[feat] Adding Metaformer support (facebookresearch#294)

CHANGELOG.md

@@ -19,6 +19,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - MLP benchmark
 - Move all triton kernels to triton v2 [#272]
 - Mem efficient attention, BW pass [#281]
+- Metaformer support [#294]
 
 ## [0.0.10] - 2022-03-14
 ### Fixed

HOWTO.md

@@ -30,6 +30,8 @@ Let's present here a couple of code snippets on how to solve a couple of questio
     - [Intro](#intro)
     - [Transformer](#transformer)
     - [In practice](#in-practice)
+  - [Hierarchical Transformers](#hierarchical-transformers)
+
 
 ## Understanding the dimension conventions
 
@@ -749,3 +751,58 @@ class xFormerStackConfig:
 [2]: Kitaev, N., Kaiser, Ł., & Levskaya, A. (2020). Reformer: The Efficient Transformer.
 
 [3]: Vaswani et al., Attention is all you need, 2017
+
+
+### Hierarchical Transformers
+
+The original Transformer proposal processes ("transforms") sequences of tokens, across possibly many layers. Crucially, the number of tokens is unchanged cross the depth of the model, and this prove to be really efficient in many domains.
+
+It seems that some domains could however benefit from an architecture more typical from CNN, where there's a tradeoff across the depth of the model in between the spatial extent (ie: number of tokens) and their expressiveness (ie: the model or embedding dimension). These architectures are handled in xformers, through the "patch_embedding" element, which translates the sequence of tokens from one layer to another.
+
+A small helper is provided to make it easier to generate matching configurations, as follows. We present in this example a truncated version of a small [Metaformer](https://arxiv.org/abs/2111.11418v1).
+
+```python
+    from xformers.factory import xFormer, xFormerConfig
+    from xformers.helpers.hierarchical_configs import (
+        BasicLayerConfig,
+        get_hierarchical_configuration,
+    )
+
+
+    base_hierarchical_configs = [
+        BasicLayerConfig(
+            embedding=64,   # the dimensions just have to match along the layers
+            attention_mechanism="scaled_dot_product",   # anything you like
+            patch_size=7,
+            stride=4,
+            padding=2,
+            seq_len=image_size * image_size // 16,
+        ),
+        BasicLayerConfig(
+            embedding=128,
+            attention_mechanism="scaled_dot_product",
+            patch_size=3,
+            stride=2,
+            padding=1,
+            seq_len=image_size * image_size // 64,
+        ),
+        BasicLayerConfig(
+            embedding=320,
+            attention_mechanism="scaled_dot_product",
+            patch_size=3,
+            stride=2,
+            padding=1,
+            seq_len=image_size * image_size // 256,
+        ),
+    ]
+
+    # Fill in the gaps in the config
+    xformer_config = get_hierarchical_configuration(
+        base_hierarchical_configs,
+        layernorm_style="pre",
+        use_rotary_embeddings=False,
+        mlp_multiplier=4,
+        dim_head=32,
+    )
+    config = xFormerConfig(xformer_config)
+```

README.md

@@ -190,6 +190,7 @@ Patrick et al., 2021](https://arxiv.org/abs/2106.05392)*
    2. transformer block benchmark
    3. [LRA](xformers/benchmarks/LRA/README.md), with SLURM suppot
 4. Programatic and sweep friendly layer and model construction
+   1. Compatible with hierarchical Transformers, like Swin or Metaformer
 5. Hackable
    1. Not using monolithic CUDA kernels, composable building blocks
    2. Using [Triton](https://triton-lang.org/) for some optimized parts, explicit, pythonic and user-accessible

docs/source/tutorials/hierarchical.rst

@@ -0,0 +1,56 @@
+Hierarchical Transformers
+=========================
+
+The original Transformer proposal processes ("transforms") sequences of tokens, across possibly many layers. Crucially, the number of tokens is unchanged cross the depth of the model, and this prove to be really efficient in many domains.
+
+It seems that some domains could however benefit from an architecture more typical from CNN, where there's a tradeoff across the depth of the model in between the spatial extent (ie: number of tokens) and their expressiveness (ie: the model or embedding dimension). These architectures are handled in xformers, through the "patch_embedding" element, which translates the sequence of tokens from one layer to another.
+
+A small helper is provided to make it easier to generate matching configurations, as follows. We present in this example a truncated version of a small Metaformer_.
+
+.. _Metaformer: https://arxiv.org/abs/2111.11418v1
+
+.. code-block:: python
+
+    from xformers.factory import xFormer, xFormerConfig
+    from xformers.helpers.hierarchical_configs import (
+        BasicLayerConfig,
+        get_hierarchical_configuration,
+    )
+
+
+    base_hierarchical_configs = [
+        BasicLayerConfig(
+            embedding=64,   # the dimensions just have to match along the layers
+            attention_mechanism="scaled_dot_product",   # anything you like
+            patch_size=7,
+            stride=4,
+            padding=2,
+            seq_len=image_size * image_size // 16,
+        ),
+        BasicLayerConfig(
+            embedding=128,
+            attention_mechanism="scaled_dot_product",
+            patch_size=3,
+            stride=2,
+            padding=1,
+            seq_len=image_size * image_size // 64,
+        ),
+        BasicLayerConfig(
+            embedding=320,
+            attention_mechanism="scaled_dot_product",
+            patch_size=3,
+            stride=2,
+            padding=1,
+            seq_len=image_size * image_size // 256,
+        ),
+    ]
+
+    # Fill in the gaps in the config
+    xformer_config = get_hierarchical_configuration(
+        base_hierarchical_configs,
+        layernorm_style="pre",
+        use_rotary_embeddings=False,
+        mlp_multiplier=4,
+        dim_head=32,
+    )
+    config = xFormerConfig(xformer_config)

docs/source/tutorials/index.rst

@@ -11,3 +11,4 @@ Tutorials
    pytorch_encoder
    reversible
    triton
+   hierarchical

examples/README.md

@@ -25,3 +25,10 @@ If your current machine does not expose enough RAM and the example reports an `O
 This is meant to be an easy introduction to using xformers in practice, mirroring closely [this Pytorch Lightning](https://pytorchlightning.github.io/lightning-tutorials/notebooks/lightning_examples/cifar10-baseline.html) tutorial. The default settings are close to this tutorial, which trains a 11M parameters ResNet on the CIFAR dataset, we train a 10.6M ViT on the same dataset. The ViT configuration is not optimal for CIFAR, since the pictures have a very small size to begin with and information is probably lost given the patches. Nevertheless you should be able to reach about 80% accuracy within about an hour on a single GPU.
 
 ![Example curves](../docs/assets/microViT.png)
+
+
+### MicroMetaformer
+
+This is very close to the MicroViT example above, but illustrating the use of a hierarchical Transformer ([Metaformer](https://arxiv.org/pdf/2111.11418.pdf)) this time, through a helper function which generates the required configuration given the pooling parameters. The suggested configuration is about 6.6M parameters big (half of a ResNet18) and trains to about 86% top-1 Cifar10 within minutes.
+
+![Example curves](../docs/assets/metaformer.png)

examples/cifarMetaformer.py

@@ -0,0 +1,184 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
+#
+# This source code is licensed under the BSD license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import pytorch_lightning as pl
+import torch
+from pl_bolts.datamodules import CIFAR10DataModule
+from pl_bolts.transforms.dataset_normalizations import cifar10_normalization
+from torch import nn
+from torchmetrics import Accuracy
+from torchvision import transforms
+
+from examples.microViT import Classifier, VisionTransformer
+from xformers.factory import xFormer, xFormerConfig
+from xformers.helpers.hierarchical_configs import (
+    BasicLayerConfig,
+    get_hierarchical_configuration,
+)
+
+
+class MetaVisionTransformer(VisionTransformer):
+    def __init__(
+        self,
+        steps,
+        learning_rate=5e-3,
+        betas=(0.9, 0.99),
+        weight_decay=0.03,
+        image_size=32,
+        num_classes=10,
+        dim=384,
+        attention="scaled_dot_product",
+        layer_norm_style="pre",
+        use_rotary_embeddings=True,
+        linear_warmup_ratio=0.1,
+        classifier=Classifier.GAP,
+    ):
+
+        super(VisionTransformer, self).__init__()
+
+        # all the inputs are saved under self.hparams (hyperparams)
+        self.save_hyperparameters()
+
+        # Generate the skeleton of our hierarchical Transformer
+
+        # This is a small poolformer configuration, adapted to the small CIFAR10 pictures (32x32)
+        # Any other related config would work,
+        # and the attention mechanisms don't have to be the same across layers
+        base_hierarchical_configs = [
+            BasicLayerConfig(
+                embedding=64,
+                attention_mechanism=attention,
+                patch_size=3,
+                stride=2,
+                padding=1,
+                seq_len=image_size * image_size // 4,
+            ),
+            BasicLayerConfig(
+                embedding=128,
+                attention_mechanism=attention,
+                patch_size=3,
+                stride=2,
+                padding=1,
+                seq_len=image_size * image_size // 16,
+            ),
+            BasicLayerConfig(
+                embedding=320,
+                attention_mechanism=attention,
+                patch_size=3,
+                stride=2,
+                padding=1,
+                seq_len=image_size * image_size // 64,
+            ),
+            BasicLayerConfig(
+                embedding=512,
+                attention_mechanism=attention,
+                patch_size=3,
+                stride=2,
+                padding=1,
+                seq_len=image_size * image_size // 256,
+            ),
+        ]
+
+        # Fill in the gaps in the config
+        xformer_config = get_hierarchical_configuration(
+            base_hierarchical_configs,
+            layernorm_style=layer_norm_style,
+            use_rotary_embeddings=use_rotary_embeddings,
+            mlp_multiplier=4,
+            dim_head=32,
+        )
+
+        # Now instantiate the metaformer trunk
+        config = xFormerConfig(xformer_config)
+        print(config)
+        self.trunk = xFormer.from_config(config)
+        print(self.trunk)
+
+        # The classifier head
+        dim = base_hierarchical_configs[-1].embedding
+        self.ln = nn.LayerNorm(dim)
+        self.head = nn.Linear(dim, num_classes)
+        self.criterion = torch.nn.CrossEntropyLoss()
+        self.val_accuracy = Accuracy()
+
+    def forward(self, x):
+        x = self.trunk(x)
+        x = self.ln(x)
+
+        if self.hparams.classifier == Classifier.TOKEN:
+            x = x[:, 0]  # only consider the token, we're classifying anyway
+        elif self.hparams.classifier == Classifier.GAP:
+            x = x.mean(dim=1)  # mean over sequence len
+
+        x = self.head(x)
+        return x
+
+
+if __name__ == "__main__":
+    pl.seed_everything(42)
+
+    # Adjust batch depending on the available memory on your machine.
+    # You can also use reversible layers to save memory
+    REF_BATCH = 512
+    BATCH = 512  # lower if not enough GPU memory
+
+    MAX_EPOCHS = 50
+    NUM_WORKERS = 4
+    GPUS = 1
+
+    train_transforms = transforms.Compose(
+        [
+            transforms.RandomCrop(32, padding=4),
+            transforms.RandomHorizontalFlip(),
+            transforms.ToTensor(),
+            cifar10_normalization(),
+        ]
+    )
+
+    test_transforms = transforms.Compose(
+        [
+            transforms.ToTensor(),
+            cifar10_normalization(),
+        ]
+    )
+
+    # We'll use a datamodule here, which already handles dataset/dataloader/sampler
+    # See https://pytorchlightning.github.io/lightning-tutorials/notebooks/lightning_examples/cifar10-baseline.html
+    # for a full tutorial
+    dm = CIFAR10DataModule(
+        data_dir="data",
+        batch_size=BATCH,
+        num_workers=NUM_WORKERS,
+        pin_memory=True,
+    )
+    dm.train_transforms = train_transforms
+    dm.test_transforms = test_transforms
+    dm.val_transforms = test_transforms
+
+    image_size = dm.size(-1)  # 32 for CIFAR
+    num_classes = dm.num_classes  # 10 for CIFAR
+
+    # compute total number of steps
+    batch_size = BATCH * GPUS
+    steps = dm.num_samples // REF_BATCH * MAX_EPOCHS
+    lm = MetaVisionTransformer(
+        steps=steps,
+        image_size=image_size,
+        num_classes=num_classes,
+        attention="scaled_dot_product",
+        layer_norm_style="pre",
+        use_rotary_embeddings=True,
+    )
+    trainer = pl.Trainer(
+        gpus=GPUS,
+        max_epochs=MAX_EPOCHS,
+        precision=16,
+        accumulate_grad_batches=REF_BATCH // BATCH,
+    )
+    trainer.fit(lm, dm)
+
+    # check the training
+    trainer.test(lm, datamodule=dm)

examples/microViT.py

@@ -171,7 +171,7 @@ def training_step(self, batch, _):
                 "train_loss": loss.mean(),
                 "learning_rate": self.lr_schedulers().get_last_lr()[0],
             },
-            step=trainer.global_step,
+            step=self.global_step,
         )
 
         return loss

tests/test_attentions.py

@@ -3,6 +3,7 @@
 # This source code is licensed under the BSD license found in the
 # LICENSE file in the root directory of this source tree.
 
+import math
 from typing import Tuple
 
 import pytest
@@ -96,6 +97,9 @@ def test_order_invariance(
     device: torch.device,
 ):
 
+    if int(math.sqrt(SEQ)) ** 2 != SEQ and attention_name == "poolling":
+        pytest.skip(f"{attention_name} requires squared sequence lengths")
+
     torch.manual_seed(42)
 
     multi_head = _get_multihead(
@@ -282,6 +286,9 @@ def test_broadcast_batch_dimension(
     device: torch.device,
     batch_sizes: Tuple[int, int, int],
 ):
+    if int(math.sqrt(SEQ)) ** 2 != SEQ and attention_name == "poolling":
+        pytest.skip(f"{attention_name} requires squared sequence lengths")
+
     Q_BATCH, K_BATCH, V_BATCH = batch_sizes
     multi_head = _get_multihead(attention_name, 0.0, 0.0, False, heads, device)