Add an example of a differentiable model (#176)

* Add an example of a differentiable model

* Fix numpy arrays warning

examples/learning_with_hrr.py

@@ -0,0 +1,207 @@
+# A partial implementation of https://arxiv.org/abs/2109.02157 
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.utils.data import Dataset, DataLoader
+import torch.optim as optim
+import torch.optim.lr_scheduler as lr_scheduler
+
+# Note: this example requires the napkinXC library: https://napkinxc.readthedocs.io/
+from napkinxc.datasets import load_dataset
+from napkinxc.measures import precision_at_k
+
+from tqdm import tqdm
+import torchhd
+from torchhd import embeddings, HRRTensor
+import torchhd.tensors
+from scipy.sparse import  vstack, lil_matrix
+import numpy as np
+
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+print("Using {} device".format(device))
+
+
+DIMENSIONS = 400 
+NUMBER_OF_EPOCHS = 1
+BATCH_SIZE = 1  
+DATASET_NAME =  "eurlex-4k" # tested on "eurlex-4k", and "Wiki10-31K"
+FC_LAYER_SIZE = 512
+
+
+def sparse_batch_collate(batch:list): 
+    """
+    Collate function which to transform scipy csr matrix to pytorch sparse tensor
+    """
+    data_batch, targets_batch = zip(*batch)
+
+    data_batch = vstack(data_batch).tocoo()
+    data_batch = torch.sparse_coo_tensor(np.array(data_batch.nonzero()), data_batch.data, data_batch.shape)
+
+    targets_batch = torch.stack(targets_batch)
+
+    return data_batch, targets_batch
+
+class multilabel_dataset(Dataset):
+    def __init__(self,x,y,n_classes) -> None:
+        self.x = x
+        self.y = y
+        self.n_classes = n_classes
+
+
+    # Define the length of the dataset.
+    def __len__(self):
+        return self.x.shape[0]
+
+    # Return a single sample from the dataset.
+    def __getitem__(self, idx):
+        labels = torch.zeros(self.n_classes, dtype=torch.int64)
+        labels[self.y[idx]] = 1.0
+        return self.x[idx], labels
+
+
+X_train, Y_train = load_dataset(DATASET_NAME, "train", verbose=True)
+X_test, Y_test = load_dataset(DATASET_NAME, "test", verbose=True)
+
+
+if DATASET_NAME == "Wiki10-31K": # Because of this issue https://github.com/mwydmuch/napkinXC/issues/18
+    X_train = lil_matrix(X_train[:,:-1])
+
+N_features = X_train.shape[1]
+N_classes   = max(max(classes) for classes in Y_train if classes != []) + 1
+
+train_dataset = multilabel_dataset(X_train,Y_train,N_classes)
+train_dataloader = DataLoader(train_dataset,BATCH_SIZE, collate_fn=sparse_batch_collate) 
+test_dataset  = multilabel_dataset(X_test,Y_test,N_classes)
+test_dataloader = DataLoader(test_dataset,collate_fn=sparse_batch_collate) 
+
+
+print("Traning on \033[1m {} \033[0m. It has {} features, and {} classes."
+      .format(DATASET_NAME,N_features,N_classes))
+
+
+# Fully Connected model for the baseline comparision 
+class FC(nn.Module):
+    def __init__(self, num_features, num_classes):
+        super(FC, self).__init__()
+        self.num_classes = num_classes
+        self.num_features = num_features
+        self.fc_layer_size = FC_LAYER_SIZE
+
+        # Network Layers
+        self.fc1 = nn.Linear(self.num_features, self.fc_layer_size)
+        self.fc2 = nn.Linear(self.fc_layer_size, self.fc_layer_size )
+        self.olayer = nn.Linear(self.fc_layer_size, self.num_classes)
+
+    def forward(self, x):
+        x = F.leaky_relu(self.fc1(x))
+        x = F.leaky_relu(self.fc2(x))
+        x = self.olayer(x)
+        return x
+
+    def pred(self, out,threshold=0.5):
+        y = F.sigmoid(out)
+        v,i = y.sort(descending=True)
+        ids = i[v>=threshold]
+        ids = ids.tolist()     
+        return ids
+
+    def loss(self,out,target):
+        loss = nn.BCEWithLogitsLoss()(out, target.type(torch.float64))
+        return loss
+
+# Modified version of FC model that returns an HRRTensor with dim << output of the FC model. 
+# It makes the model to have fewer parameters  
+class FCHRR(nn.Module):
+    def __init__(self, num_features, num_classes,dim):
+        super(FCHRR, self).__init__()
+        self.num_classes = num_classes
+        self.num_features = num_features
+        self.fc_layer_size = FC_LAYER_SIZE
+        self.dim = dim
+
+        self.classes_vec = embeddings.Random(N_classes, dim,vsa="HRR")
+        n_vec, p_vec = torchhd.HRRTensor.random(2,dim)
+        self.register_buffer("n_vec", n_vec)
+        self.register_buffer("p_vec", p_vec)
+
+        # Network Layers
+        self.fc1 = nn.Linear(self.num_features, self.fc_layer_size)
+        self.fc2 = nn.Linear(self.fc_layer_size, self.fc_layer_size )
+        self.olayer = nn.Linear(self.fc_layer_size, dim)
+
+    def forward(self, x):
+        x = F.leaky_relu(self.fc1(x))
+        x = F.leaky_relu(self.fc2(x))
+        x = self.olayer(x)
+        return x.as_subclass(HRRTensor)
+
+    def pred(self, out,threshold=0.1):
+
+        tmp_positive = self.p_vec.exact_inverse().bind(out)
+        sims = tmp_positive.cosine_similarity(self.classes_vec.weight)
+
+        v,i = sims.sort(descending=True)
+        ids = i[v>=threshold]
+        ids = ids.tolist()     
+
+        return ids
+
+    def loss(self,out,target):
+
+        loss = torch.tensor(0, dtype=torch.float32,device=device)
+
+        tmp_positives  = self.p_vec.exact_inverse().bind(out)
+        tmp_negatives = self.n_vec.exact_inverse().bind(out)
+        for i in range(target.shape[0]):
+
+            cp = self.classes_vec.weight[target[i]==1,:]
+
+            j_p = (1 - tmp_positives[i].cosine_similarity(cp)).sum()
+            j_n = tmp_negatives[i].cosine_similarity(cp.multibundle())
+
+            loss += j_p + j_n
+
+        loss /= target.shape[0]
+
+        return loss
+
+
+
+hrr_model = FCHRR(N_features,N_classes,DIMENSIONS)
+hrr_model = hrr_model.to(device)
+
+baseline_model = FC(N_features,N_classes)
+baseline_model = baseline_model.to(device)
+
+
+for model_name, model in {"HRR-FC":hrr_model,"FC":baseline_model}.items():
+    optimizer = optim.Adam(model.parameters(), lr=0.001)
+    scheduler = lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.7)
+    model.train()
+    for epoch in tqdm(range(1,NUMBER_OF_EPOCHS + 1), desc=f"{model_name} epochs",leave=False):
+
+        for samples, labels in tqdm(train_dataloader, desc="Training",leave=False):
+            samples = samples.to(device)
+            labels = labels.to(device)
+            optimizer.zero_grad()
+            out = model(samples)
+            loss = model.loss(out, labels)
+            loss.backward()
+            optimizer.step()
+
+        scheduler.step()
+
+    Y_pred = []
+    model.eval()
+    with torch.no_grad():
+        for data, target in tqdm(test_dataloader,desc="Validating",leave=False):
+            data, target = data.to(device).float(), target.to(device)
+            out = model(data) 
+            ids = model.pred(out)
+            Y_pred.append(ids)
+
+    # Calculating the P@1 metric 
+    p_at_1 = precision_at_k(Y_test, Y_pred, k=1)[0]
+    print("Result of {} model ---->  P@1 = {}".format(model_name, p_at_1))