Experimenting with tiny test models, how small can they go and be useful for regression tests?

timm/models/byobnet.py

@@ -1834,6 +1834,18 @@ model_cfgs = dict(
         stem_type='one',
         stem_chs=64,
     ),
+
+    test_tiny_resnet=ByoModelCfg(
+        blocks=(
+            ByoBlockCfg(type='basic', d=1, c=24, s=1, gs=1, br=0.25),
+            ByoBlockCfg(type='basic', d=1, c=32, s=2, gs=1, br=0.25),
+            ByoBlockCfg(type='basic', d=1, c=64, s=2, gs=1, br=0.25),
+            ByoBlockCfg(type='basic', d=1, c=128, s=2, gs=1, br=0.25),
+        ),
+        stem_chs=32,
+        stem_pool='maxpool',
+        act_layer='relu',
+    ),
 )
 
 
@@ -2035,6 +2047,11 @@ default_cfgs = generate_default_cfgs({
         crop_pct=0.9,
         first_conv=('stem.conv_kxk.0.conv', 'stem.conv_scale.conv'),
     ),
+
+    'test_tiny_byobnet.untrained': _cfgr(
+        # hf_hub_id='timm/',
+        input_size=(3, 160, 160), crop_pct=0.875, pool_size=(5, 5),
+    ),
 })
 
 
@@ -2337,3 +2354,10 @@ def mobileone_s4(pretrained=False, **kwargs) -> ByobNet:
     """
     """
     return _create_byobnet('mobileone_s4', pretrained=pretrained, **kwargs)
+
+
+@register_model
+def test_tiny_byobnet(pretrained=False, **kwargs) -> ByobNet:
+    """ Minimal test ResNet (BYOB based) model.
+    """
+    return _create_byobnet('test_tiny_byobnet', pretrained=pretrained, **kwargs)

timm/models/efficientnet.py

@@ -1056,6 +1056,31 @@ def _gen_tinynet(
     return model
 
 
+def _gen_test_efficientnet(
+        variant, channel_multiplier=1.0, depth_multiplier=1.0, pretrained=False, **kwargs):
+    """ Minimal test EfficientNet generator.
+    """
+    arch_def = [
+        ['cn_r1_k3_s1_e1_c16_skip'],
+        ['er_r1_k3_s2_e4_c24'],
+        ['er_r1_k3_s2_e4_c32'],
+        ['ir_r1_k3_s2_e4_c48_se0.25'],
+        ['ir_r1_k3_s2_e4_c64_se0.25'],
+    ]
+    round_chs_fn = partial(round_channels, multiplier=channel_multiplier, round_limit=0.)
+    model_kwargs = dict(
+        block_args=decode_arch_def(arch_def, depth_multiplier),
+        num_features=round_chs_fn(256),
+        stem_size=24,
+        round_chs_fn=round_chs_fn,
+        norm_layer=kwargs.pop('norm_layer', None) or partial(nn.BatchNorm2d, **resolve_bn_args(kwargs)),
+        act_layer=resolve_act_layer(kwargs, 'silu'),
+        **kwargs,
+    )
+    model = _create_effnet(variant, pretrained, **model_kwargs)
+    return model
+
+
 def _cfg(url='', **kwargs):
     return {
         'url': url, 'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': (7, 7),
@@ -1584,6 +1609,10 @@ default_cfgs = generate_default_cfgs({
         input_size=(3, 106, 106), pool_size=(4, 4),  # int(224 * 0.475)
         url='https://github.com/huawei-noah/CV-Backbones/releases/download/v1.2.0/tinynet_e.pth',
         hf_hub_id='timm/'),
+
+    "test_tiny_efficientnet.untrained": _cfg(
+        # hf_hub_id='timm/'
+        input_size=(3, 160, 160), pool_size=(5, 5)),
 })
 
 
@@ -2510,6 +2539,13 @@ def tinynet_e(pretrained=False, **kwargs) -> EfficientNet:
     return model
 
 
+@register_model
+def test_tiny_efficientnet(pretrained=False, **kwargs) -> EfficientNet:
+    model = _gen_test_efficientnet('test_tiny_efficientnet', pretrained=pretrained, **kwargs)
+    return model
+
+
+
 register_model_deprecations(__name__, {
     'tf_efficientnet_b0_ap': 'tf_efficientnet_b0.ap_in1k',
     'tf_efficientnet_b1_ap': 'tf_efficientnet_b1.ap_in1k',

timm/models/vision_transformer.py

@@ -1929,13 +1929,16 @@ default_cfgs = {
         hf_hub_id='timm/',
         num_classes=11821,
         input_size=(3, 256, 256), crop_pct=0.95),
-    'vit_base_patch16_reg4_gap_256': _cfg(
+    'vit_base_patch16_reg4_gap_256.untrained': _cfg(
         input_size=(3, 256, 256)),
 
-    'vit_so150m_patch16_reg4_gap_256': _cfg(
+    'vit_so150m_patch16_reg4_gap_256.untrained': _cfg(
         input_size=(3, 256, 256)),
-    'vit_so150m_patch16_reg4_map_256': _cfg(
+    'vit_so150m_patch16_reg4_map_256.untrained': _cfg(
         input_size=(3, 256, 256)),
+
+    'test_tiny_vit.untrained': _cfg(
+        input_size=(3, 160, 160), crop_pct=0.875),
 }
 
 _quick_gelu_cfgs = [
@@ -3106,6 +3109,15 @@ def vit_so150m_patch16_reg4_gap_256(pretrained: bool = False, **kwargs) -> Visio
     return model
 
 
+@register_model
+def test_tiny_vit(pretrained: bool = False, **kwargs) -> VisionTransformer:
+    """ ViT-TestTiny
+    """
+    model_args = dict(patch_size=16, embed_dim=64, depth=4, num_heads=1, mlp_ratio=3)
+    model = _create_vision_transformer('test_tiny_vit', pretrained=pretrained, **dict(model_args, **kwargs))
+    return model
+
+
 register_model_deprecations(__name__, {
     'vit_tiny_patch16_224_in21k': 'vit_tiny_patch16_224.augreg_in21k',
     'vit_small_patch32_224_in21k': 'vit_small_patch32_224.augreg_in21k',