Merge pull request #235 from pyt-team/jude/use_new_sparse_method

Replace old sparse -> dense -> sparse with new from_sparse method

test/nn/simplicial/test_san.py

@@ -6,6 +6,7 @@
 from toponetx.classes import SimplicialComplex
 
 from topomodelx.nn.simplicial.san import SAN
+from topomodelx.utils.sparse import from_sparse
 
 
 class TestSAN:
@@ -32,9 +33,7 @@ def test_forward(self):
             simplicial_complex.add_simplex(simplex)
         x_1 = torch.randn(35, 2)
         x_0 = torch.randn(15, 2)
-        incidence_0_1 = torch.from_numpy(
-            simplicial_complex.incidence_matrix(1).todense()
-        ).to_sparse()
+        incidence_0_1 = from_sparse(simplicial_complex.incidence_matrix(1))
         x = x_1 + torch.sparse.mm(incidence_0_1.T, x_0)
         in_channels = x.shape[-1]
         hidden_channels = 16
@@ -45,12 +44,8 @@ def test_forward(self):
             out_channels=out_channels,
             n_layers=1,
         )
-        laplacian_down_1 = torch.from_numpy(
-            simplicial_complex.down_laplacian_matrix(rank=1).todense()
-        ).to_sparse()
-        laplacian_up_1 = torch.from_numpy(
-            simplicial_complex.up_laplacian_matrix(rank=1).todense()
-        ).to_sparse()
+        laplacian_down_1 = from_sparse(simplicial_complex.down_laplacian_matrix(rank=1))
+        laplacian_up_1 = from_sparse(simplicial_complex.up_laplacian_matrix(rank=1))
 
         assert torch.any(
             torch.isclose(

test/nn/simplicial/test_sca_cmps.py

@@ -6,6 +6,7 @@
 from toponetx.classes import SimplicialComplex
 
 from topomodelx.nn.simplicial.sca_cmps import SCACMPS
+from topomodelx.utils.sparse import from_sparse
 
 
 class TestSCA:
@@ -38,10 +39,10 @@ def test_forward(self):
         down_lap2 = simplicial_complex.down_laplacian_matrix(rank=2)
         incidence_1t = simplicial_complex.incidence_matrix(rank=1).T
         incidence_2t = simplicial_complex.incidence_matrix(rank=2).T
-        down_lap1 = torch.from_numpy(down_lap1.todense()).to_sparse()
-        down_lap2 = torch.from_numpy(down_lap2.todense()).to_sparse()
-        incidence_1t = torch.from_numpy(incidence_1t.todense()).to_sparse()
-        incidence_2t = torch.from_numpy(incidence_2t.todense()).to_sparse()
+        down_lap1 = from_sparse(down_lap1)
+        down_lap2 = from_sparse(down_lap2)
+        incidence_1t = from_sparse(incidence_1t)
+        incidence_2t = from_sparse(incidence_2t)
         channels_list = [x_0.shape[-1], x_1.shape[-1], x_2.shape[-1]]
         complex_dim = 3
         model = SCACMPS(

test/nn/simplicial/test_sccn.py

@@ -6,6 +6,7 @@
 from toponetx.classes import SimplicialComplex
 
 from topomodelx.nn.simplicial.sccn import SCCN
+from topomodelx.utils.sparse import from_sparse
 
 
 class TestSCCN:
@@ -38,7 +39,7 @@ def test_forward(self):
         max_rank = 2
 
         def sparse_to_torch(X):
-            return torch.from_numpy(X.todense()).to_sparse()
+            return from_sparse(X)
 
         incidences = {
             f"rank_{r}": sparse_to_torch(simplicial_complex.incidence_matrix(rank=r))

test/nn/simplicial/test_sccnn.py

@@ -6,6 +6,7 @@
 from toponetx.classes import SimplicialComplex
 
 from topomodelx.nn.simplicial.sccnn import SCCNN
+from topomodelx.utils.sparse import from_sparse
 
 
 class TestSCCNN:
@@ -46,12 +47,12 @@ def test_forward(self):
         laplacian_up_1 = simplicial_complex.up_laplacian_matrix(rank=1)
         laplacian_2 = simplicial_complex.hodge_laplacian_matrix(rank=2, weight=True)
 
-        incidence_1 = torch.from_numpy(incidence_1.todense()).to_sparse()
-        incidence_2 = torch.from_numpy(incidence_2.todense()).to_sparse()
-        laplacian_0 = torch.from_numpy(laplacian_0.todense()).to_sparse()
-        laplacian_down_1 = torch.from_numpy(laplacian_down_1.todense()).to_sparse()
-        laplacian_up_1 = torch.from_numpy(laplacian_up_1.todense()).to_sparse()
-        laplacian_2 = torch.from_numpy(laplacian_2.todense()).to_sparse()
+        incidence_1 = from_sparse(incidence_1)
+        incidence_2 = from_sparse(incidence_2)
+        laplacian_0 = from_sparse(laplacian_0)
+        laplacian_down_1 = from_sparse(laplacian_down_1)
+        laplacian_up_1 = from_sparse(laplacian_up_1)
+        laplacian_2 = from_sparse(laplacian_2)
         conv_order = 2
         intermediate_channels_all = (16, 16, 16)
         out_channels_all = intermediate_channels_all

test/nn/simplicial/test_scn2.py

@@ -6,6 +6,7 @@
 from toponetx.classes import SimplicialComplex
 
 from topomodelx.nn.simplicial.scn2 import SCN2
+from topomodelx.utils.sparse import from_sparse
 
 
 class TestSCN2:
@@ -38,9 +39,9 @@ def test_forward(self):
         laplacian_1 = simplicial_complex.normalized_laplacian_matrix(rank=1)
         laplacian_2 = simplicial_complex.normalized_laplacian_matrix(rank=2)
 
-        laplacian_0 = torch.from_numpy(laplacian_0.todense()).to_sparse()
-        laplacian_1 = torch.from_numpy(laplacian_1.todense()).to_sparse()
-        laplacian_2 = torch.from_numpy(laplacian_2.todense()).to_sparse()
+        laplacian_0 = from_sparse(laplacian_0)
+        laplacian_1 = from_sparse(laplacian_1)
+        laplacian_2 = from_sparse(laplacian_2)
         in_channels_0 = x_0.shape[1]
         in_channels_1 = x_1.shape[1]
         in_channels_2 = x_2.shape[1]

test/nn/simplicial/test_scnn.py

@@ -7,6 +7,7 @@
 from toponetx.classes import SimplicialComplex
 
 from topomodelx.nn.simplicial.scnn import SCNN
+from topomodelx.utils.sparse import from_sparse
 
 
 class TestSCNN:
@@ -41,12 +42,12 @@ def test_forward(self):
         laplacian_up_1 = simplicial_complex.up_laplacian_matrix(rank=1)
         laplacian_2 = simplicial_complex.hodge_laplacian_matrix(rank=2, weight=True)
 
-        incidence_1 = torch.from_numpy(incidence_1.todense()).to_sparse()
-        incidence_2 = torch.from_numpy(incidence_2.todense()).to_sparse()
-        laplacian_0 = torch.from_numpy(laplacian_0.todense()).to_sparse()
-        laplacian_down_1 = torch.from_numpy(laplacian_down_1.todense()).to_sparse()
-        laplacian_up_1 = torch.from_numpy(laplacian_up_1.todense()).to_sparse()
-        laplacian_2 = torch.from_numpy(laplacian_2.todense()).to_sparse()
+        incidence_1 = from_sparse(incidence_1)
+        incidence_2 = from_sparse(incidence_2)
+        laplacian_0 = from_sparse(laplacian_0)
+        laplacian_down_1 = from_sparse(laplacian_down_1)
+        laplacian_up_1 = from_sparse(laplacian_up_1)
+        laplacian_2 = from_sparse(laplacian_2)
         conv_order_down = 2
         conv_order_up = 2
         intermediate_channels = 4

topomodelx/utils/sparse.py

@@ -24,4 +24,4 @@ def from_sparse(data: _csc.csc_matrix) -> torch.Tensor:
     values = torch.FloatTensor(coo.data)
     indices = torch.LongTensor(np.vstack((coo.row, coo.col)))
 
-    return torch.sparse_coo_tensor(indices, values, coo.shape)
+    return torch.sparse_coo_tensor(indices, values, coo.shape).coalesce()

tutorials/cell/can_train.ipynb

@@ -124,7 +124,8 @@
     "import torch.nn.functional as F\n",
     "\n",
     "\n",
-    "from topomodelx.nn.cell.can import CAN"
+    "from topomodelx.nn.cell.can import CAN\n",
+    "from topomodelx.utils.sparse import from_sparse"
    ]
   },
   {
@@ -253,14 +254,12 @@
     "    adjacency_0_list.append(adjacency_0)\n",
     "\n",
     "    lower_neighborhood_t = cell_complex.down_laplacian_matrix(rank=1)\n",
-    "    lower_neighborhood_t = torch.from_numpy(lower_neighborhood_t.todense()).to_sparse()\n",
+    "    lower_neighborhood_t = from_sparse(lower_neighborhood_t)\n",
     "    lower_neighborhood_list.append(lower_neighborhood_t)\n",
     "\n",
     "    try:\n",
     "        upper_neighborhood_t = cell_complex.up_laplacian_matrix(rank=1)\n",
-    "        upper_neighborhood_t = torch.from_numpy(\n",
-    "            upper_neighborhood_t.todense()\n",
-    "        ).to_sparse()\n",
+    "        upper_neighborhood_t = from_sparse(upper_neighborhood_t)\n",
     "    except:\n",
     "        upper_neighborhood_t = np.zeros(\n",
     "            (lower_neighborhood_t.shape[0], lower_neighborhood_t.shape[0])\n",

tutorials/cell/ccxn_train.ipynb

@@ -65,7 +65,8 @@
     "import toponetx.datasets as datasets\n",
     "\n",
     "\n",
-    "from topomodelx.nn.cell.ccxn import CCXN"
+    "from topomodelx.nn.cell.ccxn import CCXN\n",
+    "from topomodelx.utils.sparse import from_sparse"
    ]
   },
   {
@@ -194,8 +195,8 @@
     "\n",
     "    incidence_2_t = cell_complex.incidence_matrix(rank=2).T\n",
     "    adjacency_0 = cell_complex.adjacency_matrix(rank=0)\n",
-    "    incidence_2_t = torch.from_numpy(incidence_2_t.todense()).to_sparse()\n",
-    "    adjacency_0 = torch.from_numpy(adjacency_0.todense()).to_sparse()\n",
+    "    incidence_2_t = from_sparse(incidence_2_t)\n",
+    "    adjacency_0 = from_sparse(adjacency_0)\n",
     "    incidence_2_t_list.append(incidence_2_t)\n",
     "    adjacency_0_list.append(adjacency_0)"
    ]

tutorials/cell/cwn_train.ipynb

@@ -61,7 +61,8 @@
     "\n",
     "import toponetx.datasets as datasets\n",
     "\n",
-    "from topomodelx.nn.cell.cwn import CWN"
+    "from topomodelx.nn.cell.cwn import CWN\n",
+    "from topomodelx.utils.sparse import from_sparse"
    ]
   },
   {
@@ -222,9 +223,9 @@
     "    adjacency_1 = cell_complex.adjacency_matrix(rank=1)\n",
     "    incidence_1_t = cell_complex.incidence_matrix(rank=1).T\n",
     "\n",
-    "    incidence_2 = torch.from_numpy(incidence_2.todense()).to_sparse()\n",
-    "    adjacency_1 = torch.from_numpy(adjacency_1.todense()).to_sparse()\n",
-    "    incidence_1_t = torch.from_numpy(incidence_1_t.todense()).to_sparse()\n",
+    "    incidence_2 = from_sparse(incidence_2)\n",
+    "    adjacency_1 = from_sparse(adjacency_1)\n",
+    "    incidence_1_t = from_sparse(incidence_1_t)\n",
     "\n",
     "    incidence_2_list.append(incidence_2)\n",
     "    adjacency_1_list.append(adjacency_1)\n",

tutorials/hypergraph/allset_train.ipynb

@@ -59,7 +59,8 @@
     "from sklearn.model_selection import train_test_split\n",
     "import toponetx.datasets as datasets\n",
     "\n",
-    "from topomodelx.nn.hypergraph.allset import AllSet"
+    "from topomodelx.nn.hypergraph.allset import AllSet\n",
+    "from topomodelx.utils.sparse import from_sparse"
    ]
   },
   {
@@ -196,7 +197,7 @@
     "incidence_1_list = []\n",
     "for simplex in simplexes:\n",
     "    incidence_1 = simplex.incidence_matrix(rank=1, signed=False)\n",
-    "    # incidence_1 = torch.from_numpy(incidence_1.todense()).to_sparse()\n",
+    "    # incidence_1 = from_sparse(incidence_1)\n",
     "    # incidence_1_list.append(incidence_1)\n",
     "    hg = simplex.to_hypergraph()\n",
     "    hg_list.append(hg)\n",
@@ -205,7 +206,7 @@
     "# Extract hypergraphs incident matrices from collected hypergraphs\n",
     "for hg in hg_list:\n",
     "    incidence_1 = hg.incidence_matrix()\n",
-    "    incidence_1 = torch.from_numpy(incidence_1.todense()).to_sparse()\n",
+    "    incidence_1 = from_sparse(incidence_1)\n",
     "    incidence_1_list.append(incidence_1)"
    ]
   },

tutorials/hypergraph/allset_transformer_train.ipynb

@@ -91,6 +91,7 @@
     "from sklearn.model_selection import train_test_split\n",
     "\n",
     "from topomodelx.nn.hypergraph.allset_transformer import AllSetTransformer\n",
+    "from topomodelx.utils.sparse import from_sparse\n",
     "\n",
     "# %load_ext autoreload\n",
     "# %autoreload 2"
@@ -236,7 +237,7 @@
     "# Extract hypergraphs incident matrices from collected hypergraphs\n",
     "for hg in hg_list:\n",
     "    incidence_1 = hg.incidence_matrix()\n",
-    "    incidence_1 = torch.from_numpy(incidence_1.todense()).to_sparse()\n",
+    "    incidence_1 = from_sparse(incidence_1)\n",
     "    incidence_1_list.append(incidence_1)"
    ]
   },

tutorials/hypergraph/dhgcn_train.ipynb

@@ -26,7 +26,8 @@
     "from sklearn.model_selection import train_test_split\n",
     "\n",
     "import toponetx.datasets as datasets\n",
-    "from topomodelx.nn.hypergraph.dhgcn import DHGCN"
+    "from topomodelx.nn.hypergraph.dhgcn import DHGCN\n",
+    "from topomodelx.utils.sparse import from_sparse"
    ]
   },
   {
@@ -204,7 +205,7 @@
     "incidence_1_list = []\n",
     "for simplex in simplexes:\n",
     "    incidence_1 = simplex.incidence_matrix(rank=1, signed=False)\n",
-    "    incidence_1 = torch.from_numpy(incidence_1.todense()).to_sparse()\n",
+    "    incidence_1 = from_sparse(incidence_1)\n",
     "    incidence_1_list.append(incidence_1)\n",
     "    hg = simplex.to_hypergraph()\n",
     "    hg_list.append(hg)"

tutorials/hypergraph/hnhn_train.ipynb

@@ -44,7 +44,8 @@
     "import toponetx.datasets.graph as graph\n",
     "from topomodelx.nn.hypergraph.hnhn_layer import HNHNLayer\n",
     "import matplotlib.pyplot as plt\n",
-    "from topomodelx.nn.hypergraph.hnhn import HNHN, HNHNNetwork"
+    "from topomodelx.nn.hypergraph.hnhn import HNHN, HNHNNetwork\n",
+    "from topomodelx.utils.sparse import from_sparse"
    ]
   },
   {
@@ -119,7 +120,7 @@
    ],
    "source": [
     "incidence_1 = dataset_sim.incidence_matrix(rank=1, signed=False)\n",
-    "incidence_1 = torch.from_numpy(incidence_1.todense()).to_sparse()\n",
+    "incidence_1 = from_sparse(incidence_1)\n",
     "print(f\"The incidence matrix B1 has shape: {incidence_1.shape}.\")"
    ]
   },

tutorials/hypergraph/hypergat_train.ipynb

@@ -59,7 +59,8 @@
     "import toponetx.datasets as datasets\n",
     "from sklearn.model_selection import train_test_split\n",
     "\n",
-    "from topomodelx.nn.hypergraph.hypergat import HyperGAT"
+    "from topomodelx.nn.hypergraph.hypergat import HyperGAT\n",
+    "from topomodelx.utils.sparse import from_sparse"
    ]
   },
   {
@@ -177,7 +178,7 @@
     "# Extract hypergraphs incident matrices from collected hypergraphs\n",
     "for hg in hg_list:\n",
     "    incidence_1 = hg.incidence_matrix()\n",
-    "    incidence_1 = torch.from_numpy(incidence_1.todense()).to_sparse()\n",
+    "    incidence_1 = from_sparse(incidence_1)\n",
     "    incidence_1_list.append(incidence_1)"
    ]
   },

tutorials/hypergraph/hypersage_train.ipynb

@@ -46,7 +46,8 @@
     "import toponetx.datasets as datasets\n",
     "from sklearn.model_selection import train_test_split\n",
     "\n",
-    "from topomodelx.nn.hypergraph.hypersage import HyperSAGE"
+    "from topomodelx.nn.hypergraph.hypersage import HyperSAGE\n",
+    "from topomodelx.utils.sparse import from_sparse"
    ]
   },
   {
@@ -164,7 +165,7 @@
     "# Extract hypergraphs incident matrices from collected hypergraphs\n",
     "for hg in hg_list:\n",
     "    incidence_1 = hg.incidence_matrix()\n",
-    "    incidence_1 = torch.from_numpy(incidence_1.todense()).to_sparse()\n",
+    "    incidence_1 = from_sparse(incidence_1)\n",
     "    incidence_1_list.append(incidence_1)"
    ]
   },

tutorials/hypergraph/unigin_train.ipynb

@@ -12,7 +12,8 @@
     "from torch_geometric.datasets import TUDataset\n",
     "from torch_geometric.utils.convert import to_networkx\n",
     "from toponetx.classes.simplicial_complex import SimplicialComplex\n",
-    "from topomodelx.nn.hypergraph.unigin import UniGIN"
+    "from topomodelx.nn.hypergraph.unigin import UniGIN\n",
+    "from topomodelx.utils.sparse import from_sparse"
    ]
   },
   {
@@ -58,7 +59,7 @@
     "incidence_1_list = []\n",
     "for hg in hg_list:\n",
     "    incidence_1 = hg.incidence_matrix()\n",
-    "    incidence_1 = torch.from_numpy(incidence_1.todense()).to_sparse()\n",
+    "    incidence_1 = from_sparse(incidence_1)\n",
     "    incidence_1_list.append(incidence_1)"
    ]
   },

tutorials/simplicial/dist2cycle_train.ipynb

@@ -38,6 +38,7 @@
     "import toponetx.datasets.graph as graph\n",
     "\n",
     "from topomodelx.nn.simplicial.dist2cycle import Dist2Cycle\n",
+    "from topomodelx.utils.sparse import from_sparse\n",
     "import numpy.linalg as npla"
    ]
   },
@@ -101,8 +102,8 @@
     "incidence_1 = dataset.incidence_matrix(rank=1)\n",
     "adjacency_0 = dataset.adjacency_matrix(rank=0)\n",
     "\n",
-    "incidence_1 = torch.from_numpy(incidence_1.todense()).to_sparse()\n",
-    "adjacency_0 = torch.from_numpy(adjacency_0.todense()).to_sparse()\n",
+    "incidence_1 = from_sparse(incidence_1)\n",
+    "adjacency_0 = from_sparse(adjacency_0)\n",
     "\n",
     "print(f\"The incidence matrix B1 has shape: {incidence_1.shape}.\")\n",
     "print(f\"The adjacency matrix A0 has shape: {adjacency_0.shape}.\")"