Merge pull request #10 from histocartography/feature/model_zoo/gja

Feature/model zoo/gja

histocartography/interpretability/grad_cam.py

@@ -179,6 +179,8 @@ def _get_weights(self, class_idx, scores):
         # Backpropagate
         self._backprop(scores, class_idx)
 
+        self.backward_hook = list(reversed(self.backward_hook))
+
         # Compute alpha
         grad_2 = [f.pow(2) for f in self.backward_hook]
         grad_3 = [f.pow(3) for f in self.backward_hook]

histocartography/ml/layers/dense_gin_layer.py

@@ -85,7 +85,7 @@ def forward(self, adj, h):
             adj = adj / degree
 
         if self.add_self:
-            adj = adj + torch.eye(adj.size(1)).to(adj.device)
+            adj = adj.float() + torch.eye(adj.size(1)).to(adj.device)
 
         # adjust h dim
         if len(h.shape) < 3:

histocartography/ml/layers/gin_layer.py

@@ -148,7 +148,7 @@ def _compute_adj_lrp(self, relevance_score):
 
         adjacency_matrix = torch.clamp(self.adjacency_matrix, min=0)
         if self.agg_type == 'mean':
-            adjacency_matrix = torch.div(adjacency_matrix, self.in_degrees)
+            adjacency_matrix = torch.div(adjacency_matrix, self.in_degrees.to(adjacency_matrix.device))
         adjacency_matrix = adjacency_matrix + \
             torch.eye(self.adjacency_matrix.shape[0]).to(relevance_score.device)
         rel_unnorm = torch.mm(self.input_features, adjacency_matrix.t()) + 1e-9

histocartography/ml/layers/pna_layer.py

@@ -205,14 +205,11 @@ def forward(self, g, h):
 
         # graph and batch normalization
         if self.graph_norm:
-            if isinstance(
-                    g, dgl.DGLGraph) or isinstance(
-                    g, dgl.DGLHeteroGraph):
-                num_nodes = [g.number_of_nodes()]
-            else:
+            if hasattr(g, 'batch_num_nodes'):
                 num_nodes = g.batch_num_nodes
-            snorm_n = torch.FloatTensor(list(itertools.chain(
-                *[[np.sqrt(1 / n)] * n for n in num_nodes]))).to(h.device)
+            else:
+                num_nodes = [g.number_of_nodes()]
+            snorm_n = torch.FloatTensor(list(itertools.chain(*[[np.sqrt(1 / n)] * n for n in num_nodes]))).to(h.device)
             h = h * snorm_n[:, None]
         if self.batch_norm:
             h = self.batchnorm_h(h)

histocartography/ml/models/zoo.py

@@ -1,24 +1,21 @@
 # The models follow the naming convention:
 # <dataset_name>_<model_type>_<number_of_classes>_classes_<gnn_type>.pt
-# e.g., in bracs_tggnn_5_classes_pna:
+# e.g., in bracs_tggnn_7_classes_pna:
 # -btrained on bracs dataset
 # - with tggnn (Tissue Graph GNN) model
-# - and 5 classes
+# - and 7 classes
 # - with PNA GNN layers
 
 MODEL_NAME_TO_URL = {
     # CG-GNN
     'bracs_cggnn_3_classes_gin.pt': 'https://ibm.box.com/shared/static/pozkx0ngqjxdr34v5tpmckthts8m12u3.pt',
-    'bracs_cggnn_5_classes_pna.pt': 'https://ibm.box.com/shared/static/8yj33c4o5precry9sev5g6585vmijeso.pt',
     'bracs_cggnn_5_classes_gin.pt': 'https://ibm.box.com/shared/static/uy2xeovpo001mb1edkwms5ccz1sqhdpz.pt',
-    'bracs_cggnn_7_classes_pna.pt': 'https://ibm.box.com/shared/static/i4xixoglstzkif53rc2cm4b568ei5wnf.pt',
+    'bracs_cggnn_7_classes_pna.pt': 'https://ibm.box.com/shared/static/i4xixoglstzkif53rc2cm4b568ei5wnf.pt', 
     # TG-GNN
     'bracs_tggnn_3_classes_gin.pt': 'https://ibm.box.com/shared/static/aoogy0516lsp9vaxgw1tr9mdu5nycvvb.pt',
-    'bracs_tggnn_5_classes_pna.pt': 'https://ibm.box.com/shared/static/qvdr4j12fwo7zuute6wsgrmvxfxe1o8w.pt',
-    'bracs_tggnn_7_classes_pna.pt': 'https://ibm.box.com/shared/static/19q7kk2humvc6a8qedzg8rs5bny6qvrf.pt',
+    'bracs_tggnn_7_classes_pna.pt': 'https://ibm.box.com/shared/static/19q7kk2humvc6a8qedzg8rs5bny6qvrf.pt', 
     # HACT
-    'bracs_hact_5_classes_pna.pt': 'https://ibm.box.com/shared/static/efar14ic4mc13u5kidn6q23xzshh9oao.pt',
-    'bracs_hact_7_classes_pna.pt': 'https://ibm.box.com/shared/static/5v44c33cipdy7c2dhajkrfaywyrh2a5o.pt',
+    'bracs_hact_7_classes_pna.pt': 'https://ibm.box.com/shared/static/5v44c33cipdy7c2dhajkrfaywyrh2a5o.pt', 
 }
 
 MODEL_NAME_TO_CONFIG = {
@@ -43,31 +40,6 @@
             'hidden_dim': 128,
         }
     },
-    'bracs_cggnn_5_classes_pna.pt': {
-        'node_dim': 514,
-        'gnn_params': {
-            'layer_type': "pna_layer",
-            'output_dim': 64,
-            'num_layers': 3,
-            'readout_op': "concat",
-            'readout_type': "mean",
-            'aggregators': "mean max min std",
-            'scalers': "identity amplification attenuation",
-            'avg_d': 4,
-            'dropout': 0.,
-            'graph_norm': True,
-            'batch_norm': True,
-            'towers': 1,
-            'pretrans_layers': 1,
-            'posttrans_layers': 1,
-            'divide_input': True,
-            'residual': True,
-        },
-        'classification_params': {
-            'num_layers': 2,
-            'hidden_dim': 128,
-        }
-    },
     'bracs_cggnn_5_classes_gin.pt': {
         'node_dim': 514,
         'gnn_params': {
@@ -93,8 +65,8 @@
         'gnn_params': {
             'layer_type': "pna_layer",
             'output_dim': 64,
-            'num_layers': 5,
-            'readout_op': "concat",
+            'num_layers': 3,
+            'readout_op': "lstm",
             'readout_type': "mean",
             'aggregators': "mean max min std",
             'scalers': "identity amplification attenuation",
@@ -105,8 +77,8 @@
             'towers': 1,
             'pretrans_layers': 1,
             'posttrans_layers': 1,
-            'divide_input': True,
-            'residual': True,
+            'divide_input': False,
+            'residual': False,
         },
         'classification_params': {
             'num_layers': 2,
@@ -134,63 +106,12 @@
             'hidden_dim': 128,
         }
     },
-    'bracs_tggnn_5_classes_pna.pt': {
-        'node_dim': 514,
-        'gnn_params': {
-            'layer_type': "pna_layer",
-            'output_dim': 64,
-            'num_layers': 5,
-            'readout_op': "concat",
-            'readout_type': "mean",
-            'aggregators': "mean max min std",
-            'scalers': "identity amplification attenuation",
-            'avg_d': 4,
-            'dropout': 0.,
-            'graph_norm': True,
-            'batch_norm': True,
-            'towers': 1,
-            'pretrans_layers': 1,
-            'posttrans_layers': 1,
-            'divide_input': True,
-            'residual': True,
-        },
-        'classification_params': {
-            'num_layers': 2,
-            'hidden_dim': 128,
-        }
-    },
     'bracs_tggnn_7_classes_pna.pt': {
         'node_dim': 514,
         'gnn_params': {
             'layer_type': "pna_layer",
             'output_dim': 64,
             'num_layers': 3,
-            'readout_op': "concat",
-            'readout_type': "mean",
-            'aggregators': "mean max min std",
-            'scalers': "identity amplification attenuation",
-            'avg_d': 4,
-            'dropout': 0.,
-            'graph_norm': True,
-            'batch_norm': True,
-            'towers': 1,
-            'pretrans_layers': 1,
-            'posttrans_layers': 1,
-            'divide_input': True,
-            'residual': True,
-        },
-        'classification_params': {
-            'num_layers': 2,
-            'hidden_dim': 128,
-        }
-    },
-    # HACT
-    'bracs_hact_5_classes_pna.pt': {
-        'cg_node_dim': 514,
-        'cg_gnn_params': {
-            'layer_type': "pna_layer",
-            'output_dim': 64,
-            'num_layers': 4,
             'readout_op': "lstm",
             'readout_type': "mean",
             'aggregators': "mean max min std",
@@ -202,39 +123,21 @@
             'towers': 1,
             'pretrans_layers': 1,
             'posttrans_layers': 1,
-            'divide_input': True,
-            'residual': True,
-        },
-        'tg_node_dim': 514,
-        'tg_gnn_params': {
-            'layer_type': "pna_layer",
-            'output_dim': 64,
-            'num_layers': 4,
-            'readout_op': "lstm",
-            'readout_type': "mean",
-            'aggregators': "mean max min std",
-            'scalers': "identity amplification attenuation",
-            'avg_d': 4,
-            'dropout': 0.,
-            'graph_norm': True,
-            'batch_norm': True,
-            'towers': 1,
-            'pretrans_layers': 1,
-            'posttrans_layers': 1,
-            'divide_input': True,
-            'residual': True,
+            'divide_input': False,
+            'residual': False,
         },
         'classification_params': {
             'num_layers': 2,
             'hidden_dim': 128,
         }
     },
+    # HACT
     'bracs_hact_7_classes_pna.pt': {
         'cg_node_dim': 514,
         'cg_gnn_params': {
             'layer_type': "pna_layer",
             'output_dim': 64,
-            'num_layers': 4,
+            'num_layers': 3,
             'readout_op': "lstm",
             'readout_type': "mean",
             'aggregators': "mean max min std",
@@ -246,14 +149,14 @@
             'towers': 1,
             'pretrans_layers': 1,
             'posttrans_layers': 1,
-            'divide_input': True,
-            'residual': True,
+            'divide_input': False,
+            'residual': False,
         },
         'tg_node_dim': 514,
         'tg_gnn_params': {
             'layer_type': "pna_layer",
             'output_dim': 64,
-            'num_layers': 4,
+            'num_layers': 3,
             'readout_op': "lstm",
             'readout_type': "mean",
             'aggregators': "mean max min std",
@@ -265,8 +168,8 @@
             'towers': 1,
             'pretrans_layers': 1,
             'posttrans_layers': 1,
-            'divide_input': True,
-            'residual': True,
+            'divide_input': False,
+            'residual': False,
         },
         'classification_params': {
             'num_layers': 2,