Re-enable detectron2 model accuracy checks. (#2317)

Summary:
Detectron2 models can pass the accuracy check when the deterministic mode is disabled.

Revert abed1e7.
Fixes #2290

Pull Request resolved: #2317

Reviewed By: aaronenyeshi

Differential Revision: D58794036

Pulled By: xuzhao9

fbshipit-source-id: b0c38bd63c63b71e33f9692ad2169f8267e93ab1

torchbenchmark/models/detectron2_fasterrcnn_r_101_c4/metadata.yaml

@@ -6,5 +6,3 @@ eval_deterministic: false
 eval_nograd: true
 train_benchmark: false
 train_deterministic: false
-not_implemented:
- - test: example

torchbenchmark/models/detectron2_fasterrcnn_r_101_dc5/metadata.yaml

@@ -6,6 +6,5 @@ eval_deterministic: false
 eval_nograd: true
 not_implemented:
 - device: cpu
-- test: example
 train_benchmark: false
 train_deterministic: false

torchbenchmark/models/detectron2_fasterrcnn_r_101_fpn/metadata.yaml

@@ -6,6 +6,5 @@ eval_deterministic: false
 eval_nograd: true
 not_implemented:
 - device: cpu
-- test: example
 train_benchmark: false
 train_deterministic: false

torchbenchmark/models/detectron2_fasterrcnn_r_50_c4/metadata.yaml

@@ -6,6 +6,5 @@ eval_deterministic: false
 eval_nograd: true
 not_implemented:
 - device: cpu
-- test: example
 train_benchmark: false
 train_deterministic: false

torchbenchmark/models/detectron2_fasterrcnn_r_50_dc5/metadata.yaml

@@ -6,6 +6,5 @@ eval_deterministic: false
 eval_nograd: true
 not_implemented:
 - device: cpu
-- test: example
 train_benchmark: false
 train_deterministic: false

torchbenchmark/models/detectron2_fasterrcnn_r_50_fpn/metadata.yaml

@@ -6,6 +6,5 @@ eval_deterministic: false
 eval_nograd: true
 not_implemented:
 - device: cpu
-- test: example
 train_benchmark: false
 train_deterministic: false

torchbenchmark/models/detectron2_maskrcnn/__init__.py

@@ -58,9 +58,7 @@ class Model(BenchmarkModel):
  model_file = os.path.join(MODEL_DIR, ".data", f"{MODEL_NAME}.pkl")
  DEFAULT_TRAIN_BSIZE = 1
  DEFAULT_EVAL_BSIZE = 1
- # Skip correctness check, because the output tensor can't be verified using
- # cosine similarity or torch.close()
- SKIP_CORRECTNESS_CHECK = True
+ DISABLE_DETERMINISM = True
 
  def __init__(self, test, device, batch_size=None, extra_args=[]):
  super().__init__(

torchbenchmark/models/detectron2_maskrcnn/metadata.yaml

@@ -6,6 +6,5 @@ eval_deterministic: false
 eval_nograd: true
 not_implemented:
 - device: cpu
-- test: example
 train_benchmark: false
 train_deterministic: false

torchbenchmark/models/detectron2_maskrcnn_r_101_c4/metadata.yaml

@@ -6,6 +6,5 @@ eval_deterministic: false
 eval_nograd: true
 not_implemented:
 - device: cpu
-- test: example
 train_benchmark: false
 train_deterministic: false

torchbenchmark/models/detectron2_maskrcnn_r_101_fpn/metadata.yaml

@@ -6,6 +6,5 @@ eval_deterministic: false
 eval_nograd: true
 not_implemented:
 - device: cpu
-- test: example
 train_benchmark: false
 train_deterministic: false

torchbenchmark/models/detectron2_maskrcnn_r_50_c4/metadata.yaml

@@ -6,6 +6,5 @@ eval_deterministic: false
 eval_nograd: true
 not_implemented:
 - device: cpu
-- test: example
 train_benchmark: false
 train_deterministic: false

torchbenchmark/models/detectron2_maskrcnn_r_50_fpn/metadata.yaml

@@ -6,6 +6,5 @@ eval_deterministic: false
 eval_nograd: true
 not_implemented:
 - device: cpu
-- test: example
 train_benchmark: false
 train_deterministic: false

torchbenchmark/models/maml/__init__.py

@@ -16,9 +16,6 @@ class Model(BenchmarkModel):
  DEFAULT_EVAL_BSIZE = 1
  ALLOW_CUSTOMIZE_BSIZE = False
  CANNOT_SET_CUSTOM_OPTIMIZER = True
- # Skip correctness check, because maml runs backward and optimizer in eval()
- # Which will return non-deterministic results
- SKIP_CORRECTNESS_CHECK = True
 
  def __init__(self, test, device, batch_size=None, extra_args=[]):
  super().__init__(

torchbenchmark/util/framework/detectron2/model_factory.py

@@ -90,9 +90,7 @@ class Detectron2Model(BenchmarkModel):
  # Default batch sizes
  DEFAULT_TRAIN_BSIZE = 1
  DEFAULT_EVAL_BSIZE = 1
- # Skip correctness check, because the output tensor can't be verified using
- # cosine similarity or torch.close()
- SKIP_CORRECTNESS_CHECK = True
+ DISABLE_DETERMINISM = True
 
  def __init__(self, variant, test, device, batch_size=None, extra_args=[]):
  super().__init__(