Update several models to be compatible with TensorFlow 2.0 (#504)

* Change to tensorflow.compat.v1 for WMF model

* Change to tensorflow.compat.v1 for CDL model

* Change to tensorflow.compat.v1 for CDR model

* Change to tensorflow.compat.v1 for ConvMF model

* Fix compatibility issues for CVAE

* Update requirements.txt for NCF model

* Fix numpy dtype compatibility issue

Author: tqtg

cornac/data/dataset.py

@@ -272,7 +272,7 @@ def dok_matrix(self):
         """The user-item interaction matrix in DOK sparse format"""
         if self.__dok_matrix is None:
             self.__dok_matrix = dok_matrix(
-                (self.num_users, self.num_items), dtype=np.float32
+                (self.num_users, self.num_items), dtype='float'
             )
             for u, i, r in zip(*self.uir_tuple):
                 self.__dok_matrix[u, i] = r
@@ -364,13 +364,13 @@ def build(
             raise ValueError("data is empty after being filtered!")
 
         uir_tuple = (
-            np.asarray(u_indices, dtype=np.int),
-            np.asarray(i_indices, dtype=np.int),
-            np.asarray(r_values, dtype=np.float),
+            np.asarray(u_indices, dtype='int'),
+            np.asarray(i_indices, dtype='int'),
+            np.asarray(r_values, dtype='float'),
         )
 
         timestamps = (
-            np.fromiter((int(data[i][3]) for i in valid_idx), dtype=np.int)
+            np.fromiter((int(data[i][3]) for i in valid_idx), dtype='int')
             if fmt == "UIRT"
             else None
         )
@@ -447,7 +447,7 @@ def idx_iter(self, idx_range, batch_size=1, shuffle=False):
 
         Returns
         -------
-        iterator : batch of indices (array of np.int)
+        iterator : batch of indices (array of 'int')
 
         """
         indices = np.arange(idx_range)
@@ -481,8 +481,8 @@ def uir_iter(self, batch_size=1, shuffle=False, binary=False, num_zeros=0):
 
         Returns
         -------
-        iterator : batch of users (array of np.int), batch of items (array of np.int),
-            batch of ratings (array of np.float)
+        iterator : batch of users (array of 'int'), batch of items (array of 'int'),
+            batch of ratings (array of 'float')
 
         """
         for batch_ids in self.idx_iter(len(self.uir_tuple[0]), batch_size, shuffle):
@@ -524,8 +524,8 @@ def uij_iter(self, batch_size=1, shuffle=False, neg_sampling="uniform"):
 
         Returns
         -------
-        iterator : batch of users (array of np.int), batch of positive items (array of np.int),
-            batch of negative items (array of np.int)
+        iterator : batch of users (array of 'int'), batch of positive items (array of 'int'),
+            batch of negative items (array of 'int')
 
         """
 
@@ -562,9 +562,9 @@ def user_iter(self, batch_size=1, shuffle=False):
 
         Returns
         -------
-        iterator : batch of user indices (array of np.int)
+        iterator : batch of user indices (array of 'int')
         """
-        user_indices = np.fromiter(self.user_indices, dtype=np.int)
+        user_indices = np.fromiter(self.user_indices, dtype='int')
         for batch_ids in self.idx_iter(len(user_indices), batch_size, shuffle):
             yield user_indices[batch_ids]
 
@@ -580,9 +580,9 @@ def item_iter(self, batch_size=1, shuffle=False):
 
         Returns
         -------
-        iterator : batch of item indices (array of np.int)
+        iterator : batch of item indices (array of 'int')
         """
-        item_indices = np.fromiter(self.item_indices, np.int)
+        item_indices = np.fromiter(self.item_indices, 'int')
         for batch_ids in self.idx_iter(len(item_indices), batch_size, shuffle):
             yield item_indices[batch_ids]
 

cornac/data/graph.py

@@ -61,9 +61,9 @@ def _build_triplet(self, id_map):
             self.map_cid.append(id_map[j])
             self.val.append(v)
 
-        self.map_rid = np.asarray(self.map_rid, dtype=np.int)
-        self.map_cid = np.asarray(self.map_cid, dtype=np.int)
-        self.val = np.asarray(self.val, dtype=np.float)
+        self.map_rid = np.asarray(self.map_rid, dtype='int')
+        self.map_cid = np.asarray(self.map_cid, dtype='int')
+        self.val = np.asarray(self.val, dtype='float')
 
     def build(self, id_map=None, **kwargs):
         super().build(id_map=id_map)

cornac/data/text.py

@@ -914,7 +914,7 @@ def batch_seq(self, batch_ids, max_length=None):
         if max_length is None:
             max_length = max(len(self.sequences[mapped_id]) for mapped_id in batch_ids)
 
-        seq_mat = np.zeros((len(batch_ids), max_length), dtype=np.int)
+        seq_mat = np.zeros((len(batch_ids), max_length), dtype='int')
         for i, mapped_id in enumerate(batch_ids):
             idx_seq = self.sequences[mapped_id][:max_length]
             for j, idx in enumerate(idx_seq):

cornac/eval_methods/base_method.py

@@ -79,7 +79,7 @@ def rating_eval(model, metrics, test_set, user_based=False, verbose=False):
             miniters=100,
             total=len(u_indices),
         ),
-        dtype=np.float,
+        dtype='float',
     )
 
     gt_mat = test_set.csr_matrix
@@ -177,7 +177,7 @@ def pos_items(csr_row):
         if len(test_pos_items) == 0:
             continue
 
-        u_gt_pos = np.zeros(test_set.num_items, dtype=np.int)
+        u_gt_pos = np.zeros(test_set.num_items, dtype='int')
         u_gt_pos[test_pos_items] = 1
 
         val_pos_items = [] if val_mat is None else pos_items(val_mat.getrow(user_idx))
@@ -187,7 +187,7 @@ def pos_items(csr_row):
             else pos_items(train_mat.getrow(user_idx))
         )
 
-        u_gt_neg = np.ones(test_set.num_items, dtype=np.int)
+        u_gt_neg = np.ones(test_set.num_items, dtype='int')
         u_gt_neg[test_pos_items + val_pos_items + train_pos_items] = 0
 
         item_indices = None if exclude_unknowns else np.arange(test_set.num_items)

cornac/eval_methods/propensity_stratified_evaluation.py

@@ -87,7 +87,7 @@ def pos_items(csr_row):
         if len(test_pos_items) == 0:
             continue
 
-        u_gt_pos = np.zeros(test_set.num_items, dtype=np.float)
+        u_gt_pos = np.zeros(test_set.num_items, dtype='float')
         u_gt_pos[test_pos_items] = 1
 
         val_pos_items = [] if val_mat is None else pos_items(val_mat.getrow(user_idx))
@@ -97,7 +97,7 @@ def pos_items(csr_row):
             else pos_items(train_mat.getrow(user_idx))
         )
 
-        u_gt_neg = np.ones(test_set.num_items, dtype=np.int)
+        u_gt_neg = np.ones(test_set.num_items, dtype='int')
         u_gt_neg[test_pos_items + val_pos_items + train_pos_items] = 0
 
         item_indices = None if exclude_unknowns else np.arange(test_set.num_items)
@@ -256,7 +256,7 @@ def _estimate_propensities(self):
             item_freq[i] += 1
 
         # fit the exponential param
-        data = np.array([e for e in item_freq.values()], dtype=np.float)
+        data = np.array([e for e in item_freq.values()], dtype='float')
         results = powerlaw.Fit(data, discrete=True, fit_method="Likelihood")
         alpha = results.power_law.alpha
         fmin = results.power_law.xmin
@@ -277,7 +277,7 @@ def _build_stratified_dataset(self, test_data):
 
         # match the corresponding propensity score for each feedback
         test_props = np.array(
-            [self.props[i] for u, i, r in test_data], dtype=np.float64
+            [self.props[i] for u, i, r in test_data], dtype='float'
         )
 
         # stratify

cornac/metrics/ranking.py

@@ -437,8 +437,8 @@ def compute(self, pd_scores, gt_pos, gt_neg=None, **kwargs):
         if gt_neg is None:
             gt_neg = np.logical_not(gt_pos)
 
-        pos_scores = pd_scores[gt_pos.astype(np.bool)]
-        neg_scores = pd_scores[gt_neg.astype(np.bool)]
+        pos_scores = pd_scores[gt_pos.astype('bool')]
+        neg_scores = pd_scores[gt_neg.astype('bool')]
         ui_scores = np.repeat(pos_scores, len(neg_scores))
         uj_scores = np.tile(neg_scores, len(pos_scores))
 
@@ -476,7 +476,7 @@ def compute(self, pd_scores, gt_pos, **kwargs):
             AP score.
 
         """
-        relevant = gt_pos.astype(np.bool)
+        relevant = gt_pos.astype('bool')
         rank = rankdata(-pd_scores, "max")[relevant]
         L = rankdata(-pd_scores[relevant], "max")
         ans = (L / rank).mean()

cornac/models/cdl/recom_cdl.py

@@ -181,6 +181,8 @@ def fit(self, train_set, val_set=None):
     def _fit_cdl(self):
         import tensorflow.compat.v1 as tf
         from .cdl import Model
+        
+        tf.disable_eager_execution()
 
         R = self.train_set.csc_matrix  # csc for efficient slicing over items
         n_users, n_items = self.train_set.num_users, self.train_set.num_items

cornac/models/cdl/requirements.txt

@@ -1 +1 @@
-tensorflow==1.15.2
+tensorflow==2.12.0

cornac/models/cdr/model.py cornac/models/cdr/cdr.pycornac/models/cdr/model.py renamed to cornac/models/cdr/cdr.py

@@ -14,7 +14,7 @@
 # ============================================================================
 """Collaborative Deep Ranking model"""
 
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 
 from ..cdl.cdl import sdae
 

cornac/models/cdr/recom_cdr.py

@@ -168,8 +168,10 @@ def fit(self, train_set, val_set=None):
         return self
 
     def _fit_cdr(self):
-        import tensorflow as tf
-        from .model import Model
+        import tensorflow.compat.v1 as tf
+        from .cdr import Model
+
+        tf.disable_eager_execution()
 
         n_users = self.train_set.num_users
         n_items = self.train_set.num_items