{"notes":[{"id":"reparameterization/repvit","category":"reparameterization","legacyPath":"reparameterization/2307-repvit","legacySlug":"2307-repvit","legacyPaths":[],"title":"RepViT","subtitle":"Revisiting Mobile CNN from the ViT Perspective","author":"Z. Yuan","paperDate":"2023-07-18","publishedOn":"2024-07-19","paperTitle":"RepViT: Revisiting Mobile CNN From ViT Perspective","paperUrl":"https://arxiv.org/abs/2307.09283","bibliography":{"title":"RepViT: Revisiting Mobile CNN From ViT Perspective","authors":["Ao Wang","Hui Chen","Zijia Lin","Jungong Han","Guiguang Ding"],"url":"https://arxiv.org/abs/2307.09283","pdfUrl":"https://arxiv.org/pdf/2307.09283","codeUrl":null,"preprint":{"id":"2307.09283","submittedOn":"2023-07-18","url":"https://arxiv.org/abs/2307.09283"},"publication":{"venue":"CVPR","year":2024,"kind":"conference","url":"https://arxiv.org/abs/2307.09283"},"citation":"@misc{reparameterizationrepvit2023,\n  title = {{RepViT: Revisiting Mobile CNN From ViT Perspective}},\n  author = {Ao Wang and Hui Chen and Zijia Lin and Jungong Han and Guiguang Ding},\n  year = {2023},\n  eprint = {2307.09283},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2307.09283}\n}\n","citationPath":"/content/papers/citations/reparameterization/repvit.583fd659fe6d.bib"},"thumbnail":{"src":"/content/papers/assets/f501c7d5eb209527b5401ba7607cd34496d2dda89aeb1848aa9927e0d0f713eb.jpg","width":1224,"height":704},"locale":"en","path":"/content/papers/en/reparameterization/repvit.8b88345cbf24.md","headings":[{"id":"problem-definition","text":"Problem Definition","depth":2},{"id":"solution-approach","text":"Solution Approach","depth":2},{"id":"redesigning-the-architecture","text":"Redesigning the Architecture","depth":3},{"id":"reparameterizing-the-structure","text":"Reparameterizing the Structure","depth":3},{"id":"reducing-expansion-ratio-and-increasing-width","text":"Reducing Expansion Ratio and Increasing Width","depth":3},{"id":"adjusting-the-stem-structure","text":"Adjusting the Stem Structure","depth":3},{"id":"adjusting-the-downsampling-structure","text":"Adjusting the Downsampling Structure","depth":3},{"id":"simple-classifier","text":"Simple Classifier","depth":3},{"id":"adjusting-stage-ratios","text":"Adjusting Stage Ratios","depth":3},{"id":"kernel-size-selection","text":"Kernel Size Selection","depth":3},{"id":"squeeze-and-excitation-se-layer-decision","text":"Squeeze-and-Excitation (SE) Layer Decision","depth":3},{"id":"discussion","text":"Discussion","depth":2},{"id":"performance-on-imagenet-1k","text":"Performance on ImageNet-1K","depth":3},{"id":"ablation-study---reparameterized-structure","text":"Ablation Study - Reparameterized Structure","depth":3},{"id":"ablation-study---impact-of-se-layers","text":"Ablation Study - Impact of SE Layers","depth":3},{"id":"conclusion","text":"Conclusion","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"reparameterization/fastvit","relation":"comparison","topic":"Fast visual backbones for deployment"},{"id":"reparameterization/mobileone","relation":"comparison","topic":"Fast visual backbones for deployment"}]},{"id":"reparameterization/fastvit","category":"reparameterization","legacyPath":"reparameterization/2303-fastvit","legacySlug":"2303-fastvit","legacyPaths":[],"title":"FastViT","subtitle":"Reparameterized ViT Experiments","author":"Z. Yuan","paperDate":"2023-03-24","publishedOn":"2024-07-16","paperTitle":"FastViT: A Fast Hybrid Vision Transformer using Structural Reparameterization","paperUrl":"https://arxiv.org/abs/2303.14189","bibliography":{"title":"FastViT: A Fast Hybrid Vision Transformer using Structural Reparameterization","authors":["Pavan Kumar Anasosalu Vasu","James Gabriel","Jeff Zhu","Oncel Tuzel","Anurag Ranjan"],"url":"https://arxiv.org/abs/2303.14189","pdfUrl":"https://arxiv.org/pdf/2303.14189","codeUrl":null,"preprint":{"id":"2303.14189","submittedOn":"2023-03-24","url":"https://arxiv.org/abs/2303.14189"},"publication":{"venue":"ICCV","year":2023,"kind":"conference","url":"https://arxiv.org/abs/2303.14189"},"citation":"@misc{reparameterizationfastvit2023,\n  title = {{FastViT: A Fast Hybrid Vision Transformer using Structural Reparameterization}},\n  author = {Pavan Kumar Anasosalu Vasu and James Gabriel and Jeff Zhu and Oncel Tuzel and Anurag Ranjan},\n  year = {2023},\n  eprint = {2303.14189},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2303.14189}\n}\n","citationPath":"/content/papers/citations/reparameterization/fastvit.0b2741f99b2b.bib"},"thumbnail":{"src":"/content/papers/assets/095c905b22b0b0e80a3fdb19b68680885648cdd70aa6ff754434e08a3a1aa78a.jpg","width":1618,"height":1038},"locale":"en","path":"/content/papers/en/reparameterization/fastvit.78850b950b9b.md","headings":[{"id":"problem-definition","text":"Problem Definition","depth":2},{"id":"solution","text":"Solution","depth":2},{"id":"model-architecture","text":"Model Architecture","depth":3},{"id":"model-configurations","text":"Model Configurations","depth":3},{"id":"training-and-evaluation-strategy","text":"Training and Evaluation Strategy","depth":3},{"id":"discussion","text":"Discussion","depth":2},{"id":"performance-on-imagenet","text":"Performance on ImageNet","depth":3},{"id":"conclusion","text":"Conclusion","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"lightweight/mobilevit","relation":"comparison","topic":"Combining lightweight convolutions and attention"},{"id":"lightweight/mobile-former","relation":"comparison","topic":"Combining lightweight convolutions and attention"},{"id":"reparameterization/repvit","relation":"comparison","topic":"Fast visual backbones for deployment"}]},{"id":"reparameterization/mobileone","category":"reparameterization","legacyPath":"reparameterization/2206-mobileone","legacySlug":"2206-mobileone","legacyPaths":[],"title":"MobileOne","subtitle":"One Millisecond Optimization","author":"Z. Yuan","paperDate":"2022-06-08","publishedOn":"2024-07-16","paperTitle":"MobileOne: An Improved One millisecond Mobile Backbone","paperUrl":"https://arxiv.org/abs/2206.04040","bibliography":{"title":"MobileOne: An Improved One millisecond Mobile Backbone","authors":["Pavan Kumar Anasosalu Vasu","James Gabriel","Jeff Zhu","Oncel Tuzel","Anurag Ranjan"],"url":"https://arxiv.org/abs/2206.04040","pdfUrl":"https://arxiv.org/pdf/2206.04040","codeUrl":null,"preprint":{"id":"2206.04040","submittedOn":"2022-06-08","url":"https://arxiv.org/abs/2206.04040"},"publication":{"venue":"CVPR","year":2023,"kind":"conference","url":"https://arxiv.org/abs/2206.04040"},"citation":"@misc{reparameterizationmobileone2022,\n  title = {{MobileOne: An Improved One millisecond Mobile Backbone}},\n  author = {Pavan Kumar Anasosalu Vasu and James Gabriel and Jeff Zhu and Oncel Tuzel and Anurag Ranjan},\n  year = {2022},\n  eprint = {2206.04040},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2206.04040}\n}\n","citationPath":"/content/papers/citations/reparameterization/mobileone.cbade423cf61.bib"},"thumbnail":{"src":"/content/papers/assets/f1d3676416f16224fbdba82b7113b2c4894d48fee9149dd2c8bc61fe781b22e2.jpg","width":840,"height":976},"locale":"en","path":"/content/papers/en/reparameterization/mobileone.247ee1e1848d.md","headings":[{"id":"problem-definition","text":"Problem Definition","depth":2},{"id":"problem-solving","text":"Problem Solving","depth":2},{"id":"bottleneck-analysis---activation-functions","text":"Bottleneck Analysis - Activation Functions","depth":3},{"id":"bottleneck-analysis---architecture","text":"Bottleneck Analysis - Architecture","depth":3},{"id":"model-architecture","text":"Model Architecture","depth":3},{"id":"model-configuration","text":"Model Configuration","depth":3},{"id":"discussion","text":"Discussion","depth":2},{"id":"imagenet-performance","text":"ImageNet Performance","depth":3},{"id":"conclusion","text":"Conclusion","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"reparameterization/dbb","relation":"background","topic":"Merging training branches for inference"},{"id":"reparameterization/fastvit","relation":"comparison","topic":"Fast visual backbones for deployment"},{"id":"reparameterization/repvit","relation":"comparison","topic":"Fast visual backbones for deployment"}]}],"text":"Research on reparameterization is thriving and gradually spreading to different model architectures.\n\nThis paper aims to redesign MobileNet-V3. Unlike the previous papers we looked at, this time we start from the ViT perspective. Structural reparameterization is used inside selected blocks; the entire original MobileNet-V3 is not simply converted into an equivalent network.\n\n## Problem Definition\n\nResearch on lightweight vision models is primarily divided into two camps:\n\n1. **CNN-Based Research**: Represented by models like the MobileNet series, which reduce the number of parameters through depthwise separable convolutions and enhance the model's expressiveness using inverted residual bottleneck structures.\n2. **ViT-Based Research**: Represented by models like MobileViT, Mobile-Former, and FastViT, which combine convolution and attention in different ways; the separation of token and channel mixing can be analyzed from the MetaFormer perspective.\n\n\n> [!TIP]\n> \n> For those unfamiliar with MetaFormer, you might find these references useful:\n> \n> - [**[21.11] PoolFormer: You Need a Meta!**](/blog/papers/vision-transformers/poolformer)\n> - [**[22.10] CAFormer: MetaFormer User Manual**](/blog/papers/vision-transformers/caformer)\n\n\n---\n\nIs there an insurmountable gap between these two camps?\n\nThis paper aims to bridge this gap by redesigning MobileNet-V3 and rethinking it from the ViT perspective.\n\n## Solution Approach\n\n### Redesigning the Architecture\n\n![model arch](/content/papers/assets/06ed7a72d8ae57a70a82c3e9c172ed7d3e5c4d78a288431fa70a50cb20ac57a2.jpg)\n\nThe main target is measured latency on an iPhone 12. Parameter count and FLOPs still appear in the later comparison tables.\n\nThe goal is to be fast, faster, and extremely fast!\n\n---\n\nWe start with MobileNet-V3 and take inspiration from recent lightweight ViTs, which typically use DeiT's training methods.\n\nSpecifically, the training methods used include:\n\n- AdamW optimizer\n- Cosine learning rate decay; the architectural experiments train for 300 epochs, while the main results also include 450-epoch runs\n- RegNetY-16GF as the teacher for knowledge distillation in the main comparison table; the step-by-step architectural experiments in the figure do not use distillation\n- Mixup, AutoAugment and Random Erasing for data augmentation, together with Label Smoothing\n\nThe subsequent architectural changes follow this training setup; whether distillation is used depends on the particular experiment table. Latency is measured on iPhone 12 with iOS 16, Core ML compilation and batch size 1. The MobileNetV3-L baseline also uses GELU for compiler compatibility.\n\n\n> [!TIP]\n> \n> For those unfamiliar with DeiT, you might find this reference helpful:\n> \n> - [**[20.12] DeiT: Distillation ViT**](/blog/papers/vision-transformers/deit)\n\n\n### Reparameterizing the Structure\n\n![reparameterization](/content/papers/assets/f501c7d5eb209527b5401ba7607cd34496d2dda89aeb1848aa9927e0d0f713eb.jpg)\n\nNext, the structure of MobileNet-V3 is adjusted.\n\nMetaFormer offers a useful decomposition of a block into two parts, rather than a unique explanation of Transformer performance:\n\n1. **Token-level information exchange**, which mixes spatial positions and can be local or global.\n2. **Feature-level information exchange**, which corresponds to channel-level information exchange in images.\n\nIn the original MobileNet structure, these two aspects are coupled together:\n\n- First, $1 \\times 1$ convolutions handle \"channel\" information exchange.\n- Next, $3 \\times 3$ depthwise convolutions handle \"local spatial\" information exchange.\n- Followed by a $SE$ layer, which handles both \"global and channel\" information exchange.\n- Finally, another $1 \\times 1$ convolution handles \"channel\" information exchange.\n\nThe authors separate these aspects in the new structure:\n\n- Start with \"local spatial\" information exchange using a $3 \\times 3$ depthwise convolution. During training it has an additional $1 \\times 1$ depthwise branch and a shortcut; for inference these are fused into one $3 \\times 3$ convolution.\n- Then apply the $SE$ layer for \"global and channel\" information exchange.\n- Conclude with two $1 \\times 1$ convolutions forming the channel mixer.\n\nTogether, these changes reduce measured latency from 1.01 ms to 0.81 ms, about 20%. The reciprocal throughput increase would be about 25%.\n\n---\n\nThe speed increases, but this step lowers ImageNet Top-1 accuracy from 71.5% to 68.3%.\n\nNext, measures are taken to recover the lost accuracy.\n\n### Reducing Expansion Ratio and Increasing Width\n\nIn typical ViTs, the expansion ratio in the channel mixer is often set to 4, making the hidden dimensions in the feed-forward network (FFN) four times the input dimensions; MobileNetV3-L varies from about 2.3 to 6, with 6 common in its last two stages. Wider channel mixers increase inference time.\n\nPrevious research has shown that the FFN contains a lot of redundant information, and a lower expansion ratio is often sufficient.\n\n- [**[21.04] LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference**](https://arxiv.org/abs/2104.01136)\n- [**[21.04] All Tokens Matter: Token Labeling for Training Better Vision Transformers**](https://arxiv.org/abs/2104.10858)\n- [**[22.12] SAViT: Structure-Aware Vision Transformer Pruning via Collaborative Optimization**](https://proceedings.neurips.cc/paper_files/paper/2022/hash/3b11c5cc84b6da2838db348b37dbd1a2-Abstract-Conference.html)\n- [**[23.10] NViT: Vision Transformer Compression and Parameter Redistribution**](https://openreview.net/forum?id=LzBBxCg-xpa)\n\nThe authors decide to use an expansion ratio of 2 while widening the stages to 48, 96, 192 and 384 channels to restore expressiveness.\n\n---\n\nThis adjustment recovers accuracy and surpasses the original MobileNet-V3, achieving 73.5% at 0.89 ms latency.\n\n### Adjusting the Stem Structure\n\n![stem](/content/papers/assets/c27fa0f49ed2077a500847497081431c9e2d8839787b01928b393c88998e0476.jpg)\n\nViTs typically use a patchify operation as the backbone, dividing the input image into non-overlapping patches, with simple stem structures often using a large $16 \\times 16$ convolution. In hierarchical ViTs, a $4 \\times 4$ convolution is used.\n\nIn contrast, MobileNet-V3 has a complex stem structure with $3 \\times 3$ convolutions, depthwise separable convolutions, and inverted residual bottlenecks, as shown in (a).\n\nSince the stem structure processes input images at the highest resolution, its complexity can create significant speed bottlenecks on mobile devices. As a trade-off, MobileNet-V3's stem structure uses a small number of filters (only 16 channels), limiting its expressiveness.\n\nTo address this, the authors use early convolutions with two $3 \\times 3$ convolutions and a stride of 2 to reduce computation while increasing the number of filters to enhance expressiveness, as shown in (b).\n\n### Adjusting the Downsampling Structure\n\n![downsampling](/content/papers/assets/0bbcdd3a8a12c880757e93ad6909726532c09b8ff73f9227236f7b64f8052956.jpg)\n\nMobileNet-V3 downsamples with a stride-2 depthwise convolution inside an inverted bottleneck. While this design is efficient, it may lack sufficient network depth, leading to information loss and negatively impacting performance, as shown in (c).\n\nTo address this, the authors use a more complex downsampling structure, as shown in (d). This structure includes a stride-2 depthwise convolution for spatial downsampling and a 1×1 convolution for adjusting channels, a RepViT module at the front, and an FFN module after $1 \\times 1$ convolutions to retain more information.\n\n### Simple Classifier\n\n![classifier](/content/papers/assets/957796ecd5fbf7b05f72605921e4e6f5a2867f80685b466f5e34e8ac31cbc231.jpg)\n\nIn lightweight ViTs, the classifier typically consists of global average pooling and a linear layer, which is very fast on mobile devices.\n\nIn contrast, MobileNet-V3's classifier includes a $1 \\times 1$ convolution and an additional linear layer to expand features into a higher-dimensional space, as shown in (e). While this design is crucial for MobileNet-V3, it increases inference time on mobile devices.\n\nThe authors abandon MobileNet-V3’s design and opt for a simple global average pooling and linear layer combination for the final architecture.\n\n---\n\nThis design brings inference time back to 0.77 ms, but accuracy drops again!\n\n### Adjusting Stage Ratios\n\nSeveral architectures allocate more blocks to the third stage, but no single ratio is universal. EfficientFormer-L2, for example, uses 1:1:3:1.5.\n\n- [**[19.05] On Network Design Spaces for Visual Recognition**](https://arxiv.org/abs/1905.13214)\n- [**[20.03] Designing Network Design Spaces**](https://arxiv.org/abs/2003.13678)\n\nRecent studies suggest that more aggressive ratios might benefit smaller models. For example, Conv2Former-T and Conv2Former-S use ratios of 1:1:4:1 and 1:1:8:1, respectively.\n\n- [**[22.11] Conv2Former: A Simple Transformer-Style ConvNet for Visual Recognition**](https://arxiv.org/abs/2211.11943)\n\nThe authors adopt a 1:1:7:1 ratio with deeper stages of 2:2:14:2, achieving a deeper network layout.\n\nThis adjustment finally recovers the lost accuracy, reaching 76.9% at 0.91 ms latency.\n\n### Kernel Size Selection\n\nIn previous research, larger kernel convolutions (e.g., $5 \\times 5$ or $7 \\times 7$) in small networks have been effective in improving model expressiveness.\n\nIn the tested mobile/compiler environment, larger kernels have higher computation and memory costs and less optimization than 3×3 kernels. This is not a universal limitation of mobile hardware.\n\nTo ensure inference efficiency on mobile devices, the authors prioritize $3 \\times 3$ convolutions across all modules.\n\nThis keeps accuracy at 76.9% while reducing inference time from 0.91 to 0.89 ms.\n\n### Squeeze-and-Excitation (SE) Layer Decision\n\nThe SE module has been widely used in various network architectures to enhance model expressiveness.\n\n- **However, SE modules are quite slow!**\n\nTo balance the benefits of SE modules with speed, the authors use them sparingly:\n\n- **Applying SE modules selectively in the 1st, 3rd, 5th... blocks of each stage.**\n\n## Discussion\n\n### Performance on ImageNet-1K\n\n![imagenet performance](/content/papers/assets/a770f33217932f0a48bec6aef799f169e1586ea530b44ef21e981a47355fd38e.jpg)\n\n![imagenet performance 1](/content/papers/assets/e3dafd77d254d4b972394924f9fa3ebe7771e6d8380180323e6ad1c3fa671d3f.jpg)\n\nExperiments were conducted using standard 224×224 pixel images for training and testing, evaluating the performance of various models with different training epochs (300 or 450 epochs).\n\nKey results include:\n\n1. **Model Performance Comparison**:\n\n   - Among the 300-epoch models near 0.9 ms in the table, RepViT-M0.9 reaches 78.7% Top-1 accuracy, 3.0 and 2.0 percentage points above EfficientFormerV2-S0 at 75.7% and FastViT-T8 at 76.7%, respectively. RepViT-M1.1 reaches 80.7% at 1.1 ms after 300 epochs, 1.7 points above EfficientFormerV2-S1 at 79.0%.\n\n2. **Latency and Accuracy**:\n\n   - RepViT-M1.0 has 1.0 ms latency on the tested iPhone 12. Its Top-1 accuracy is 80.0% after 300 epochs and 80.3% after 450 epochs; only the latter exceeds 80%.\n   - The largest model, RepViT-M2.3, reaches 83.3% / 83.7% Top-1 accuracy after 300 / 450 epochs at 2.3 ms latency.\n\n3. **Impact of Knowledge Distillation**:\n\n   Even without knowledge distillation (in the table below), RepViT has higher accuracy in several comparisons at nearby latencies. For example, RepViT-M1.0 reaches 78.6% at 1.0 ms, 2.7 percentage points above MobileOne-S1 at 0.9 ms. For larger models, RepViT-M2.3 is 1.1 points above PoolFormer-S36, while latency falls from 3.5 to 2.3 ms, a reduction of about 34.3%.\n\n   ![distillation](/content/papers/assets/8d905761d2f2e6bd1a9395b3e9a51992a8cf446dd6bb262c948e27008dbe3c38.jpg)\n\n### Ablation Study - Reparameterized Structure\n\n![ablation1](/content/papers/assets/9a94ef0a0eaa584e4fac4ec9cb6cf2cec0c6ef6c7b4a0cff23cdfc369693bd71.jpg)\n\nTo validate the effectiveness of the reparameterized structure in RepViT, the authors conducted ablation studies by removing the multi-branch topology during training on ImageNet-1K.\n\nRemoving the training-time structural reparameterization (SR) branches lowers Top-1 accuracy by 0.27, 0.20 and 0.20 percentage points for M0.9, M1.5 and M2.3, respectively. This table compares training topologies, not the latency of branches left in the deployed model.\n\n### Ablation Study - Impact of SE Layers\n\n![ablation2](/content/papers/assets/1d0a7823c37ade2a61a9588214ae12e81637b08ca94640f095f8b06ba3b8fcd3.jpg)\n\nTo verify the advantage of using SE layers in a cross-block manner, the authors compared scenarios with all SE layers removed and where SE layers are used in each block.\n\nFor M0.9, alternating SE gives 78.74% accuracy at 0.87 ms, nearly matching SE in every block at 78.75% while improving on its 0.92 ms latency. For M1.5, both placements reach 82.29%, while latency falls from 1.58 to 1.52 ms with alternating SE. Removing SE altogether is faster again, but accuracy is lower.\n\n## Conclusion\n\nThe authors describe RepViT-M1.0 as an early lightweight model reaching 80% ImageNet-1K accuracy at about 1.0 ms on iPhone 12. This is a historical result under the paper’s conditions; application preprocessing and postprocessing require separate measurements.\n\nThe authors have explored efficient designs for lightweight CNNs, combining innovative lightweight ViT structures to introduce the RepViT series.\n\nThis lightweight CNN model, designed for resource-constrained mobile devices, demonstrates superior performance in various vision tasks. RepViT not only outperforms current state-of-the-art lightweight ViTs and CNNs in accuracy but also excels in latency.\n\nThe authors hope RepViT will become a solid baseline for lightweight model design and inspire further research and innovation in this field.\n\n\n> [!TIP]\n> \n> While RepViT's parameter size may be a concern, many modern mobile devices have ample capacity. However, model size remains a critical factor for many clients, which can limit its adoption.\n> \n> If your development scenario is not constrained by model size and prioritizes accuracy and inference time, RepViT is worth evaluating on the target device, including latency, memory use and model file size.\n","imageSizes":{"/content/papers/assets/06ed7a72d8ae57a70a82c3e9c172ed7d3e5c4d78a288431fa70a50cb20ac57a2.jpg":{"width":1224,"height":996},"/content/papers/assets/f501c7d5eb209527b5401ba7607cd34496d2dda89aeb1848aa9927e0d0f713eb.jpg":{"width":1224,"height":704},"/content/papers/assets/c27fa0f49ed2077a500847497081431c9e2d8839787b01928b393c88998e0476.jpg":{"width":2360,"height":1430},"/content/papers/assets/0bbcdd3a8a12c880757e93ad6909726532c09b8ff73f9227236f7b64f8052956.jpg":{"width":2365,"height":1430},"/content/papers/assets/957796ecd5fbf7b05f72605921e4e6f5a2867f80685b466f5e34e8ac31cbc231.jpg":{"width":2365,"height":1430},"/content/papers/assets/a770f33217932f0a48bec6aef799f169e1586ea530b44ef21e981a47355fd38e.jpg":{"width":1224,"height":752},"/content/papers/assets/e3dafd77d254d4b972394924f9fa3ebe7771e6d8380180323e6ad1c3fa671d3f.jpg":{"width":1176,"height":860},"/content/papers/assets/8d905761d2f2e6bd1a9395b3e9a51992a8cf446dd6bb262c948e27008dbe3c38.jpg":{"width":984,"height":752},"/content/papers/assets/9a94ef0a0eaa584e4fac4ec9cb6cf2cec0c6ef6c7b4a0cff23cdfc369693bd71.jpg":{"width":1090,"height":244},"/content/papers/assets/1d0a7823c37ade2a61a9588214ae12e81637b08ca94640f095f8b06ba3b8fcd3.jpg":{"width":940,"height":280}}}
