mark unit tests as working, skip failing unit test (pytorch#12313)

Summary:
* enabled fp16 tests for test_torch

* enable fp16 tests for test_nn

* enabled multilabelmargin loss for fp16

* removed skip for test_pdist_empty_col

* Enable test_nn tests that pass with compiler fixes etc.

* Enable test_legacy_nn tests that pass with compiler fixes etc.

ezyang bddppq
Pull Request resolved: pytorch#12313

Differential Revision: D10189922

Pulled By: bddppq

fbshipit-source-id: a5592817c04b14e355cb062d42ebea406f0c92b6

Author: iotamudelta

test/common_nn.py

@@ -41,7 +41,6 @@ def get_weight(m):
         constructor_args=(10, 8),
         input_size=(4, 10),
         reference_fn=lambda i, p: torch.mm(i, p[0].t()) + p[1].view(1, -1).expand(4, 8),
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='Linear',
@@ -103,35 +102,30 @@ def get_weight(m):
         constructor_args=(1,),
         input_size=(10, 20),
         reference_fn=lambda i, _: torch.exp(i).div(torch.exp(i).sum(1, True).expand(10, 20)),
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='Softmax2d',
         input_size=(1, 3, 10, 20),
         reference_fn=lambda i, _: torch.exp(i).div(torch.exp(i).sum(1, False)),
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='LogSoftmax',
         constructor_args=(1,),
         input_size=(10, 20),
         reference_fn=lambda i, _: torch.exp(i).div_(torch.exp(i).sum(1, True).expand(10, 20)).log_(),
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='LogSoftmax',
         constructor_args=(1,),
         input_size=(1, 3, 10, 20),
         reference_fn=lambda i, _: torch.exp(i).div_(torch.exp(i).sum(1, False)).log_(),
         desc='multiparam',
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='ELU',
         constructor_args=(2.,),
         input_size=(3, 2, 5),
         reference_fn=lambda x, _: torch.where(x >= 0, x, 2 * (x.exp() - 1)),
-        test_cuda=(not TEST_WITH_ROCM),
     ),
     # TODO: reference function
     dict(
@@ -204,7 +198,6 @@ def get_weight(m):
         input_size=(2, 3, 4),
         desc='1d_multiparam',
         reference_fn=lambda i, p: torch.clamp(i, min=0) + torch.clamp(i, max=0) * p[0][0],
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='PReLU',
@@ -218,7 +211,6 @@ def get_weight(m):
         input_size=(2, 3, 4, 5),
         desc='2d_multiparam',
         reference_fn=lambda i, p: torch.clamp(i, min=0) + torch.clamp(i, max=0) * p[0][0],
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='PReLU',
@@ -232,31 +224,26 @@ def get_weight(m):
         input_size=(2, 3, 4, 5, 6),
         desc='3d_multiparam',
         reference_fn=lambda i, p: torch.clamp(i, min=0) + torch.clamp(i, max=0) * p[0][0],
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='Softsign',
         input_size=(3, 2, 5),
         reference_fn=lambda i, _: i.div(1 + torch.abs(i)),
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='Softmin',
         constructor_args=(1,),
         input_size=(10, 20),
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='Softmin',
         constructor_args=(1,),
         input_size=(2, 3, 5, 10),
         desc='multidim',
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='Tanhshrink',
         input_size=(2, 3, 4, 5),
-        test_cuda=(not TEST_WITH_ROCM)
     ),
 ]
 
@@ -573,7 +560,6 @@ def ctcloss_reference(log_probs, targets, input_lengths, target_lengths, blank=0
         reference_fn=lambda i, t, m:
             kldivloss_reference(i, t, get_reduction(m)),
         check_sum_reduction=True,
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='MSELoss',
@@ -590,7 +576,6 @@ def ctcloss_reference(log_probs, targets, input_lengths, target_lengths, blank=0
         reference_fn=lambda i, t, m: -(t * i.log() + (1 - t) * (1 - i).log()).sum() /
             (i.numel() if get_reduction(m) else 1),
         check_gradgrad=False,
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='BCELoss',
@@ -601,7 +586,6 @@ def ctcloss_reference(log_probs, targets, input_lengths, target_lengths, blank=0
             (i.numel() if get_reduction(m) else 1),
         desc='weights',
         check_gradgrad=False,
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='CrossEntropyLoss',
@@ -660,7 +644,6 @@ def ctcloss_reference(log_probs, targets, input_lengths, target_lengths, blank=0
         target_fn=lambda: torch.rand(5, 10).mul(2).floor(),
         reference_fn=lambda i, t, m: -(t * i.sigmoid().log() + (1 - t) * (-i).sigmoid().log()).sum() / i.numel(),
         check_gradgrad=False,
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='MultiMarginLoss',
@@ -759,7 +742,6 @@ def ctcloss_reference(log_probs, targets, input_lengths, target_lengths, blank=0
         reference_fn=lambda i, t, m:
             marginrankingloss_reference(i[0], i[1], t, reduction=get_reduction(m)),
         check_sum_reduction=True,
-        test_cuda=(not TEST_WITH_ROCM)
     ),
     dict(
         module_name='MarginRankingLoss',
@@ -770,7 +752,6 @@ def ctcloss_reference(log_probs, targets, input_lengths, target_lengths, blank=0
             marginrankingloss_reference(i[0], i[1], t, margin=0.5, reduction=get_reduction(m)),
         desc='margin',
         check_sum_reduction=True,
-        test_cuda=(not TEST_WITH_ROCM)
     ),
 ]