):

examples/graphhd.py

@@ -121,7 +121,7 @@ def forward(self, x):
 
     model.classify.weight[:] = F.normalize(model.classify.weight)
 
-accuracy = torchmetrics.Accuracy()
+accuracy = torchmetrics.Accuracy("multiclass", num_classes=graphs.num_classes)
 
 with torch.no_grad():
     for samples in tqdm(test_ld, desc="Testing"):

examples/language_recognition.py

@@ -74,7 +74,8 @@ def forward(self, x):
         return logit
 
 
-model = Model(len(train_ds.classes), NUM_TOKENS)
+num_classes = len(train_ds.classes)
+model = Model(num_classes, NUM_TOKENS)
 model = model.to(device)
 
 with torch.no_grad():
@@ -87,7 +88,7 @@ def forward(self, x):
 
     model.classify.weight[:] = F.normalize(model.classify.weight)
 
-accuracy = torchmetrics.Accuracy()
+accuracy = torchmetrics.Accuracy("multiclass", num_classes=num_classes)
 
 with torch.no_grad():
     for samples, labels in tqdm(test_ld, desc="Testing"):

examples/mnist.py

@@ -53,7 +53,8 @@ def forward(self, x):
         return logit
 
 
-model = Model(len(train_ds.classes), IMG_SIZE)
+num_classes = len(train_ds.classes)
+model = Model(num_classes, IMG_SIZE)
 model = model.to(device)
 
 with torch.no_grad():
@@ -66,8 +67,7 @@ def forward(self, x):
 
     model.classify.weight[:] = F.normalize(model.classify.weight)
 
-accuracy = torchmetrics.Accuracy()
-
+accuracy = torchmetrics.Accuracy("multiclass", num_classes=num_classes)
 
 with torch.no_grad():
     for samples, labels in tqdm(test_ld, desc="Testing"):

examples/mnist_hugging_face.py

@@ -6,8 +6,10 @@
 
 # Note: this example requires the torchmetrics library: https://torchmetrics.readthedocs.io
 import torchmetrics
-from tqdm import tqdm
+
+# Note: this example requires the accelerate library: https://github.com/huggingface/accelerate
 from accelerate import Accelerator
+from tqdm import tqdm
 
 import torchhd
 from torchhd import embeddings
@@ -54,7 +56,8 @@ def forward(self, x):
         return logit
 
 
-model = Model(len(train_ds.classes), IMG_SIZE)
+num_classes = len(train_ds.classes)
+model = Model(num_classes, IMG_SIZE)
 model.to(device)
 
 model, train_ld, test_ld = accelerator.prepare(model, train_ld, test_ld)
@@ -69,8 +72,7 @@ def forward(self, x):
 
     model.classify.weight[:] = F.normalize(model.classify.weight)
 
-accuracy = torchmetrics.Accuracy()
-
+accuracy = torchmetrics.Accuracy("multiclass", num_classes=num_classes)
 
 with torch.no_grad():
     for samples, labels in tqdm(test_ld, desc="Testing"):

examples/mnist_nonlinear.py

@@ -28,6 +28,7 @@
 test_ds = MNIST("../data", train=False, transform=transform, download=True)
 test_ld = torch.utils.data.DataLoader(test_ds, batch_size=BATCH_SIZE, shuffle=False)
 
+
 class Model(nn.Module):
     def __init__(self, num_classes, size):
         super(Model, self).__init__()
@@ -50,6 +51,7 @@ def forward(self, x):
         return logit
 
 
+num_classes = len(train_ds.classes)
 model = Model(len(train_ds.classes), IMG_SIZE)
 model = model.to(device)
 
@@ -63,8 +65,7 @@ def forward(self, x):
 
     model.classify.weight[:] = F.normalize(model.classify.weight)
 
-accuracy = torchmetrics.Accuracy()
-
+accuracy = torchmetrics.Accuracy("multiclass", num_classes=num_classes)
 
 with torch.no_grad():
     for samples, labels in tqdm(test_ld, desc="Testing"):

examples/mnist_torch_lightning.py

@@ -6,8 +6,10 @@
 
 # Note: this example requires the torchmetrics library: https://torchmetrics.readthedocs.io
 import torchmetrics
-from tqdm import tqdm
+
+# Note: this example requires the pytorch-lightning library: https://www.pytorchlightning.ai
 import pytorch_lightning as pl
+from tqdm import tqdm
 
 import torchhd
 from torchhd import embeddings
@@ -60,7 +62,8 @@ def configure_optimizers(self):
         return
 
 
-model = Model(len(train_ds.classes), IMG_SIZE)
+num_classes = len(train_ds.classes)
+model = Model(num_classes, IMG_SIZE)
 trainer = pl.Trainer(
     accelerator="cpu",
     devices=1,
@@ -77,8 +80,7 @@ def configure_optimizers(self):
 
     model.classify.weight[:] = F.normalize(model.classify.weight)
 
-accuracy = torchmetrics.Accuracy()
-
+accuracy = torchmetrics.Accuracy("multiclass", num_classes=num_classes)
 
 with torch.no_grad():
     for samples, labels in tqdm(test_ld, desc="Testing"):

examples/random_projection.py

@@ -58,16 +58,13 @@ def __init__(self, num_classes, size):
         self.target = embeddings.Level(
             500, DIMENSIONS, low=MIN_TEMPERATURE, high=MAX_TEMPERATURE
         )
-        self.project = embeddings.Projection(size, DIMENSIONS)
-        self.bias = nn.parameter.Parameter(torch.empty(DIMENSIONS), requires_grad=False)
-        self.bias.data.uniform_(0, 2 * math.pi)
+        self.project = embeddings.Sinusoid(size, DIMENSIONS)
 
         self.regression = nn.Linear(DIMENSIONS, num_classes, bias=False)
         self.regression.weight.data.fill_(0.0)
 
     def encode(self, x):
-        enc = self.project(x)
-        sample_hv = torch.cos(enc + self.bias) * torch.sin(enc)
+        sample_hv = self.project(x)
         return torchhd.hard_quantize(sample_hv)
 
     def forward(self, x):

examples/reghd.py

@@ -18,7 +18,7 @@
 DIMENSIONS = 10000  # number of hypervector dimensions
 NUM_FEATURES = 5  # number of features in dataset
 
-ds = AirfoilSelfNoise("../data", download=False)
+ds = AirfoilSelfNoise("../data", download=True)
 
 # Get necessary statistics for data and target transform
 STD_DEVS = ds.data.std(0)
@@ -57,14 +57,10 @@ def __init__(self, num_classes, size):
 
         self.lr = 0.00001
         self.M = torch.zeros(1, DIMENSIONS)
-        self.project = embeddings.Projection(size, DIMENSIONS)
-        self.project.weight.data.normal_(0, 1)
-        self.bias = nn.parameter.Parameter(torch.empty(DIMENSIONS), requires_grad=False)
-        self.bias.data.uniform_(0, 2 * math.pi)
+        self.project = embeddings.Sinusoid(size, DIMENSIONS)
 
     def encode(self, x):
-        enc = self.project(x)
-        sample_hv = torch.cos(enc + self.bias) * torch.sin(enc)
+        sample_hv = self.project(x)
         return torchhd.hard_quantize(sample_hv)
 
     def model_update(self, x, y):