make the threshold for acurracy more precise (#17194)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17194

we found that there is a per row absolute error due to int8 quant
and a relative error table-wide in case fp16 is used

Reviewed By: csummersea

Differential Revision: D14113353

fbshipit-source-id: c7065aa9d15c453c2e5609f421ad0155145af889

1 files changed, 23 insertions, 13 deletions

diff --git a/caffe2/python/lengths_reducer_fused_8bit_rowwise_ops_test.py b/caffe2/python/lengths_reducer_fused_8bit_rowwise_ops_test.py
index 8f36dc02b7..8835092366 100644
--- a/caffe2/python/lengths_reducer_fused_8bit_rowwise_ops_test.py
+++ b/caffe2/python/lengths_reducer_fused_8bit_rowwise_ops_test.py
@@ -7,18 +7,28 @@ from caffe2.python import core, workspace
 from hypothesis import given
 
 
-def compare_rowwise(emb_orig, emb_reconstructed):
+def compare_rowwise(emb_orig, emb_reconstructed, fp16):
+    # there is an absolute error introduced per row through int8 quantization
+    # and a relative error introduced when quantizing back from fp32 to fp16
     assert(emb_orig.shape == emb_reconstructed.shape)
-    range = np.amax(emb_orig, axis=1) - np.amin(emb_orig, axis=1)
-    # TOOO: figure out the right threshold, this has to do with the
-    # fact that the data types are float16, in float32, it should be /1.9
-    threshold = range / 255.0 / 1.5
-    diff = np.amax(np.abs(emb_orig - emb_reconstructed), axis=1)
-    n_violated = ((threshold - diff) < 0).sum()
-    if n_violated > 0:
-        print(n_violated, threshold, diff, threshold < diff, emb_orig,
-              emb_reconstructed, emb_orig - emb_reconstructed)
-    assert(n_violated == 0)
+    rtol = 1e-8
+    if fp16:
+        rtol = 1e-3
+    erange = np.amax(emb_orig, axis=1) - np.amin(emb_orig, axis=1)
+
+    threshold = erange / 255.0 / 1.9
+
+    for i in range(emb_orig.shape[0]):
+        r_orig = emb_orig[i, :]
+        r_reconstructed = emb_reconstructed[i, :]
+
+        isclose = np.isclose(r_orig, r_reconstructed, atol=threshold[i], rtol=rtol)
+        n_violated = isclose.size - isclose.sum()
+
+        if n_violated > 0:
+            print(isclose, threshold[i])
+            print(i, r_orig, r_reconstructed, threshold[i], r_orig - r_reconstructed)
+        assert(n_violated == 0)
 
 
 class TestLengthsReducerOpsFused8BitRowwise(hu.HypothesisTestCase):
@@ -102,7 +112,7 @@ class TestLengthsReducerOpsFused8BitRowwise(hu.HypothesisTestCase):
 
         dequantized_data = workspace.FetchBlob("dequantized_data")
         np.testing.assert_array_almost_equal(input_data, workspace.FetchBlob("input_data"))
-        compare_rowwise(input_data, dequantized_data)
+        compare_rowwise(input_data, dequantized_data, fp16)
 
         sum_reference = workspace.FetchBlob("sum_reference")
         sum_quantized = workspace.FetchBlob("sum_quantized")
@@ -179,7 +189,7 @@ class TestLengthsReducerOpsFused8BitRowwise(hu.HypothesisTestCase):
 
         dequantized_data = workspace.FetchBlob("dequantized_data")
         np.testing.assert_array_almost_equal(input_data, workspace.FetchBlob("input_data"))
-        compare_rowwise(input_data, dequantized_data)
+        compare_rowwise(input_data, dequantized_data, fp16)
 
         mean_reference = workspace.FetchBlob("mean_reference")
         mean_quantized = workspace.FetchBlob("mean_quantized")