facebook/mask2former-swin-tiny-coco-instance

Original47.4Mparameters

Fixture8.0Mparameters

Layer stacks4compared

Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

Original · 4

Fixture · 4

Kind	Original	Fixture
SwinPatchMerging	3	3	✓ kept
no SwinPatchMerging	1	1	✓ kept

Original · 3

Fixture · 3

Kind	Original	Fixture
Conv2d	3	3	✓ kept

Original · 6

Fixture · 2

Kind	Original	Fixture
FormerPixelDecoderEncoderLayer	6	2	✓ kept

Original · 9

Fixture · 1

Kind	Original	Fixture
FormerMaskedAttentionDecoderLayer	9	1	✓ kept

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original

Fixture

Component	Original	Share	Fixture	Share
model.pixel_level_module.encoder.swin.encoder.layers	27.5M	58.0%	4.9M	60.7%
model.transformer_module.decoder.layers	14.2M	30.0%	659.3K	8.2%
model.pixel_level_module.decoder.encoder.layers	4.4M	9.3%	1.4M	17.3%
model.pixel_level_module	691.3K	1.5%	680.0K	8.4%
model.pixel_level_module.decoder.input_projections	346.4K	0.7%	231.7K	2.9%
model.transformer_module	249.9K	0.5%	191.0K	2.4%
class_predictor	20.8K	0.0%	10.4K	0.1%

Config	Property	Original	Fixture
model	model_type	mask2former	mask2former	✓ kept
backbone_config	model_type	swin	swin	✓ kept
backbone_config	tie_word_embeddings	True	True	✓ kept
backbone_config	hidden_act	gelu	gelu	✓ kept
backbone_config	image_size	224	224	✓ kept
backbone_config	patch_size	4	4	✓ kept
backbone_config	num_channels	3	3	✓ kept

Instances of each class in the built model (prefix Mask2 omitted).

Class	Original	Fixture
Linear	169	49
LayerNorm	73	25
Dropout	13	5
GELUActivation	12	4
ReLU	12	4
SwinAttention	12	4
SwinLayer	12	4
SwinMLP	12	4
SwinRelativePositionBias	12	4
SwinDropPath	11	3
FormerAttention	9	1
FormerMaskedAttentionDecoderLayer	9	1
ModuleList	9	9
MultiheadAttention	9	1
NonDynamicallyQuantizableLinear	9	1
Sequential	8	5
Conv2d	7	10
FormerPixelDecoderEncoderLayer	6	2
FormerPixelDecoderEncoderMultiscaleDeformableAttention	6	2
GroupNorm	5	5
SwinStage	4	4
Embedding	3	3
FormerPredictionBlock	3	3
SwinPatchMerging	3	3
Identity	2	2
FormerSinePositionEmbedding	2	2
FormerForUniversalSegmentation	1	1
FormerHungarianMatcher	1	1
FormerLoss	1	1
FormerMLPPredictionHead	1	1
FormerMaskPredictor	1	1
FormerMaskedAttentionDecoder	1	1
FormerModel	1	1
FormerPixelDecoder	1	1
FormerPixelDecoderEncoderOnly	1	1
FormerPixelLevelModule	1	1
FormerTransformerModule	1	1
ModuleDict	1	1
SwinBackbone	1	1
SwinEmbeddings	1	1
SwinEncoder	1	1
SwinModel	1	1
SwinPatchEmbeddings	1	1

9 fields changed, 78 unchanged.