d2cec09baa
* Duplicate swiftformer * Convert SwiftFormerPatchEmbedding * Convert SwiftFormerEmbeddings * Convert TFSwiftFormerMlp * Convert TFSwiftFormerConvEncoder * Convert TFSwiftFormerLocalRepresentation * convert TFSwiftFormerEncoderBlock * Convert SwiftFormerStage * Convert SwiftFormerEncoder * Add TFSWiftFormerPreTrainedModel * Convert SwiftFormerForImageClassification * Add kwargs and start drop path * Fix syntax * Change Model class name * Add TFSwiftFormer to __init__ * Duplicate test_modeling_swiftformer * First test conversions * Change require_torch to require_tf * Add exports to swiftformer __init__ * Add TFSwiftFormerModel wrapper * Fix __init__ and run black * Remove docstring from MainLayer, fix padding * Use keras.layers.Activation on keras.Sequential * Fix swiftformer exports * Fix activation layer from config * Remove post_inits * Use tf.keras.layers.ZeroPadding2D * Convert torch normalize * Change tf test input shape * Fix softmax and reduce_sum * Convert expand_dims and repeat * Add missing reshape and tranpose * Simplify TFSwiftFormerEncoderBlock.call * Fix mismatch in patch embeddings * Fix expected output shape to match channels last * Fix swiftformer typo * Disable test_onnx * Fix TFSwiftFormerForImageClassification call * Add unpack inputs * Convert flatten(2).mean(-1) * Change vision dummy inputs (to be reviewed) * Change test_forward_signature to use .call * Fix @unpack_inputs * Set return_tensors="tf" and rename class * Rename wrongly named patch_embeddings layer * Add serving_output and change dummy_input shape * Make dimensions BCHW and transpose inside embedding layer * Change SwiftFormerEncoderBlock * Fix ruff problems * Add image size to swiftformer config * Change tranpose to MainLayer and use -1 for reshape * Remove serving_outputs and dummy_inputs * Remove test_initialization test from tf model * Make Sequential component a separate layer * Fix layers' names * Tranpose encoder outputs * Fix tests and check if hidden states is not None * Fix TFSwiftFormerForImageClassification * Run make fixup * Run make fix-copies * Update modeling_tf_auto * Update docs * Fix modeling auto mapping * Update modelint_tf_swiftformer docs * Fill image_size doc and type * Add reduction=None to loss computation * Update docs * make style * Debug: Delete the tip to see if that changes anything * Re-add tip * Remove add_code_sample_docstrings * Remove unused import * Get the debug to actually tell us the problem it has with the docs * Try a substitution to match the PyTorch file? * Add swiftformer to ignore list * Add build() methods * Update copyright year Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Remove FIXME comment * Remove from_pt * Update copyright year Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Rename one-letter variables * Remove FIXMEs related to momentum * Remove old TODO comment * Remove outstanding FIXME comments * Get dropout rate from config * Add specific dropout config for MLP * Add convencoder dropout to config * Pass config to SwiftFormerDropPath layer * Fix drop_path variable name and add Adapted from comment * Run ruff * Removed copied from comment * Run fix copies * Change drop_path to identity to match pt * Cleanup build() methods and move to new keras imports * Update docs/source/en/model_doc/swiftformer.md Co-authored-by: Matt <Rocketknight1@users.noreply.github.com> * Raise error if drop_path_rate > 0.0 * Apply suggestions from code review Replace (self.dim), with self.dim, Co-authored-by: Matt <Rocketknight1@users.noreply.github.com> * Remove drop_path function * Add training to TFSwiftFormerEncoder * Set self.built = True last Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Should have been added to previous commit Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Apply suggestions from code review Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Change default_feature_extractor to default_image_processor Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Import Keras from modeling_tf_utils * Remove relative import * Run ruff --fix * Move import keras to tf_available * Add copied from comment to test_forward_signature * Reduce batch size and num_labels * Extract loss logic to hf_compute_loss * Run ruff format --------- Co-authored-by: Matt <rocketknight1@gmail.com> Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>
3.5 KiB
SwiftFormer
Overview
The SwiftFormer model was proposed in SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications by Abdelrahman Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, Fahad Shahbaz Khan.
The SwiftFormer paper introduces a novel efficient additive attention mechanism that effectively replaces the quadratic matrix multiplication operations in the self-attention computation with linear element-wise multiplications. A series of models called 'SwiftFormer' is built based on this, which achieves state-of-the-art performance in terms of both accuracy and mobile inference speed. Even their small variant achieves 78.5% top-1 ImageNet1K accuracy with only 0.8 ms latency on iPhone 14, which is more accurate and 2× faster compared to MobileViT-v2.
The abstract from the paper is the following:
Self-attention has become a defacto choice for capturing global context in various vision applications. However, its quadratic computational complexity with respect to image resolution limits its use in real-time applications, especially for deployment on resource-constrained mobile devices. Although hybrid approaches have been proposed to combine the advantages of convolutions and self-attention for a better speed-accuracy trade-off, the expensive matrix multiplication operations in self-attention remain a bottleneck. In this work, we introduce a novel efficient additive attention mechanism that effectively replaces the quadratic matrix multiplication operations with linear element-wise multiplications. Our design shows that the key-value interaction can be replaced with a linear layer without sacrificing any accuracy. Unlike previous state-of-the-art methods, our efficient formulation of self-attention enables its usage at all stages of the network. Using our proposed efficient additive attention, we build a series of models called "SwiftFormer" which achieves state-of-the-art performance in terms of both accuracy and mobile inference speed. Our small variant achieves 78.5% top-1 ImageNet-1K accuracy with only 0.8 ms latency on iPhone 14, which is more accurate and 2x faster compared to MobileViT-v2.
This model was contributed by shehan97. The TensorFlow version was contributed by joaocmd. The original code can be found here.
SwiftFormerConfig
autodoc SwiftFormerConfig
SwiftFormerModel
autodoc SwiftFormerModel - forward
SwiftFormerForImageClassification
autodoc SwiftFormerForImageClassification - forward
TFSwiftFormerModel
autodoc TFSwiftFormerModel - call
TFSwiftFormerForImageClassification
autodoc TFSwiftFormerForImageClassification - call