Use torch.testing.assert_close instead of approx

The tests outputs `RuntimeError: Can't call numpy() on Tensor that
requires grad. Use tensor.detach().numpy() instead.` for torch above
v1.4. The error seems to came from trying to convert tensor to scalar.

Replace the pytest approx() with `torch.testing.assert_close()`
to let torch handle the conversion.
Use `assert_allclose()` for torch versions before 1.9.0.

requirements-test.txt

@@ -4,3 +4,4 @@
 torch
 pytest>=8,<9
 pytest-cov>=2,<3
+packaging

tests/test_crf.py

@@ -2,10 +2,10 @@
 import math
 import random
 
-from pytest import approx
 import pytest
 import torch
 import torch.nn as nn
+from packaging.version import Version
 
 from torchcrf import CRF
 
@@ -116,7 +116,7 @@ def test_works_with_mask(self):
             denominator = math.log(sum(math.exp(s) for s in all_scores))
             manual_llh += numerator - denominator
 
-        assert llh.item() == approx(manual_llh)
+        assert_close(llh, manual_llh)
         llh.backward()  # ensure gradients can be computed
 
     def test_works_without_mask(self):
@@ -130,7 +130,7 @@ def test_works_without_mask(self):
         # No mask means the mask is all ones
         llh_mask = crf(emissions, tags, mask=torch.ones_like(tags).byte())
 
-        assert llh_no_mask.item() == approx(llh_mask.item())
+        assert_close(llh_no_mask, llh_mask)
 
     def test_batched_loss(self):
         crf = make_crf()
@@ -154,7 +154,7 @@ def test_batched_loss(self):
             # shape: ()
             total_llh += crf(emissions_, tags_)
 
-        assert llh.item() == approx(total_llh.item())
+        assert_close(llh, total_llh)
 
     def test_reduction_none(self):
         crf = make_crf()
@@ -186,8 +186,7 @@ def test_reduction_none(self):
             denominator = math.log(sum(math.exp(s) for s in all_scores))
             manual_llh.append(numerator - denominator)
 
-        for llh_, manual_llh_ in zip(llh, manual_llh):
-            assert llh_.item() == approx(manual_llh_)
+        assert_close(llh, torch.tensor(manual_llh))
 
     def test_reduction_mean(self):
         crf = make_crf()
@@ -219,7 +218,7 @@ def test_reduction_mean(self):
             denominator = math.log(sum(math.exp(s) for s in all_scores))
             manual_llh += numerator - denominator
 
-        assert llh.item() == approx(manual_llh / batch_size)
+        assert_close(llh, manual_llh / batch_size)
 
     def test_reduction_token_mean(self):
         crf = make_crf()
@@ -258,7 +257,7 @@ def test_reduction_token_mean(self):
             manual_llh += numerator - denominator
             n_tokens += seq_len
 
-        assert llh.item() == approx(manual_llh / n_tokens)
+        assert_close(llh, manual_llh / n_tokens)
 
     def test_batch_first(self):
         crf = make_crf()
@@ -281,7 +280,7 @@ def test_batch_first(self):
         tags = tags.transpose(0, 1)
         llh_bf = crf_bf(emissions, tags)
 
-        assert llh.item() == approx(llh_bf.item())
+        assert_close(llh, llh_bf)
 
     def test_emissions_has_bad_number_of_dimension(self):
         emissions = torch.randn(1, 2)
@@ -467,3 +466,13 @@ def test_first_timestep_mask_is_not_all_on(self):
         with pytest.raises(ValueError) as excinfo:
             crf.decode(emissions, mask=mask)
         assert 'mask of the first timestep must all be on' in str(excinfo.value)
+
+
+if Version(torch.__version__) >= Version("1.9.0"):
+
+    def assert_close(actual, expected):
+        torch.testing.assert_close(actual, expected, atol=1e-12, rtol=1e-6)
+else:
+
+    def assert_close(actual, expected):
+        torch.testing.assert_allclose(actual, expected, atol=1e-12, rtol=1e-6)