[WIP] Add support for sparse arrays in fit, predict and evaluate. (ke…

…ras-team#8930)

* Support of sparse arrays as input in fit, predict and evaluate.

* Fixed small out of range bug.

* Fixed formating.

* Removed the conversion loop if training from symbolic tensors.

* Now we check for the whole feed instead of just inputs for the conversion.

* Improved tests.

* Refactoring tests.

* Duplicated tests to check sparse placeholder.

* Fixed bug. In predict(), the variable ins only have inputs.

* Fixed pep8.

keras/engine/training.py

@@ -10,6 +10,7 @@
 from keras.utils import Sequence
 from keras.utils import GeneratorEnqueuer
 from keras.utils import OrderedEnqueuer
+from scipy.sparse import issparse
 
 try:
     import queue
@@ -1157,6 +1158,13 @@ def _fit_loop(self, f, ins, out_labels=None, batch_size=None,
         for cbk in callbacks:
             cbk.validation_data = val_ins
 
+        # To prevent a slowdown, we find beforehand the arrays that need conversion.
+        feed = self._feed_inputs + self._feed_targets + self._feed_sample_weights
+        indices_for_conversion_to_dense = []
+        for i in range(len(feed)):
+            if issparse(ins[i]) and not K.is_sparse(feed[i]):
+                indices_for_conversion_to_dense.append(i)
+
         for epoch in range(initial_epoch, epochs):
             callbacks.on_epoch_begin(epoch)
             epoch_logs = {}
@@ -1210,6 +1218,9 @@ def _fit_loop(self, f, ins, out_labels=None, batch_size=None,
                     batch_logs['batch'] = batch_index
                     batch_logs['size'] = len(batch_ids)
                     callbacks.on_batch_begin(batch_index, batch_logs)
+                    for i in indices_for_conversion_to_dense:
+                        ins_batch[i] = ins_batch[i].toarray()
+
                     outs = f(ins_batch)
                     if not isinstance(outs, list):
                         outs = [outs]
@@ -1261,6 +1272,12 @@ def _predict_loop(self, f, ins, batch_size=32, verbose=0, steps=None):
                 progbar = Progbar(target=steps)
             else:
                 progbar = Progbar(target=num_samples)
+
+        indices_for_conversion_to_dense = []
+        for i in range(len(self._feed_inputs)):
+            if issparse(ins[i]) and not K.is_sparse(self._feed_inputs[i]):
+                indices_for_conversion_to_dense.append(i)
+
         if steps is not None:
             # Step-based predictions.
             # Since we do not know how many samples
@@ -1296,6 +1313,9 @@ def _predict_loop(self, f, ins, batch_size=32, verbose=0, steps=None):
                     ins_batch = _slice_arrays(ins[:-1], batch_ids) + [ins[-1]]
                 else:
                     ins_batch = _slice_arrays(ins, batch_ids)
+                for i in indices_for_conversion_to_dense:
+                    ins_batch[i] = ins_batch[i].toarray()
+
                 batch_outs = f(ins_batch)
                 if not isinstance(batch_outs, list):
                     batch_outs = [batch_outs]
@@ -1339,6 +1359,14 @@ def _test_loop(self, f, ins, batch_size=None, verbose=0, steps=None):
                 progbar = Progbar(target=steps)
             else:
                 progbar = Progbar(target=num_samples)
+
+        # To prevent a slowdown, we find beforehand the arrays that need conversion.
+        feed = self._feed_inputs + self._feed_targets + self._feed_sample_weights
+        indices_for_conversion_to_dense = []
+        for i in range(len(feed)):
+            if issparse(ins[i]) and not K.is_sparse(feed[i]):
+                indices_for_conversion_to_dense.append(i)
+
         if steps is not None:
             for step in range(steps):
                 batch_outs = f(ins)
@@ -1366,6 +1394,8 @@ def _test_loop(self, f, ins, batch_size=None, verbose=0, steps=None):
                     ins_batch = _slice_arrays(ins[:-1], batch_ids) + [ins[-1]]
                 else:
                     ins_batch = _slice_arrays(ins, batch_ids)
+                for i in indices_for_conversion_to_dense:
+                    ins_batch[i] = ins_batch[i].toarray()
 
                 batch_outs = f(ins_batch)
                 if isinstance(batch_outs, list):

tests/keras/engine/test_training.py

@@ -485,16 +485,83 @@ def gen_data(batch_sz):
     assert all(['Sequence' not in str(w_.message) for w_ in w]), 'A warning was raised for Sequence.'
 
 
+@keras_test
+def test_sparse_input_target_fit():
+    test_inputs = [sparse.random(6, 3, density=0.25).tocsr() for _ in range(2)]
+    test_outputs = [sparse.random(6, i, density=0.25).tocsr() for i in range(3, 5)]
+    in1 = Input(shape=(3,))
+    in2 = Input(shape=(3,))
+    out1 = Dropout(0.5, name='dropout')(in1)
+    out2 = Dense(4, name='dense_1')(in2)
+    model = Model([in1, in2], [out1, out2])
+    model.compile('rmsprop', 'mse')
+    model.fit(test_inputs, test_outputs, epochs=1, batch_size=2, validation_split=0.2)
+
+
+@keras_test
+def test_sparse_input_target_evaluate():
+    test_inputs = [sparse.random(6, 3, density=0.25).tocsr() for _ in range(2)]
+    test_outputs = [sparse.random(6, i, density=0.25).tocsr() for i in range(3, 5)]
+    in1 = Input(shape=(3,))
+    in2 = Input(shape=(3,))
+    out1 = Dropout(0.5, name='dropout')(in1)
+    out2 = Dense(4, name='dense_1')(in2)
+    model = Model([in1, in2], [out1, out2])
+    model.compile('rmsprop', 'mse')
+    model.evaluate(test_inputs, test_outputs, batch_size=2)
+
+
+@keras_test
+def test_sparse_input_predict():
+    test_inputs = [sparse.random(6, 3, density=0.25).tocsr() for _ in range(2)]
+    in1 = Input(shape=(3,))
+    in2 = Input(shape=(3,))
+    out1 = Dropout(0.5, name='dropout')(in1)
+    out2 = Dense(4, name='dense_1')(in2)
+    model = Model([in1, in2], [out1, out2])
+    model.compile('rmsprop', 'mse')
+    model.predict(test_inputs, batch_size=2)
+
+
+@pytest.mark.skipif(K.backend() != 'tensorflow', reason='sparse operations supported only by TensorFlow')
+@keras_test
+def test_sparse_placeholder_fit():
+    test_inputs = [sparse.random(6, 3, density=0.25).tocsr() for _ in range(2)]
+    test_outputs = [sparse.random(6, i, density=0.25).tocsr() for i in range(3, 5)]
+    in1 = Input(shape=(3,))
+    in2 = Input(shape=(3,), sparse=True)
+    out1 = Dropout(0.5, name='dropout')(in1)
+    out2 = Dense(4, name='dense_1')(in2)
+    model = Model([in1, in2], [out1, out2])
+    model.compile('rmsprop', 'mse')
+    model.fit(test_inputs, test_outputs, epochs=1, batch_size=2, validation_split=0.2)
+
+
+@pytest.mark.skipif(K.backend() != 'tensorflow', reason='sparse operations supported only by TensorFlow')
+@keras_test
+def test_sparse_placeholder_evaluate():
+    test_inputs = [sparse.random(6, 3, density=0.25).tocsr() for _ in range(2)]
+    test_outputs = [sparse.random(6, i, density=0.25).tocsr() for i in range(3, 5)]
+    in1 = Input(shape=(3,))
+    in2 = Input(shape=(3,), sparse=True)
+    out1 = Dropout(0.5, name='dropout')(in1)
+    out2 = Dense(4, name='dense_1')(in2)
+    model = Model([in1, in2], [out1, out2])
+    model.compile('rmsprop', 'mse')
+    model.evaluate(test_inputs, test_outputs, batch_size=2)
+
+
 @pytest.mark.skipif(K.backend() != 'tensorflow', reason='sparse operations supported only by TensorFlow')
 @keras_test
-def test_sparse_input_validation_split():
-    test_input = sparse.random(6, 3, density=0.25).tocsr()
-    in1 = Input(shape=(3,), sparse=True)
-    out1 = Dense(4)(in1)
-    test_output = np.random.random((6, 4))
-    model = Model(in1, out1)
+def test_sparse_placeholder_predict():
+    test_inputs = [sparse.random(6, 3, density=0.25).tocsr() for _ in range(2)]
+    in1 = Input(shape=(3,))
+    in2 = Input(shape=(3,), sparse=True)
+    out1 = Dropout(0.5, name='dropout')(in1)
+    out2 = Dense(4, name='dense_1')(in2)
+    model = Model([in1, in2], [out1, out2])
     model.compile('rmsprop', 'mse')
-    model.fit(test_input, test_output, epochs=1, batch_size=2, validation_split=0.2)
+    model.predict(test_inputs, batch_size=2)
 
 
 @keras_test