Update README.md
Browse files
README.md
CHANGED
|
@@ -105,27 +105,6 @@ this table.
|
|
| 105 |
|
| 106 |
The BERT-CNN row above chooses among all nine candidates. The two-candidate setting is shown below.
|
| 107 |
|
| 108 |
-
Gains over Best Beam, with paired bootstrap 95% intervals over images:
|
| 109 |
-
|
| 110 |
-
| Metric | Gain |
|
| 111 |
-
| --- | --- |
|
| 112 |
-
| CIDEr | **+0.0325 [+0.021, +0.044]** |
|
| 113 |
-
| SPICE | +0.0068 [+0.005, +0.009] |
|
| 114 |
-
| BERTScore | +0.0007 [+0.0004, +0.0010] |
|
| 115 |
-
|
| 116 |
-
Two retrainings with seeds 43 and 44 gave CIDEr 1.1418 and 1.1415. Both are
|
| 117 |
-
significantly above Best Beam.
|
| 118 |
-
|
| 119 |
-
**Two-candidate setting (first two beams only).** On the same 5,000 images, the retrained
|
| 120 |
-
BERT-CNN reaches CIDEr 1.1317 here, close to the paper's reported 1.131, but falls below
|
| 121 |
-
Best Beam with nine candidates. This model improves on Best Beam in both settings.
|
| 122 |
-
|
| 123 |
-
| Method | CIDEr, 2 candidates | CIDEr, 9 candidates |
|
| 124 |
-
| --- | ---: | ---: |
|
| 125 |
-
| Best Beam | 1.1150 | 1.1150 |
|
| 126 |
-
| Retrained PyTorch BERT-CNN | 1.1317 | 1.0855 |
|
| 127 |
-
| **This model** | **1.1383** | **1.1475** |
|
| 128 |
-
|
| 129 |
**Ablations on the same benchmark (CIDEr):**
|
| 130 |
|
| 131 |
| Variant | CIDEr |
|
|
|
|
| 105 |
|
| 106 |
The BERT-CNN row above chooses among all nine candidates. The two-candidate setting is shown below.
|
| 107 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 108 |
**Ablations on the same benchmark (CIDEr):**
|
| 109 |
|
| 110 |
| Variant | CIDEr |
|