Merge pull request huggingface#637 from rwightman/levit_visformer_rednet

LeVit, Visformer, RedNet/Involution models and layers

Author: rwightman

.github/workflows/tests.yml

@@ -17,8 +17,8 @@ jobs:
       matrix:
         os: [ubuntu-latest, macOS-latest]
         python: ['3.8']
-        torch: ['1.8.0']
-        torchvision: ['0.9.0']
+        torch: ['1.8.1']
+        torchvision: ['0.9.1']
     runs-on: ${{ matrix.os }}
 
     steps:

README.md

@@ -23,6 +23,14 @@ I'm fortunate to be able to dedicate significant time and money of my own suppor
 
 ## What's New
 
+### May 25, 2021
+* Add LeViT, Visformer, ConViT (PR by Aman Arora), Twins (PR by paper authors) transformer models
+* Add ResMLP and gMLP MLP vision models to the existing MLP Mixer impl
+* Fix a number of torchscript issues with various vision transformer models
+* Cleanup input_size/img_size override handling and improve testing / test coverage for all vision transformer and MLP models
+* More flexible pos embedding resize (non-square) for ViT and TnT. Thanks [Alexander Soare](https://github.com/alexander-soare)
+* Add `efficientnetv2_rw_m` model and weights (started training before official code). 84.8 top-1, 53M params.
+
 ### May 14, 2021
 * Add EfficientNet-V2 official model defs w/ ported weights from official [Tensorflow/Keras](https://github.com/google/automl/tree/master/efficientnetv2) impl.
   * 1k trained variants: `tf_efficientnetv2_s/m/l`
@@ -166,30 +174,6 @@ I'm fortunate to be able to dedicate significant time and money of my own suppor
 * Misc fixes for SiLU ONNX export, default_cfg missing from Feature extraction models, Linear layer w/ AMP + torchscript
 * PyPi release @ 0.3.2 (needed by EfficientDet)
 
-### Oct 30, 2020
-* Test with PyTorch 1.7 and fix a small top-n metric view vs reshape issue.
-* Convert newly added 224x224 Vision Transformer weights from official JAX repo. 81.8 top-1 for B/16, 83.1 L/16.
-* Support PyTorch 1.7 optimized, native SiLU (aka Swish) activation. Add mapping to 'silu' name, custom swish will eventually be deprecated.
-* Fix regression for loading pretrained classifier via direct model entrypoint functions. Didn't impact create_model() factory usage.
-* PyPi release @ 0.3.0 version!
-
-### Oct 26, 2020
-* Update Vision Transformer models to be compatible with official code release at https://github.com/google-research/vision_transformer
-* Add Vision Transformer weights (ImageNet-21k pretrain) for 384x384 base and large models converted from official jax impl
-  * ViT-B/16 - 84.2
-  * ViT-B/32 - 81.7
-  * ViT-L/16 - 85.2
-  * ViT-L/32 - 81.5
-
-### Oct 21, 2020
-* Weights added for Vision Transformer (ViT) models. 77.86 top-1 for 'small' and 79.35 for 'base'. Thanks to [Christof](https://www.kaggle.com/christofhenkel) for training the base model w/ lots of GPUs.
-
-### Oct 13, 2020
-* Initial impl of Vision Transformer models. Both patch and hybrid (CNN backbone) variants. Currently trying to train...
-* Adafactor and AdaHessian (FP32 only, no AMP) optimizers
-* EdgeTPU-M (`efficientnet_em`) model trained in PyTorch, 79.3 top-1
-* Pip release, doc updates pending a few more changes...
-
 
 ## Introduction
 
@@ -207,6 +191,7 @@ A full version of the list below with source links can be found in the [document
 * Bottleneck Transformers - https://arxiv.org/abs/2101.11605
 * CaiT (Class-Attention in Image Transformers) - https://arxiv.org/abs/2103.17239
 * CoaT (Co-Scale Conv-Attentional Image Transformers) - https://arxiv.org/abs/2104.06399
+* ConViT (Soft Convolutional Inductive Biases Vision Transformers)- https://arxiv.org/abs/2103.10697
 * CspNet (Cross-Stage Partial Networks) - https://arxiv.org/abs/1911.11929
 * DeiT (Vision Transformer) - https://arxiv.org/abs/2012.12877
 * DenseNet - https://arxiv.org/abs/1608.06993
@@ -224,13 +209,15 @@ A full version of the list below with source links can be found in the [document
     * MobileNet-V2 - https://arxiv.org/abs/1801.04381
     * Single-Path NAS - https://arxiv.org/abs/1904.02877
 * GhostNet - https://arxiv.org/abs/1911.11907
+* gMLP - https://arxiv.org/abs/2105.08050
 * GPU-Efficient Networks - https://arxiv.org/abs/2006.14090
 * Halo Nets - https://arxiv.org/abs/2103.12731
 * HardCoRe-NAS - https://arxiv.org/abs/2102.11646
 * HRNet - https://arxiv.org/abs/1908.07919
 * Inception-V3 - https://arxiv.org/abs/1512.00567
 * Inception-ResNet-V2 and Inception-V4 - https://arxiv.org/abs/1602.07261
 * Lambda Networks - https://arxiv.org/abs/2102.08602
+* LeViT (Vision Transformer in ConvNet's Clothing) - https://arxiv.org/abs/2104.01136
 * MLP-Mixer - https://arxiv.org/abs/2105.01601
 * MobileNet-V3 (MBConvNet w/ Efficient Head) - https://arxiv.org/abs/1905.02244
 * NASNet-A - https://arxiv.org/abs/1707.07012
@@ -240,6 +227,7 @@ A full version of the list below with source links can be found in the [document
 * Pooling-based Vision Transformer (PiT) - https://arxiv.org/abs/2103.16302
 * RegNet - https://arxiv.org/abs/2003.13678
 * RepVGG - https://arxiv.org/abs/2101.03697
+* ResMLP - https://arxiv.org/abs/2105.03404
 * ResNet/ResNeXt
     * ResNet (v1b/v1.5) - https://arxiv.org/abs/1512.03385
     * ResNeXt - https://arxiv.org/abs/1611.05431
@@ -257,6 +245,7 @@ A full version of the list below with source links can be found in the [document
 * Swin Transformer - https://arxiv.org/abs/2103.14030
 * Transformer-iN-Transformer (TNT) - https://arxiv.org/abs/2103.00112
 * TResNet - https://arxiv.org/abs/2003.13630
+* Twins (Spatial Attention in Vision Transformers) - https://arxiv.org/pdf/2104.13840.pdf
 * Vision Transformer - https://arxiv.org/abs/2010.11929
 * VovNet V2 and V1 - https://arxiv.org/abs/1911.06667
 * Xception - https://arxiv.org/abs/1610.02357

docs/archived_changes.md

@@ -1,5 +1,29 @@
 # Archived Changes
 
+### Oct 30, 2020
+* Test with PyTorch 1.7 and fix a small top-n metric view vs reshape issue.
+* Convert newly added 224x224 Vision Transformer weights from official JAX repo. 81.8 top-1 for B/16, 83.1 L/16.
+* Support PyTorch 1.7 optimized, native SiLU (aka Swish) activation. Add mapping to 'silu' name, custom swish will eventually be deprecated.
+* Fix regression for loading pretrained classifier via direct model entrypoint functions. Didn't impact create_model() factory usage.
+* PyPi release @ 0.3.0 version!
+
+### Oct 26, 2020
+* Update Vision Transformer models to be compatible with official code release at https://github.com/google-research/vision_transformer
+* Add Vision Transformer weights (ImageNet-21k pretrain) for 384x384 base and large models converted from official jax impl
+  * ViT-B/16 - 84.2
+  * ViT-B/32 - 81.7
+  * ViT-L/16 - 85.2
+  * ViT-L/32 - 81.5
+
+### Oct 21, 2020
+* Weights added for Vision Transformer (ViT) models. 77.86 top-1 for 'small' and 79.35 for 'base'. Thanks to [Christof](https://www.kaggle.com/christofhenkel) for training the base model w/ lots of GPUs.
+
+### Oct 13, 2020
+* Initial impl of Vision Transformer models. Both patch and hybrid (CNN backbone) variants. Currently trying to train...
+* Adafactor and AdaHessian (FP32 only, no AMP) optimizers
+* EdgeTPU-M (`efficientnet_em`) model trained in PyTorch, 79.3 top-1
+* Pip release, doc updates pending a few more changes...
+
 ### Sept 18, 2020
 * New ResNet 'D' weights. 72.7 (top-1) ResNet-18-D, 77.1 ResNet-34-D, 80.5 ResNet-50-D
 * Added a few untrained defs for other ResNet models (66D, 101D, 152D, 200/200D)

docs/changes.md

@@ -1,5 +1,33 @@
 # Recent Changes
 
+### May 25, 2021
+* Add LeViT, Visformer, Convit (PR by Aman Arora), Twins (PR by paper authors) transformer models
+* Cleanup input_size/img_size override handling and testing for all vision transformer models
+* Add `efficientnetv2_rw_m` model and weights (started training before official code). 84.8 top-1, 53M params.
+
+### May 14, 2021
+* Add EfficientNet-V2 official model defs w/ ported weights from official [Tensorflow/Keras](https://github.com/google/automl/tree/master/efficientnetv2) impl.
+  * 1k trained variants: `tf_efficientnetv2_s/m/l`
+  * 21k trained variants: `tf_efficientnetv2_s/m/l_in21k`
+  * 21k pretrained -> 1k fine-tuned: `tf_efficientnetv2_s/m/l_in21ft1k`
+  * v2 models w/ v1 scaling: `tf_efficientnetv2_b0` through `b3`
+  * Rename my prev V2 guess `efficientnet_v2s` -> `efficientnetv2_rw_s`
+  * Some blank `efficientnetv2_*` models in-place for future native PyTorch training
+
+### May 5, 2021
+* Add MLP-Mixer models and port pretrained weights from [Google JAX impl](https://github.com/google-research/vision_transformer/tree/linen)
+* Add CaiT models and pretrained weights from [FB](https://github.com/facebookresearch/deit)
+* Add ResNet-RS models and weights from [TF](https://github.com/tensorflow/tpu/tree/master/models/official/resnet/resnet_rs). Thanks [Aman Arora](https://github.com/amaarora)
+* Add CoaT models and weights. Thanks [Mohammed Rizin](https://github.com/morizin)
+* Add new ImageNet-21k weights & finetuned weights for TResNet, MobileNet-V3, ViT models. Thanks [mrT](https://github.com/mrT23)
+* Add GhostNet models and weights. Thanks [Kai Han](https://github.com/iamhankai)
+* Update ByoaNet attention modles
+   * Improve SA module inits
+   * Hack together experimental stand-alone Swin based attn module and `swinnet`
+   * Consistent '26t' model defs for experiments.
+* Add improved Efficientnet-V2S (prelim model def) weights. 83.8 top-1.
+* WandB logging support
+
 ### April 13, 2021
 * Add Swin Transformer models and weights from https://github.com/microsoft/Swin-Transformer
 

tests/test_models.py

@@ -16,7 +16,8 @@
 
 # transformer models don't support many of the spatial / feature based model functionalities
 NON_STD_FILTERS = [
-    'vit_*', 'tnt_*', 'pit_*', 'swin_*', 'coat_*', 'cait_*', '*mixer_*', 'gmlp_*', 'resmlp_*', 'twins_*', 'convit_*']
+    'vit_*', 'tnt_*', 'pit_*', 'swin_*', 'coat_*', 'cait_*', '*mixer_*', 'gmlp_*', 'resmlp_*', 'twins_*',
+    'convit_*', 'levit*', 'visformer*']
 NUM_NON_STD = len(NON_STD_FILTERS)
 
 # exclude models that cause specific test failures
@@ -25,29 +26,56 @@
     EXCLUDE_FILTERS = [
         '*efficientnet_l2*', '*resnext101_32x48d', '*in21k', '*152x4_bitm', '*101x3_bitm',
         '*nfnet_f3*', '*nfnet_f4*', '*nfnet_f5*', '*nfnet_f6*', '*nfnet_f7*', 
-        '*resnetrs350*', '*resnetrs420*'] + NON_STD_FILTERS
+        '*resnetrs350*', '*resnetrs420*']
 else:
-    EXCLUDE_FILTERS = NON_STD_FILTERS
+    EXCLUDE_FILTERS = []
 
-MAX_FWD_SIZE = 384
-MAX_BWD_SIZE = 128
-MAX_FWD_FEAT_SIZE = 448
+TARGET_FWD_SIZE = MAX_FWD_SIZE = 384
+TARGET_BWD_SIZE = 128
+MAX_BWD_SIZE = 320
+MAX_FWD_OUT_SIZE = 448
+TARGET_JIT_SIZE = 128
+MAX_JIT_SIZE = 320
+TARGET_FFEAT_SIZE = 96
+MAX_FFEAT_SIZE = 256
+
+
+def _get_input_size(model=None, model_name='', target=None):
+    if model is None:
+        assert model_name, "One of model or model_name must be provided"
+        input_size = get_model_default_value(model_name, 'input_size')
+        fixed_input_size = get_model_default_value(model_name, 'fixed_input_size')
+        min_input_size = get_model_default_value(model_name, 'min_input_size')
+    else:
+        default_cfg = model.default_cfg
+        input_size = default_cfg['input_size']
+        fixed_input_size = default_cfg.get('fixed_input_size', None)
+        min_input_size = default_cfg.get('min_input_size', None)
+    assert input_size is not None
+
+    if fixed_input_size:
+        return input_size
+
+    if min_input_size:
+        if target and max(input_size) > target:
+            input_size = min_input_size
+    else:
+        if target and max(input_size) > target:
+            input_size = tuple([min(x, target) for x in input_size])
+    return input_size
 
 
 @pytest.mark.timeout(120)
-@pytest.mark.parametrize('model_name', list_models(exclude_filters=EXCLUDE_FILTERS[:-NUM_NON_STD]))
+@pytest.mark.parametrize('model_name', list_models(exclude_filters=EXCLUDE_FILTERS))
 @pytest.mark.parametrize('batch_size', [1])
 def test_model_forward(model_name, batch_size):
     """Run a single forward pass with each model"""
     model = create_model(model_name, pretrained=False)
     model.eval()
 
-    input_size = model.default_cfg['input_size']
-    if any([x > MAX_FWD_SIZE for x in input_size]):
-        if is_model_default_key(model_name, 'fixed_input_size'):
-            pytest.skip("Fixed input size model > limit.")
-        # cap forward test at max res 384 * 384 to keep resource down
-        input_size = tuple([min(x, MAX_FWD_SIZE) for x in input_size])
+    input_size = _get_input_size(model=model, target=TARGET_FWD_SIZE)
+    if max(input_size) > MAX_FWD_SIZE:
+        pytest.skip("Fixed input size model > limit.")
     inputs = torch.randn((batch_size, *input_size))
     outputs = model(inputs)
 
@@ -56,26 +84,22 @@ def test_model_forward(model_name, batch_size):
 
 
 @pytest.mark.timeout(120)
-@pytest.mark.parametrize('model_name', list_models(exclude_filters=EXCLUDE_FILTERS))
+@pytest.mark.parametrize('model_name', list_models(exclude_filters=EXCLUDE_FILTERS, name_matches_cfg=True))
 @pytest.mark.parametrize('batch_size', [2])
 def test_model_backward(model_name, batch_size):
     """Run a single forward pass with each model"""
+    input_size = _get_input_size(model_name=model_name, target=TARGET_BWD_SIZE)
+    if max(input_size) > MAX_BWD_SIZE:
+        pytest.skip("Fixed input size model > limit.")
+
     model = create_model(model_name, pretrained=False, num_classes=42)
     num_params = sum([x.numel() for x in model.parameters()])
-    model.eval()
-
-    input_size = model.default_cfg['input_size']
-    if not is_model_default_key(model_name, 'fixed_input_size'):
-        min_input_size = get_model_default_value(model_name, 'min_input_size')
-        if min_input_size is not None:
-            input_size = min_input_size
-        else:
-            if any([x > MAX_BWD_SIZE for x in input_size]):
-                # cap backward test at 128 * 128 to keep resource usage down
-                input_size = tuple([min(x, MAX_BWD_SIZE) for x in input_size])
+    model.train()
 
     inputs = torch.randn((batch_size, *input_size))
     outputs = model(inputs)
+    if isinstance(outputs, tuple):
+        outputs = torch.cat(outputs)
     outputs.mean().backward()
     for n, x in model.named_parameters():
         assert x.grad is not None, f'No gradient for {n}'
@@ -100,10 +124,10 @@ def test_model_default_cfgs(model_name, batch_size):
     pool_size = cfg['pool_size']
     input_size = model.default_cfg['input_size']
 
-    if all([x <= MAX_FWD_FEAT_SIZE for x in input_size]) and \
+    if all([x <= MAX_FWD_OUT_SIZE for x in input_size]) and \
             not any([fnmatch.fnmatch(model_name, x) for x in EXCLUDE_FILTERS]):
         # output sizes only checked if default res <= 448 * 448 to keep resource down
-        input_size = tuple([min(x, MAX_FWD_FEAT_SIZE) for x in input_size])
+        input_size = tuple([min(x, MAX_FWD_OUT_SIZE) for x in input_size])
         input_tensor = torch.randn((batch_size, *input_size))
 
         # test forward_features (always unpooled)
@@ -154,26 +178,25 @@ def test_model_features_pretrained(model_name, batch_size):
 
 EXCLUDE_JIT_FILTERS = [
     '*iabn*', 'tresnet*',  # models using inplace abn unlikely to ever be scriptable
-    'dla*', 'hrnet*', 'ghostnet*', # hopefully fix at some point
+    'dla*', 'hrnet*', 'ghostnet*',  # hopefully fix at some point
+    'vit_large_*', 'vit_huge_*',
 ]
 
 
 @pytest.mark.timeout(120)
-@pytest.mark.parametrize('model_name', list_models(exclude_filters=EXCLUDE_FILTERS + EXCLUDE_JIT_FILTERS))
+@pytest.mark.parametrize(
+    'model_name', list_models(exclude_filters=EXCLUDE_FILTERS + EXCLUDE_JIT_FILTERS, name_matches_cfg=True))
 @pytest.mark.parametrize('batch_size', [1])
 def test_model_forward_torchscript(model_name, batch_size):
     """Run a single forward pass with each model"""
+    input_size = _get_input_size(model_name=model_name, target=TARGET_JIT_SIZE)
+    if max(input_size) > MAX_JIT_SIZE:
+        pytest.skip("Fixed input size model > limit.")
+
     with set_scriptable(True):
         model = create_model(model_name, pretrained=False)
     model.eval()
 
-    if has_model_default_key(model_name, 'fixed_input_size'):
-        input_size = get_model_default_value(model_name, 'input_size')
-    elif has_model_default_key(model_name, 'min_input_size'):
-        input_size = get_model_default_value(model_name, 'min_input_size')
-    else:
-        input_size = (3, 128, 128)  # jit compile is already a bit slow and we've tested normal res already...
-
     model = torch.jit.script(model)
     outputs = model(torch.randn((batch_size, *input_size)))
 
@@ -183,7 +206,7 @@ def test_model_forward_torchscript(model_name, batch_size):
 
 EXCLUDE_FEAT_FILTERS = [
     '*pruned*',  # hopefully fix at some point
-]
+] + NON_STD_FILTERS
 if 'GITHUB_ACTIONS' in os.environ:  # and 'Linux' in platform.system():
     # GitHub Linux runner is slower and hits memory limits sooner than MacOS, exclude bigger models
     EXCLUDE_FEAT_FILTERS += ['*resnext101_32x32d', '*resnext101_32x16d']
@@ -199,12 +222,9 @@ def test_model_forward_features(model_name, batch_size):
     expected_channels = model.feature_info.channels()
     assert len(expected_channels) >= 4  # all models here should have at least 4 feature levels by default, some 5 or 6
 
-    if has_model_default_key(model_name, 'fixed_input_size'):
-        input_size = get_model_default_value(model_name, 'input_size')
-    elif has_model_default_key(model_name, 'min_input_size'):
-        input_size = get_model_default_value(model_name, 'min_input_size')
-    else:
-        input_size = (3, 96, 96)  # jit compile is already a bit slow and we've tested normal res already...
+    input_size = _get_input_size(model=model, target=TARGET_FFEAT_SIZE)
+    if max(input_size) > MAX_FFEAT_SIZE:
+        pytest.skip("Fixed input size model > limit.")
 
     outputs = model(torch.randn((batch_size, *input_size)))
     assert len(expected_channels) == len(outputs)

timm/data/parsers/parser_tfds.py

@@ -25,8 +25,8 @@
 
 
 MAX_TP_SIZE = 8  # maximum TF threadpool size, only doing jpeg decodes and queuing activities
-SHUFFLE_SIZE = 16834  # samples to shuffle in DS queue
-PREFETCH_SIZE = 4096  # samples to prefetch
+SHUFFLE_SIZE = 20480  # samples to shuffle in DS queue
+PREFETCH_SIZE = 2048  # samples to prefetch
 
 
 def even_split_indices(split, n, num_samples):
@@ -144,14 +144,16 @@ def _lazy_init(self):
         ds = self.builder.as_dataset(
             split=self.subsplit or self.split, shuffle_files=self.shuffle, read_config=read_config)
         # avoid overloading threading w/ combo fo TF ds threads + PyTorch workers
-        ds.options().experimental_threading.private_threadpool_size = max(1, MAX_TP_SIZE // num_workers)
-        ds.options().experimental_threading.max_intra_op_parallelism = 1
+        options = tf.data.Options()
+        options.experimental_threading.private_threadpool_size = max(1, MAX_TP_SIZE // num_workers)
+        options.experimental_threading.max_intra_op_parallelism = 1
+        ds = ds.with_options(options)
         if self.is_training or self.repeats > 1:
             # to prevent excessive drop_last batch behaviour w/ IterableDatasets
             # see warnings at https://pytorch.org/docs/stable/data.html#multi-process-data-loading
             ds = ds.repeat()  # allow wrap around and break iteration manually
         if self.shuffle:
-            ds = ds.shuffle(min(self.num_samples // self._num_pipelines, SHUFFLE_SIZE), seed=0)
+            ds = ds.shuffle(min(self.num_samples, SHUFFLE_SIZE) // self._num_pipelines, seed=0)
         ds = ds.prefetch(min(self.num_samples // self._num_pipelines, PREFETCH_SIZE))
         self.ds = tfds.as_numpy(ds)
 

timm/models/__init__.py

@@ -16,6 +16,8 @@
 from .inception_resnet_v2 import *
 from .inception_v3 import *
 from .inception_v4 import *
+from .levit import *
+#from .levit import *
 from .mlp_mixer import *
 from .mobilenetv3 import *
 from .nasnet import *
@@ -35,6 +37,7 @@
 from .tnt import *
 from .tresnet import *
 from .vgg import *
+from .visformer import *
 from .vision_transformer import *
 from .vision_transformer_hybrid import *
 from .vovnet import *