support num_flatten_dims=-1 of API fc. (#22634)

* support num_flatten_dims=-1 of API fc. test=develop

* fix name of class Test* and add CUDAPlace test. test=develop

python/paddle/fluid/layers/nn.py

@@ -326,6 +326,8 @@ def fc(input,
     mul_results = []
     for input_var, param_attr in helper.iter_inputs_and_params():
         input_shape = input_var.shape
+        if num_flatten_dims == -1:
+            num_flatten_dims = len(input_shape) - 1
         param_shape = [
             reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
         ] + [size]

python/paddle/fluid/tests/unittests/test_fc_op.py

@@ -16,7 +16,9 @@ import unittest
 import numpy as np
 from op_test import OpTest
 import paddle.fluid as fluid
-from paddle.fluid import Program, program_guard
+from paddle.fluid import Program, program_guard, core
+
+SEED = 2020
 
 
 def fc_refer(matrix, with_bias, with_relu=False):
@@ -131,6 +133,35 @@ class TestFCOpWithPadding(TestFCOp):
         self.matrix = MatrixGenerate(1, 4, 3, 128, 128, 2)
 
 
+class TestFcOp_NumFlattenDims_NegOne(unittest.TestCase):
+    def test_api(self):
+        startup_program = Program()
+        main_program = Program()
+        startup_program.random_seed = SEED
+        main_program.random_seed = SEED
+
+        with program_guard(main_program, startup_program):
+            input = np.random.random([2, 2, 25]).astype("float32")
+            x = fluid.layers.data(
+                name="x",
+                shape=[2, 2, 25],
+                append_batch_size=False,
+                dtype="float32")
+
+            out_1 = fluid.layers.fc(input=x, size=1, num_flatten_dims=-1)
+            out_2 = fluid.layers.fc(input=x, size=1, num_flatten_dims=2)
+
+        place = fluid.CPUPlace() if not core.is_compiled_with_cuda(
+        ) else fluid.CUDAPlace(0)
+        exe = fluid.Executor(place=place)
+        exe.run(startup_program)
+        res_1, res_2 = exe.run(main_program,
+                               feed={"x": input},
+                               fetch_list=[out_1, out_2])
+
+        assert np.array_equal(res_1, res_2)
+
+
 class TestFCOpError(unittest.TestCase):
     def test_errors(self):
         with program_guard(Program(), Program()):