{ "pdf_info": [ { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 78, 502, 115 ], "lines": [ { "bbox": [ 106, 79, 505, 97 ], "spans": [ { "bbox": [ 106, 79, 505, 97 ], "score": 1.0, "content": "CONDITIONAL POSITIONAL ENCODINGS FOR VISION", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 98, 231, 117 ], "spans": [ { "bbox": [ 106, 98, 231, 117 ], "score": 1.0, "content": "TRANSFORMERS", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 112, 134, 452, 180 ], "lines": [ { "bbox": [ 111, 133, 455, 148 ], "spans": [ { "bbox": [ 111, 133, 163, 148 ], "score": 1.0, "content": "Xiangxiang", "type": "text" }, { "bbox": [ 164, 135, 187, 146 ], "score": 0.75, "content": "\\mathbf { C h u ^ { 1 } }", "type": "inline_equation" }, { "bbox": [ 188, 133, 455, 148 ], "score": 1.0, "content": ", Zhi Tian1, Bo Zhang1, Xinlong Wang2, Chunhua Shen3∗", "type": "text" } ], "index": 2 }, { "bbox": [ 110, 143, 411, 162 ], "spans": [ { "bbox": [ 110, 143, 411, 162 ], "score": 1.0, "content": "1 Meituan Inc. 2 Beijing Academy of AI 3 Zhejiang University, China", "type": "text" } ], "index": 3 }, { "bbox": [ 112, 158, 380, 171 ], "spans": [ { "bbox": [ 112, 158, 380, 171 ], "score": 1.0, "content": "{chuxiangxiang, tianzhi02, zhangbo97}@meituan.com,", "type": "text" } ], "index": 4 }, { "bbox": [ 112, 170, 330, 181 ], "spans": [ { "bbox": [ 112, 170, 330, 181 ], "score": 1.0, "content": "xinlong.wang96@gmail.com, chunhua@me.com", "type": "text" } ], "index": 5 } ], "index": 3.5 }, { "type": "title", "bbox": [ 278, 209, 333, 221 ], "lines": [ { "bbox": [ 276, 209, 335, 222 ], "spans": [ { "bbox": [ 276, 209, 335, 222 ], "score": 1.0, "content": "ABSTRACT", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 143, 235, 468, 377 ], "lines": [ { "bbox": [ 142, 235, 470, 248 ], "spans": [ { "bbox": [ 142, 235, 470, 248 ], "score": 1.0, "content": "We propose a conditional positional encoding (CPE) scheme for vision Trans-", "type": "text" } ], "index": 7 }, { "bbox": [ 142, 246, 469, 258 ], "spans": [ { "bbox": [ 142, 246, 469, 258 ], "score": 1.0, "content": "formers (Dosovitskiy et al., 2021; Touvron et al., 2020). Unlike previous fixed", "type": "text" } ], "index": 8 }, { "bbox": [ 141, 257, 470, 270 ], "spans": [ { "bbox": [ 141, 257, 470, 270 ], "score": 1.0, "content": "or learnable positional encodings that are predefined and independent of input to-", "type": "text" } ], "index": 9 }, { "bbox": [ 141, 268, 470, 280 ], "spans": [ { "bbox": [ 141, 268, 470, 280 ], "score": 1.0, "content": "kens, CPE is dynamically generated and conditioned on the local neighborhood of", "type": "text" } ], "index": 10 }, { "bbox": [ 141, 278, 470, 291 ], "spans": [ { "bbox": [ 141, 278, 470, 291 ], "score": 1.0, "content": "the input tokens. As a result, CPE can easily generalize to the input sequences that", "type": "text" } ], "index": 11 }, { "bbox": [ 141, 290, 470, 302 ], "spans": [ { "bbox": [ 141, 290, 470, 302 ], "score": 1.0, "content": "are longer than what the model has ever seen during the training. Besides, CPE can", "type": "text" } ], "index": 12 }, { "bbox": [ 141, 301, 469, 313 ], "spans": [ { "bbox": [ 141, 301, 469, 313 ], "score": 1.0, "content": "keep the desired translation equivalence in vision tasks, resulting in improved per-", "type": "text" } ], "index": 13 }, { "bbox": [ 141, 312, 470, 324 ], "spans": [ { "bbox": [ 141, 312, 470, 324 ], "score": 1.0, "content": "formance. We implement CPE with a simple Position Encoding Generator (PEG)", "type": "text" } ], "index": 14 }, { "bbox": [ 141, 323, 470, 335 ], "spans": [ { "bbox": [ 141, 323, 470, 335 ], "score": 1.0, "content": "to get seamlessly incorporated into the current Transformer framework. Built on", "type": "text" } ], "index": 15 }, { "bbox": [ 141, 333, 470, 346 ], "spans": [ { "bbox": [ 141, 333, 470, 346 ], "score": 1.0, "content": "PEG, we present Conditional Position encoding Vision Transformer (CPVT). We", "type": "text" } ], "index": 16 }, { "bbox": [ 141, 344, 469, 357 ], "spans": [ { "bbox": [ 141, 344, 469, 357 ], "score": 1.0, "content": "demonstrate that CPVT has visually similar attention maps compared to those", "type": "text" } ], "index": 17 }, { "bbox": [ 141, 355, 470, 368 ], "spans": [ { "bbox": [ 141, 355, 470, 368 ], "score": 1.0, "content": "with learned positional encodings and delivers outperforming results. Our Code", "type": "text" } ], "index": 18 }, { "bbox": [ 142, 367, 321, 378 ], "spans": [ { "bbox": [ 142, 367, 321, 378 ], "score": 1.0, "content": "is available at: https://git.io/CPVT.", "type": "text" } ], "index": 19 } ], "index": 13 }, { "type": "title", "bbox": [ 108, 400, 205, 413 ], "lines": [ { "bbox": [ 105, 399, 208, 416 ], "spans": [ { "bbox": [ 105, 399, 208, 416 ], "score": 1.0, "content": "1 INTRODUCTION", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 426, 505, 503 ], "lines": [ { "bbox": [ 105, 426, 505, 438 ], "spans": [ { "bbox": [ 105, 426, 505, 438 ], "score": 1.0, "content": "Recently, Transformers (Vaswani et al., 2017) have been viewed as a strong alternative to Convolu-", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 437, 505, 450 ], "spans": [ { "bbox": [ 105, 437, 505, 450 ], "score": 1.0, "content": "tional Neural Networks (CNNs) in visual recognition tasks such as classification (Dosovitskiy et al.,", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 447, 505, 460 ], "spans": [ { "bbox": [ 105, 447, 505, 460 ], "score": 1.0, "content": "2021) and detection (Carion et al., 2020; Zhu et al., 2021). Unlike the convolution operation in", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 459, 505, 471 ], "spans": [ { "bbox": [ 106, 459, 505, 471 ], "score": 1.0, "content": "CNNs, which has a limited receptive field, the self-attention mechanism in the Transformers can", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 470, 506, 483 ], "spans": [ { "bbox": [ 105, 470, 506, 483 ], "score": 1.0, "content": "capture the long-distance information and dynamically adapt the receptive field according to the", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 480, 505, 493 ], "spans": [ { "bbox": [ 105, 480, 505, 493 ], "score": 1.0, "content": "image content. Consequently, Transformers are considered more flexible and powerful than CNNs,", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 493, 363, 504 ], "spans": [ { "bbox": [ 106, 493, 363, 504 ], "score": 1.0, "content": "being promising to achieve more progress in visual recognition.", "type": "text" } ], "index": 27 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 509, 505, 618 ], "lines": [ { "bbox": [ 106, 509, 505, 521 ], "spans": [ { "bbox": [ 106, 509, 505, 521 ], "score": 1.0, "content": "However, the self-attention operation in Transformers is permutation-invariant, which discards the", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 520, 505, 532 ], "spans": [ { "bbox": [ 106, 520, 505, 532 ], "score": 1.0, "content": "order of the tokens in an input sequence. To mitigate this issue, previous works (Vaswani et al.,", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 529, 506, 544 ], "spans": [ { "bbox": [ 105, 529, 506, 544 ], "score": 1.0, "content": "2017; Dosovitskiy et al., 2021) add the absolute positional encodings to each input token (see Fig-", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 542, 505, 554 ], "spans": [ { "bbox": [ 106, 542, 505, 554 ], "score": 1.0, "content": "ure 1a), which enables order-awareness. The positional encoding can either be learnable or fixed", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 552, 505, 565 ], "spans": [ { "bbox": [ 105, 552, 505, 565 ], "score": 1.0, "content": "with sinusoidal functions of different frequencies. Despite being effective, these positional encod-", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 563, 506, 577 ], "spans": [ { "bbox": [ 105, 563, 506, 577 ], "score": 1.0, "content": "ings seriously harm the flexibility of the Transformers, hampering their broader applications. Taking", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 574, 505, 587 ], "spans": [ { "bbox": [ 105, 574, 505, 587 ], "score": 1.0, "content": "the learnable version as an example, the encodings are often a vector of equal length to the input", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 585, 506, 599 ], "spans": [ { "bbox": [ 105, 585, 506, 599 ], "score": 1.0, "content": "sequence, which are jointly updated with the network weights during training. As a result, the length", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 596, 505, 609 ], "spans": [ { "bbox": [ 105, 596, 505, 609 ], "score": 1.0, "content": "and the value of the positional encodings are fixed once trained. During testing, it causes difficulties", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 608, 375, 621 ], "spans": [ { "bbox": [ 105, 608, 375, 621 ], "score": 1.0, "content": "of handling the sequences longer than the ones in the training data.", "type": "text" } ], "index": 37 } ], "index": 32.5 }, { "type": "text", "bbox": [ 107, 624, 504, 712 ], "lines": [ { "bbox": [ 105, 624, 505, 637 ], "spans": [ { "bbox": [ 105, 624, 505, 637 ], "score": 1.0, "content": "The inability to adapt to longer input sequences during testing greatly limits the range of general-", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 635, 505, 648 ], "spans": [ { "bbox": [ 105, 635, 505, 648 ], "score": 1.0, "content": "ization. For instance, in vision tasks like object detection, we expect the model can be applied to the", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 647, 505, 659 ], "spans": [ { "bbox": [ 106, 647, 505, 659 ], "score": 1.0, "content": "images of any size during inference, which might be much larger than the training images. A possi-", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 657, 505, 670 ], "spans": [ { "bbox": [ 105, 657, 505, 670 ], "score": 1.0, "content": "ble remedy is to use bicubic interpolation to upsample the positional encodings to the target length,", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 669, 506, 681 ], "spans": [ { "bbox": [ 106, 669, 506, 681 ], "score": 1.0, "content": "but it degrades the performance without fine-tuning as later shown in our experiments. For vision", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 679, 506, 692 ], "spans": [ { "bbox": [ 105, 679, 506, 692 ], "score": 1.0, "content": "in general, we expect that the models be translation-equivariant. For example, the output feature", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 691, 505, 703 ], "spans": [ { "bbox": [ 106, 691, 505, 703 ], "score": 1.0, "content": "maps of CNNs shift accordingly as the target objects are moved in the input images. However, the", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 700, 506, 714 ], "spans": [ { "bbox": [ 105, 700, 506, 714 ], "score": 1.0, "content": "absolute positional encoding scheme might break the translation equivalence because it adds unique", "type": "text" } ], "index": 45 } ], "index": 41.5 } ], "page_idx": 0, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 119, 722, 204, 731 ], "lines": [ { "bbox": [ 118, 719, 205, 734 ], "spans": [ { "bbox": [ 118, 719, 205, 734 ], "score": 1.0, "content": "∗Corresponding author.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 25, 293, 38 ], "spans": [ { "bbox": [ 106, 25, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 308, 760 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "1", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 78, 502, 115 ], "lines": [ { "bbox": [ 106, 79, 505, 97 ], "spans": [ { "bbox": [ 106, 79, 505, 97 ], "score": 1.0, "content": "CONDITIONAL POSITIONAL ENCODINGS FOR VISION", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 98, 231, 117 ], "spans": [ { "bbox": [ 106, 98, 231, 117 ], "score": 1.0, "content": "TRANSFORMERS", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 112, 134, 452, 180 ], "lines": [ { "bbox": [ 111, 133, 455, 148 ], "spans": [ { "bbox": [ 111, 133, 163, 148 ], "score": 1.0, "content": "Xiangxiang", "type": "text" }, { "bbox": [ 164, 135, 187, 146 ], "score": 0.75, "content": "\\mathbf { C h u ^ { 1 } }", "type": "inline_equation" }, { "bbox": [ 188, 133, 455, 148 ], "score": 1.0, "content": ", Zhi Tian1, Bo Zhang1, Xinlong Wang2, Chunhua Shen3∗", "type": "text" } ], "index": 2 }, { "bbox": [ 110, 143, 411, 162 ], "spans": [ { "bbox": [ 110, 143, 411, 162 ], "score": 1.0, "content": "1 Meituan Inc. 2 Beijing Academy of AI 3 Zhejiang University, China", "type": "text" } ], "index": 3 }, { "bbox": [ 112, 158, 380, 171 ], "spans": [ { "bbox": [ 112, 158, 380, 171 ], "score": 1.0, "content": "{chuxiangxiang, tianzhi02, zhangbo97}@meituan.com,", "type": "text" } ], "index": 4 }, { "bbox": [ 112, 170, 330, 181 ], "spans": [ { "bbox": [ 112, 170, 330, 181 ], "score": 1.0, "content": "xinlong.wang96@gmail.com, chunhua@me.com", "type": "text" } ], "index": 5 } ], "index": 3.5, "bbox_fs": [ 110, 133, 455, 181 ] }, { "type": "title", "bbox": [ 278, 209, 333, 221 ], "lines": [ { "bbox": [ 276, 209, 335, 222 ], "spans": [ { "bbox": [ 276, 209, 335, 222 ], "score": 1.0, "content": "ABSTRACT", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 143, 235, 468, 377 ], "lines": [ { "bbox": [ 142, 235, 470, 248 ], "spans": [ { "bbox": [ 142, 235, 470, 248 ], "score": 1.0, "content": "We propose a conditional positional encoding (CPE) scheme for vision Trans-", "type": "text" } ], "index": 7 }, { "bbox": [ 142, 246, 469, 258 ], "spans": [ { "bbox": [ 142, 246, 469, 258 ], "score": 1.0, "content": "formers (Dosovitskiy et al., 2021; Touvron et al., 2020). Unlike previous fixed", "type": "text" } ], "index": 8 }, { "bbox": [ 141, 257, 470, 270 ], "spans": [ { "bbox": [ 141, 257, 470, 270 ], "score": 1.0, "content": "or learnable positional encodings that are predefined and independent of input to-", "type": "text" } ], "index": 9 }, { "bbox": [ 141, 268, 470, 280 ], "spans": [ { "bbox": [ 141, 268, 470, 280 ], "score": 1.0, "content": "kens, CPE is dynamically generated and conditioned on the local neighborhood of", "type": "text" } ], "index": 10 }, { "bbox": [ 141, 278, 470, 291 ], "spans": [ { "bbox": [ 141, 278, 470, 291 ], "score": 1.0, "content": "the input tokens. As a result, CPE can easily generalize to the input sequences that", "type": "text" } ], "index": 11 }, { "bbox": [ 141, 290, 470, 302 ], "spans": [ { "bbox": [ 141, 290, 470, 302 ], "score": 1.0, "content": "are longer than what the model has ever seen during the training. Besides, CPE can", "type": "text" } ], "index": 12 }, { "bbox": [ 141, 301, 469, 313 ], "spans": [ { "bbox": [ 141, 301, 469, 313 ], "score": 1.0, "content": "keep the desired translation equivalence in vision tasks, resulting in improved per-", "type": "text" } ], "index": 13 }, { "bbox": [ 141, 312, 470, 324 ], "spans": [ { "bbox": [ 141, 312, 470, 324 ], "score": 1.0, "content": "formance. We implement CPE with a simple Position Encoding Generator (PEG)", "type": "text" } ], "index": 14 }, { "bbox": [ 141, 323, 470, 335 ], "spans": [ { "bbox": [ 141, 323, 470, 335 ], "score": 1.0, "content": "to get seamlessly incorporated into the current Transformer framework. Built on", "type": "text" } ], "index": 15 }, { "bbox": [ 141, 333, 470, 346 ], "spans": [ { "bbox": [ 141, 333, 470, 346 ], "score": 1.0, "content": "PEG, we present Conditional Position encoding Vision Transformer (CPVT). We", "type": "text" } ], "index": 16 }, { "bbox": [ 141, 344, 469, 357 ], "spans": [ { "bbox": [ 141, 344, 469, 357 ], "score": 1.0, "content": "demonstrate that CPVT has visually similar attention maps compared to those", "type": "text" } ], "index": 17 }, { "bbox": [ 141, 355, 470, 368 ], "spans": [ { "bbox": [ 141, 355, 470, 368 ], "score": 1.0, "content": "with learned positional encodings and delivers outperforming results. Our Code", "type": "text" } ], "index": 18 }, { "bbox": [ 142, 367, 321, 378 ], "spans": [ { "bbox": [ 142, 367, 321, 378 ], "score": 1.0, "content": "is available at: https://git.io/CPVT.", "type": "text" } ], "index": 19 } ], "index": 13, "bbox_fs": [ 141, 235, 470, 378 ] }, { "type": "title", "bbox": [ 108, 400, 205, 413 ], "lines": [ { "bbox": [ 105, 399, 208, 416 ], "spans": [ { "bbox": [ 105, 399, 208, 416 ], "score": 1.0, "content": "1 INTRODUCTION", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 426, 505, 503 ], "lines": [ { "bbox": [ 105, 426, 505, 438 ], "spans": [ { "bbox": [ 105, 426, 505, 438 ], "score": 1.0, "content": "Recently, Transformers (Vaswani et al., 2017) have been viewed as a strong alternative to Convolu-", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 437, 505, 450 ], "spans": [ { "bbox": [ 105, 437, 505, 450 ], "score": 1.0, "content": "tional Neural Networks (CNNs) in visual recognition tasks such as classification (Dosovitskiy et al.,", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 447, 505, 460 ], "spans": [ { "bbox": [ 105, 447, 505, 460 ], "score": 1.0, "content": "2021) and detection (Carion et al., 2020; Zhu et al., 2021). Unlike the convolution operation in", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 459, 505, 471 ], "spans": [ { "bbox": [ 106, 459, 505, 471 ], "score": 1.0, "content": "CNNs, which has a limited receptive field, the self-attention mechanism in the Transformers can", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 470, 506, 483 ], "spans": [ { "bbox": [ 105, 470, 506, 483 ], "score": 1.0, "content": "capture the long-distance information and dynamically adapt the receptive field according to the", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 480, 505, 493 ], "spans": [ { "bbox": [ 105, 480, 505, 493 ], "score": 1.0, "content": "image content. Consequently, Transformers are considered more flexible and powerful than CNNs,", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 493, 363, 504 ], "spans": [ { "bbox": [ 106, 493, 363, 504 ], "score": 1.0, "content": "being promising to achieve more progress in visual recognition.", "type": "text" } ], "index": 27 } ], "index": 24, "bbox_fs": [ 105, 426, 506, 504 ] }, { "type": "text", "bbox": [ 107, 509, 505, 618 ], "lines": [ { "bbox": [ 106, 509, 505, 521 ], "spans": [ { "bbox": [ 106, 509, 505, 521 ], "score": 1.0, "content": "However, the self-attention operation in Transformers is permutation-invariant, which discards the", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 520, 505, 532 ], "spans": [ { "bbox": [ 106, 520, 505, 532 ], "score": 1.0, "content": "order of the tokens in an input sequence. To mitigate this issue, previous works (Vaswani et al.,", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 529, 506, 544 ], "spans": [ { "bbox": [ 105, 529, 506, 544 ], "score": 1.0, "content": "2017; Dosovitskiy et al., 2021) add the absolute positional encodings to each input token (see Fig-", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 542, 505, 554 ], "spans": [ { "bbox": [ 106, 542, 505, 554 ], "score": 1.0, "content": "ure 1a), which enables order-awareness. The positional encoding can either be learnable or fixed", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 552, 505, 565 ], "spans": [ { "bbox": [ 105, 552, 505, 565 ], "score": 1.0, "content": "with sinusoidal functions of different frequencies. Despite being effective, these positional encod-", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 563, 506, 577 ], "spans": [ { "bbox": [ 105, 563, 506, 577 ], "score": 1.0, "content": "ings seriously harm the flexibility of the Transformers, hampering their broader applications. Taking", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 574, 505, 587 ], "spans": [ { "bbox": [ 105, 574, 505, 587 ], "score": 1.0, "content": "the learnable version as an example, the encodings are often a vector of equal length to the input", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 585, 506, 599 ], "spans": [ { "bbox": [ 105, 585, 506, 599 ], "score": 1.0, "content": "sequence, which are jointly updated with the network weights during training. As a result, the length", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 596, 505, 609 ], "spans": [ { "bbox": [ 105, 596, 505, 609 ], "score": 1.0, "content": "and the value of the positional encodings are fixed once trained. During testing, it causes difficulties", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 608, 375, 621 ], "spans": [ { "bbox": [ 105, 608, 375, 621 ], "score": 1.0, "content": "of handling the sequences longer than the ones in the training data.", "type": "text" } ], "index": 37 } ], "index": 32.5, "bbox_fs": [ 105, 509, 506, 621 ] }, { "type": "text", "bbox": [ 107, 624, 504, 712 ], "lines": [ { "bbox": [ 105, 624, 505, 637 ], "spans": [ { "bbox": [ 105, 624, 505, 637 ], "score": 1.0, "content": "The inability to adapt to longer input sequences during testing greatly limits the range of general-", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 635, 505, 648 ], "spans": [ { "bbox": [ 105, 635, 505, 648 ], "score": 1.0, "content": "ization. For instance, in vision tasks like object detection, we expect the model can be applied to the", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 647, 505, 659 ], "spans": [ { "bbox": [ 106, 647, 505, 659 ], "score": 1.0, "content": "images of any size during inference, which might be much larger than the training images. A possi-", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 657, 505, 670 ], "spans": [ { "bbox": [ 105, 657, 505, 670 ], "score": 1.0, "content": "ble remedy is to use bicubic interpolation to upsample the positional encodings to the target length,", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 669, 506, 681 ], "spans": [ { "bbox": [ 106, 669, 506, 681 ], "score": 1.0, "content": "but it degrades the performance without fine-tuning as later shown in our experiments. For vision", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 679, 506, 692 ], "spans": [ { "bbox": [ 105, 679, 506, 692 ], "score": 1.0, "content": "in general, we expect that the models be translation-equivariant. For example, the output feature", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 691, 505, 703 ], "spans": [ { "bbox": [ 106, 691, 505, 703 ], "score": 1.0, "content": "maps of CNNs shift accordingly as the target objects are moved in the input images. However, the", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 700, 506, 714 ], "spans": [ { "bbox": [ 105, 700, 506, 714 ], "score": 1.0, "content": "absolute positional encoding scheme might break the translation equivalence because it adds unique", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 345, 505, 358 ], "spans": [ { "bbox": [ 105, 345, 505, 358 ], "score": 1.0, "content": "positional encodings to each token (or each image patch). One may overcome the issue with rela-", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 106, 357, 505, 369 ], "spans": [ { "bbox": [ 106, 357, 505, 369 ], "score": 1.0, "content": "tive positional encodings as in (Shaw et al., 2018). However, relative positional encodings not only", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 105, 367, 505, 380 ], "spans": [ { "bbox": [ 105, 367, 505, 380 ], "score": 1.0, "content": "come with extra computational costs, but also require modifying the implementation of the standard", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 105, 378, 505, 391 ], "spans": [ { "bbox": [ 105, 378, 505, 391 ], "score": 1.0, "content": "Transformers. Last but not least, the relative positional encodings cannot work equally well as the", "type": "text", "cross_page": true } ], "index": 11 }, { "bbox": [ 105, 389, 505, 402 ], "spans": [ { "bbox": [ 105, 389, 505, 402 ], "score": 1.0, "content": "absolute ones, because the image recognition task still requires absolute position information (Islam", "type": "text", "cross_page": true } ], "index": 12 }, { "bbox": [ 105, 401, 378, 412 ], "spans": [ { "bbox": [ 105, 401, 378, 412 ], "score": 1.0, "content": "et al., 2020), which the relative positional encodings fail to provide.", "type": "text", "cross_page": true } ], "index": 13 } ], "index": 41.5, "bbox_fs": [ 105, 624, 506, 714 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 124, 80, 473, 256 ], "blocks": [ { "type": "image_body", "bbox": [ 124, 80, 473, 256 ], "group_id": 0, "lines": [ { "bbox": [ 124, 80, 473, 256 ], "spans": [ { "bbox": [ 124, 80, 473, 256 ], "score": 0.973, "type": "image", "image_path": "0c4b2513e8c89ba2d4113b63f9bdb7b99d31471e828098a2ee34f19b9c920289.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 124, 80, 473, 138.66666666666666 ], "spans": [], "index": 0 }, { "bbox": [ 124, 138.66666666666666, 473, 197.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 124, 197.33333333333331, 473, 255.99999999999997 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 265, 504, 321 ], "group_id": 0, "lines": [ { "bbox": [ 106, 265, 504, 279 ], "spans": [ { "bbox": [ 106, 265, 504, 279 ], "score": 1.0, "content": "Figure 1. Vision Transformers: (a) ViT (Dosovitskiy et al., 2021) with explicit 1D learnable posi-", "type": "text" } ], "index": 3 }, { "bbox": [ 107, 277, 504, 289 ], "spans": [ { "bbox": [ 107, 277, 504, 289 ], "score": 1.0, "content": "tional encodings (PE) (b) CPVT with conditional positional encoding from the proposed Position", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 287, 505, 300 ], "spans": [ { "bbox": [ 106, 287, 505, 300 ], "score": 1.0, "content": "Encoding Generator (PEG) plugin, which is the default choice. (c) CPVT-GAP without class token", "type": "text" } ], "index": 5 }, { "bbox": [ 107, 299, 505, 311 ], "spans": [ { "bbox": [ 107, 299, 505, 311 ], "score": 1.0, "content": "(cls), but with global average pooling (GAP) over all items in the sequence. Note that GAP is a", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 308, 297, 323 ], "spans": [ { "bbox": [ 106, 308, 297, 323 ], "score": 1.0, "content": "bonus version which has boosted performance.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 }, { "type": "text", "bbox": [ 107, 345, 505, 412 ], "lines": [ { "bbox": [ 105, 345, 505, 358 ], "spans": [ { "bbox": [ 105, 345, 505, 358 ], "score": 1.0, "content": "positional encodings to each token (or each image patch). One may overcome the issue with rela-", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 357, 505, 369 ], "spans": [ { "bbox": [ 106, 357, 505, 369 ], "score": 1.0, "content": "tive positional encodings as in (Shaw et al., 2018). However, relative positional encodings not only", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 367, 505, 380 ], "spans": [ { "bbox": [ 105, 367, 505, 380 ], "score": 1.0, "content": "come with extra computational costs, but also require modifying the implementation of the standard", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 378, 505, 391 ], "spans": [ { "bbox": [ 105, 378, 505, 391 ], "score": 1.0, "content": "Transformers. Last but not least, the relative positional encodings cannot work equally well as the", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 389, 505, 402 ], "spans": [ { "bbox": [ 105, 389, 505, 402 ], "score": 1.0, "content": "absolute ones, because the image recognition task still requires absolute position information (Islam", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 401, 378, 412 ], "spans": [ { "bbox": [ 105, 401, 378, 412 ], "score": 1.0, "content": "et al., 2020), which the relative positional encodings fail to provide.", "type": "text" } ], "index": 13 } ], "index": 10.5 }, { "type": "text", "bbox": [ 106, 417, 505, 495 ], "lines": [ { "bbox": [ 105, 416, 505, 430 ], "spans": [ { "bbox": [ 105, 416, 505, 430 ], "score": 1.0, "content": "In this work, we advocate a novel positional encoding (PE) scheme to incorporate the position", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 428, 505, 441 ], "spans": [ { "bbox": [ 105, 428, 505, 441 ], "score": 1.0, "content": "information into Transformers. Unlike the predefined and input-agnostic positional encodings used", "type": "text" } ], "index": 15 }, { "bbox": [ 104, 438, 506, 453 ], "spans": [ { "bbox": [ 104, 438, 506, 453 ], "score": 1.0, "content": "in previous works (Dosovitskiy et al., 2021; Vaswani et al., 2017; Shaw et al., 2018), the proposed", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 450, 506, 464 ], "spans": [ { "bbox": [ 105, 450, 506, 464 ], "score": 1.0, "content": "PE is dynamically generated and conditioned on the local neighborhood of input tokens. Thus, our", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 461, 505, 474 ], "spans": [ { "bbox": [ 105, 461, 505, 474 ], "score": 1.0, "content": "positional encodings can change along with the input size and try to keep translation equivalence. We", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 472, 505, 485 ], "spans": [ { "bbox": [ 105, 472, 505, 485 ], "score": 1.0, "content": "demonstrate that the vision transformers (Dosovitskiy et al., 2021; Touvron et al., 2020) with our new", "type": "text" } ], "index": 19 }, { "bbox": [ 104, 482, 506, 497 ], "spans": [ { "bbox": [ 104, 482, 506, 497 ], "score": 1.0, "content": "PE (i.e. CPVT, see Figure 1c) achieve even better performance. We summarize our contributions as,", "type": "text" } ], "index": 20 } ], "index": 17 }, { "type": "text", "bbox": [ 134, 513, 505, 579 ], "lines": [ { "bbox": [ 132, 512, 506, 526 ], "spans": [ { "bbox": [ 132, 512, 506, 526 ], "score": 1.0, "content": "• We propose a novel positional encoding (PE) scheme, termed conditional position encod-", "type": "text" } ], "index": 21 }, { "bbox": [ 140, 523, 506, 536 ], "spans": [ { "bbox": [ 140, 523, 506, 536 ], "score": 1.0, "content": "ings (CPE). CPE is dynamically generated with Positional Encoding Generators (PEG) and", "type": "text" } ], "index": 22 }, { "bbox": [ 141, 536, 506, 548 ], "spans": [ { "bbox": [ 141, 536, 506, 548 ], "score": 1.0, "content": "can be effortlessly implemented by the modern deep learning frameworks (Paszke et al.,", "type": "text" } ], "index": 23 }, { "bbox": [ 141, 545, 506, 559 ], "spans": [ { "bbox": [ 141, 545, 506, 559 ], "score": 1.0, "content": "2019; Abadi et al., 2016; Chen et al., 2015), requiring no changes to the current Trans-", "type": "text" } ], "index": 24 }, { "bbox": [ 141, 557, 506, 569 ], "spans": [ { "bbox": [ 141, 557, 506, 569 ], "score": 1.0, "content": "former APIs. Through an in-depth analysis and thorough experimentations, we unveil that", "type": "text" } ], "index": 25 }, { "bbox": [ 141, 568, 482, 581 ], "spans": [ { "bbox": [ 141, 568, 482, 581 ], "score": 1.0, "content": "this design affords both absolute and relative encoding yet it goes above and beyond.", "type": "text" } ], "index": 26 } ], "index": 23.5 }, { "type": "text", "bbox": [ 134, 593, 504, 626 ], "lines": [ { "bbox": [ 133, 592, 506, 606 ], "spans": [ { "bbox": [ 133, 592, 506, 606 ], "score": 1.0, "content": "• As opposed to widely-used absolute positional encodings, CPE can provide a kind of", "type": "text" } ], "index": 27 }, { "bbox": [ 142, 604, 505, 616 ], "spans": [ { "bbox": [ 142, 604, 505, 616 ], "score": 1.0, "content": "stronger explicit bias towards the translation equivalence which is important to improve", "type": "text" } ], "index": 28 }, { "bbox": [ 142, 615, 279, 627 ], "spans": [ { "bbox": [ 142, 615, 279, 627 ], "score": 1.0, "content": "the performance of Transformers.", "type": "text" } ], "index": 29 } ], "index": 28 }, { "type": "text", "bbox": [ 134, 640, 505, 673 ], "lines": [ { "bbox": [ 133, 639, 506, 653 ], "spans": [ { "bbox": [ 133, 639, 506, 653 ], "score": 1.0, "content": "• Built on CPE, we propose Conditional Position encoding Vision Transformer (CPVT). It", "type": "text" } ], "index": 30 }, { "bbox": [ 141, 651, 505, 663 ], "spans": [ { "bbox": [ 141, 651, 505, 663 ], "score": 1.0, "content": "achieves better performance than previous vison transformers (Dosovitskiy et al., 2021;", "type": "text" } ], "index": 31 }, { "bbox": [ 142, 662, 231, 675 ], "spans": [ { "bbox": [ 142, 662, 231, 675 ], "score": 1.0, "content": "Touvron et al., 2020).", "type": "text" } ], "index": 32 } ], "index": 31 }, { "type": "text", "bbox": [ 135, 687, 504, 732 ], "lines": [ { "bbox": [ 133, 687, 505, 701 ], "spans": [ { "bbox": [ 133, 687, 505, 701 ], "score": 1.0, "content": "• CPE can well generalize to arbitrary input resolutions, which are required in many impor-", "type": "text" } ], "index": 33 }, { "bbox": [ 141, 699, 505, 711 ], "spans": [ { "bbox": [ 141, 699, 505, 711 ], "score": 1.0, "content": "tant downstream tasks such as segmentation and detection. Through experiments we show", "type": "text" } ], "index": 34 }, { "bbox": [ 141, 709, 505, 722 ], "spans": [ { "bbox": [ 141, 709, 505, 722 ], "score": 1.0, "content": "that CPE can boost the segmentation and detection performance for pyramid transformers", "type": "text" } ], "index": 35 }, { "bbox": [ 141, 721, 312, 733 ], "spans": [ { "bbox": [ 141, 721, 312, 733 ], "score": 1.0, "content": "like (Wang et al., 2021) by a clear margin.", "type": "text" } ], "index": 36 } ], "index": 34.5 } ], "page_idx": 1, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 309, 760 ], "lines": [ { "bbox": [ 301, 750, 310, 763 ], "spans": [ { "bbox": [ 301, 750, 310, 763 ], "score": 1.0, "content": "", "type": "text", "height": 13, "width": 9 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 124, 80, 473, 256 ], "blocks": [ { "type": "image_body", "bbox": [ 124, 80, 473, 256 ], "group_id": 0, "lines": [ { "bbox": [ 124, 80, 473, 256 ], "spans": [ { "bbox": [ 124, 80, 473, 256 ], "score": 0.973, "type": "image", "image_path": "0c4b2513e8c89ba2d4113b63f9bdb7b99d31471e828098a2ee34f19b9c920289.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 124, 80, 473, 138.66666666666666 ], "spans": [], "index": 0 }, { "bbox": [ 124, 138.66666666666666, 473, 197.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 124, 197.33333333333331, 473, 255.99999999999997 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 265, 504, 321 ], "group_id": 0, "lines": [ { "bbox": [ 106, 265, 504, 279 ], "spans": [ { "bbox": [ 106, 265, 504, 279 ], "score": 1.0, "content": "Figure 1. Vision Transformers: (a) ViT (Dosovitskiy et al., 2021) with explicit 1D learnable posi-", "type": "text" } ], "index": 3 }, { "bbox": [ 107, 277, 504, 289 ], "spans": [ { "bbox": [ 107, 277, 504, 289 ], "score": 1.0, "content": "tional encodings (PE) (b) CPVT with conditional positional encoding from the proposed Position", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 287, 505, 300 ], "spans": [ { "bbox": [ 106, 287, 505, 300 ], "score": 1.0, "content": "Encoding Generator (PEG) plugin, which is the default choice. (c) CPVT-GAP without class token", "type": "text" } ], "index": 5 }, { "bbox": [ 107, 299, 505, 311 ], "spans": [ { "bbox": [ 107, 299, 505, 311 ], "score": 1.0, "content": "(cls), but with global average pooling (GAP) over all items in the sequence. Note that GAP is a", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 308, 297, 323 ], "spans": [ { "bbox": [ 106, 308, 297, 323 ], "score": 1.0, "content": "bonus version which has boosted performance.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 }, { "type": "text", "bbox": [ 107, 345, 505, 412 ], "lines": [], "index": 10.5, "bbox_fs": [ 105, 345, 505, 412 ], "lines_deleted": true }, { "type": "text", "bbox": [ 106, 417, 505, 495 ], "lines": [ { "bbox": [ 105, 416, 505, 430 ], "spans": [ { "bbox": [ 105, 416, 505, 430 ], "score": 1.0, "content": "In this work, we advocate a novel positional encoding (PE) scheme to incorporate the position", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 428, 505, 441 ], "spans": [ { "bbox": [ 105, 428, 505, 441 ], "score": 1.0, "content": "information into Transformers. Unlike the predefined and input-agnostic positional encodings used", "type": "text" } ], "index": 15 }, { "bbox": [ 104, 438, 506, 453 ], "spans": [ { "bbox": [ 104, 438, 506, 453 ], "score": 1.0, "content": "in previous works (Dosovitskiy et al., 2021; Vaswani et al., 2017; Shaw et al., 2018), the proposed", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 450, 506, 464 ], "spans": [ { "bbox": [ 105, 450, 506, 464 ], "score": 1.0, "content": "PE is dynamically generated and conditioned on the local neighborhood of input tokens. Thus, our", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 461, 505, 474 ], "spans": [ { "bbox": [ 105, 461, 505, 474 ], "score": 1.0, "content": "positional encodings can change along with the input size and try to keep translation equivalence. We", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 472, 505, 485 ], "spans": [ { "bbox": [ 105, 472, 505, 485 ], "score": 1.0, "content": "demonstrate that the vision transformers (Dosovitskiy et al., 2021; Touvron et al., 2020) with our new", "type": "text" } ], "index": 19 }, { "bbox": [ 104, 482, 506, 497 ], "spans": [ { "bbox": [ 104, 482, 506, 497 ], "score": 1.0, "content": "PE (i.e. CPVT, see Figure 1c) achieve even better performance. We summarize our contributions as,", "type": "text" } ], "index": 20 }, { "bbox": [ 132, 512, 506, 526 ], "spans": [ { "bbox": [ 132, 512, 506, 526 ], "score": 1.0, "content": "• We propose a novel positional encoding (PE) scheme, termed conditional position encod-", "type": "text" } ], "index": 21 }, { "bbox": [ 140, 523, 506, 536 ], "spans": [ { "bbox": [ 140, 523, 506, 536 ], "score": 1.0, "content": "ings (CPE). CPE is dynamically generated with Positional Encoding Generators (PEG) and", "type": "text" } ], "index": 22 }, { "bbox": [ 141, 536, 506, 548 ], "spans": [ { "bbox": [ 141, 536, 506, 548 ], "score": 1.0, "content": "can be effortlessly implemented by the modern deep learning frameworks (Paszke et al.,", "type": "text" } ], "index": 23 }, { "bbox": [ 141, 545, 506, 559 ], "spans": [ { "bbox": [ 141, 545, 506, 559 ], "score": 1.0, "content": "2019; Abadi et al., 2016; Chen et al., 2015), requiring no changes to the current Trans-", "type": "text" } ], "index": 24 }, { "bbox": [ 141, 557, 506, 569 ], "spans": [ { "bbox": [ 141, 557, 506, 569 ], "score": 1.0, "content": "former APIs. Through an in-depth analysis and thorough experimentations, we unveil that", "type": "text" } ], "index": 25 }, { "bbox": [ 141, 568, 482, 581 ], "spans": [ { "bbox": [ 141, 568, 482, 581 ], "score": 1.0, "content": "this design affords both absolute and relative encoding yet it goes above and beyond.", "type": "text" } ], "index": 26 } ], "index": 17, "bbox_fs": [ 104, 416, 506, 497 ] }, { "type": "text", "bbox": [ 134, 513, 505, 579 ], "lines": [], "index": 23.5, "bbox_fs": [ 132, 512, 506, 581 ], "lines_deleted": true }, { "type": "text", "bbox": [ 134, 593, 504, 626 ], "lines": [ { "bbox": [ 133, 592, 506, 606 ], "spans": [ { "bbox": [ 133, 592, 506, 606 ], "score": 1.0, "content": "• As opposed to widely-used absolute positional encodings, CPE can provide a kind of", "type": "text" } ], "index": 27 }, { "bbox": [ 142, 604, 505, 616 ], "spans": [ { "bbox": [ 142, 604, 505, 616 ], "score": 1.0, "content": "stronger explicit bias towards the translation equivalence which is important to improve", "type": "text" } ], "index": 28 }, { "bbox": [ 142, 615, 279, 627 ], "spans": [ { "bbox": [ 142, 615, 279, 627 ], "score": 1.0, "content": "the performance of Transformers.", "type": "text" } ], "index": 29 } ], "index": 28, "bbox_fs": [ 133, 592, 506, 627 ] }, { "type": "text", "bbox": [ 134, 640, 505, 673 ], "lines": [ { "bbox": [ 133, 639, 506, 653 ], "spans": [ { "bbox": [ 133, 639, 506, 653 ], "score": 1.0, "content": "• Built on CPE, we propose Conditional Position encoding Vision Transformer (CPVT). It", "type": "text" } ], "index": 30 }, { "bbox": [ 141, 651, 505, 663 ], "spans": [ { "bbox": [ 141, 651, 505, 663 ], "score": 1.0, "content": "achieves better performance than previous vison transformers (Dosovitskiy et al., 2021;", "type": "text" } ], "index": 31 }, { "bbox": [ 142, 662, 231, 675 ], "spans": [ { "bbox": [ 142, 662, 231, 675 ], "score": 1.0, "content": "Touvron et al., 2020).", "type": "text" } ], "index": 32 } ], "index": 31, "bbox_fs": [ 133, 639, 506, 675 ] }, { "type": "text", "bbox": [ 135, 687, 504, 732 ], "lines": [ { "bbox": [ 133, 687, 505, 701 ], "spans": [ { "bbox": [ 133, 687, 505, 701 ], "score": 1.0, "content": "• CPE can well generalize to arbitrary input resolutions, which are required in many impor-", "type": "text" } ], "index": 33 }, { "bbox": [ 141, 699, 505, 711 ], "spans": [ { "bbox": [ 141, 699, 505, 711 ], "score": 1.0, "content": "tant downstream tasks such as segmentation and detection. Through experiments we show", "type": "text" } ], "index": 34 }, { "bbox": [ 141, 709, 505, 722 ], "spans": [ { "bbox": [ 141, 709, 505, 722 ], "score": 1.0, "content": "that CPE can boost the segmentation and detection performance for pyramid transformers", "type": "text" } ], "index": 35 }, { "bbox": [ 141, 721, 312, 733 ], "spans": [ { "bbox": [ 141, 721, 312, 733 ], "score": 1.0, "content": "like (Wang et al., 2021) by a clear margin.", "type": "text" } ], "index": 36 } ], "index": 34.5, "bbox_fs": [ 133, 687, 505, 733 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 81, 211, 93 ], "lines": [ { "bbox": [ 104, 79, 213, 96 ], "spans": [ { "bbox": [ 104, 79, 213, 96 ], "score": 1.0, "content": "2 RELATED WORK", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 106, 504, 150 ], "lines": [ { "bbox": [ 105, 106, 505, 119 ], "spans": [ { "bbox": [ 105, 106, 505, 119 ], "score": 1.0, "content": "Since self-attention itself is permutation-equivariant (see A), positional encodings are commonly", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 117, 506, 130 ], "spans": [ { "bbox": [ 105, 117, 506, 130 ], "score": 1.0, "content": "employed to incorporate the order of sequences (Vaswani et al., 2017). The positional encodings", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 128, 505, 140 ], "spans": [ { "bbox": [ 105, 128, 505, 140 ], "score": 1.0, "content": "can either be fixed or learnable, while either being absolute or relative. Vision transformers follow", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 139, 377, 151 ], "spans": [ { "bbox": [ 105, 139, 377, 151 ], "score": 1.0, "content": "the same fashion to imbue the network with positional information.", "type": "text" } ], "index": 4 } ], "index": 2.5 }, { "type": "text", "bbox": [ 107, 162, 505, 217 ], "lines": [ { "bbox": [ 105, 161, 505, 175 ], "spans": [ { "bbox": [ 105, 161, 505, 175 ], "score": 1.0, "content": "Absolute Positional Encoding. The absolute positional encoding is the most widely used. In", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 174, 505, 186 ], "spans": [ { "bbox": [ 106, 174, 505, 186 ], "score": 1.0, "content": "the original transformer (Vaswani et al., 2017), the encodings are generated with the sinusoidal", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 185, 505, 196 ], "spans": [ { "bbox": [ 106, 185, 505, 196 ], "score": 1.0, "content": "functions of different frequencies and then they are added to the inputs. Alternatively, the positional", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 196, 505, 208 ], "spans": [ { "bbox": [ 106, 196, 505, 208 ], "score": 1.0, "content": "encodings can be learnable, where they are implemented with a fixed-dimension matrix/tensor and", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 207, 329, 219 ], "spans": [ { "bbox": [ 105, 207, 329, 219 ], "score": 1.0, "content": "jointly updated with the model’s parameters with SGD.", "type": "text" } ], "index": 9 } ], "index": 7 }, { "type": "text", "bbox": [ 107, 230, 504, 307 ], "lines": [ { "bbox": [ 105, 230, 506, 243 ], "spans": [ { "bbox": [ 105, 230, 506, 243 ], "score": 1.0, "content": "Relative Positional Encoding. The relative position encoding (Shaw et al., 2018) considers dis-", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 242, 505, 253 ], "spans": [ { "bbox": [ 106, 242, 505, 253 ], "score": 1.0, "content": "tances between the tokens in the input sequence. Compared to the absolute ones, the relative posi-", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 253, 505, 264 ], "spans": [ { "bbox": [ 106, 253, 505, 264 ], "score": 1.0, "content": "tional encodings can be translation-equivariant and can naturally handle the sequences longer than", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 263, 506, 276 ], "spans": [ { "bbox": [ 105, 263, 506, 276 ], "score": 1.0, "content": "the longest sequences during training (i.e., being inductive). A 2-D relative position encoding is", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 273, 506, 286 ], "spans": [ { "bbox": [ 105, 273, 506, 286 ], "score": 1.0, "content": "proposed for image classification in (Bello et al., 2019), showing superiority to 2D sinusoidal em-", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 285, 505, 297 ], "spans": [ { "bbox": [ 105, 285, 505, 297 ], "score": 1.0, "content": "beddings. The relative positional encoding is further improved in XLNet (Yang et al., 2019b) and", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 295, 336, 308 ], "spans": [ { "bbox": [ 105, 295, 336, 308 ], "score": 1.0, "content": "DeBERTa (He et al., 2020), showing better performance.", "type": "text" } ], "index": 16 } ], "index": 13 }, { "type": "text", "bbox": [ 107, 319, 504, 375 ], "lines": [ { "bbox": [ 106, 319, 505, 333 ], "spans": [ { "bbox": [ 106, 319, 505, 333 ], "score": 1.0, "content": "Other forms. Complex-value embeddings (Wang et al., 2019) are an extension to model global", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 329, 505, 343 ], "spans": [ { "bbox": [ 105, 329, 505, 343 ], "score": 1.0, "content": "absolute encodings and show improvement. RoFormer (Su et al., 2021) utilizes a rotary position em-", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 342, 505, 353 ], "spans": [ { "bbox": [ 105, 342, 505, 353 ], "score": 1.0, "content": "bedding to encode both absolute and relative position information for text classification. FLOATER", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 352, 506, 365 ], "spans": [ { "bbox": [ 105, 352, 506, 365 ], "score": 1.0, "content": "(Liu et al., 2020) proposes a novel continuous dynamical model to capture position encodings. It is", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 363, 504, 376 ], "spans": [ { "bbox": [ 105, 363, 504, 376 ], "score": 1.0, "content": "not limited by the maximum sequence length during training, meanwhile being parameter-efficient.", "type": "text" } ], "index": 21 } ], "index": 19 }, { "type": "text", "bbox": [ 108, 386, 504, 420 ], "lines": [ { "bbox": [ 106, 387, 505, 399 ], "spans": [ { "bbox": [ 106, 387, 505, 399 ], "score": 1.0, "content": "Similar designs to CPE. Convolutions are used to model local relations in ASR and machine", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 397, 505, 411 ], "spans": [ { "bbox": [ 105, 397, 505, 411 ], "score": 1.0, "content": "translation (Gulati et al., 2020; Mohamed et al., 2019; Yang et al., 2019a; Yu et al., 2018). However,", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 408, 413, 421 ], "spans": [ { "bbox": [ 106, 408, 413, 421 ], "score": 1.0, "content": "they are mainly limited to 1D signals. We instead process 2D vision images.", "type": "text" } ], "index": 24 } ], "index": 23 }, { "type": "title", "bbox": [ 106, 435, 470, 449 ], "lines": [ { "bbox": [ 104, 434, 472, 451 ], "spans": [ { "bbox": [ 104, 434, 472, 451 ], "score": 1.0, "content": "3 VISION TRANSFORMER WITH CONDITIONAL POSITION ENCODINGS", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "title", "bbox": [ 107, 460, 187, 472 ], "lines": [ { "bbox": [ 105, 460, 189, 474 ], "spans": [ { "bbox": [ 105, 460, 189, 474 ], "score": 1.0, "content": "3.1 MOTIVATION", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 107, 481, 505, 559 ], "lines": [ { "bbox": [ 105, 481, 505, 493 ], "spans": [ { "bbox": [ 105, 481, 297, 493 ], "score": 1.0, "content": "In vision transformers, an input image of size", "type": "text" }, { "bbox": [ 298, 482, 332, 492 ], "score": 0.91, "content": "H \\times W", "type": "inline_equation" }, { "bbox": [ 333, 481, 457, 493 ], "score": 1.0, "content": "is split into patches with size", "type": "text" }, { "bbox": [ 457, 482, 486, 492 ], "score": 0.91, "content": "S \\times S", "type": "inline_equation" }, { "bbox": [ 486, 481, 505, 493 ], "score": 1.0, "content": ", the", "type": "text" } ], "index": 27 }, { "bbox": [ 102, 488, 507, 507 ], "spans": [ { "bbox": [ 102, 488, 190, 507 ], "score": 1.0, "content": "number of patches is", "type": "text" }, { "bbox": [ 191, 491, 236, 505 ], "score": 0.92, "content": "\\begin{array} { r } { \\dot { N } = \\frac { H \\dot { W } \\mathbb { 1 } } { S ^ { 2 } } } \\end{array}", "type": "inline_equation" }, { "bbox": [ 236, 488, 507, 507 ], "score": 1.0, "content": ". The patches are added with the same number of learnable absolute", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 503, 505, 516 ], "spans": [ { "bbox": [ 105, 503, 505, 516 ], "score": 1.0, "content": "positional encoding vectors. In this work, we argue that the positional encodings used here have", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 515, 504, 527 ], "spans": [ { "bbox": [ 106, 515, 504, 527 ], "score": 1.0, "content": "two issues. First, it prevents the model from handling the sequences longer than the learnable PE.", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 525, 505, 538 ], "spans": [ { "bbox": [ 105, 525, 505, 538 ], "score": 1.0, "content": "Second, it makes the model not translation-equivariant because a unique positional encoding vector", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 537, 505, 549 ], "spans": [ { "bbox": [ 105, 537, 505, 549 ], "score": 1.0, "content": "is added to every one patch. The translation equivalence plays an important role in classification", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 547, 493, 561 ], "spans": [ { "bbox": [ 105, 547, 493, 561 ], "score": 1.0, "content": "because we hope the networks’ responses changes accordingly as the object moves in the image.", "type": "text" } ], "index": 33 } ], "index": 30 }, { "type": "text", "bbox": [ 107, 564, 505, 641 ], "lines": [ { "bbox": [ 105, 563, 505, 577 ], "spans": [ { "bbox": [ 105, 563, 505, 577 ], "score": 1.0, "content": "One may note that the first issue can be remedied by removing the positional encodings since except", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 576, 505, 588 ], "spans": [ { "bbox": [ 105, 576, 505, 588 ], "score": 1.0, "content": "for the positional encodings, all other components (e.g., MHSA and FFN) of the vision transformer", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 586, 505, 599 ], "spans": [ { "bbox": [ 105, 586, 505, 599 ], "score": 1.0, "content": "can directly be applied to longer sequences. However, this solution severely deteriorates the perfor-", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 597, 505, 610 ], "spans": [ { "bbox": [ 105, 597, 505, 610 ], "score": 1.0, "content": "mance. This is understandable because the order of the input sequence is an important clue and the", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 608, 505, 621 ], "spans": [ { "bbox": [ 105, 608, 505, 621 ], "score": 1.0, "content": "model has no way to extract the order without the positional encodings. The experiment results on", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 618, 506, 632 ], "spans": [ { "bbox": [ 105, 618, 506, 632 ], "score": 1.0, "content": "ImageNet are shown in Table 1. By removing the positional encodings, DeiT-tiny’s performance on", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 630, 329, 643 ], "spans": [ { "bbox": [ 106, 630, 260, 643 ], "score": 1.0, "content": "ImageNet dramatically degrades from", "type": "text" }, { "bbox": [ 260, 630, 287, 641 ], "score": 0.85, "content": "7 2 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 288, 630, 298, 643 ], "score": 1.0, "content": "to", "type": "text" }, { "bbox": [ 299, 630, 325, 641 ], "score": 0.86, "content": "6 8 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 326, 630, 329, 643 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 40 } ], "index": 37 }, { "type": "text", "bbox": [ 107, 647, 505, 713 ], "lines": [ { "bbox": [ 105, 646, 505, 660 ], "spans": [ { "bbox": [ 105, 646, 505, 660 ], "score": 1.0, "content": "Second, in DeiT (Touvron et al., 2020), they show that we can interpolate the position encodings", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 658, 505, 670 ], "spans": [ { "bbox": [ 105, 658, 505, 670 ], "score": 1.0, "content": "to make them have the same length of the longer sequences. However, this method requires fine-", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 669, 505, 681 ], "spans": [ { "bbox": [ 106, 669, 505, 681 ], "score": 1.0, "content": "tuning the model a few more epochs, otherwise the performance will remarkably drop, as shown", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 678, 505, 693 ], "spans": [ { "bbox": [ 105, 678, 505, 693 ], "score": 1.0, "content": "in Table 1. This goes contrary to what we would expect. With the higher-resolution inputs, we", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 691, 505, 704 ], "spans": [ { "bbox": [ 105, 691, 505, 704 ], "score": 1.0, "content": "often expect a remarkable performance improvement without any fine-tuning. Finally, the relative", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 702, 505, 714 ], "spans": [ { "bbox": [ 105, 702, 505, 714 ], "score": 1.0, "content": "position encodings (Shaw et al., 2018; Bello et al., 2019) can cope with both the aforementioned", "type": "text" } ], "index": 46 } ], "index": 43.5 } ], "page_idx": 2, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 120, 721, 291, 732 ], "lines": [ { "bbox": [ 119, 718, 292, 734 ], "spans": [ { "bbox": [ 119, 718, 123, 734 ], "score": 1.0, "content": "1", "type": "text" }, { "bbox": [ 123, 721, 132, 730 ], "score": 0.57, "content": "H", "type": "inline_equation" }, { "bbox": [ 133, 718, 235, 734 ], "score": 1.0, "content": "and W shall be divisible by", "type": "text" }, { "bbox": [ 235, 722, 242, 730 ], "score": 0.75, "content": "S", "type": "inline_equation" }, { "bbox": [ 242, 718, 292, 734 ], "score": 1.0, "content": ", respectively.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 25, 293, 38 ], "spans": [ { "bbox": [ 106, 25, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 760 ], "lines": [ { "bbox": [ 301, 750, 310, 762 ], "spans": [ { "bbox": [ 301, 750, 310, 762 ], "score": 1.0, "content": "3", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 81, 211, 93 ], "lines": [ { "bbox": [ 104, 79, 213, 96 ], "spans": [ { "bbox": [ 104, 79, 213, 96 ], "score": 1.0, "content": "2 RELATED WORK", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 106, 504, 150 ], "lines": [ { "bbox": [ 105, 106, 505, 119 ], "spans": [ { "bbox": [ 105, 106, 505, 119 ], "score": 1.0, "content": "Since self-attention itself is permutation-equivariant (see A), positional encodings are commonly", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 117, 506, 130 ], "spans": [ { "bbox": [ 105, 117, 506, 130 ], "score": 1.0, "content": "employed to incorporate the order of sequences (Vaswani et al., 2017). The positional encodings", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 128, 505, 140 ], "spans": [ { "bbox": [ 105, 128, 505, 140 ], "score": 1.0, "content": "can either be fixed or learnable, while either being absolute or relative. Vision transformers follow", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 139, 377, 151 ], "spans": [ { "bbox": [ 105, 139, 377, 151 ], "score": 1.0, "content": "the same fashion to imbue the network with positional information.", "type": "text" } ], "index": 4 } ], "index": 2.5, "bbox_fs": [ 105, 106, 506, 151 ] }, { "type": "text", "bbox": [ 107, 162, 505, 217 ], "lines": [ { "bbox": [ 105, 161, 505, 175 ], "spans": [ { "bbox": [ 105, 161, 505, 175 ], "score": 1.0, "content": "Absolute Positional Encoding. The absolute positional encoding is the most widely used. In", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 174, 505, 186 ], "spans": [ { "bbox": [ 106, 174, 505, 186 ], "score": 1.0, "content": "the original transformer (Vaswani et al., 2017), the encodings are generated with the sinusoidal", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 185, 505, 196 ], "spans": [ { "bbox": [ 106, 185, 505, 196 ], "score": 1.0, "content": "functions of different frequencies and then they are added to the inputs. Alternatively, the positional", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 196, 505, 208 ], "spans": [ { "bbox": [ 106, 196, 505, 208 ], "score": 1.0, "content": "encodings can be learnable, where they are implemented with a fixed-dimension matrix/tensor and", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 207, 329, 219 ], "spans": [ { "bbox": [ 105, 207, 329, 219 ], "score": 1.0, "content": "jointly updated with the model’s parameters with SGD.", "type": "text" } ], "index": 9 } ], "index": 7, "bbox_fs": [ 105, 161, 505, 219 ] }, { "type": "text", "bbox": [ 107, 230, 504, 307 ], "lines": [ { "bbox": [ 105, 230, 506, 243 ], "spans": [ { "bbox": [ 105, 230, 506, 243 ], "score": 1.0, "content": "Relative Positional Encoding. The relative position encoding (Shaw et al., 2018) considers dis-", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 242, 505, 253 ], "spans": [ { "bbox": [ 106, 242, 505, 253 ], "score": 1.0, "content": "tances between the tokens in the input sequence. Compared to the absolute ones, the relative posi-", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 253, 505, 264 ], "spans": [ { "bbox": [ 106, 253, 505, 264 ], "score": 1.0, "content": "tional encodings can be translation-equivariant and can naturally handle the sequences longer than", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 263, 506, 276 ], "spans": [ { "bbox": [ 105, 263, 506, 276 ], "score": 1.0, "content": "the longest sequences during training (i.e., being inductive). A 2-D relative position encoding is", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 273, 506, 286 ], "spans": [ { "bbox": [ 105, 273, 506, 286 ], "score": 1.0, "content": "proposed for image classification in (Bello et al., 2019), showing superiority to 2D sinusoidal em-", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 285, 505, 297 ], "spans": [ { "bbox": [ 105, 285, 505, 297 ], "score": 1.0, "content": "beddings. The relative positional encoding is further improved in XLNet (Yang et al., 2019b) and", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 295, 336, 308 ], "spans": [ { "bbox": [ 105, 295, 336, 308 ], "score": 1.0, "content": "DeBERTa (He et al., 2020), showing better performance.", "type": "text" } ], "index": 16 } ], "index": 13, "bbox_fs": [ 105, 230, 506, 308 ] }, { "type": "text", "bbox": [ 107, 319, 504, 375 ], "lines": [ { "bbox": [ 106, 319, 505, 333 ], "spans": [ { "bbox": [ 106, 319, 505, 333 ], "score": 1.0, "content": "Other forms. Complex-value embeddings (Wang et al., 2019) are an extension to model global", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 329, 505, 343 ], "spans": [ { "bbox": [ 105, 329, 505, 343 ], "score": 1.0, "content": "absolute encodings and show improvement. RoFormer (Su et al., 2021) utilizes a rotary position em-", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 342, 505, 353 ], "spans": [ { "bbox": [ 105, 342, 505, 353 ], "score": 1.0, "content": "bedding to encode both absolute and relative position information for text classification. FLOATER", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 352, 506, 365 ], "spans": [ { "bbox": [ 105, 352, 506, 365 ], "score": 1.0, "content": "(Liu et al., 2020) proposes a novel continuous dynamical model to capture position encodings. It is", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 363, 504, 376 ], "spans": [ { "bbox": [ 105, 363, 504, 376 ], "score": 1.0, "content": "not limited by the maximum sequence length during training, meanwhile being parameter-efficient.", "type": "text" } ], "index": 21 } ], "index": 19, "bbox_fs": [ 105, 319, 506, 376 ] }, { "type": "text", "bbox": [ 108, 386, 504, 420 ], "lines": [ { "bbox": [ 106, 387, 505, 399 ], "spans": [ { "bbox": [ 106, 387, 505, 399 ], "score": 1.0, "content": "Similar designs to CPE. Convolutions are used to model local relations in ASR and machine", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 397, 505, 411 ], "spans": [ { "bbox": [ 105, 397, 505, 411 ], "score": 1.0, "content": "translation (Gulati et al., 2020; Mohamed et al., 2019; Yang et al., 2019a; Yu et al., 2018). However,", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 408, 413, 421 ], "spans": [ { "bbox": [ 106, 408, 413, 421 ], "score": 1.0, "content": "they are mainly limited to 1D signals. We instead process 2D vision images.", "type": "text" } ], "index": 24 } ], "index": 23, "bbox_fs": [ 105, 387, 505, 421 ] }, { "type": "title", "bbox": [ 106, 435, 470, 449 ], "lines": [ { "bbox": [ 104, 434, 472, 451 ], "spans": [ { "bbox": [ 104, 434, 472, 451 ], "score": 1.0, "content": "3 VISION TRANSFORMER WITH CONDITIONAL POSITION ENCODINGS", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "title", "bbox": [ 107, 460, 187, 472 ], "lines": [ { "bbox": [ 105, 460, 189, 474 ], "spans": [ { "bbox": [ 105, 460, 189, 474 ], "score": 1.0, "content": "3.1 MOTIVATION", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 107, 481, 505, 559 ], "lines": [ { "bbox": [ 105, 481, 505, 493 ], "spans": [ { "bbox": [ 105, 481, 297, 493 ], "score": 1.0, "content": "In vision transformers, an input image of size", "type": "text" }, { "bbox": [ 298, 482, 332, 492 ], "score": 0.91, "content": "H \\times W", "type": "inline_equation" }, { "bbox": [ 333, 481, 457, 493 ], "score": 1.0, "content": "is split into patches with size", "type": "text" }, { "bbox": [ 457, 482, 486, 492 ], "score": 0.91, "content": "S \\times S", "type": "inline_equation" }, { "bbox": [ 486, 481, 505, 493 ], "score": 1.0, "content": ", the", "type": "text" } ], "index": 27 }, { "bbox": [ 102, 488, 507, 507 ], "spans": [ { "bbox": [ 102, 488, 190, 507 ], "score": 1.0, "content": "number of patches is", "type": "text" }, { "bbox": [ 191, 491, 236, 505 ], "score": 0.92, "content": "\\begin{array} { r } { \\dot { N } = \\frac { H \\dot { W } \\mathbb { 1 } } { S ^ { 2 } } } \\end{array}", "type": "inline_equation" }, { "bbox": [ 236, 488, 507, 507 ], "score": 1.0, "content": ". The patches are added with the same number of learnable absolute", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 503, 505, 516 ], "spans": [ { "bbox": [ 105, 503, 505, 516 ], "score": 1.0, "content": "positional encoding vectors. In this work, we argue that the positional encodings used here have", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 515, 504, 527 ], "spans": [ { "bbox": [ 106, 515, 504, 527 ], "score": 1.0, "content": "two issues. First, it prevents the model from handling the sequences longer than the learnable PE.", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 525, 505, 538 ], "spans": [ { "bbox": [ 105, 525, 505, 538 ], "score": 1.0, "content": "Second, it makes the model not translation-equivariant because a unique positional encoding vector", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 537, 505, 549 ], "spans": [ { "bbox": [ 105, 537, 505, 549 ], "score": 1.0, "content": "is added to every one patch. The translation equivalence plays an important role in classification", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 547, 493, 561 ], "spans": [ { "bbox": [ 105, 547, 493, 561 ], "score": 1.0, "content": "because we hope the networks’ responses changes accordingly as the object moves in the image.", "type": "text" } ], "index": 33 } ], "index": 30, "bbox_fs": [ 102, 481, 507, 561 ] }, { "type": "text", "bbox": [ 107, 564, 505, 641 ], "lines": [ { "bbox": [ 105, 563, 505, 577 ], "spans": [ { "bbox": [ 105, 563, 505, 577 ], "score": 1.0, "content": "One may note that the first issue can be remedied by removing the positional encodings since except", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 576, 505, 588 ], "spans": [ { "bbox": [ 105, 576, 505, 588 ], "score": 1.0, "content": "for the positional encodings, all other components (e.g., MHSA and FFN) of the vision transformer", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 586, 505, 599 ], "spans": [ { "bbox": [ 105, 586, 505, 599 ], "score": 1.0, "content": "can directly be applied to longer sequences. However, this solution severely deteriorates the perfor-", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 597, 505, 610 ], "spans": [ { "bbox": [ 105, 597, 505, 610 ], "score": 1.0, "content": "mance. This is understandable because the order of the input sequence is an important clue and the", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 608, 505, 621 ], "spans": [ { "bbox": [ 105, 608, 505, 621 ], "score": 1.0, "content": "model has no way to extract the order without the positional encodings. The experiment results on", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 618, 506, 632 ], "spans": [ { "bbox": [ 105, 618, 506, 632 ], "score": 1.0, "content": "ImageNet are shown in Table 1. By removing the positional encodings, DeiT-tiny’s performance on", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 630, 329, 643 ], "spans": [ { "bbox": [ 106, 630, 260, 643 ], "score": 1.0, "content": "ImageNet dramatically degrades from", "type": "text" }, { "bbox": [ 260, 630, 287, 641 ], "score": 0.85, "content": "7 2 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 288, 630, 298, 643 ], "score": 1.0, "content": "to", "type": "text" }, { "bbox": [ 299, 630, 325, 641 ], "score": 0.86, "content": "6 8 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 326, 630, 329, 643 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 40 } ], "index": 37, "bbox_fs": [ 105, 563, 506, 643 ] }, { "type": "text", "bbox": [ 107, 647, 505, 713 ], "lines": [ { "bbox": [ 105, 646, 505, 660 ], "spans": [ { "bbox": [ 105, 646, 505, 660 ], "score": 1.0, "content": "Second, in DeiT (Touvron et al., 2020), they show that we can interpolate the position encodings", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 658, 505, 670 ], "spans": [ { "bbox": [ 105, 658, 505, 670 ], "score": 1.0, "content": "to make them have the same length of the longer sequences. However, this method requires fine-", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 669, 505, 681 ], "spans": [ { "bbox": [ 106, 669, 505, 681 ], "score": 1.0, "content": "tuning the model a few more epochs, otherwise the performance will remarkably drop, as shown", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 678, 505, 693 ], "spans": [ { "bbox": [ 105, 678, 505, 693 ], "score": 1.0, "content": "in Table 1. This goes contrary to what we would expect. With the higher-resolution inputs, we", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 691, 505, 704 ], "spans": [ { "bbox": [ 105, 691, 505, 704 ], "score": 1.0, "content": "often expect a remarkable performance improvement without any fine-tuning. Finally, the relative", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 702, 505, 714 ], "spans": [ { "bbox": [ 105, 702, 505, 714 ], "score": 1.0, "content": "position encodings (Shaw et al., 2018; Bello et al., 2019) can cope with both the aforementioned", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 196, 505, 209 ], "spans": [ { "bbox": [ 105, 196, 505, 209 ], "score": 1.0, "content": "issues. However, the relative positional encoding cannot provide absolute position information,", "type": "text", "cross_page": true } ], "index": 6 }, { "bbox": [ 106, 207, 505, 219 ], "spans": [ { "bbox": [ 106, 207, 505, 219 ], "score": 1.0, "content": "which is also important to the classification performance (Islam et al., 2020). As shown in Table 1,", "type": "text", "cross_page": true } ], "index": 7 }, { "bbox": [ 106, 218, 459, 231 ], "spans": [ { "bbox": [ 106, 218, 383, 231 ], "score": 1.0, "content": "the model with relative position encodings has inferior performance (", "type": "text", "cross_page": true }, { "bbox": [ 383, 218, 410, 229 ], "score": 0.85, "content": "7 0 . 5 \\%", "type": "inline_equation", "cross_page": true }, { "bbox": [ 410, 218, 425, 231 ], "score": 1.0, "content": "vs.", "type": "text", "cross_page": true }, { "bbox": [ 425, 218, 453, 229 ], "score": 0.86, "content": "7 2 . 2 \\%", "type": "inline_equation", "cross_page": true }, { "bbox": [ 453, 218, 459, 231 ], "score": 1.0, "content": ").", "type": "text", "cross_page": true } ], "index": 8 } ], "index": 43.5, "bbox_fs": [ 105, 646, 505, 714 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 116, 128, 495, 186 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 80, 504, 114 ], "group_id": 0, "lines": [ { "bbox": [ 106, 79, 505, 94 ], "spans": [ { "bbox": [ 106, 79, 505, 94 ], "score": 1.0, "content": "Table 1. Comparison of various positional encoding (PE) strategies tested on ImageNet valida-", "type": "text" } ], "index": 0 }, { "bbox": [ 107, 91, 504, 104 ], "spans": [ { "bbox": [ 107, 91, 504, 104 ], "score": 1.0, "content": "tion set in terms of the top-1 accuracy. Removing the positional encodings greatly damages the", "type": "text" } ], "index": 1 }, { "bbox": [ 107, 103, 486, 115 ], "spans": [ { "bbox": [ 107, 103, 486, 115 ], "score": 1.0, "content": "performance. The relative positional encodings have inferior performance to the absolute ones", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "table_body", "bbox": [ 116, 128, 495, 186 ], "group_id": 0, "lines": [ { "bbox": [ 116, 128, 495, 186 ], "spans": [ { "bbox": [ 116, 128, 495, 186 ], "score": 0.977, "html": "
ModelEncodingTop-1@224(%)Top-1@384(%)
DeiT-tiny (Touvron et al., 2020)X68.268.6
DeiT-tiny (Touvron et al., 2020)learnable72.271.2
DeiT-tiny (Touvron et al., 2020)sin-cos72.370.8
DeiT-tiny2D RPE (Shaw et al., 2018)70.569.8
", "type": "table", "image_path": "777658109f20f506564abf78728d41e9469175d72aa4de071291272306f43b7e.jpg" } ] } ], "index": 4, "virtual_lines": [ { "bbox": [ 116, 128, 495, 147.33333333333334 ], "spans": [], "index": 3 }, { "bbox": [ 116, 147.33333333333334, 495, 166.66666666666669 ], "spans": [], "index": 4 }, { "bbox": [ 116, 166.66666666666669, 495, 186.00000000000003 ], "spans": [], "index": 5 } ] } ], "index": 2.5 }, { "type": "text", "bbox": [ 108, 196, 505, 230 ], "lines": [ { "bbox": [ 105, 196, 505, 209 ], "spans": [ { "bbox": [ 105, 196, 505, 209 ], "score": 1.0, "content": "issues. However, the relative positional encoding cannot provide absolute position information,", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 207, 505, 219 ], "spans": [ { "bbox": [ 106, 207, 505, 219 ], "score": 1.0, "content": "which is also important to the classification performance (Islam et al., 2020). As shown in Table 1,", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 218, 459, 231 ], "spans": [ { "bbox": [ 106, 218, 383, 231 ], "score": 1.0, "content": "the model with relative position encodings has inferior performance (", "type": "text" }, { "bbox": [ 383, 218, 410, 229 ], "score": 0.85, "content": "7 0 . 5 \\%", "type": "inline_equation" }, { "bbox": [ 410, 218, 425, 231 ], "score": 1.0, "content": "vs.", "type": "text" }, { "bbox": [ 425, 218, 453, 229 ], "score": 0.86, "content": "7 2 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 453, 218, 459, 231 ], "score": 1.0, "content": ").", "type": "text" } ], "index": 8 } ], "index": 7 }, { "type": "title", "bbox": [ 108, 243, 303, 254 ], "lines": [ { "bbox": [ 105, 241, 304, 256 ], "spans": [ { "bbox": [ 105, 241, 304, 256 ], "score": 1.0, "content": "3.2 CONDITIONAL POSITIONAL ENCODINGS", "type": "text" } ], "index": 9 } ], "index": 9 }, { "type": "text", "bbox": [ 111, 263, 488, 275 ], "lines": [ { "bbox": [ 108, 262, 490, 278 ], "spans": [ { "bbox": [ 108, 262, 490, 278 ], "score": 1.0, "content": "We argue that a successful positional encoding for vision tasks should meet these requirements,", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 126, 284, 505, 347 ], "lines": [ { "bbox": [ 125, 284, 504, 298 ], "spans": [ { "bbox": [ 125, 284, 504, 298 ], "score": 1.0, "content": "(1) Making the input sequence permutation-variant and providing stronger explicit bias to-", "type": "text" } ], "index": 11 }, { "bbox": [ 142, 296, 272, 307 ], "spans": [ { "bbox": [ 142, 296, 272, 307 ], "score": 1.0, "content": "wards translation-equivariance.", "type": "text" } ], "index": 12 }, { "bbox": [ 125, 309, 488, 324 ], "spans": [ { "bbox": [ 125, 309, 488, 324 ], "score": 1.0, "content": "(2) Being inductive and able to handle the sequences longer than the ones during training.", "type": "text" } ], "index": 13 }, { "bbox": [ 125, 324, 505, 338 ], "spans": [ { "bbox": [ 125, 324, 505, 338 ], "score": 1.0, "content": "(3) Having the ability to provide the absolute position to a certain degree. This is important to", "type": "text" } ], "index": 14 }, { "bbox": [ 142, 336, 342, 348 ], "spans": [ { "bbox": [ 142, 336, 342, 348 ], "score": 1.0, "content": "the performance as shown in (Islam et al., 2020).", "type": "text" } ], "index": 15 } ], "index": 13 }, { "type": "text", "bbox": [ 106, 356, 505, 455 ], "lines": [ { "bbox": [ 105, 356, 506, 369 ], "spans": [ { "bbox": [ 105, 356, 506, 369 ], "score": 1.0, "content": "In this work, we find that characterizing the local relationship by positional encodings is sufficient", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 367, 506, 381 ], "spans": [ { "bbox": [ 105, 367, 506, 381 ], "score": 1.0, "content": "to meet all of the above. First, it is permutation-variant because the permutation of input sequences", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 379, 504, 390 ], "spans": [ { "bbox": [ 106, 379, 504, 390 ], "score": 1.0, "content": "also affects the order in some local neighborhoods. However, translation of an object in an input im-", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 388, 505, 402 ], "spans": [ { "bbox": [ 105, 388, 505, 402 ], "score": 1.0, "content": "age does not change the order in its local neighborhood, i.e., translation-equivariant (see Section A).", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 400, 506, 413 ], "spans": [ { "bbox": [ 106, 400, 506, 413 ], "score": 1.0, "content": "Second, the model can easily generalize to longer sequences since only the local neighborhoods of a", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 410, 505, 424 ], "spans": [ { "bbox": [ 105, 410, 505, 424 ], "score": 1.0, "content": "token are involved. Besides, if the absolute position of any input token is known, the absolute posi-", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 422, 505, 434 ], "spans": [ { "bbox": [ 106, 422, 505, 434 ], "score": 1.0, "content": "tion of all the other tokens can be inferred by the mutual relation between input tokens. We will show", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 433, 505, 446 ], "spans": [ { "bbox": [ 106, 433, 505, 446 ], "score": 1.0, "content": "that the tokens on the borders can be aware of their absolute positions due to the commonly-used", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 442, 168, 458 ], "spans": [ { "bbox": [ 105, 442, 168, 458 ], "score": 1.0, "content": "zero paddings.", "type": "text" } ], "index": 24 } ], "index": 20 }, { "type": "text", "bbox": [ 106, 461, 210, 548 ], "lines": [ { "bbox": [ 106, 460, 212, 474 ], "spans": [ { "bbox": [ 106, 460, 212, 474 ], "score": 1.0, "content": "Therefore, we propose", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 472, 212, 484 ], "spans": [ { "bbox": [ 106, 472, 212, 484 ], "score": 1.0, "content": "positional encoding gen-", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 483, 211, 495 ], "spans": [ { "bbox": [ 106, 483, 211, 495 ], "score": 1.0, "content": "erators (PEG) to dynam-", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 493, 212, 506 ], "spans": [ { "bbox": [ 105, 493, 212, 506 ], "score": 1.0, "content": "ically produce the posi-", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 504, 212, 516 ], "spans": [ { "bbox": [ 105, 504, 212, 516 ], "score": 1.0, "content": "tional encodings condi-", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 515, 212, 528 ], "spans": [ { "bbox": [ 105, 515, 212, 528 ], "score": 1.0, "content": "tioned on the local neigh-", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 526, 212, 539 ], "spans": [ { "bbox": [ 105, 526, 212, 539 ], "score": 1.0, "content": "borhood of an input to-", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 536, 127, 550 ], "spans": [ { "bbox": [ 105, 536, 127, 550 ], "score": 1.0, "content": "ken.", "type": "text" } ], "index": 32 } ], "index": 28.5 }, { "type": "text", "bbox": [ 106, 554, 212, 610 ], "lines": [ { "bbox": [ 105, 552, 213, 567 ], "spans": [ { "bbox": [ 105, 552, 152, 566 ], "score": 1.0, "content": "Positional", "type": "text" }, { "bbox": [ 167, 552, 213, 567 ], "score": 1.0, "content": "Encoding", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 564, 213, 577 ], "spans": [ { "bbox": [ 105, 564, 156, 577 ], "score": 1.0, "content": "Generator.", "type": "text" }, { "bbox": [ 172, 564, 213, 577 ], "score": 1.0, "content": "PEG is", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 575, 212, 588 ], "spans": [ { "bbox": [ 105, 575, 212, 588 ], "score": 1.0, "content": "illustrated in Figure 2.", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 586, 212, 599 ], "spans": [ { "bbox": [ 105, 586, 212, 599 ], "score": 1.0, "content": "To condition on the", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 598, 212, 611 ], "spans": [ { "bbox": [ 105, 598, 212, 611 ], "score": 1.0, "content": "local neighbors, we first", "type": "text" } ], "index": 42 } ], "index": 35 }, { "type": "image", "bbox": [ 223, 461, 502, 557 ], "blocks": [ { "type": "image_body", "bbox": [ 223, 461, 502, 557 ], "group_id": 0, "lines": [ { "bbox": [ 223, 461, 502, 557 ], "spans": [ { "bbox": [ 223, 461, 502, 557 ], "score": 0.969, "type": "image", "image_path": "e12ed07e330e1cbad3fc807996e87af006eec2275384ddc7e13bc09ba9a2aac7.jpg" } ] } ], "index": 38, "virtual_lines": [ { "bbox": [ 223, 461, 502, 493.0 ], "spans": [], "index": 37 }, { "bbox": [ 223, 493.0, 502, 525.0 ], "spans": [], "index": 38 }, { "bbox": [ 223, 525.0, 502, 557.0 ], "spans": [], "index": 39 } ] }, { "type": "image_caption", "bbox": [ 218, 567, 504, 590 ], "group_id": 0, "lines": [ { "bbox": [ 218, 565, 505, 580 ], "spans": [ { "bbox": [ 218, 565, 505, 580 ], "score": 1.0, "content": "Figure 2. Schematic illustration of Positional Encoding Generator", "type": "text" } ], "index": 40 }, { "bbox": [ 219, 578, 480, 590 ], "spans": [ { "bbox": [ 219, 578, 271, 590 ], "score": 1.0, "content": "(PEG). Note", "type": "text" }, { "bbox": [ 271, 578, 278, 588 ], "score": 0.74, "content": "d", "type": "inline_equation" }, { "bbox": [ 279, 578, 370, 590 ], "score": 1.0, "content": "is the embedding size,", "type": "text" }, { "bbox": [ 370, 578, 380, 588 ], "score": 0.8, "content": "N", "type": "inline_equation" }, { "bbox": [ 381, 578, 480, 590 ], "score": 1.0, "content": "is the number of tokens.", "type": "text" } ], "index": 41 } ], "index": 40.5 } ], "index": 39.25 }, { "type": "text", "bbox": [ 106, 609, 505, 676 ], "lines": [ { "bbox": [ 104, 606, 507, 623 ], "spans": [ { "bbox": [ 104, 606, 257, 623 ], "score": 1.0, "content": "reshape the flattened input sequence", "type": "text" }, { "bbox": [ 257, 608, 322, 619 ], "score": 0.92, "content": "X \\in \\mathbb { R } ^ { B \\times N \\times C }", "type": "inline_equation" }, { "bbox": [ 323, 606, 394, 623 ], "score": 1.0, "content": "of DeiT back to", "type": "text" }, { "bbox": [ 394, 608, 477, 620 ], "score": 0.92, "content": "X ^ { \\prime } \\in \\mathbb { R } ^ { B \\times H \\times W \\times C }", "type": "inline_equation" }, { "bbox": [ 477, 606, 507, 623 ], "score": 1.0, "content": "in the", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 619, 505, 632 ], "spans": [ { "bbox": [ 105, 619, 305, 632 ], "score": 1.0, "content": "2-D image space. Then, a function (denoted by", "type": "text" }, { "bbox": [ 306, 621, 315, 631 ], "score": 0.82, "content": "\\mathcal { F }", "type": "inline_equation" }, { "bbox": [ 316, 619, 505, 632 ], "score": 1.0, "content": "in Figure 2) is repeatedly applied to the local", "type": "text" } ], "index": 44 }, { "bbox": [ 102, 626, 507, 647 ], "spans": [ { "bbox": [ 102, 626, 141, 647 ], "score": 1.0, "content": "patch in", "type": "text" }, { "bbox": [ 142, 631, 155, 641 ], "score": 0.88, "content": "X ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 155, 626, 349, 647 ], "score": 1.0, "content": "to produce the conditional positional encodings", "type": "text" }, { "bbox": [ 350, 630, 406, 641 ], "score": 0.91, "content": "E ^ { \\tilde { B } \\times H \\times \\tilde { W } \\times C }", "type": "inline_equation" }, { "bbox": [ 406, 626, 507, 647 ], "score": 1.0, "content": ". PEG can be efficiently", "type": "text" } ], "index": 45 }, { "bbox": [ 104, 639, 506, 667 ], "spans": [ { "bbox": [ 104, 639, 306, 667 ], "score": 1.0, "content": "implemented with a 2-D convolution with kernel zero paddings here are important to make the mo", "type": "text" }, { "bbox": [ 306, 643, 313, 653 ], "score": 0.65, "content": "k", "type": "inline_equation" }, { "bbox": [ 314, 639, 317, 667 ], "score": 1.0, "content": "l", "type": "text" }, { "bbox": [ 317, 642, 346, 654 ], "score": 0.82, "content": "( k \\geq 3 )", "type": "inline_equation" }, { "bbox": [ 347, 639, 367, 667 ], "score": 1.0, "content": "and re of", "type": "text" }, { "bbox": [ 368, 641, 385, 655 ], "score": 0.91, "content": "\\frac { k - 1 } { 2 }", "type": "inline_equation" }, { "bbox": [ 386, 639, 478, 667 ], "score": 1.0, "content": "zero paddings. Note tabsolute positions, and", "type": "text" }, { "bbox": [ 488, 639, 506, 667 ], "score": 1.0, "content": "thecan", "type": "text" } ], "index": 46 }, { "bbox": [ 478, 654, 488, 663 ], "spans": [ { "bbox": [ 478, 654, 488, 663 ], "score": 0.83, "content": "\\mathcal { F }", "type": "inline_equation" } ], "index": 47 }, { "bbox": [ 105, 663, 408, 677 ], "spans": [ { "bbox": [ 105, 663, 408, 677 ], "score": 1.0, "content": "be of various forms such as various types of convolutions and many others.", "type": "text" } ], "index": 48 } ], "index": 45.5 }, { "type": "title", "bbox": [ 108, 689, 405, 700 ], "lines": [ { "bbox": [ 105, 688, 406, 702 ], "spans": [ { "bbox": [ 105, 688, 406, 702 ], "score": 1.0, "content": "3.3 CONDITIONAL POSITIONAL ENCODING VISION TRANSFORMERS", "type": "text" } ], "index": 49 } ], "index": 49 }, { "type": "text", "bbox": [ 107, 709, 501, 732 ], "lines": [ { "bbox": [ 106, 709, 503, 722 ], "spans": [ { "bbox": [ 106, 709, 503, 722 ], "score": 1.0, "content": "Built on the conditional positional encodings, we propose our Conditional Positional Encoding Vi-", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 720, 503, 733 ], "spans": [ { "bbox": [ 105, 720, 503, 733 ], "score": 1.0, "content": "sion Transformers (CPVT). Except that our positional encodings are conditional, we exactly follow", "type": "text" } ], "index": 51 } ], "index": 50.5 } ], "page_idx": 3, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 759 ], "lines": [] } ], "para_blocks": [ { "type": "table", "bbox": [ 116, 128, 495, 186 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 80, 504, 114 ], "group_id": 0, "lines": [ { "bbox": [ 106, 79, 505, 94 ], "spans": [ { "bbox": [ 106, 79, 505, 94 ], "score": 1.0, "content": "Table 1. Comparison of various positional encoding (PE) strategies tested on ImageNet valida-", "type": "text" } ], "index": 0 }, { "bbox": [ 107, 91, 504, 104 ], "spans": [ { "bbox": [ 107, 91, 504, 104 ], "score": 1.0, "content": "tion set in terms of the top-1 accuracy. Removing the positional encodings greatly damages the", "type": "text" } ], "index": 1 }, { "bbox": [ 107, 103, 486, 115 ], "spans": [ { "bbox": [ 107, 103, 486, 115 ], "score": 1.0, "content": "performance. The relative positional encodings have inferior performance to the absolute ones", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "table_body", "bbox": [ 116, 128, 495, 186 ], "group_id": 0, "lines": [ { "bbox": [ 116, 128, 495, 186 ], "spans": [ { "bbox": [ 116, 128, 495, 186 ], "score": 0.977, "html": "
ModelEncodingTop-1@224(%)Top-1@384(%)
DeiT-tiny (Touvron et al., 2020)X68.268.6
DeiT-tiny (Touvron et al., 2020)learnable72.271.2
DeiT-tiny (Touvron et al., 2020)sin-cos72.370.8
DeiT-tiny2D RPE (Shaw et al., 2018)70.569.8
", "type": "table", "image_path": "777658109f20f506564abf78728d41e9469175d72aa4de071291272306f43b7e.jpg" } ] } ], "index": 4, "virtual_lines": [ { "bbox": [ 116, 128, 495, 147.33333333333334 ], "spans": [], "index": 3 }, { "bbox": [ 116, 147.33333333333334, 495, 166.66666666666669 ], "spans": [], "index": 4 }, { "bbox": [ 116, 166.66666666666669, 495, 186.00000000000003 ], "spans": [], "index": 5 } ] } ], "index": 2.5 }, { "type": "text", "bbox": [ 108, 196, 505, 230 ], "lines": [], "index": 7, "bbox_fs": [ 105, 196, 505, 231 ], "lines_deleted": true }, { "type": "title", "bbox": [ 108, 243, 303, 254 ], "lines": [ { "bbox": [ 105, 241, 304, 256 ], "spans": [ { "bbox": [ 105, 241, 304, 256 ], "score": 1.0, "content": "3.2 CONDITIONAL POSITIONAL ENCODINGS", "type": "text" } ], "index": 9 } ], "index": 9 }, { "type": "text", "bbox": [ 111, 263, 488, 275 ], "lines": [ { "bbox": [ 108, 262, 490, 278 ], "spans": [ { "bbox": [ 108, 262, 490, 278 ], "score": 1.0, "content": "We argue that a successful positional encoding for vision tasks should meet these requirements,", "type": "text" } ], "index": 10 } ], "index": 10, "bbox_fs": [ 108, 262, 490, 278 ] }, { "type": "list", "bbox": [ 126, 284, 505, 347 ], "lines": [ { "bbox": [ 125, 284, 504, 298 ], "spans": [ { "bbox": [ 125, 284, 504, 298 ], "score": 1.0, "content": "(1) Making the input sequence permutation-variant and providing stronger explicit bias to-", "type": "text" } ], "index": 11, "is_list_start_line": true }, { "bbox": [ 142, 296, 272, 307 ], "spans": [ { "bbox": [ 142, 296, 272, 307 ], "score": 1.0, "content": "wards translation-equivariance.", "type": "text" } ], "index": 12, "is_list_end_line": true }, { "bbox": [ 125, 309, 488, 324 ], "spans": [ { "bbox": [ 125, 309, 488, 324 ], "score": 1.0, "content": "(2) Being inductive and able to handle the sequences longer than the ones during training.", "type": "text" } ], "index": 13, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 125, 324, 505, 338 ], "spans": [ { "bbox": [ 125, 324, 505, 338 ], "score": 1.0, "content": "(3) Having the ability to provide the absolute position to a certain degree. This is important to", "type": "text" } ], "index": 14, "is_list_start_line": true }, { "bbox": [ 142, 336, 342, 348 ], "spans": [ { "bbox": [ 142, 336, 342, 348 ], "score": 1.0, "content": "the performance as shown in (Islam et al., 2020).", "type": "text" } ], "index": 15, "is_list_end_line": true } ], "index": 13, "bbox_fs": [ 125, 284, 505, 348 ] }, { "type": "text", "bbox": [ 106, 356, 505, 455 ], "lines": [ { "bbox": [ 105, 356, 506, 369 ], "spans": [ { "bbox": [ 105, 356, 506, 369 ], "score": 1.0, "content": "In this work, we find that characterizing the local relationship by positional encodings is sufficient", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 367, 506, 381 ], "spans": [ { "bbox": [ 105, 367, 506, 381 ], "score": 1.0, "content": "to meet all of the above. First, it is permutation-variant because the permutation of input sequences", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 379, 504, 390 ], "spans": [ { "bbox": [ 106, 379, 504, 390 ], "score": 1.0, "content": "also affects the order in some local neighborhoods. However, translation of an object in an input im-", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 388, 505, 402 ], "spans": [ { "bbox": [ 105, 388, 505, 402 ], "score": 1.0, "content": "age does not change the order in its local neighborhood, i.e., translation-equivariant (see Section A).", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 400, 506, 413 ], "spans": [ { "bbox": [ 106, 400, 506, 413 ], "score": 1.0, "content": "Second, the model can easily generalize to longer sequences since only the local neighborhoods of a", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 410, 505, 424 ], "spans": [ { "bbox": [ 105, 410, 505, 424 ], "score": 1.0, "content": "token are involved. Besides, if the absolute position of any input token is known, the absolute posi-", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 422, 505, 434 ], "spans": [ { "bbox": [ 106, 422, 505, 434 ], "score": 1.0, "content": "tion of all the other tokens can be inferred by the mutual relation between input tokens. We will show", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 433, 505, 446 ], "spans": [ { "bbox": [ 106, 433, 505, 446 ], "score": 1.0, "content": "that the tokens on the borders can be aware of their absolute positions due to the commonly-used", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 442, 168, 458 ], "spans": [ { "bbox": [ 105, 442, 168, 458 ], "score": 1.0, "content": "zero paddings.", "type": "text" } ], "index": 24 } ], "index": 20, "bbox_fs": [ 105, 356, 506, 458 ] }, { "type": "text", "bbox": [ 106, 461, 210, 548 ], "lines": [ { "bbox": [ 106, 460, 212, 474 ], "spans": [ { "bbox": [ 106, 460, 212, 474 ], "score": 1.0, "content": "Therefore, we propose", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 472, 212, 484 ], "spans": [ { "bbox": [ 106, 472, 212, 484 ], "score": 1.0, "content": "positional encoding gen-", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 483, 211, 495 ], "spans": [ { "bbox": [ 106, 483, 211, 495 ], "score": 1.0, "content": "erators (PEG) to dynam-", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 493, 212, 506 ], "spans": [ { "bbox": [ 105, 493, 212, 506 ], "score": 1.0, "content": "ically produce the posi-", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 504, 212, 516 ], "spans": [ { "bbox": [ 105, 504, 212, 516 ], "score": 1.0, "content": "tional encodings condi-", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 515, 212, 528 ], "spans": [ { "bbox": [ 105, 515, 212, 528 ], "score": 1.0, "content": "tioned on the local neigh-", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 526, 212, 539 ], "spans": [ { "bbox": [ 105, 526, 212, 539 ], "score": 1.0, "content": "borhood of an input to-", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 536, 127, 550 ], "spans": [ { "bbox": [ 105, 536, 127, 550 ], "score": 1.0, "content": "ken.", "type": "text" } ], "index": 32 } ], "index": 28.5, "bbox_fs": [ 105, 460, 212, 550 ] }, { "type": "text", "bbox": [ 106, 554, 212, 610 ], "lines": [ { "bbox": [ 105, 552, 213, 567 ], "spans": [ { "bbox": [ 105, 552, 152, 566 ], "score": 1.0, "content": "Positional", "type": "text" }, { "bbox": [ 167, 552, 213, 567 ], "score": 1.0, "content": "Encoding", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 564, 213, 577 ], "spans": [ { "bbox": [ 105, 564, 156, 577 ], "score": 1.0, "content": "Generator.", "type": "text" }, { "bbox": [ 172, 564, 213, 577 ], "score": 1.0, "content": "PEG is", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 575, 212, 588 ], "spans": [ { "bbox": [ 105, 575, 212, 588 ], "score": 1.0, "content": "illustrated in Figure 2.", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 586, 212, 599 ], "spans": [ { "bbox": [ 105, 586, 212, 599 ], "score": 1.0, "content": "To condition on the", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 598, 212, 611 ], "spans": [ { "bbox": [ 105, 598, 212, 611 ], "score": 1.0, "content": "local neighbors, we first", "type": "text" } ], "index": 42 } ], "index": 35, "bbox_fs": [ 105, 552, 213, 611 ] }, { "type": "image", "bbox": [ 223, 461, 502, 557 ], "blocks": [ { "type": "image_body", "bbox": [ 223, 461, 502, 557 ], "group_id": 0, "lines": [ { "bbox": [ 223, 461, 502, 557 ], "spans": [ { "bbox": [ 223, 461, 502, 557 ], "score": 0.969, "type": "image", "image_path": "e12ed07e330e1cbad3fc807996e87af006eec2275384ddc7e13bc09ba9a2aac7.jpg" } ] } ], "index": 38, "virtual_lines": [ { "bbox": [ 223, 461, 502, 493.0 ], "spans": [], "index": 37 }, { "bbox": [ 223, 493.0, 502, 525.0 ], "spans": [], "index": 38 }, { "bbox": [ 223, 525.0, 502, 557.0 ], "spans": [], "index": 39 } ] }, { "type": "image_caption", "bbox": [ 218, 567, 504, 590 ], "group_id": 0, "lines": [ { "bbox": [ 218, 565, 505, 580 ], "spans": [ { "bbox": [ 218, 565, 505, 580 ], "score": 1.0, "content": "Figure 2. Schematic illustration of Positional Encoding Generator", "type": "text" } ], "index": 40 }, { "bbox": [ 219, 578, 480, 590 ], "spans": [ { "bbox": [ 219, 578, 271, 590 ], "score": 1.0, "content": "(PEG). Note", "type": "text" }, { "bbox": [ 271, 578, 278, 588 ], "score": 0.74, "content": "d", "type": "inline_equation" }, { "bbox": [ 279, 578, 370, 590 ], "score": 1.0, "content": "is the embedding size,", "type": "text" }, { "bbox": [ 370, 578, 380, 588 ], "score": 0.8, "content": "N", "type": "inline_equation" }, { "bbox": [ 381, 578, 480, 590 ], "score": 1.0, "content": "is the number of tokens.", "type": "text" } ], "index": 41 } ], "index": 40.5 } ], "index": 39.25 }, { "type": "text", "bbox": [ 106, 609, 505, 676 ], "lines": [ { "bbox": [ 104, 606, 507, 623 ], "spans": [ { "bbox": [ 104, 606, 257, 623 ], "score": 1.0, "content": "reshape the flattened input sequence", "type": "text" }, { "bbox": [ 257, 608, 322, 619 ], "score": 0.92, "content": "X \\in \\mathbb { R } ^ { B \\times N \\times C }", "type": "inline_equation" }, { "bbox": [ 323, 606, 394, 623 ], "score": 1.0, "content": "of DeiT back to", "type": "text" }, { "bbox": [ 394, 608, 477, 620 ], "score": 0.92, "content": "X ^ { \\prime } \\in \\mathbb { R } ^ { B \\times H \\times W \\times C }", "type": "inline_equation" }, { "bbox": [ 477, 606, 507, 623 ], "score": 1.0, "content": "in the", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 619, 505, 632 ], "spans": [ { "bbox": [ 105, 619, 305, 632 ], "score": 1.0, "content": "2-D image space. Then, a function (denoted by", "type": "text" }, { "bbox": [ 306, 621, 315, 631 ], "score": 0.82, "content": "\\mathcal { F }", "type": "inline_equation" }, { "bbox": [ 316, 619, 505, 632 ], "score": 1.0, "content": "in Figure 2) is repeatedly applied to the local", "type": "text" } ], "index": 44 }, { "bbox": [ 102, 626, 507, 647 ], "spans": [ { "bbox": [ 102, 626, 141, 647 ], "score": 1.0, "content": "patch in", "type": "text" }, { "bbox": [ 142, 631, 155, 641 ], "score": 0.88, "content": "X ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 155, 626, 349, 647 ], "score": 1.0, "content": "to produce the conditional positional encodings", "type": "text" }, { "bbox": [ 350, 630, 406, 641 ], "score": 0.91, "content": "E ^ { \\tilde { B } \\times H \\times \\tilde { W } \\times C }", "type": "inline_equation" }, { "bbox": [ 406, 626, 507, 647 ], "score": 1.0, "content": ". PEG can be efficiently", "type": "text" } ], "index": 45 }, { "bbox": [ 104, 639, 506, 667 ], "spans": [ { "bbox": [ 104, 639, 306, 667 ], "score": 1.0, "content": "implemented with a 2-D convolution with kernel zero paddings here are important to make the mo", "type": "text" }, { "bbox": [ 306, 643, 313, 653 ], "score": 0.65, "content": "k", "type": "inline_equation" }, { "bbox": [ 314, 639, 317, 667 ], "score": 1.0, "content": "l", "type": "text" }, { "bbox": [ 317, 642, 346, 654 ], "score": 0.82, "content": "( k \\geq 3 )", "type": "inline_equation" }, { "bbox": [ 347, 639, 367, 667 ], "score": 1.0, "content": "and re of", "type": "text" }, { "bbox": [ 368, 641, 385, 655 ], "score": 0.91, "content": "\\frac { k - 1 } { 2 }", "type": "inline_equation" }, { "bbox": [ 386, 639, 478, 667 ], "score": 1.0, "content": "zero paddings. Note tabsolute positions, and", "type": "text" }, { "bbox": [ 488, 639, 506, 667 ], "score": 1.0, "content": "thecan", "type": "text" } ], "index": 46 }, { "bbox": [ 478, 654, 488, 663 ], "spans": [ { "bbox": [ 478, 654, 488, 663 ], "score": 0.83, "content": "\\mathcal { F }", "type": "inline_equation" } ], "index": 47 }, { "bbox": [ 105, 663, 408, 677 ], "spans": [ { "bbox": [ 105, 663, 408, 677 ], "score": 1.0, "content": "be of various forms such as various types of convolutions and many others.", "type": "text" } ], "index": 48 } ], "index": 45.5, "bbox_fs": [ 102, 606, 507, 677 ] }, { "type": "title", "bbox": [ 108, 689, 405, 700 ], "lines": [ { "bbox": [ 105, 688, 406, 702 ], "spans": [ { "bbox": [ 105, 688, 406, 702 ], "score": 1.0, "content": "3.3 CONDITIONAL POSITIONAL ENCODING VISION TRANSFORMERS", "type": "text" } ], "index": 49 } ], "index": 49 }, { "type": "text", "bbox": [ 107, 709, 501, 732 ], "lines": [ { "bbox": [ 106, 709, 503, 722 ], "spans": [ { "bbox": [ 106, 709, 503, 722 ], "score": 1.0, "content": "Built on the conditional positional encodings, we propose our Conditional Positional Encoding Vi-", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 720, 503, 733 ], "spans": [ { "bbox": [ 105, 720, 503, 733 ], "score": 1.0, "content": "sion Transformers (CPVT). Except that our positional encodings are conditional, we exactly follow", "type": "text" } ], "index": 51 } ], "index": 50.5, "bbox_fs": [ 105, 709, 503, 733 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 504, 126 ], "lines": [ { "bbox": [ 105, 82, 506, 95 ], "spans": [ { "bbox": [ 105, 82, 506, 95 ], "score": 1.0, "content": "ViT and DeiT to design our vision transformers and we also have three sizes CPVT-Ti, CPVT-S and", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 92, 506, 107 ], "spans": [ { "bbox": [ 105, 92, 506, 107 ], "score": 1.0, "content": "CPVT-B. Similar to the original positional encodings in DeiT, the conditional positional encodings", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 105, 506, 118 ], "spans": [ { "bbox": [ 105, 105, 506, 118 ], "score": 1.0, "content": "are also added to the input sequence, as shown in Figure 1 (b). In CPVT, the position where PEG is", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 115, 454, 128 ], "spans": [ { "bbox": [ 105, 115, 454, 128 ], "score": 1.0, "content": "applied is also important to the performance, which will be studied in the experiments.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "text", "bbox": [ 107, 132, 505, 187 ], "lines": [ { "bbox": [ 105, 131, 506, 145 ], "spans": [ { "bbox": [ 105, 131, 506, 145 ], "score": 1.0, "content": "In addition, both DeiT and ViT utilize an extra learnable class token to perform classification (i.e.,", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 144, 505, 156 ], "spans": [ { "bbox": [ 106, 144, 505, 156 ], "score": 1.0, "content": "cls token shown in Figure 1 (a) and (b)). By design, the class token is not translation-invariant,", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 153, 505, 168 ], "spans": [ { "bbox": [ 105, 153, 505, 168 ], "score": 1.0, "content": "although it can learn to be so. A simple alternative is to directly replace it with a global average", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 165, 505, 178 ], "spans": [ { "bbox": [ 105, 165, 505, 178 ], "score": 1.0, "content": "pooling (GAP), which is inherently translation-invariant, resulting in our CVPT-GAP. Together with", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 175, 403, 190 ], "spans": [ { "bbox": [ 105, 175, 403, 190 ], "score": 1.0, "content": "CPE, CVPT-GAP achieves much better image classification performance.", "type": "text" } ], "index": 8 } ], "index": 6 }, { "type": "title", "bbox": [ 108, 205, 200, 218 ], "lines": [ { "bbox": [ 105, 205, 201, 219 ], "spans": [ { "bbox": [ 105, 205, 201, 219 ], "score": 1.0, "content": "4 EXPERIMENTS", "type": "text" } ], "index": 9 } ], "index": 9 }, { "type": "title", "bbox": [ 107, 231, 160, 243 ], "lines": [ { "bbox": [ 105, 230, 160, 244 ], "spans": [ { "bbox": [ 105, 230, 160, 244 ], "score": 1.0, "content": "4.1 SETUP", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 107, 252, 505, 297 ], "lines": [ { "bbox": [ 104, 251, 506, 267 ], "spans": [ { "bbox": [ 104, 251, 506, 267 ], "score": 1.0, "content": "Datasets. Following DeiT (Touvron et al., 2020), we use ILSVRC-2012 ImageNet dataset (Deng", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 264, 505, 276 ], "spans": [ { "bbox": [ 105, 264, 505, 276 ], "score": 1.0, "content": "et al., 2009) with 1K classes and 1.3M images to train all our models. We report the results on the", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 275, 505, 286 ], "spans": [ { "bbox": [ 106, 275, 505, 286 ], "score": 1.0, "content": "validation set with 50K images. Unlike ViT (Dosovitskiy et al., 2021), we do not use the much", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 285, 330, 298 ], "spans": [ { "bbox": [ 105, 285, 330, 298 ], "score": 1.0, "content": "larger undisclosed JFT-300M dataset (Sun et al., 2017).", "type": "text" } ], "index": 14 } ], "index": 12.5 }, { "type": "text", "bbox": [ 107, 302, 504, 346 ], "lines": [ { "bbox": [ 105, 302, 504, 315 ], "spans": [ { "bbox": [ 105, 302, 504, 315 ], "score": 1.0, "content": "Model variants. We have three models with various sizes to adapt to various computing scenarios.", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 313, 506, 326 ], "spans": [ { "bbox": [ 105, 313, 506, 326 ], "score": 1.0, "content": "The detailed settings are shown in Table 9 (see B.1). All experiments in this paper are performed on", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 324, 506, 337 ], "spans": [ { "bbox": [ 105, 324, 506, 337 ], "score": 1.0, "content": "Tesla V100 machines. Training the tiny model for 300 epochs takes about 1.3 days on a single node", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 335, 463, 348 ], "spans": [ { "bbox": [ 105, 335, 463, 348 ], "score": 1.0, "content": "with 8 V100 GPU cards. CPVT-S and CPVT-B take about 1.6 and 2.5 days, respectively.", "type": "text" } ], "index": 18 } ], "index": 16.5 }, { "type": "text", "bbox": [ 107, 352, 505, 407 ], "lines": [ { "bbox": [ 106, 352, 505, 364 ], "spans": [ { "bbox": [ 106, 352, 505, 364 ], "score": 1.0, "content": "Training details All the models (except for CPVT-B) are trained for 300 epochs with a global batch", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 362, 505, 376 ], "spans": [ { "bbox": [ 105, 362, 505, 376 ], "score": 1.0, "content": "size of 2048 on Tesla V100 machines using AdamW optimizer (Loshchilov & Hutter, 2019). We", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 373, 505, 388 ], "spans": [ { "bbox": [ 105, 373, 505, 388 ], "score": 1.0, "content": "do not tune the hyper-parameters and strictly comply with the settings in DeiT (Touvron et al.,", "type": "text" } ], "index": 21 }, { "bbox": [ 104, 383, 506, 399 ], "spans": [ { "bbox": [ 104, 383, 315, 399 ], "score": 1.0, "content": "2020). The learning rate is scaled with this formula", "type": "text" }, { "bbox": [ 316, 385, 448, 397 ], "score": 0.85, "content": "l r _ { \\mathrm { s c a l e } } = 0 . 0 0 0 5 { \\cdot } \\mathrm { B a t c h S i z e } _ { \\mathrm { g l o b a l } } / { \\scriptstyle 5 1 2 }", "type": "inline_equation" }, { "bbox": [ 448, 383, 506, 399 ], "score": 1.0, "content": ". The detailed", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 397, 234, 408 ], "spans": [ { "bbox": [ 105, 397, 234, 408 ], "score": 1.0, "content": "hyperparameters are in the B.2.", "type": "text" } ], "index": 23 } ], "index": 21 }, { "type": "title", "bbox": [ 107, 423, 323, 434 ], "lines": [ { "bbox": [ 105, 421, 325, 435 ], "spans": [ { "bbox": [ 105, 421, 325, 435 ], "score": 1.0, "content": "4.2 GENERALIZATION TO HIGHER RESOLUTIONS", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 443, 505, 543 ], "lines": [ { "bbox": [ 106, 443, 505, 457 ], "spans": [ { "bbox": [ 106, 443, 505, 457 ], "score": 1.0, "content": "As mentioned before, our proposed PEG can directly generalize to larger image sizes without any", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 455, 506, 468 ], "spans": [ { "bbox": [ 105, 455, 400, 468 ], "score": 1.0, "content": "fine-tuning. We confirm this here by evaluating the models trained with", "type": "text" }, { "bbox": [ 400, 455, 444, 466 ], "score": 0.9, "content": "2 2 4 \\times 2 2 4", "type": "inline_equation" }, { "bbox": [ 445, 455, 506, 468 ], "score": 1.0, "content": "images on the", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 465, 505, 478 ], "spans": [ { "bbox": [ 106, 466, 150, 477 ], "score": 0.86, "content": "3 8 4 \\times 3 8 4", "type": "inline_equation" }, { "bbox": [ 150, 465, 153, 478 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 154, 466, 197, 477 ], "score": 0.77, "content": "4 4 8 \\times 4 4 8", "type": "inline_equation" }, { "bbox": [ 197, 465, 200, 478 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 200, 466, 244, 477 ], "score": 0.73, "content": "5 1 2 \\times 5 1 2", "type": "inline_equation" }, { "bbox": [ 244, 465, 505, 478 ], "score": 1.0, "content": "images, respectively. The results are shown in Table 2. With the", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 477, 504, 488 ], "spans": [ { "bbox": [ 106, 477, 151, 487 ], "score": 0.87, "content": "3 8 4 \\times 3 8 4", "type": "inline_equation" }, { "bbox": [ 152, 477, 476, 488 ], "score": 1.0, "content": "input images, the DeiT-tiny with learnable positional encodings degrades from", "type": "text" }, { "bbox": [ 477, 477, 504, 487 ], "score": 0.86, "content": "7 2 . 2 \\%", "type": "inline_equation" } ], "index": 28 }, { "bbox": [ 105, 488, 506, 501 ], "spans": [ { "bbox": [ 105, 488, 117, 501 ], "score": 1.0, "content": "to", "type": "text" }, { "bbox": [ 117, 488, 144, 498 ], "score": 0.86, "content": "7 1 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 145, 488, 421, 501 ], "score": 1.0, "content": ". When equipped with sine encoding, the tiny model degrades from", "type": "text" }, { "bbox": [ 421, 488, 449, 498 ], "score": 0.87, "content": "7 2 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 449, 488, 461, 501 ], "score": 1.0, "content": "to", "type": "text" }, { "bbox": [ 461, 488, 488, 498 ], "score": 0.9, "content": "7 0 . 8 \\%", "type": "inline_equation" }, { "bbox": [ 488, 488, 506, 501 ], "score": 1.0, "content": ". In", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 498, 505, 512 ], "spans": [ { "bbox": [ 105, 498, 505, 512 ], "score": 1.0, "content": "constrat, our CPVT model with the proposed PEG can directly process the larger input images, and", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 509, 505, 523 ], "spans": [ { "bbox": [ 105, 509, 271, 523 ], "score": 1.0, "content": "CPVT-Ti’s performance is boosted from", "type": "text" }, { "bbox": [ 271, 510, 298, 520 ], "score": 0.87, "content": "7 3 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 299, 509, 310, 523 ], "score": 1.0, "content": "to", "type": "text" }, { "bbox": [ 311, 510, 338, 520 ], "score": 0.87, "content": "7 4 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 338, 509, 406, 523 ], "score": 1.0, "content": "when applied to", "type": "text" }, { "bbox": [ 407, 510, 451, 520 ], "score": 0.9, "content": "3 8 4 \\times 3 8 4", "type": "inline_equation" }, { "bbox": [ 451, 509, 505, 523 ], "score": 1.0, "content": "images. Our", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 520, 505, 533 ], "spans": [ { "bbox": [ 105, 520, 253, 533 ], "score": 1.0, "content": "CPVT-Ti outperforms DeiT-tiny by", "type": "text" }, { "bbox": [ 254, 521, 276, 531 ], "score": 0.85, "content": "3 . 0 \\%", "type": "inline_equation" }, { "bbox": [ 276, 520, 505, 533 ], "score": 1.0, "content": ". This gap continues to increase as the input resolution", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 531, 145, 546 ], "spans": [ { "bbox": [ 105, 531, 145, 546 ], "score": 1.0, "content": "enlarges.", "type": "text" } ], "index": 33 } ], "index": 29 }, { "type": "table", "bbox": [ 126, 591, 486, 706 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 555, 503, 578 ], "group_id": 0, "lines": [ { "bbox": [ 107, 554, 504, 567 ], "spans": [ { "bbox": [ 107, 554, 504, 567 ], "score": 1.0, "content": "Table 2. Direct evaluation on other resolutions without fine-tuning. The models are trained on", "type": "text" } ], "index": 34 }, { "bbox": [ 108, 566, 453, 577 ], "spans": [ { "bbox": [ 108, 566, 147, 577 ], "score": 0.88, "content": "2 2 4 \\times 2 2 4", "type": "inline_equation" }, { "bbox": [ 147, 566, 288, 577 ], "score": 1.0, "content": ". A simple PEG of a single layer of", "type": "text" }, { "bbox": [ 288, 566, 307, 577 ], "score": 0.87, "content": "3 \\times 3", "type": "inline_equation" }, { "bbox": [ 308, 566, 453, 577 ], "score": 1.0, "content": "depth-wise convolution is used here", "type": "text" } ], "index": 35 } ], "index": 34.5 }, { "type": "table_body", "bbox": [ 126, 591, 486, 706 ], "group_id": 0, "lines": [ { "bbox": [ 126, 591, 486, 706 ], "spans": [ { "bbox": [ 126, 591, 486, 706 ], "score": 0.986, "html": "
ModelParams160(%)224(%)384(%)448(%)512(%)
DeiT-tinyDeiT-tiny (sin)DeiT-tiny (no pos)CPVT-TiCPVT-Ti ‡6M6M6M6M6M65.665.262.166.8(+1.2)67.7 (+2.1)72.272.368.272.4(+0.2)73.4(+1.2)71.270.868.673.2(+2.0)74.2(+3.0)68.868.268.471.8(+3.0)72.6(+3.8)65.965.165.070.3(+4.4)70.8(+4.9)
DeiT-smallCPVT-S22M22M75.676.1(+0.5)79.979.978.180.4(+1.5)75.978.6(+2.7)72.676.8(+4.2)
DeiT-baseCPVT-B86M86M79.180.5(+1.4)81.881.9(+0.1)79.782.3(+2.6)79.882.4(+2.6)78.281.0(+2.8)
", "type": "table", "image_path": "ab1cbd1190803aa7fd6c2b7ced714aed46988bc72d1faf1b2b651896176f78f0.jpg" } ] } ], "index": 37, "virtual_lines": [ { "bbox": [ 126, 591, 486, 629.3333333333334 ], "spans": [], "index": 36 }, { "bbox": [ 126, 629.3333333333334, 486, 667.6666666666667 ], "spans": [], "index": 37 }, { "bbox": [ 126, 667.6666666666667, 486, 706.0000000000001 ], "spans": [], "index": 38 } ] }, { "type": "table_footnote", "bbox": [ 174, 708, 437, 719 ], "group_id": 0, "lines": [ { "bbox": [ 173, 706, 437, 720 ], "spans": [ { "bbox": [ 173, 706, 437, 720 ], "score": 1.0, "content": "‡: Insert one PEG each after the first encoder till the fifth encoder", "type": "text" } ], "index": 39 } ], "index": 39 } ], "index": 37 } ], "page_idx": 4, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 308, 760 ], "lines": [ { "bbox": [ 302, 750, 309, 763 ], "spans": [ { "bbox": [ 302, 750, 309, 763 ], "score": 1.0, "content": "5", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 82, 504, 126 ], "lines": [ { "bbox": [ 105, 82, 506, 95 ], "spans": [ { "bbox": [ 105, 82, 506, 95 ], "score": 1.0, "content": "ViT and DeiT to design our vision transformers and we also have three sizes CPVT-Ti, CPVT-S and", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 92, 506, 107 ], "spans": [ { "bbox": [ 105, 92, 506, 107 ], "score": 1.0, "content": "CPVT-B. Similar to the original positional encodings in DeiT, the conditional positional encodings", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 105, 506, 118 ], "spans": [ { "bbox": [ 105, 105, 506, 118 ], "score": 1.0, "content": "are also added to the input sequence, as shown in Figure 1 (b). In CPVT, the position where PEG is", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 115, 454, 128 ], "spans": [ { "bbox": [ 105, 115, 454, 128 ], "score": 1.0, "content": "applied is also important to the performance, which will be studied in the experiments.", "type": "text" } ], "index": 3 } ], "index": 1.5, "bbox_fs": [ 105, 82, 506, 128 ] }, { "type": "text", "bbox": [ 107, 132, 505, 187 ], "lines": [ { "bbox": [ 105, 131, 506, 145 ], "spans": [ { "bbox": [ 105, 131, 506, 145 ], "score": 1.0, "content": "In addition, both DeiT and ViT utilize an extra learnable class token to perform classification (i.e.,", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 144, 505, 156 ], "spans": [ { "bbox": [ 106, 144, 505, 156 ], "score": 1.0, "content": "cls token shown in Figure 1 (a) and (b)). By design, the class token is not translation-invariant,", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 153, 505, 168 ], "spans": [ { "bbox": [ 105, 153, 505, 168 ], "score": 1.0, "content": "although it can learn to be so. A simple alternative is to directly replace it with a global average", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 165, 505, 178 ], "spans": [ { "bbox": [ 105, 165, 505, 178 ], "score": 1.0, "content": "pooling (GAP), which is inherently translation-invariant, resulting in our CVPT-GAP. Together with", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 175, 403, 190 ], "spans": [ { "bbox": [ 105, 175, 403, 190 ], "score": 1.0, "content": "CPE, CVPT-GAP achieves much better image classification performance.", "type": "text" } ], "index": 8 } ], "index": 6, "bbox_fs": [ 105, 131, 506, 190 ] }, { "type": "title", "bbox": [ 108, 205, 200, 218 ], "lines": [ { "bbox": [ 105, 205, 201, 219 ], "spans": [ { "bbox": [ 105, 205, 201, 219 ], "score": 1.0, "content": "4 EXPERIMENTS", "type": "text" } ], "index": 9 } ], "index": 9 }, { "type": "title", "bbox": [ 107, 231, 160, 243 ], "lines": [ { "bbox": [ 105, 230, 160, 244 ], "spans": [ { "bbox": [ 105, 230, 160, 244 ], "score": 1.0, "content": "4.1 SETUP", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 107, 252, 505, 297 ], "lines": [ { "bbox": [ 104, 251, 506, 267 ], "spans": [ { "bbox": [ 104, 251, 506, 267 ], "score": 1.0, "content": "Datasets. Following DeiT (Touvron et al., 2020), we use ILSVRC-2012 ImageNet dataset (Deng", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 264, 505, 276 ], "spans": [ { "bbox": [ 105, 264, 505, 276 ], "score": 1.0, "content": "et al., 2009) with 1K classes and 1.3M images to train all our models. We report the results on the", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 275, 505, 286 ], "spans": [ { "bbox": [ 106, 275, 505, 286 ], "score": 1.0, "content": "validation set with 50K images. Unlike ViT (Dosovitskiy et al., 2021), we do not use the much", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 285, 330, 298 ], "spans": [ { "bbox": [ 105, 285, 330, 298 ], "score": 1.0, "content": "larger undisclosed JFT-300M dataset (Sun et al., 2017).", "type": "text" } ], "index": 14 } ], "index": 12.5, "bbox_fs": [ 104, 251, 506, 298 ] }, { "type": "text", "bbox": [ 107, 302, 504, 346 ], "lines": [ { "bbox": [ 105, 302, 504, 315 ], "spans": [ { "bbox": [ 105, 302, 504, 315 ], "score": 1.0, "content": "Model variants. We have three models with various sizes to adapt to various computing scenarios.", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 313, 506, 326 ], "spans": [ { "bbox": [ 105, 313, 506, 326 ], "score": 1.0, "content": "The detailed settings are shown in Table 9 (see B.1). All experiments in this paper are performed on", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 324, 506, 337 ], "spans": [ { "bbox": [ 105, 324, 506, 337 ], "score": 1.0, "content": "Tesla V100 machines. Training the tiny model for 300 epochs takes about 1.3 days on a single node", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 335, 463, 348 ], "spans": [ { "bbox": [ 105, 335, 463, 348 ], "score": 1.0, "content": "with 8 V100 GPU cards. CPVT-S and CPVT-B take about 1.6 and 2.5 days, respectively.", "type": "text" } ], "index": 18 } ], "index": 16.5, "bbox_fs": [ 105, 302, 506, 348 ] }, { "type": "text", "bbox": [ 107, 352, 505, 407 ], "lines": [ { "bbox": [ 106, 352, 505, 364 ], "spans": [ { "bbox": [ 106, 352, 505, 364 ], "score": 1.0, "content": "Training details All the models (except for CPVT-B) are trained for 300 epochs with a global batch", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 362, 505, 376 ], "spans": [ { "bbox": [ 105, 362, 505, 376 ], "score": 1.0, "content": "size of 2048 on Tesla V100 machines using AdamW optimizer (Loshchilov & Hutter, 2019). We", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 373, 505, 388 ], "spans": [ { "bbox": [ 105, 373, 505, 388 ], "score": 1.0, "content": "do not tune the hyper-parameters and strictly comply with the settings in DeiT (Touvron et al.,", "type": "text" } ], "index": 21 }, { "bbox": [ 104, 383, 506, 399 ], "spans": [ { "bbox": [ 104, 383, 315, 399 ], "score": 1.0, "content": "2020). The learning rate is scaled with this formula", "type": "text" }, { "bbox": [ 316, 385, 448, 397 ], "score": 0.85, "content": "l r _ { \\mathrm { s c a l e } } = 0 . 0 0 0 5 { \\cdot } \\mathrm { B a t c h S i z e } _ { \\mathrm { g l o b a l } } / { \\scriptstyle 5 1 2 }", "type": "inline_equation" }, { "bbox": [ 448, 383, 506, 399 ], "score": 1.0, "content": ". The detailed", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 397, 234, 408 ], "spans": [ { "bbox": [ 105, 397, 234, 408 ], "score": 1.0, "content": "hyperparameters are in the B.2.", "type": "text" } ], "index": 23 } ], "index": 21, "bbox_fs": [ 104, 352, 506, 408 ] }, { "type": "title", "bbox": [ 107, 423, 323, 434 ], "lines": [ { "bbox": [ 105, 421, 325, 435 ], "spans": [ { "bbox": [ 105, 421, 325, 435 ], "score": 1.0, "content": "4.2 GENERALIZATION TO HIGHER RESOLUTIONS", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 443, 505, 543 ], "lines": [ { "bbox": [ 106, 443, 505, 457 ], "spans": [ { "bbox": [ 106, 443, 505, 457 ], "score": 1.0, "content": "As mentioned before, our proposed PEG can directly generalize to larger image sizes without any", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 455, 506, 468 ], "spans": [ { "bbox": [ 105, 455, 400, 468 ], "score": 1.0, "content": "fine-tuning. We confirm this here by evaluating the models trained with", "type": "text" }, { "bbox": [ 400, 455, 444, 466 ], "score": 0.9, "content": "2 2 4 \\times 2 2 4", "type": "inline_equation" }, { "bbox": [ 445, 455, 506, 468 ], "score": 1.0, "content": "images on the", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 465, 505, 478 ], "spans": [ { "bbox": [ 106, 466, 150, 477 ], "score": 0.86, "content": "3 8 4 \\times 3 8 4", "type": "inline_equation" }, { "bbox": [ 150, 465, 153, 478 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 154, 466, 197, 477 ], "score": 0.77, "content": "4 4 8 \\times 4 4 8", "type": "inline_equation" }, { "bbox": [ 197, 465, 200, 478 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 200, 466, 244, 477 ], "score": 0.73, "content": "5 1 2 \\times 5 1 2", "type": "inline_equation" }, { "bbox": [ 244, 465, 505, 478 ], "score": 1.0, "content": "images, respectively. The results are shown in Table 2. With the", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 477, 504, 488 ], "spans": [ { "bbox": [ 106, 477, 151, 487 ], "score": 0.87, "content": "3 8 4 \\times 3 8 4", "type": "inline_equation" }, { "bbox": [ 152, 477, 476, 488 ], "score": 1.0, "content": "input images, the DeiT-tiny with learnable positional encodings degrades from", "type": "text" }, { "bbox": [ 477, 477, 504, 487 ], "score": 0.86, "content": "7 2 . 2 \\%", "type": "inline_equation" } ], "index": 28 }, { "bbox": [ 105, 488, 506, 501 ], "spans": [ { "bbox": [ 105, 488, 117, 501 ], "score": 1.0, "content": "to", "type": "text" }, { "bbox": [ 117, 488, 144, 498 ], "score": 0.86, "content": "7 1 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 145, 488, 421, 501 ], "score": 1.0, "content": ". When equipped with sine encoding, the tiny model degrades from", "type": "text" }, { "bbox": [ 421, 488, 449, 498 ], "score": 0.87, "content": "7 2 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 449, 488, 461, 501 ], "score": 1.0, "content": "to", "type": "text" }, { "bbox": [ 461, 488, 488, 498 ], "score": 0.9, "content": "7 0 . 8 \\%", "type": "inline_equation" }, { "bbox": [ 488, 488, 506, 501 ], "score": 1.0, "content": ". In", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 498, 505, 512 ], "spans": [ { "bbox": [ 105, 498, 505, 512 ], "score": 1.0, "content": "constrat, our CPVT model with the proposed PEG can directly process the larger input images, and", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 509, 505, 523 ], "spans": [ { "bbox": [ 105, 509, 271, 523 ], "score": 1.0, "content": "CPVT-Ti’s performance is boosted from", "type": "text" }, { "bbox": [ 271, 510, 298, 520 ], "score": 0.87, "content": "7 3 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 299, 509, 310, 523 ], "score": 1.0, "content": "to", "type": "text" }, { "bbox": [ 311, 510, 338, 520 ], "score": 0.87, "content": "7 4 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 338, 509, 406, 523 ], "score": 1.0, "content": "when applied to", "type": "text" }, { "bbox": [ 407, 510, 451, 520 ], "score": 0.9, "content": "3 8 4 \\times 3 8 4", "type": "inline_equation" }, { "bbox": [ 451, 509, 505, 523 ], "score": 1.0, "content": "images. Our", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 520, 505, 533 ], "spans": [ { "bbox": [ 105, 520, 253, 533 ], "score": 1.0, "content": "CPVT-Ti outperforms DeiT-tiny by", "type": "text" }, { "bbox": [ 254, 521, 276, 531 ], "score": 0.85, "content": "3 . 0 \\%", "type": "inline_equation" }, { "bbox": [ 276, 520, 505, 533 ], "score": 1.0, "content": ". This gap continues to increase as the input resolution", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 531, 145, 546 ], "spans": [ { "bbox": [ 105, 531, 145, 546 ], "score": 1.0, "content": "enlarges.", "type": "text" } ], "index": 33 } ], "index": 29, "bbox_fs": [ 105, 443, 506, 546 ] }, { "type": "table", "bbox": [ 126, 591, 486, 706 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 555, 503, 578 ], "group_id": 0, "lines": [ { "bbox": [ 107, 554, 504, 567 ], "spans": [ { "bbox": [ 107, 554, 504, 567 ], "score": 1.0, "content": "Table 2. Direct evaluation on other resolutions without fine-tuning. The models are trained on", "type": "text" } ], "index": 34 }, { "bbox": [ 108, 566, 453, 577 ], "spans": [ { "bbox": [ 108, 566, 147, 577 ], "score": 0.88, "content": "2 2 4 \\times 2 2 4", "type": "inline_equation" }, { "bbox": [ 147, 566, 288, 577 ], "score": 1.0, "content": ". A simple PEG of a single layer of", "type": "text" }, { "bbox": [ 288, 566, 307, 577 ], "score": 0.87, "content": "3 \\times 3", "type": "inline_equation" }, { "bbox": [ 308, 566, 453, 577 ], "score": 1.0, "content": "depth-wise convolution is used here", "type": "text" } ], "index": 35 } ], "index": 34.5 }, { "type": "table_body", "bbox": [ 126, 591, 486, 706 ], "group_id": 0, "lines": [ { "bbox": [ 126, 591, 486, 706 ], "spans": [ { "bbox": [ 126, 591, 486, 706 ], "score": 0.986, "html": "
ModelParams160(%)224(%)384(%)448(%)512(%)
DeiT-tinyDeiT-tiny (sin)DeiT-tiny (no pos)CPVT-TiCPVT-Ti ‡6M6M6M6M6M65.665.262.166.8(+1.2)67.7 (+2.1)72.272.368.272.4(+0.2)73.4(+1.2)71.270.868.673.2(+2.0)74.2(+3.0)68.868.268.471.8(+3.0)72.6(+3.8)65.965.165.070.3(+4.4)70.8(+4.9)
DeiT-smallCPVT-S22M22M75.676.1(+0.5)79.979.978.180.4(+1.5)75.978.6(+2.7)72.676.8(+4.2)
DeiT-baseCPVT-B86M86M79.180.5(+1.4)81.881.9(+0.1)79.782.3(+2.6)79.882.4(+2.6)78.281.0(+2.8)
", "type": "table", "image_path": "ab1cbd1190803aa7fd6c2b7ced714aed46988bc72d1faf1b2b651896176f78f0.jpg" } ] } ], "index": 37, "virtual_lines": [ { "bbox": [ 126, 591, 486, 629.3333333333334 ], "spans": [], "index": 36 }, { "bbox": [ 126, 629.3333333333334, 486, 667.6666666666667 ], "spans": [], "index": 37 }, { "bbox": [ 126, 667.6666666666667, 486, 706.0000000000001 ], "spans": [], "index": 38 } ] }, { "type": "table_footnote", "bbox": [ 174, 708, 437, 719 ], "group_id": 0, "lines": [ { "bbox": [ 173, 706, 437, 720 ], "spans": [ { "bbox": [ 173, 706, 437, 720 ], "score": 1.0, "content": "‡: Insert one PEG each after the first encoder till the fifth encoder", "type": "text" } ], "index": 39 } ], "index": 39 } ], "index": 37 } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 82, 309, 93 ], "lines": [ { "bbox": [ 106, 82, 310, 95 ], "spans": [ { "bbox": [ 106, 82, 310, 95 ], "score": 1.0, "content": "4.3 CPVT WITH GLOBAL AVERAGE POOLING", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 106, 102, 505, 235 ], "lines": [ { "bbox": [ 106, 103, 506, 115 ], "spans": [ { "bbox": [ 106, 103, 506, 115 ], "score": 1.0, "content": "By design, the proposed PEG is translation-equivariant (ignore paddings). Thus, if we further use", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 113, 506, 127 ], "spans": [ { "bbox": [ 105, 113, 506, 127 ], "score": 1.0, "content": "the translation-invariant global average pooling (GAP) instead of the cls token before the final", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 126, 505, 137 ], "spans": [ { "bbox": [ 106, 126, 505, 137 ], "score": 1.0, "content": "classification layer of CPVT. CPVT can be translation-invariant, which should be beneficial to the", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 135, 505, 149 ], "spans": [ { "bbox": [ 105, 135, 505, 149 ], "score": 1.0, "content": "ImageNet classification task. Note the using GAP here results in even less computation complexity", "type": "text" } ], "index": 4 }, { "bbox": [ 104, 146, 505, 160 ], "spans": [ { "bbox": [ 104, 146, 505, 160 ], "score": 1.0, "content": "because we do not need to compute the attention interaction between the class token and the image", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 158, 505, 170 ], "spans": [ { "bbox": [ 105, 158, 427, 170 ], "score": 1.0, "content": "patches. As shown in Table 3, using GAP here can boost CPVT by more than", "type": "text" }, { "bbox": [ 427, 158, 443, 168 ], "score": 0.86, "content": "1 \\%", "type": "inline_equation" }, { "bbox": [ 443, 158, 505, 170 ], "score": 1.0, "content": ". For example,", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 169, 505, 182 ], "spans": [ { "bbox": [ 105, 169, 265, 182 ], "score": 1.0, "content": "equipping CPVT-Ti with GAP obtains", "type": "text" }, { "bbox": [ 266, 169, 293, 180 ], "score": 0.87, "content": "7 4 . 9 \\%", "type": "inline_equation" }, { "bbox": [ 293, 169, 505, 182 ], "score": 1.0, "content": "top-1 accuracy on the ImageNet validation dataset,", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 180, 505, 192 ], "spans": [ { "bbox": [ 106, 180, 294, 192 ], "score": 1.0, "content": "which outperforms DeiT-tiny by a large margin", "type": "text" }, { "bbox": [ 295, 180, 328, 191 ], "score": 0.89, "content": "( + 2 . 7 \\% )", "type": "inline_equation" }, { "bbox": [ 328, 180, 505, 192 ], "score": 1.0, "content": ". Moreover, it even exceeds DeiT-tiny model", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 191, 504, 203 ], "spans": [ { "bbox": [ 106, 191, 171, 203 ], "score": 1.0, "content": "with distillation", "type": "text" }, { "bbox": [ 172, 191, 204, 202 ], "score": 0.86, "content": "( 7 4 . 5 \\% )", "type": "inline_equation" }, { "bbox": [ 204, 191, 482, 203 ], "score": 1.0, "content": ". In contrast, DeiT with GAP cannot gain so much improvement (only", "type": "text" }, { "bbox": [ 482, 191, 504, 202 ], "score": 0.86, "content": "0 . 4 \\%", "type": "inline_equation" } ], "index": 9 }, { "bbox": [ 105, 202, 505, 214 ], "spans": [ { "bbox": [ 105, 202, 505, 214 ], "score": 1.0, "content": "as shown in Table 3) because the original learnable absolute PE is not translation-equivariant and", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 212, 505, 225 ], "spans": [ { "bbox": [ 105, 212, 505, 225 ], "score": 1.0, "content": "thus GAP with the PE is not translation-invariant. Given the superior performance, we hope our", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 224, 347, 236 ], "spans": [ { "bbox": [ 106, 224, 347, 236 ], "score": 1.0, "content": "model can be a strong PE alternative in vision transformers.", "type": "text" } ], "index": 12 } ], "index": 6.5 }, { "type": "table", "bbox": [ 155, 281, 456, 397 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 244, 503, 267 ], "group_id": 0, "lines": [ { "bbox": [ 106, 242, 504, 258 ], "spans": [ { "bbox": [ 106, 242, 504, 258 ], "score": 1.0, "content": "Table 3. Performance comparison of Class Token (CLT) and global average pooling (GAP) on", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 255, 321, 268 ], "spans": [ { "bbox": [ 106, 255, 321, 268 ], "score": 1.0, "content": "ImageNet. CPVT’s can be further boosted with GAP", "type": "text" } ], "index": 14 } ], "index": 13.5 }, { "type": "table_body", "bbox": [ 155, 281, 456, 397 ], "group_id": 0, "lines": [ { "bbox": [ 155, 281, 456, 397 ], "spans": [ { "bbox": [ 155, 281, 456, 397 ], "score": 0.982, "html": "
ModelHeadParamsTop-1 Acc(%)Top-5 Acc(%）
DeiT-tiny (Touvron et al., 2020)DeiT-tinyCPVT-Ti tCPVT-Ti tCLTGAPCLTGAP6M6M6M6M72.272.673.474.991.091.291.892.6
DeiT-small (Touvron et al.,2020)DeiT-smallCPVT-S ‡CPVT-S tCLTGAP22M22M23M79.980.280.595.095.295.295.7
CLT
GAP23M81.5
", "type": "table", "image_path": "3dc1c296f9299bd138b8b334ecbaa6aa75e5f72905c2e70be90bfa6573173918.jpg" } ] } ], "index": 16, "virtual_lines": [ { "bbox": [ 155, 281, 456, 319.6666666666667 ], "spans": [], "index": 15 }, { "bbox": [ 155, 319.6666666666667, 456, 358.33333333333337 ], "spans": [], "index": 16 }, { "bbox": [ 155, 358.33333333333337, 456, 397.00000000000006 ], "spans": [], "index": 17 } ] }, { "type": "table_footnote", "bbox": [ 187, 399, 422, 410 ], "group_id": 0, "lines": [ { "bbox": [ 186, 398, 424, 411 ], "spans": [ { "bbox": [ 186, 398, 424, 411 ], "score": 1.0, "content": "‡: Insert one PEG each after the first encoder till the fifth encoder", "type": "text" } ], "index": 18 } ], "index": 18 } ], "index": 16 }, { "type": "title", "bbox": [ 107, 430, 226, 442 ], "lines": [ { "bbox": [ 105, 430, 227, 443 ], "spans": [ { "bbox": [ 105, 430, 227, 443 ], "score": 1.0, "content": "4.4 COMPLEXITY OF PEG", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "text", "bbox": [ 106, 456, 505, 534 ], "lines": [ { "bbox": [ 105, 456, 505, 469 ], "spans": [ { "bbox": [ 105, 456, 291, 469 ], "score": 1.0, "content": "Few Parameters. Given the model dimension", "type": "text" }, { "bbox": [ 292, 457, 298, 467 ], "score": 0.77, "content": "d", "type": "inline_equation" }, { "bbox": [ 299, 456, 505, 469 ], "score": 1.0, "content": ", the extra number of parameters introduced by PEG", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 466, 506, 480 ], "spans": [ { "bbox": [ 105, 466, 115, 480 ], "score": 1.0, "content": "is", "type": "text" }, { "bbox": [ 115, 467, 152, 478 ], "score": 0.93, "content": "d \\times l \\times k ^ { 2 }", "type": "inline_equation" }, { "bbox": [ 153, 466, 205, 480 ], "score": 1.0, "content": "if we choose", "type": "text" }, { "bbox": [ 205, 468, 210, 477 ], "score": 0.3, "content": "l", "type": "inline_equation" }, { "bbox": [ 210, 466, 355, 480 ], "score": 1.0, "content": "depth-wise convolutions with kernel", "type": "text" }, { "bbox": [ 356, 468, 362, 477 ], "score": 0.76, "content": "k", "type": "inline_equation" }, { "bbox": [ 363, 466, 404, 480 ], "score": 1.0, "content": ". If we use", "type": "text" }, { "bbox": [ 404, 468, 409, 477 ], "score": 0.35, "content": "l", "type": "inline_equation" }, { "bbox": [ 409, 466, 506, 480 ], "score": 1.0, "content": "separable convolutions,", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 477, 505, 491 ], "spans": [ { "bbox": [ 105, 477, 186, 491 ], "score": 1.0, "content": "this value becomes", "type": "text" }, { "bbox": [ 186, 477, 235, 490 ], "score": 0.92, "content": "l ( d ^ { 2 } + k ^ { 2 } d )", "type": "inline_equation" }, { "bbox": [ 235, 477, 267, 491 ], "score": 1.0, "content": ". When", "type": "text" }, { "bbox": [ 267, 478, 294, 488 ], "score": 0.92, "content": "k = 3", "type": "inline_equation" }, { "bbox": [ 294, 477, 312, 491 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 313, 478, 336, 488 ], "score": 0.89, "content": "l = 1", "type": "inline_equation" }, { "bbox": [ 337, 477, 383, 491 ], "score": 1.0, "content": ", CPVT-Ti", "type": "text" }, { "bbox": [ 383, 478, 421, 489 ], "score": 0.85, "content": "d = 1 9 2 ,", "type": "inline_equation" }, { "bbox": [ 421, 477, 505, 491 ], "score": 1.0, "content": ") brings about 1, 728", "type": "text" } ], "index": 22 }, { "bbox": [ 104, 489, 504, 501 ], "spans": [ { "bbox": [ 104, 489, 405, 501 ], "score": 1.0, "content": "parameters. Note that DeiT-tiny utilizes learnable position encodings with", "type": "text" }, { "bbox": [ 405, 489, 504, 500 ], "score": 0.9, "content": "1 9 2 \\times 1 4 \\times 1 4 = 3 7 6 3 2", "type": "inline_equation" } ], "index": 23 }, { "bbox": [ 104, 499, 506, 513 ], "spans": [ { "bbox": [ 104, 499, 506, 513 ], "score": 1.0, "content": "parameters. Therefore, CPVT-Ti has 35, 904 fewer number of parameters than DeiT-tiny. Even", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 511, 506, 524 ], "spans": [ { "bbox": [ 106, 511, 383, 524 ], "score": 1.0, "content": "using 4 layers of separable convolutions, CPVT-Ti introduces only", "type": "text" }, { "bbox": [ 383, 511, 480, 522 ], "score": 0.88, "content": "3 8 9 5 2 - 3 7 6 3 2 = 9 6 0", "type": "inline_equation" }, { "bbox": [ 480, 511, 506, 524 ], "score": 1.0, "content": "more", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 522, 460, 535 ], "spans": [ { "bbox": [ 105, 522, 309, 535 ], "score": 1.0, "content": "parameters, which is negelectable compared to the", "type": "text" }, { "bbox": [ 310, 522, 333, 533 ], "score": 0.3, "content": "5 . 7 \\mathbf { M }", "type": "inline_equation" }, { "bbox": [ 334, 522, 460, 535 ], "score": 1.0, "content": "model parameters of DeiT-tiny.", "type": "text" } ], "index": 26 } ], "index": 23 }, { "type": "text", "bbox": [ 107, 543, 505, 577 ], "lines": [ { "bbox": [ 105, 543, 504, 556 ], "spans": [ { "bbox": [ 105, 543, 204, 556 ], "score": 1.0, "content": "FLOPs. As for FLOPs,", "type": "text" }, { "bbox": [ 205, 545, 209, 554 ], "score": 0.63, "content": "l", "type": "inline_equation" }, { "bbox": [ 209, 543, 248, 556 ], "score": 1.0, "content": "layers of", "type": "text" }, { "bbox": [ 248, 544, 272, 555 ], "score": 0.9, "content": "k \\times k", "type": "inline_equation" }, { "bbox": [ 273, 543, 415, 556 ], "score": 1.0, "content": "depth-wise convolutions possesses", "type": "text" }, { "bbox": [ 415, 543, 504, 555 ], "score": 0.91, "content": "1 4 \\times 1 4 \\times d \\times l \\times k ^ { 2 }", "type": "inline_equation" } ], "index": 27 }, { "bbox": [ 105, 555, 505, 567 ], "spans": [ { "bbox": [ 105, 555, 333, 567 ], "score": 1.0, "content": "FLOPS. Taking the tiny model for example, it involves", "type": "text" }, { "bbox": [ 334, 555, 441, 566 ], "score": 0.91, "content": "1 9 6 \\times 1 9 2 \\times 9 = 0 . 3 4 M", "type": "inline_equation" }, { "bbox": [ 441, 555, 505, 567 ], "score": 1.0, "content": "FLOPS for the", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 566, 486, 577 ], "spans": [ { "bbox": [ 105, 566, 155, 577 ], "score": 1.0, "content": "simple case", "type": "text" }, { "bbox": [ 155, 566, 180, 576 ], "score": 0.91, "content": "k = 3", "type": "inline_equation" }, { "bbox": [ 181, 566, 198, 577 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 199, 567, 221, 576 ], "score": 0.9, "content": "l = 1", "type": "inline_equation" }, { "bbox": [ 221, 566, 486, 577 ], "score": 1.0, "content": ", which is neglectable because the model has 2.1G FLOPs in total.", "type": "text" } ], "index": 29 } ], "index": 28 }, { "type": "title", "bbox": [ 108, 590, 257, 601 ], "lines": [ { "bbox": [ 106, 590, 259, 603 ], "spans": [ { "bbox": [ 106, 590, 259, 603 ], "score": 1.0, "content": "4.5 PERFORMANCE COMPARISON", "type": "text" } ], "index": 30 } ], "index": 30 }, { "type": "text", "bbox": [ 107, 610, 296, 731 ], "lines": [ { "bbox": [ 106, 610, 297, 622 ], "spans": [ { "bbox": [ 106, 610, 297, 622 ], "score": 1.0, "content": "We evaluate the performance of CPVT models", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 622, 297, 633 ], "spans": [ { "bbox": [ 105, 622, 297, 633 ], "score": 1.0, "content": "on the ImageNet validation dataset and report", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 632, 296, 645 ], "spans": [ { "bbox": [ 106, 632, 296, 645 ], "score": 1.0, "content": "the results in Table 4. Compared with DeiT,", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 642, 297, 657 ], "spans": [ { "bbox": [ 106, 642, 297, 657 ], "score": 1.0, "content": "CPVT models have much better top-1 accuracy", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 655, 297, 667 ], "spans": [ { "bbox": [ 106, 655, 297, 667 ], "score": 1.0, "content": "with similar throughputs. Our models can en-", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 666, 297, 678 ], "spans": [ { "bbox": [ 106, 666, 297, 678 ], "score": 1.0, "content": "joy performance improvement when inputs are", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 677, 297, 688 ], "spans": [ { "bbox": [ 105, 677, 297, 688 ], "score": 1.0, "content": "upscaled without fine-tuning, while DeiT de-", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 687, 297, 700 ], "spans": [ { "bbox": [ 105, 687, 297, 700 ], "score": 1.0, "content": "grades as discussed in Table 2, see also Fig-", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 699, 297, 711 ], "spans": [ { "bbox": [ 106, 699, 297, 711 ], "score": 1.0, "content": "ure 3 for a clear comparison. Noticeably, Our", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 709, 297, 721 ], "spans": [ { "bbox": [ 106, 709, 297, 721 ], "score": 1.0, "content": "model with GAP marked a new state-of-the-art", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 720, 206, 732 ], "spans": [ { "bbox": [ 106, 720, 206, 732 ], "score": 1.0, "content": "for vision Transformers.", "type": "text" } ], "index": 41 } ], "index": 36 }, { "type": "image", "bbox": [ 309, 581, 498, 689 ], "blocks": [ { "type": "image_body", "bbox": [ 309, 581, 498, 689 ], "group_id": 0, "lines": [ { "bbox": [ 309, 581, 498, 689 ], "spans": [ { "bbox": [ 309, 581, 498, 689 ], "score": 0.97, "type": "image", "image_path": "c53aeceb62b3945f870e2289f3bd7dc81107062bd6e3b4e704ad90e2320543d6.jpg" } ] } ], "index": 46, "virtual_lines": [ { "bbox": [ 309, 581, 498, 593.0 ], "spans": [], "index": 42 }, { "bbox": [ 309, 593.0, 498, 605.0 ], "spans": [], "index": 43 }, { "bbox": [ 309, 605.0, 498, 617.0 ], "spans": [], "index": 44 }, { "bbox": [ 309, 617.0, 498, 629.0 ], "spans": [], "index": 45 }, { "bbox": [ 309, 629.0, 498, 641.0 ], "spans": [], "index": 46 }, { "bbox": [ 309, 641.0, 498, 653.0 ], "spans": [], "index": 47 }, { "bbox": [ 309, 653.0, 498, 665.0 ], "spans": [], "index": 48 }, { "bbox": [ 309, 665.0, 498, 677.0 ], "spans": [], "index": 49 }, { "bbox": [ 309, 677.0, 498, 689.0 ], "spans": [], "index": 50 } ] }, { "type": "image_caption", "bbox": [ 303, 703, 505, 748 ], "group_id": 0, "lines": [ { "bbox": [ 303, 703, 505, 715 ], "spans": [ { "bbox": [ 303, 703, 505, 715 ], "score": 1.0, "content": "Figure 3. Comparison of CPVT and DeiT models", "type": "text" } ], "index": 51 }, { "bbox": [ 303, 714, 504, 726 ], "spans": [ { "bbox": [ 303, 714, 479, 726 ], "score": 1.0, "content": "under various configurations. Note CPVT", "type": "text" }, { "bbox": [ 480, 714, 504, 725 ], "score": 0.53, "content": "@ 3 8 4", "type": "inline_equation" } ], "index": 52 }, { "bbox": [ 303, 725, 505, 737 ], "spans": [ { "bbox": [ 303, 725, 505, 737 ], "score": 1.0, "content": "has improved performance. More PEGs can result", "type": "text" } ], "index": 53 }, { "bbox": [ 303, 736, 491, 748 ], "spans": [ { "bbox": [ 303, 736, 491, 748 ], "score": 1.0, "content": "in better performance. CPVT-GAP is the best.", "type": "text" } ], "index": 54 } ], "index": 52.5 } ], "index": 49.25 } ], "page_idx": 5, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 25, 293, 38 ], "spans": [ { "bbox": [ 106, 25, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "6", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 82, 309, 93 ], "lines": [ { "bbox": [ 106, 82, 310, 95 ], "spans": [ { "bbox": [ 106, 82, 310, 95 ], "score": 1.0, "content": "4.3 CPVT WITH GLOBAL AVERAGE POOLING", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 106, 102, 505, 235 ], "lines": [ { "bbox": [ 106, 103, 506, 115 ], "spans": [ { "bbox": [ 106, 103, 506, 115 ], "score": 1.0, "content": "By design, the proposed PEG is translation-equivariant (ignore paddings). Thus, if we further use", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 113, 506, 127 ], "spans": [ { "bbox": [ 105, 113, 506, 127 ], "score": 1.0, "content": "the translation-invariant global average pooling (GAP) instead of the cls token before the final", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 126, 505, 137 ], "spans": [ { "bbox": [ 106, 126, 505, 137 ], "score": 1.0, "content": "classification layer of CPVT. CPVT can be translation-invariant, which should be beneficial to the", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 135, 505, 149 ], "spans": [ { "bbox": [ 105, 135, 505, 149 ], "score": 1.0, "content": "ImageNet classification task. Note the using GAP here results in even less computation complexity", "type": "text" } ], "index": 4 }, { "bbox": [ 104, 146, 505, 160 ], "spans": [ { "bbox": [ 104, 146, 505, 160 ], "score": 1.0, "content": "because we do not need to compute the attention interaction between the class token and the image", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 158, 505, 170 ], "spans": [ { "bbox": [ 105, 158, 427, 170 ], "score": 1.0, "content": "patches. As shown in Table 3, using GAP here can boost CPVT by more than", "type": "text" }, { "bbox": [ 427, 158, 443, 168 ], "score": 0.86, "content": "1 \\%", "type": "inline_equation" }, { "bbox": [ 443, 158, 505, 170 ], "score": 1.0, "content": ". For example,", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 169, 505, 182 ], "spans": [ { "bbox": [ 105, 169, 265, 182 ], "score": 1.0, "content": "equipping CPVT-Ti with GAP obtains", "type": "text" }, { "bbox": [ 266, 169, 293, 180 ], "score": 0.87, "content": "7 4 . 9 \\%", "type": "inline_equation" }, { "bbox": [ 293, 169, 505, 182 ], "score": 1.0, "content": "top-1 accuracy on the ImageNet validation dataset,", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 180, 505, 192 ], "spans": [ { "bbox": [ 106, 180, 294, 192 ], "score": 1.0, "content": "which outperforms DeiT-tiny by a large margin", "type": "text" }, { "bbox": [ 295, 180, 328, 191 ], "score": 0.89, "content": "( + 2 . 7 \\% )", "type": "inline_equation" }, { "bbox": [ 328, 180, 505, 192 ], "score": 1.0, "content": ". Moreover, it even exceeds DeiT-tiny model", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 191, 504, 203 ], "spans": [ { "bbox": [ 106, 191, 171, 203 ], "score": 1.0, "content": "with distillation", "type": "text" }, { "bbox": [ 172, 191, 204, 202 ], "score": 0.86, "content": "( 7 4 . 5 \\% )", "type": "inline_equation" }, { "bbox": [ 204, 191, 482, 203 ], "score": 1.0, "content": ". In contrast, DeiT with GAP cannot gain so much improvement (only", "type": "text" }, { "bbox": [ 482, 191, 504, 202 ], "score": 0.86, "content": "0 . 4 \\%", "type": "inline_equation" } ], "index": 9 }, { "bbox": [ 105, 202, 505, 214 ], "spans": [ { "bbox": [ 105, 202, 505, 214 ], "score": 1.0, "content": "as shown in Table 3) because the original learnable absolute PE is not translation-equivariant and", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 212, 505, 225 ], "spans": [ { "bbox": [ 105, 212, 505, 225 ], "score": 1.0, "content": "thus GAP with the PE is not translation-invariant. Given the superior performance, we hope our", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 224, 347, 236 ], "spans": [ { "bbox": [ 106, 224, 347, 236 ], "score": 1.0, "content": "model can be a strong PE alternative in vision transformers.", "type": "text" } ], "index": 12 } ], "index": 6.5, "bbox_fs": [ 104, 103, 506, 236 ] }, { "type": "table", "bbox": [ 155, 281, 456, 397 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 244, 503, 267 ], "group_id": 0, "lines": [ { "bbox": [ 106, 242, 504, 258 ], "spans": [ { "bbox": [ 106, 242, 504, 258 ], "score": 1.0, "content": "Table 3. Performance comparison of Class Token (CLT) and global average pooling (GAP) on", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 255, 321, 268 ], "spans": [ { "bbox": [ 106, 255, 321, 268 ], "score": 1.0, "content": "ImageNet. CPVT’s can be further boosted with GAP", "type": "text" } ], "index": 14 } ], "index": 13.5 }, { "type": "table_body", "bbox": [ 155, 281, 456, 397 ], "group_id": 0, "lines": [ { "bbox": [ 155, 281, 456, 397 ], "spans": [ { "bbox": [ 155, 281, 456, 397 ], "score": 0.982, "html": "
ModelHeadParamsTop-1 Acc(%)Top-5 Acc(%）
DeiT-tiny (Touvron et al., 2020)DeiT-tinyCPVT-Ti tCPVT-Ti tCLTGAPCLTGAP6M6M6M6M72.272.673.474.991.091.291.892.6
DeiT-small (Touvron et al.,2020)DeiT-smallCPVT-S ‡CPVT-S tCLTGAP22M22M23M79.980.280.595.095.295.295.7
CLT
GAP23M81.5
", "type": "table", "image_path": "3dc1c296f9299bd138b8b334ecbaa6aa75e5f72905c2e70be90bfa6573173918.jpg" } ] } ], "index": 16, "virtual_lines": [ { "bbox": [ 155, 281, 456, 319.6666666666667 ], "spans": [], "index": 15 }, { "bbox": [ 155, 319.6666666666667, 456, 358.33333333333337 ], "spans": [], "index": 16 }, { "bbox": [ 155, 358.33333333333337, 456, 397.00000000000006 ], "spans": [], "index": 17 } ] }, { "type": "table_footnote", "bbox": [ 187, 399, 422, 410 ], "group_id": 0, "lines": [ { "bbox": [ 186, 398, 424, 411 ], "spans": [ { "bbox": [ 186, 398, 424, 411 ], "score": 1.0, "content": "‡: Insert one PEG each after the first encoder till the fifth encoder", "type": "text" } ], "index": 18 } ], "index": 18 } ], "index": 16 }, { "type": "title", "bbox": [ 107, 430, 226, 442 ], "lines": [ { "bbox": [ 105, 430, 227, 443 ], "spans": [ { "bbox": [ 105, 430, 227, 443 ], "score": 1.0, "content": "4.4 COMPLEXITY OF PEG", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "text", "bbox": [ 106, 456, 505, 534 ], "lines": [ { "bbox": [ 105, 456, 505, 469 ], "spans": [ { "bbox": [ 105, 456, 291, 469 ], "score": 1.0, "content": "Few Parameters. Given the model dimension", "type": "text" }, { "bbox": [ 292, 457, 298, 467 ], "score": 0.77, "content": "d", "type": "inline_equation" }, { "bbox": [ 299, 456, 505, 469 ], "score": 1.0, "content": ", the extra number of parameters introduced by PEG", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 466, 506, 480 ], "spans": [ { "bbox": [ 105, 466, 115, 480 ], "score": 1.0, "content": "is", "type": "text" }, { "bbox": [ 115, 467, 152, 478 ], "score": 0.93, "content": "d \\times l \\times k ^ { 2 }", "type": "inline_equation" }, { "bbox": [ 153, 466, 205, 480 ], "score": 1.0, "content": "if we choose", "type": "text" }, { "bbox": [ 205, 468, 210, 477 ], "score": 0.3, "content": "l", "type": "inline_equation" }, { "bbox": [ 210, 466, 355, 480 ], "score": 1.0, "content": "depth-wise convolutions with kernel", "type": "text" }, { "bbox": [ 356, 468, 362, 477 ], "score": 0.76, "content": "k", "type": "inline_equation" }, { "bbox": [ 363, 466, 404, 480 ], "score": 1.0, "content": ". If we use", "type": "text" }, { "bbox": [ 404, 468, 409, 477 ], "score": 0.35, "content": "l", "type": "inline_equation" }, { "bbox": [ 409, 466, 506, 480 ], "score": 1.0, "content": "separable convolutions,", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 477, 505, 491 ], "spans": [ { "bbox": [ 105, 477, 186, 491 ], "score": 1.0, "content": "this value becomes", "type": "text" }, { "bbox": [ 186, 477, 235, 490 ], "score": 0.92, "content": "l ( d ^ { 2 } + k ^ { 2 } d )", "type": "inline_equation" }, { "bbox": [ 235, 477, 267, 491 ], "score": 1.0, "content": ". When", "type": "text" }, { "bbox": [ 267, 478, 294, 488 ], "score": 0.92, "content": "k = 3", "type": "inline_equation" }, { "bbox": [ 294, 477, 312, 491 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 313, 478, 336, 488 ], "score": 0.89, "content": "l = 1", "type": "inline_equation" }, { "bbox": [ 337, 477, 383, 491 ], "score": 1.0, "content": ", CPVT-Ti", "type": "text" }, { "bbox": [ 383, 478, 421, 489 ], "score": 0.85, "content": "d = 1 9 2 ,", "type": "inline_equation" }, { "bbox": [ 421, 477, 505, 491 ], "score": 1.0, "content": ") brings about 1, 728", "type": "text" } ], "index": 22 }, { "bbox": [ 104, 489, 504, 501 ], "spans": [ { "bbox": [ 104, 489, 405, 501 ], "score": 1.0, "content": "parameters. Note that DeiT-tiny utilizes learnable position encodings with", "type": "text" }, { "bbox": [ 405, 489, 504, 500 ], "score": 0.9, "content": "1 9 2 \\times 1 4 \\times 1 4 = 3 7 6 3 2", "type": "inline_equation" } ], "index": 23 }, { "bbox": [ 104, 499, 506, 513 ], "spans": [ { "bbox": [ 104, 499, 506, 513 ], "score": 1.0, "content": "parameters. Therefore, CPVT-Ti has 35, 904 fewer number of parameters than DeiT-tiny. Even", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 511, 506, 524 ], "spans": [ { "bbox": [ 106, 511, 383, 524 ], "score": 1.0, "content": "using 4 layers of separable convolutions, CPVT-Ti introduces only", "type": "text" }, { "bbox": [ 383, 511, 480, 522 ], "score": 0.88, "content": "3 8 9 5 2 - 3 7 6 3 2 = 9 6 0", "type": "inline_equation" }, { "bbox": [ 480, 511, 506, 524 ], "score": 1.0, "content": "more", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 522, 460, 535 ], "spans": [ { "bbox": [ 105, 522, 309, 535 ], "score": 1.0, "content": "parameters, which is negelectable compared to the", "type": "text" }, { "bbox": [ 310, 522, 333, 533 ], "score": 0.3, "content": "5 . 7 \\mathbf { M }", "type": "inline_equation" }, { "bbox": [ 334, 522, 460, 535 ], "score": 1.0, "content": "model parameters of DeiT-tiny.", "type": "text" } ], "index": 26 } ], "index": 23, "bbox_fs": [ 104, 456, 506, 535 ] }, { "type": "text", "bbox": [ 107, 543, 505, 577 ], "lines": [ { "bbox": [ 105, 543, 504, 556 ], "spans": [ { "bbox": [ 105, 543, 204, 556 ], "score": 1.0, "content": "FLOPs. As for FLOPs,", "type": "text" }, { "bbox": [ 205, 545, 209, 554 ], "score": 0.63, "content": "l", "type": "inline_equation" }, { "bbox": [ 209, 543, 248, 556 ], "score": 1.0, "content": "layers of", "type": "text" }, { "bbox": [ 248, 544, 272, 555 ], "score": 0.9, "content": "k \\times k", "type": "inline_equation" }, { "bbox": [ 273, 543, 415, 556 ], "score": 1.0, "content": "depth-wise convolutions possesses", "type": "text" }, { "bbox": [ 415, 543, 504, 555 ], "score": 0.91, "content": "1 4 \\times 1 4 \\times d \\times l \\times k ^ { 2 }", "type": "inline_equation" } ], "index": 27 }, { "bbox": [ 105, 555, 505, 567 ], "spans": [ { "bbox": [ 105, 555, 333, 567 ], "score": 1.0, "content": "FLOPS. Taking the tiny model for example, it involves", "type": "text" }, { "bbox": [ 334, 555, 441, 566 ], "score": 0.91, "content": "1 9 6 \\times 1 9 2 \\times 9 = 0 . 3 4 M", "type": "inline_equation" }, { "bbox": [ 441, 555, 505, 567 ], "score": 1.0, "content": "FLOPS for the", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 566, 486, 577 ], "spans": [ { "bbox": [ 105, 566, 155, 577 ], "score": 1.0, "content": "simple case", "type": "text" }, { "bbox": [ 155, 566, 180, 576 ], "score": 0.91, "content": "k = 3", "type": "inline_equation" }, { "bbox": [ 181, 566, 198, 577 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 199, 567, 221, 576 ], "score": 0.9, "content": "l = 1", "type": "inline_equation" }, { "bbox": [ 221, 566, 486, 577 ], "score": 1.0, "content": ", which is neglectable because the model has 2.1G FLOPs in total.", "type": "text" } ], "index": 29 } ], "index": 28, "bbox_fs": [ 105, 543, 505, 577 ] }, { "type": "title", "bbox": [ 108, 590, 257, 601 ], "lines": [ { "bbox": [ 106, 590, 259, 603 ], "spans": [ { "bbox": [ 106, 590, 259, 603 ], "score": 1.0, "content": "4.5 PERFORMANCE COMPARISON", "type": "text" } ], "index": 30 } ], "index": 30 }, { "type": "text", "bbox": [ 107, 610, 296, 731 ], "lines": [ { "bbox": [ 106, 610, 297, 622 ], "spans": [ { "bbox": [ 106, 610, 297, 622 ], "score": 1.0, "content": "We evaluate the performance of CPVT models", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 622, 297, 633 ], "spans": [ { "bbox": [ 105, 622, 297, 633 ], "score": 1.0, "content": "on the ImageNet validation dataset and report", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 632, 296, 645 ], "spans": [ { "bbox": [ 106, 632, 296, 645 ], "score": 1.0, "content": "the results in Table 4. Compared with DeiT,", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 642, 297, 657 ], "spans": [ { "bbox": [ 106, 642, 297, 657 ], "score": 1.0, "content": "CPVT models have much better top-1 accuracy", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 655, 297, 667 ], "spans": [ { "bbox": [ 106, 655, 297, 667 ], "score": 1.0, "content": "with similar throughputs. Our models can en-", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 666, 297, 678 ], "spans": [ { "bbox": [ 106, 666, 297, 678 ], "score": 1.0, "content": "joy performance improvement when inputs are", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 677, 297, 688 ], "spans": [ { "bbox": [ 105, 677, 297, 688 ], "score": 1.0, "content": "upscaled without fine-tuning, while DeiT de-", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 687, 297, 700 ], "spans": [ { "bbox": [ 105, 687, 297, 700 ], "score": 1.0, "content": "grades as discussed in Table 2, see also Fig-", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 699, 297, 711 ], "spans": [ { "bbox": [ 106, 699, 297, 711 ], "score": 1.0, "content": "ure 3 for a clear comparison. Noticeably, Our", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 709, 297, 721 ], "spans": [ { "bbox": [ 106, 709, 297, 721 ], "score": 1.0, "content": "model with GAP marked a new state-of-the-art", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 720, 206, 732 ], "spans": [ { "bbox": [ 106, 720, 206, 732 ], "score": 1.0, "content": "for vision Transformers.", "type": "text" } ], "index": 41 } ], "index": 36, "bbox_fs": [ 105, 610, 297, 732 ] }, { "type": "image", "bbox": [ 309, 581, 498, 689 ], "blocks": [ { "type": "image_body", "bbox": [ 309, 581, 498, 689 ], "group_id": 0, "lines": [ { "bbox": [ 309, 581, 498, 689 ], "spans": [ { "bbox": [ 309, 581, 498, 689 ], "score": 0.97, "type": "image", "image_path": "c53aeceb62b3945f870e2289f3bd7dc81107062bd6e3b4e704ad90e2320543d6.jpg" } ] } ], "index": 46, "virtual_lines": [ { "bbox": [ 309, 581, 498, 593.0 ], "spans": [], "index": 42 }, { "bbox": [ 309, 593.0, 498, 605.0 ], "spans": [], "index": 43 }, { "bbox": [ 309, 605.0, 498, 617.0 ], "spans": [], "index": 44 }, { "bbox": [ 309, 617.0, 498, 629.0 ], "spans": [], "index": 45 }, { "bbox": [ 309, 629.0, 498, 641.0 ], "spans": [], "index": 46 }, { "bbox": [ 309, 641.0, 498, 653.0 ], "spans": [], "index": 47 }, { "bbox": [ 309, 653.0, 498, 665.0 ], "spans": [], "index": 48 }, { "bbox": [ 309, 665.0, 498, 677.0 ], "spans": [], "index": 49 }, { "bbox": [ 309, 677.0, 498, 689.0 ], "spans": [], "index": 50 } ] }, { "type": "image_caption", "bbox": [ 303, 703, 505, 748 ], "group_id": 0, "lines": [ { "bbox": [ 303, 703, 505, 715 ], "spans": [ { "bbox": [ 303, 703, 505, 715 ], "score": 1.0, "content": "Figure 3. Comparison of CPVT and DeiT models", "type": "text" } ], "index": 51 }, { "bbox": [ 303, 714, 504, 726 ], "spans": [ { "bbox": [ 303, 714, 479, 726 ], "score": 1.0, "content": "under various configurations. Note CPVT", "type": "text" }, { "bbox": [ 480, 714, 504, 725 ], "score": 0.53, "content": "@ 3 8 4", "type": "inline_equation" } ], "index": 52 }, { "bbox": [ 303, 725, 505, 737 ], "spans": [ { "bbox": [ 303, 725, 505, 737 ], "score": 1.0, "content": "has improved performance. More PEGs can result", "type": "text" } ], "index": 53 }, { "bbox": [ 303, 736, 491, 748 ], "spans": [ { "bbox": [ 303, 736, 491, 748 ], "score": 1.0, "content": "in better performance. CPVT-GAP is the best.", "type": "text" } ], "index": 54 } ], "index": 52.5 } ], "index": 49.25 } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 120, 112, 491, 416 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 80, 502, 103 ], "group_id": 0, "lines": [ { "bbox": [ 106, 79, 504, 93 ], "spans": [ { "bbox": [ 106, 79, 504, 93 ], "score": 1.0, "content": "Table 4. Comparison with ConvNets and Transformers on ImageNet and ImageNet Real (Beyer", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 90, 452, 104 ], "spans": [ { "bbox": [ 106, 90, 452, 104 ], "score": 1.0, "content": "et al., 2020). CPVT have much better performance compared with prior Transformers", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table_body", "bbox": [ 120, 112, 491, 416 ], "group_id": 0, "lines": [ { "bbox": [ 120, 112, 491, 416 ], "spans": [ { "bbox": [ 120, 112, 491, 416 ], "score": 0.986, "html": "
ModelsParams(M) InputInput|throughput*ImNettop-1 %Realtop-1 %
ResNet-50 (He et al., 2016)ResNet-101 (He et al., 2016)ResNet-152 (He et al.,2016)RegNetY-4GF (Radosavovic et al.,2020)EfficientNet-BO (Tan &Le,2019)EfficientNet-B1 (Tan&Le,2019)EfficientNet-B2 (Tan &Le,2019)EfficientNet-B3(Tan&Le,2019)EfficientNet-B4 (Tan& Le,2019)2545602122421226.176.282.5
224222422242753.6526.41156.7753.677.483.7
78.380.084.186.4
5891219522422694.377.183.5
240²1662.51255.7732.179.184.9
260²300280.181.685.986.8
3802349.482.988.0
ViT-B/16 (Dosovitskiy et al., 2021)ViT-L/1686307384285.927.377.976.511
3842
DeiT-tiny w/o PE (Touvron et al., 2020)DeiT-tiny (Touvron et al.,2020)DeiT-tiny (sine)CPVT-Ti tCPVT-Ti-GAP‡66666224222422242224²22422536.52536.52536.52500.72520.168.272.272.373.474.9180.180.381.382.5
DeiT-tiny (Touvron et al., 2020)CPVT-Tim66224222422536.52500.774.575.982.183.0
DeiT-small (Touvron et al., 2020)CPVT-S $CPVT-S-GAP‡22232322422242224²940.4930.5942.379.980.585.786.086.6
81.5
DeiT-base (Touvron et al.,2020)CPVT-B ‡CPVT-B-GAP‡862242292.381.886.7
888822422242285.5290.282.387.087.7
82.7
", "type": "table", "image_path": "d3a58c8cebdf84cdf2503d3f0b23c663985798b43d4e76a1b1f0608cbe091018.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 120, 112, 491, 213.33333333333331 ], "spans": [], "index": 2 }, { "bbox": [ 120, 213.33333333333331, 491, 314.66666666666663 ], "spans": [], "index": 3 }, { "bbox": [ 120, 314.66666666666663, 491, 415.99999999999994 ], "spans": [], "index": 4 } ] }, { "type": "table_footnote", "bbox": [ 135, 418, 398, 449 ], "group_id": 0, "lines": [ { "bbox": [ 135, 416, 399, 429 ], "spans": [ { "bbox": [ 135, 416, 399, 429 ], "score": 1.0, "content": "?: Measured in img/s on a 16GB V100 GPU as in (Touvron et al., 2020).", "type": "text" } ], "index": 5 }, { "bbox": [ 134, 426, 374, 439 ], "spans": [ { "bbox": [ 134, 426, 374, 439 ], "score": 1.0, "content": "‡: Insert one PEG each after the first encoder till the fifth encoder", "type": "text" } ], "index": 6 }, { "bbox": [ 135, 436, 378, 450 ], "spans": [ { "bbox": [ 135, 436, 378, 450 ], "score": 1.0, "content": "⚗ : trained with hard distillation using RegNetY-160 as the teacher.", "type": "text" } ], "index": 7 } ], "index": 6 } ], "index": 3 }, { "type": "text", "bbox": [ 108, 468, 504, 502 ], "lines": [ { "bbox": [ 106, 468, 506, 482 ], "spans": [ { "bbox": [ 106, 468, 506, 482 ], "score": 1.0, "content": "We further train CPVT-Ti and DeiT-tiny using the aforementioned training settings plus the hard", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 479, 506, 493 ], "spans": [ { "bbox": [ 106, 479, 506, 493 ], "score": 1.0, "content": "distillation proposed in (Touvron et al., 2020). Specifically, we use RegNetY-160 (Radosavovic", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 491, 428, 503 ], "spans": [ { "bbox": [ 106, 491, 275, 503 ], "score": 1.0, "content": "et al., 2020) as the teacher. CPVT obtains", "type": "text" }, { "bbox": [ 275, 491, 302, 501 ], "score": 0.87, "content": "7 5 . 9 \\%", "type": "inline_equation" }, { "bbox": [ 302, 491, 402, 503 ], "score": 1.0, "content": ", exceeding DeiT-tiny by", "type": "text" }, { "bbox": [ 402, 491, 424, 502 ], "score": 0.84, "content": "1 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 424, 491, 428, 503 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 10 } ], "index": 9 }, { "type": "title", "bbox": [ 107, 515, 354, 527 ], "lines": [ { "bbox": [ 105, 514, 356, 528 ], "spans": [ { "bbox": [ 105, 514, 356, 528 ], "score": 1.0, "content": "4.6 PEG ON PYRAMID TRANSFORMER ARCHITECTURES", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 107, 536, 505, 569 ], "lines": [ { "bbox": [ 105, 535, 505, 549 ], "spans": [ { "bbox": [ 105, 535, 505, 549 ], "score": 1.0, "content": "PVT (Wang et al., 2021) is a vision transformer with the multi-stage design like ResNet (He et al.,", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 545, 505, 561 ], "spans": [ { "bbox": [ 105, 545, 505, 561 ], "score": 1.0, "content": "2016). Swin (Liu et al., 2021) is a follow-up work and comes with higher performance. We apply", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 557, 350, 572 ], "spans": [ { "bbox": [ 105, 557, 350, 572 ], "score": 1.0, "content": "our method on both to demonstrate its generalization ability.", "type": "text" } ], "index": 14 } ], "index": 13 }, { "type": "text", "bbox": [ 107, 582, 505, 648 ], "lines": [ { "bbox": [ 105, 581, 505, 594 ], "spans": [ { "bbox": [ 105, 581, 505, 594 ], "score": 1.0, "content": "ImageNet classification. Specifically, we remove its learnable PE and apply our PEG in position", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 592, 505, 605 ], "spans": [ { "bbox": [ 105, 592, 505, 605 ], "score": 1.0, "content": "0 of each stage with a GAP head. We use the same training settings to make a fair comparison and", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 602, 506, 618 ], "spans": [ { "bbox": [ 105, 602, 422, 618 ], "score": 1.0, "content": "show the results in Table 13. Our method can significantly boost PVT-tiny by", "type": "text" }, { "bbox": [ 423, 604, 444, 614 ], "score": 0.85, "content": "3 . 1 \\%", "type": "inline_equation" }, { "bbox": [ 445, 602, 506, 618 ], "score": 1.0, "content": "and Swin-tiny", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 614, 506, 627 ], "spans": [ { "bbox": [ 105, 614, 119, 627 ], "score": 1.0, "content": "by", "type": "text" }, { "bbox": [ 120, 615, 146, 626 ], "score": 0.86, "content": "1 . 1 5 \\%", "type": "inline_equation" }, { "bbox": [ 147, 614, 506, 627 ], "score": 1.0, "content": "on ImageNet (c.f. B.5). We also evaluate the performance of PEG on some downstream", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 626, 505, 638 ], "spans": [ { "bbox": [ 105, 626, 505, 638 ], "score": 1.0, "content": "semantic segmentation and object detection tasks (see B.6). Note these tasks usually handle the", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 637, 503, 649 ], "spans": [ { "bbox": [ 106, 637, 503, 649 ], "score": 1.0, "content": "various input resolutions as the training because multi-scale data augmentation is extensively used.", "type": "text" } ], "index": 20 } ], "index": 17.5 }, { "type": "title", "bbox": [ 108, 664, 218, 677 ], "lines": [ { "bbox": [ 105, 662, 221, 679 ], "spans": [ { "bbox": [ 105, 662, 221, 679 ], "score": 1.0, "content": "5 ABLATION STUDY", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "title", "bbox": [ 107, 689, 334, 700 ], "lines": [ { "bbox": [ 106, 689, 335, 701 ], "spans": [ { "bbox": [ 106, 689, 335, 701 ], "score": 1.0, "content": "5.1 POSITIONAL ENCODING OR MERELY A HYBRID?", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 107, 709, 504, 732 ], "lines": [ { "bbox": [ 105, 709, 504, 722 ], "spans": [ { "bbox": [ 105, 709, 504, 722 ], "score": 1.0, "content": "One might suspect that the PEG’s improvement comes from the extra learnable parameters intro-", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 719, 506, 734 ], "spans": [ { "bbox": [ 105, 719, 506, 734 ], "score": 1.0, "content": "duced by the convolutional layers in PEG, instead of the local relationship retained by PEG. One", "type": "text" } ], "index": 24 } ], "index": 23.5 } ], "page_idx": 6, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 303, 751, 309, 759 ], "lines": [ { "bbox": [ 302, 750, 309, 762 ], "spans": [ { "bbox": [ 302, 750, 309, 762 ], "score": 1.0, "content": "7", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 120, 112, 491, 416 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 80, 502, 103 ], "group_id": 0, "lines": [ { "bbox": [ 106, 79, 504, 93 ], "spans": [ { "bbox": [ 106, 79, 504, 93 ], "score": 1.0, "content": "Table 4. Comparison with ConvNets and Transformers on ImageNet and ImageNet Real (Beyer", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 90, 452, 104 ], "spans": [ { "bbox": [ 106, 90, 452, 104 ], "score": 1.0, "content": "et al., 2020). CPVT have much better performance compared with prior Transformers", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table_body", "bbox": [ 120, 112, 491, 416 ], "group_id": 0, "lines": [ { "bbox": [ 120, 112, 491, 416 ], "spans": [ { "bbox": [ 120, 112, 491, 416 ], "score": 0.986, "html": "
ModelsParams(M) InputInput|throughput*ImNettop-1 %Realtop-1 %
ResNet-50 (He et al., 2016)ResNet-101 (He et al., 2016)ResNet-152 (He et al.,2016)RegNetY-4GF (Radosavovic et al.,2020)EfficientNet-BO (Tan &Le,2019)EfficientNet-B1 (Tan&Le,2019)EfficientNet-B2 (Tan &Le,2019)EfficientNet-B3(Tan&Le,2019)EfficientNet-B4 (Tan& Le,2019)2545602122421226.176.282.5
224222422242753.6526.41156.7753.677.483.7
78.380.084.186.4
5891219522422694.377.183.5
240²1662.51255.7732.179.184.9
260²300280.181.685.986.8
3802349.482.988.0
ViT-B/16 (Dosovitskiy et al., 2021)ViT-L/1686307384285.927.377.976.511
3842
DeiT-tiny w/o PE (Touvron et al., 2020)DeiT-tiny (Touvron et al.,2020)DeiT-tiny (sine)CPVT-Ti tCPVT-Ti-GAP‡66666224222422242224²22422536.52536.52536.52500.72520.168.272.272.373.474.9180.180.381.382.5
DeiT-tiny (Touvron et al., 2020)CPVT-Tim66224222422536.52500.774.575.982.183.0
DeiT-small (Touvron et al., 2020)CPVT-S $CPVT-S-GAP‡22232322422242224²940.4930.5942.379.980.585.786.086.6
81.5
DeiT-base (Touvron et al.,2020)CPVT-B ‡CPVT-B-GAP‡862242292.381.886.7
888822422242285.5290.282.387.087.7
82.7
", "type": "table", "image_path": "d3a58c8cebdf84cdf2503d3f0b23c663985798b43d4e76a1b1f0608cbe091018.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 120, 112, 491, 213.33333333333331 ], "spans": [], "index": 2 }, { "bbox": [ 120, 213.33333333333331, 491, 314.66666666666663 ], "spans": [], "index": 3 }, { "bbox": [ 120, 314.66666666666663, 491, 415.99999999999994 ], "spans": [], "index": 4 } ] }, { "type": "table_footnote", "bbox": [ 135, 418, 398, 449 ], "group_id": 0, "lines": [ { "bbox": [ 135, 416, 399, 429 ], "spans": [ { "bbox": [ 135, 416, 399, 429 ], "score": 1.0, "content": "?: Measured in img/s on a 16GB V100 GPU as in (Touvron et al., 2020).", "type": "text" } ], "index": 5 }, { "bbox": [ 134, 426, 374, 439 ], "spans": [ { "bbox": [ 134, 426, 374, 439 ], "score": 1.0, "content": "‡: Insert one PEG each after the first encoder till the fifth encoder", "type": "text" } ], "index": 6 }, { "bbox": [ 135, 436, 378, 450 ], "spans": [ { "bbox": [ 135, 436, 378, 450 ], "score": 1.0, "content": "⚗ : trained with hard distillation using RegNetY-160 as the teacher.", "type": "text" } ], "index": 7 } ], "index": 6 } ], "index": 3 }, { "type": "text", "bbox": [ 108, 468, 504, 502 ], "lines": [ { "bbox": [ 106, 468, 506, 482 ], "spans": [ { "bbox": [ 106, 468, 506, 482 ], "score": 1.0, "content": "We further train CPVT-Ti and DeiT-tiny using the aforementioned training settings plus the hard", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 479, 506, 493 ], "spans": [ { "bbox": [ 106, 479, 506, 493 ], "score": 1.0, "content": "distillation proposed in (Touvron et al., 2020). Specifically, we use RegNetY-160 (Radosavovic", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 491, 428, 503 ], "spans": [ { "bbox": [ 106, 491, 275, 503 ], "score": 1.0, "content": "et al., 2020) as the teacher. CPVT obtains", "type": "text" }, { "bbox": [ 275, 491, 302, 501 ], "score": 0.87, "content": "7 5 . 9 \\%", "type": "inline_equation" }, { "bbox": [ 302, 491, 402, 503 ], "score": 1.0, "content": ", exceeding DeiT-tiny by", "type": "text" }, { "bbox": [ 402, 491, 424, 502 ], "score": 0.84, "content": "1 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 424, 491, 428, 503 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 10 } ], "index": 9, "bbox_fs": [ 106, 468, 506, 503 ] }, { "type": "title", "bbox": [ 107, 515, 354, 527 ], "lines": [ { "bbox": [ 105, 514, 356, 528 ], "spans": [ { "bbox": [ 105, 514, 356, 528 ], "score": 1.0, "content": "4.6 PEG ON PYRAMID TRANSFORMER ARCHITECTURES", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 107, 536, 505, 569 ], "lines": [ { "bbox": [ 105, 535, 505, 549 ], "spans": [ { "bbox": [ 105, 535, 505, 549 ], "score": 1.0, "content": "PVT (Wang et al., 2021) is a vision transformer with the multi-stage design like ResNet (He et al.,", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 545, 505, 561 ], "spans": [ { "bbox": [ 105, 545, 505, 561 ], "score": 1.0, "content": "2016). Swin (Liu et al., 2021) is a follow-up work and comes with higher performance. We apply", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 557, 350, 572 ], "spans": [ { "bbox": [ 105, 557, 350, 572 ], "score": 1.0, "content": "our method on both to demonstrate its generalization ability.", "type": "text" } ], "index": 14 } ], "index": 13, "bbox_fs": [ 105, 535, 505, 572 ] }, { "type": "text", "bbox": [ 107, 582, 505, 648 ], "lines": [ { "bbox": [ 105, 581, 505, 594 ], "spans": [ { "bbox": [ 105, 581, 505, 594 ], "score": 1.0, "content": "ImageNet classification. Specifically, we remove its learnable PE and apply our PEG in position", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 592, 505, 605 ], "spans": [ { "bbox": [ 105, 592, 505, 605 ], "score": 1.0, "content": "0 of each stage with a GAP head. We use the same training settings to make a fair comparison and", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 602, 506, 618 ], "spans": [ { "bbox": [ 105, 602, 422, 618 ], "score": 1.0, "content": "show the results in Table 13. Our method can significantly boost PVT-tiny by", "type": "text" }, { "bbox": [ 423, 604, 444, 614 ], "score": 0.85, "content": "3 . 1 \\%", "type": "inline_equation" }, { "bbox": [ 445, 602, 506, 618 ], "score": 1.0, "content": "and Swin-tiny", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 614, 506, 627 ], "spans": [ { "bbox": [ 105, 614, 119, 627 ], "score": 1.0, "content": "by", "type": "text" }, { "bbox": [ 120, 615, 146, 626 ], "score": 0.86, "content": "1 . 1 5 \\%", "type": "inline_equation" }, { "bbox": [ 147, 614, 506, 627 ], "score": 1.0, "content": "on ImageNet (c.f. B.5). We also evaluate the performance of PEG on some downstream", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 626, 505, 638 ], "spans": [ { "bbox": [ 105, 626, 505, 638 ], "score": 1.0, "content": "semantic segmentation and object detection tasks (see B.6). Note these tasks usually handle the", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 637, 503, 649 ], "spans": [ { "bbox": [ 106, 637, 503, 649 ], "score": 1.0, "content": "various input resolutions as the training because multi-scale data augmentation is extensively used.", "type": "text" } ], "index": 20 } ], "index": 17.5, "bbox_fs": [ 105, 581, 506, 649 ] }, { "type": "title", "bbox": [ 108, 664, 218, 677 ], "lines": [ { "bbox": [ 105, 662, 221, 679 ], "spans": [ { "bbox": [ 105, 662, 221, 679 ], "score": 1.0, "content": "5 ABLATION STUDY", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "title", "bbox": [ 107, 689, 334, 700 ], "lines": [ { "bbox": [ 106, 689, 335, 701 ], "spans": [ { "bbox": [ 106, 689, 335, 701 ], "score": 1.0, "content": "5.1 POSITIONAL ENCODING OR MERELY A HYBRID?", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 107, 709, 504, 732 ], "lines": [ { "bbox": [ 105, 709, 504, 722 ], "spans": [ { "bbox": [ 105, 709, 504, 722 ], "score": 1.0, "content": "One might suspect that the PEG’s improvement comes from the extra learnable parameters intro-", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 719, 506, 734 ], "spans": [ { "bbox": [ 105, 719, 506, 734 ], "score": 1.0, "content": "duced by the convolutional layers in PEG, instead of the local relationship retained by PEG. One", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 82, 505, 96 ], "spans": [ { "bbox": [ 106, 82, 505, 96 ], "score": 1.0, "content": "way to test the function of PEG is only adding it when calculating Q and K in the attention layer,", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 106, 94, 504, 106 ], "spans": [ { "bbox": [ 106, 94, 437, 106 ], "score": 1.0, "content": "so that only the positional information of PEG is passed through. We can achieve", "type": "text", "cross_page": true }, { "bbox": [ 438, 94, 465, 104 ], "score": 0.87, "content": "7 1 . 3 \\%", "type": "inline_equation", "cross_page": true }, { "bbox": [ 465, 94, 504, 106 ], "score": 1.0, "content": "top-1 ac-", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 105, 104, 505, 117 ], "spans": [ { "bbox": [ 105, 104, 453, 117 ], "score": 1.0, "content": "curacy on ImageNet with DeiT-tiny. This is significantly better than DeiT-tiny w/o PE", "type": "text", "cross_page": true }, { "bbox": [ 454, 105, 487, 116 ], "score": 0.85, "content": "( 6 8 . 2 \\% )", "type": "inline_equation", "cross_page": true }, { "bbox": [ 487, 104, 505, 117 ], "score": 1.0, "content": "and", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 105, 114, 506, 129 ], "spans": [ { "bbox": [ 105, 114, 290, 129 ], "score": 1.0, "content": "is similar to the one with PEG on Q, K and V", "type": "text", "cross_page": true }, { "bbox": [ 290, 115, 324, 127 ], "score": 0.81, "content": "( 7 2 . 4 \\% )", "type": "inline_equation", "cross_page": true }, { "bbox": [ 324, 114, 506, 129 ], "score": 1.0, "content": ", which suggests that PEG mainly serves as a", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 105, 127, 222, 140 ], "spans": [ { "bbox": [ 105, 127, 222, 140 ], "score": 1.0, "content": "positional encoding scheme.", "type": "text", "cross_page": true } ], "index": 4 } ], "index": 23.5, "bbox_fs": [ 105, 709, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 106, 82, 505, 138 ], "lines": [ { "bbox": [ 106, 82, 505, 96 ], "spans": [ { "bbox": [ 106, 82, 505, 96 ], "score": 1.0, "content": "way to test the function of PEG is only adding it when calculating Q and K in the attention layer,", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 94, 504, 106 ], "spans": [ { "bbox": [ 106, 94, 437, 106 ], "score": 1.0, "content": "so that only the positional information of PEG is passed through. We can achieve", "type": "text" }, { "bbox": [ 438, 94, 465, 104 ], "score": 0.87, "content": "7 1 . 3 \\%", "type": "inline_equation" }, { "bbox": [ 465, 94, 504, 106 ], "score": 1.0, "content": "top-1 ac-", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 104, 505, 117 ], "spans": [ { "bbox": [ 105, 104, 453, 117 ], "score": 1.0, "content": "curacy on ImageNet with DeiT-tiny. This is significantly better than DeiT-tiny w/o PE", "type": "text" }, { "bbox": [ 454, 105, 487, 116 ], "score": 0.85, "content": "( 6 8 . 2 \\% )", "type": "inline_equation" }, { "bbox": [ 487, 104, 505, 117 ], "score": 1.0, "content": "and", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 114, 506, 129 ], "spans": [ { "bbox": [ 105, 114, 290, 129 ], "score": 1.0, "content": "is similar to the one with PEG on Q, K and V", "type": "text" }, { "bbox": [ 290, 115, 324, 127 ], "score": 0.81, "content": "( 7 2 . 4 \\% )", "type": "inline_equation" }, { "bbox": [ 324, 114, 506, 129 ], "score": 1.0, "content": ", which suggests that PEG mainly serves as a", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 127, 222, 140 ], "spans": [ { "bbox": [ 105, 127, 222, 140 ], "score": 1.0, "content": "positional encoding scheme.", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "text", "bbox": [ 107, 144, 296, 275 ], "lines": [ { "bbox": [ 106, 143, 297, 155 ], "spans": [ { "bbox": [ 106, 143, 297, 155 ], "score": 1.0, "content": "We also design another experiment to remove", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 154, 296, 167 ], "spans": [ { "bbox": [ 106, 154, 276, 167 ], "score": 1.0, "content": "this concern. By randomly-initializing a", "type": "text" }, { "bbox": [ 276, 154, 296, 165 ], "score": 0.86, "content": "3 \\times 3", "type": "inline_equation" } ], "index": 6 }, { "bbox": [ 105, 164, 296, 178 ], "spans": [ { "bbox": [ 105, 164, 296, 178 ], "score": 1.0, "content": "PEG and fixing its weights during the train-", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 176, 297, 189 ], "spans": [ { "bbox": [ 105, 176, 186, 189 ], "score": 1.0, "content": "ing, we can obtain", "type": "text" }, { "bbox": [ 186, 176, 214, 187 ], "score": 0.87, "content": "7 1 . 3 \\%", "type": "inline_equation" }, { "bbox": [ 214, 176, 297, 189 ], "score": 1.0, "content": "accuracy (Table 5),", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 187, 297, 199 ], "spans": [ { "bbox": [ 105, 187, 195, 199 ], "score": 1.0, "content": "which is much higher", "type": "text" }, { "bbox": [ 196, 187, 228, 199 ], "score": 0.88, "content": "( 3 . 1 \\% \\uparrow )", "type": "inline_equation" }, { "bbox": [ 228, 187, 297, 199 ], "score": 1.0, "content": "than DeiT with-", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 198, 297, 209 ], "spans": [ { "bbox": [ 106, 198, 153, 209 ], "score": 1.0, "content": "out any PE", "type": "text" }, { "bbox": [ 154, 199, 186, 209 ], "score": 0.87, "content": "( 6 8 . 2 \\% )", "type": "inline_equation" }, { "bbox": [ 186, 198, 297, 209 ], "score": 1.0, "content": ". Since the weights of PEG", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 209, 297, 221 ], "spans": [ { "bbox": [ 106, 209, 297, 221 ], "score": 1.0, "content": "are fixed and the performance improvement can", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 220, 297, 232 ], "spans": [ { "bbox": [ 106, 220, 297, 232 ], "score": 1.0, "content": "only be due to the introduced position informa-", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 230, 296, 243 ], "spans": [ { "bbox": [ 106, 230, 296, 243 ], "score": 1.0, "content": "tion. On the contrary, when we exhaustively", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 241, 297, 254 ], "spans": [ { "bbox": [ 106, 241, 297, 254 ], "score": 1.0, "content": "use 12 convolutional layers (kernel size being", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 252, 297, 266 ], "spans": [ { "bbox": [ 106, 252, 297, 266 ], "score": 1.0, "content": "1, i.e., not producing local relationship) to re-", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 263, 297, 276 ], "spans": [ { "bbox": [ 106, 263, 297, 276 ], "score": 1.0, "content": "place the PEG, these layers have much more", "type": "text" } ], "index": 16 } ], "index": 10.5 }, { "type": "table", "bbox": [ 304, 179, 510, 259 ], "blocks": [ { "type": "table_caption", "bbox": [ 304, 141, 504, 165 ], "group_id": 0, "lines": [ { "bbox": [ 303, 140, 504, 154 ], "spans": [ { "bbox": [ 303, 140, 504, 154 ], "score": 1.0, "content": "Table 5. Positional encoding rather than added pa-", "type": "text" } ], "index": 17 }, { "bbox": [ 303, 154, 456, 165 ], "spans": [ { "bbox": [ 303, 154, 456, 165 ], "score": 1.0, "content": "rameters gives the most improvement", "type": "text" } ], "index": 18 } ], "index": 17.5 }, { "type": "table_body", "bbox": [ 304, 179, 510, 259 ], "group_id": 0, "lines": [ { "bbox": [ 304, 179, 510, 259 ], "spans": [ { "bbox": [ 304, 179, 510, 259 ], "score": 0.976, "html": "
KernelStyleParams (M)Top-1 Acc (%)
none 35.6868.2
3fixed (random init)5.6871.3
1(12 ×)fixed (learned init)5.6872.3
learnable6.1368.6
3learnable5.6872.4
", "type": "table", "image_path": "11790728028cd9dbe8edf5d3fdf3853d3c301a8b464af0a0b87bab1c3bc3cc46.jpg" } ] } ], "index": 21.5, "virtual_lines": [ { "bbox": [ 304, 179, 510, 192.33333333333334 ], "spans": [], "index": 19 }, { "bbox": [ 304, 192.33333333333334, 510, 205.66666666666669 ], "spans": [], "index": 20 }, { "bbox": [ 304, 205.66666666666669, 510, 219.00000000000003 ], "spans": [], "index": 21 }, { "bbox": [ 304, 219.00000000000003, 510, 232.33333333333337 ], "spans": [], "index": 22 }, { "bbox": [ 304, 232.33333333333337, 510, 245.6666666666667 ], "spans": [], "index": 23 }, { "bbox": [ 304, 245.6666666666667, 510, 259.00000000000006 ], "spans": [], "index": 24 } ] }, { "type": "table_footnote", "bbox": [ 102, 275, 475, 286 ], "group_id": 0, "lines": [ { "bbox": [ 106, 273, 472, 288 ], "spans": [ { "bbox": [ 106, 273, 411, 288 ], "score": 1.0, "content": "learnable parameters than PEG. However, it only boosts the performance by", "type": "text" }, { "bbox": [ 412, 275, 434, 286 ], "score": 0.86, "content": "0 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 434, 273, 445, 288 ], "score": 1.0, "content": "to", "type": "text" }, { "bbox": [ 445, 275, 472, 285 ], "score": 0.88, "content": "6 8 . 6 \\%", "type": "inline_equation" } ], "index": 25 } ], "index": 25 } ], "index": 21.5 }, { "type": "text", "bbox": [ 107, 291, 505, 336 ], "lines": [ { "bbox": [ 105, 291, 505, 304 ], "spans": [ { "bbox": [ 105, 291, 505, 304 ], "score": 1.0, "content": "Another interesting finding is that fixing a learned PEG also helps training. When we initialize with", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 303, 505, 315 ], "spans": [ { "bbox": [ 105, 303, 505, 315 ], "score": 1.0, "content": "a learned PEG instead of the random values and train the tiny version of the model from scratch", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 313, 505, 326 ], "spans": [ { "bbox": [ 105, 313, 336, 326 ], "score": 1.0, "content": "while keeping the PEG fixed, the model can also achieve", "type": "text" }, { "bbox": [ 337, 314, 364, 324 ], "score": 0.87, "content": "7 2 . 3 \\%", "type": "inline_equation" }, { "bbox": [ 364, 313, 505, 326 ], "score": 1.0, "content": "top-1 accuracy on ImageNet. This", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 325, 281, 337 ], "spans": [ { "bbox": [ 105, 325, 244, 337 ], "score": 1.0, "content": "is very close to the learnable PEG", "type": "text" }, { "bbox": [ 245, 325, 277, 336 ], "score": 0.86, "content": "( 7 2 . 4 \\% )", "type": "inline_equation" }, { "bbox": [ 277, 325, 281, 337 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 29 } ], "index": 27.5 }, { "type": "title", "bbox": [ 107, 352, 237, 363 ], "lines": [ { "bbox": [ 105, 351, 239, 365 ], "spans": [ { "bbox": [ 105, 351, 239, 365 ], "score": 1.0, "content": "5.2 PEG POSITION IN CPVT", "type": "text" } ], "index": 30 } ], "index": 30 }, { "type": "text", "bbox": [ 107, 374, 505, 418 ], "lines": [ { "bbox": [ 105, 374, 505, 387 ], "spans": [ { "bbox": [ 105, 374, 505, 387 ], "score": 1.0, "content": "We also experiment by varying the position of the PEG in the model. Table 6 (left) presents the", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 384, 505, 397 ], "spans": [ { "bbox": [ 105, 384, 505, 397 ], "score": 1.0, "content": "ablations for variable positions (denoted as PosIdx) based on the tiny model. We consider the input", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 396, 505, 408 ], "spans": [ { "bbox": [ 105, 396, 505, 408 ], "score": 1.0, "content": "of the first encoder by index -1. Therefore, position 0 is the output of the first encoder block. PEG", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 407, 363, 420 ], "spans": [ { "bbox": [ 105, 407, 215, 420 ], "score": 1.0, "content": "shows strong performance", "type": "text" }, { "bbox": [ 216, 407, 255, 418 ], "score": 0.85, "content": "( \\sim 7 2 . 4 \\% )", "type": "inline_equation" }, { "bbox": [ 255, 407, 363, 420 ], "score": 1.0, "content": "when it is placed at [0, 3].", "type": "text" } ], "index": 34 } ], "index": 32.5 }, { "type": "text", "bbox": [ 106, 423, 505, 501 ], "lines": [ { "bbox": [ 105, 423, 505, 437 ], "spans": [ { "bbox": [ 105, 423, 505, 437 ], "score": 1.0, "content": "Note that positioning the PEG at 0 can have much better performance than positioning it at -1 (i.e.,", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 435, 505, 446 ], "spans": [ { "bbox": [ 106, 435, 505, 446 ], "score": 1.0, "content": "before the first encoder), as shown in Table 6 (left). We observe that the difference between the two", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 446, 505, 458 ], "spans": [ { "bbox": [ 106, 446, 505, 458 ], "score": 1.0, "content": "situations is they have different receptive fields. Specifically, the former has a global field while the", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 457, 505, 470 ], "spans": [ { "bbox": [ 105, 457, 505, 470 ], "score": 1.0, "content": "latter can only see a local area. Hence, they are supposed to work similarly well if we enlarge the", "type": "text" } ], "index": 38 }, { "bbox": [ 104, 466, 505, 482 ], "spans": [ { "bbox": [ 104, 466, 505, 482 ], "score": 1.0, "content": "convolution’s kernel size. To verify our hypothesis, we use a quite large kernel size 27 with a padding", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 479, 506, 491 ], "spans": [ { "bbox": [ 105, 479, 506, 491 ], "score": 1.0, "content": "size 13 at position -1, whose result is reported in Table 6 (right). It achieves similar performance to", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 490, 400, 502 ], "spans": [ { "bbox": [ 106, 490, 239, 502 ], "score": 1.0, "content": "the one positioning the PEG at 0", "type": "text" }, { "bbox": [ 240, 490, 271, 501 ], "score": 0.8, "content": "( 7 2 . 5 \\% )", "type": "inline_equation" }, { "bbox": [ 272, 490, 400, 502 ], "score": 1.0, "content": ", which verifies our assumption.", "type": "text" } ], "index": 41 } ], "index": 38 }, { "type": "table", "bbox": [ 115, 540, 262, 619 ], "blocks": [ { "type": "table_caption", "bbox": [ 123, 513, 487, 526 ], "group_id": 1, "lines": [ { "bbox": [ 122, 512, 487, 528 ], "spans": [ { "bbox": [ 122, 512, 487, 528 ], "score": 1.0, "content": "Table 6. Comparison of different plugin positions (left) and kernels (right) using DeiT-tiny", "type": "text" } ], "index": 42 } ], "index": 42 }, { "type": "table_body", "bbox": [ 115, 540, 262, 619 ], "group_id": 1, "lines": [ { "bbox": [ 115, 540, 262, 619 ], "spans": [ { "bbox": [ 115, 540, 262, 619 ], "score": 0.958, "html": "
PosIdxTop-1 (%)Top-5 (%)
none -168.288.7
0 370.690.2
72.491.2
72.391.1
71.790.8
6 1069.089.1
", "type": "table", "image_path": "0481d99dee42f3e65399bd2ca0e28796db6647911c0f5a0ee152c022cd181da9.jpg" } ] } ], "index": 44.0, "virtual_lines": [ { "bbox": [ 115, 540, 262, 579.5 ], "spans": [], "index": 43 }, { "bbox": [ 115, 579.5, 262, 619.0 ], "spans": [], "index": 45 } ] } ], "index": 43.0 }, { "type": "table", "bbox": [ 267, 562, 496, 597 ], "blocks": [ { "type": "table_body", "bbox": [ 267, 562, 496, 597 ], "group_id": 2, "lines": [ { "bbox": [ 267, 562, 496, 597 ], "spans": [ { "bbox": [ 267, 562, 496, 597 ], "score": 0.96, "html": "
PosIdxkernelParamsTop-1 (%)Top-5 (%)
-13×35.7M70.690.2
-127×275.8M72.591.3
", "type": "table", "image_path": "8f8db5718255214be93af4acfd1c2a152c0682df2fbe744bcae9d8859becad24.jpg" } ] } ], "index": 45.0, "virtual_lines": [ { "bbox": [ 267, 562, 496, 579.5 ], "spans": [], "index": 44 }, { "bbox": [ 267, 579.5, 496, 597.0 ], "spans": [], "index": 46 } ] } ], "index": 45.0 }, { "type": "title", "bbox": [ 106, 649, 358, 660 ], "lines": [ { "bbox": [ 105, 649, 359, 662 ], "spans": [ { "bbox": [ 105, 649, 359, 662 ], "score": 1.0, "content": "5.3 COMPARISONS WITH OTHER POSITIONAL ENCODINGS", "type": "text" } ], "index": 47 } ], "index": 47 }, { "type": "text", "bbox": [ 107, 670, 505, 704 ], "lines": [ { "bbox": [ 106, 670, 505, 684 ], "spans": [ { "bbox": [ 106, 670, 505, 684 ], "score": 1.0, "content": "We compare PEG with other commonly used encodings: absolute positional encoding (e.g. sinu-", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 681, 505, 695 ], "spans": [ { "bbox": [ 105, 681, 505, 695 ], "score": 1.0, "content": "soidal (Vaswani et al., 2017)), relative positional encoding (RPE) (Shaw et al., 2018) and learnable", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 693, 421, 704 ], "spans": [ { "bbox": [ 105, 693, 421, 704 ], "score": 1.0, "content": "encoding (LE) (Devlin et al., 2019; Radford et al., 2018), as shown in Table 7.", "type": "text" } ], "index": 50 } ], "index": 49 }, { "type": "text", "bbox": [ 107, 709, 504, 732 ], "lines": [ { "bbox": [ 106, 708, 505, 721 ], "spans": [ { "bbox": [ 106, 708, 180, 721 ], "score": 1.0, "content": "DeiT-tiny obtains", "type": "text" }, { "bbox": [ 180, 709, 208, 720 ], "score": 0.88, "content": "7 2 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 208, 708, 505, 721 ], "score": 1.0, "content": "with the learnable absolute PE. We experiment with the 2-D sinusoidal", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 720, 505, 733 ], "spans": [ { "bbox": [ 105, 720, 505, 733 ], "score": 1.0, "content": "encodings and it achieves on-par performance. For RPE, we follow (Shaw et al., 2018) and set the", "type": "text" } ], "index": 52 } ], "index": 51.5 } ], "page_idx": 7, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 308, 760 ], "lines": [ { "bbox": [ 300, 750, 309, 761 ], "spans": [ { "bbox": [ 300, 750, 309, 761 ], "score": 1.0, "content": "8", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 106, 82, 505, 138 ], "lines": [], "index": 2, "bbox_fs": [ 105, 82, 506, 140 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 144, 296, 275 ], "lines": [ { "bbox": [ 106, 143, 297, 155 ], "spans": [ { "bbox": [ 106, 143, 297, 155 ], "score": 1.0, "content": "We also design another experiment to remove", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 154, 296, 167 ], "spans": [ { "bbox": [ 106, 154, 276, 167 ], "score": 1.0, "content": "this concern. By randomly-initializing a", "type": "text" }, { "bbox": [ 276, 154, 296, 165 ], "score": 0.86, "content": "3 \\times 3", "type": "inline_equation" } ], "index": 6 }, { "bbox": [ 105, 164, 296, 178 ], "spans": [ { "bbox": [ 105, 164, 296, 178 ], "score": 1.0, "content": "PEG and fixing its weights during the train-", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 176, 297, 189 ], "spans": [ { "bbox": [ 105, 176, 186, 189 ], "score": 1.0, "content": "ing, we can obtain", "type": "text" }, { "bbox": [ 186, 176, 214, 187 ], "score": 0.87, "content": "7 1 . 3 \\%", "type": "inline_equation" }, { "bbox": [ 214, 176, 297, 189 ], "score": 1.0, "content": "accuracy (Table 5),", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 187, 297, 199 ], "spans": [ { "bbox": [ 105, 187, 195, 199 ], "score": 1.0, "content": "which is much higher", "type": "text" }, { "bbox": [ 196, 187, 228, 199 ], "score": 0.88, "content": "( 3 . 1 \\% \\uparrow )", "type": "inline_equation" }, { "bbox": [ 228, 187, 297, 199 ], "score": 1.0, "content": "than DeiT with-", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 198, 297, 209 ], "spans": [ { "bbox": [ 106, 198, 153, 209 ], "score": 1.0, "content": "out any PE", "type": "text" }, { "bbox": [ 154, 199, 186, 209 ], "score": 0.87, "content": "( 6 8 . 2 \\% )", "type": "inline_equation" }, { "bbox": [ 186, 198, 297, 209 ], "score": 1.0, "content": ". Since the weights of PEG", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 209, 297, 221 ], "spans": [ { "bbox": [ 106, 209, 297, 221 ], "score": 1.0, "content": "are fixed and the performance improvement can", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 220, 297, 232 ], "spans": [ { "bbox": [ 106, 220, 297, 232 ], "score": 1.0, "content": "only be due to the introduced position informa-", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 230, 296, 243 ], "spans": [ { "bbox": [ 106, 230, 296, 243 ], "score": 1.0, "content": "tion. On the contrary, when we exhaustively", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 241, 297, 254 ], "spans": [ { "bbox": [ 106, 241, 297, 254 ], "score": 1.0, "content": "use 12 convolutional layers (kernel size being", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 252, 297, 266 ], "spans": [ { "bbox": [ 106, 252, 297, 266 ], "score": 1.0, "content": "1, i.e., not producing local relationship) to re-", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 263, 297, 276 ], "spans": [ { "bbox": [ 106, 263, 297, 276 ], "score": 1.0, "content": "place the PEG, these layers have much more", "type": "text" } ], "index": 16 } ], "index": 10.5, "bbox_fs": [ 105, 143, 297, 276 ] }, { "type": "table", "bbox": [ 304, 179, 510, 259 ], "blocks": [ { "type": "table_caption", "bbox": [ 304, 141, 504, 165 ], "group_id": 0, "lines": [ { "bbox": [ 303, 140, 504, 154 ], "spans": [ { "bbox": [ 303, 140, 504, 154 ], "score": 1.0, "content": "Table 5. Positional encoding rather than added pa-", "type": "text" } ], "index": 17 }, { "bbox": [ 303, 154, 456, 165 ], "spans": [ { "bbox": [ 303, 154, 456, 165 ], "score": 1.0, "content": "rameters gives the most improvement", "type": "text" } ], "index": 18 } ], "index": 17.5 }, { "type": "table_body", "bbox": [ 304, 179, 510, 259 ], "group_id": 0, "lines": [ { "bbox": [ 304, 179, 510, 259 ], "spans": [ { "bbox": [ 304, 179, 510, 259 ], "score": 0.976, "html": "
KernelStyleParams (M)Top-1 Acc (%)
none 35.6868.2
3fixed (random init)5.6871.3
1(12 ×)fixed (learned init)5.6872.3
learnable6.1368.6
3learnable5.6872.4
", "type": "table", "image_path": "11790728028cd9dbe8edf5d3fdf3853d3c301a8b464af0a0b87bab1c3bc3cc46.jpg" } ] } ], "index": 21.5, "virtual_lines": [ { "bbox": [ 304, 179, 510, 192.33333333333334 ], "spans": [], "index": 19 }, { "bbox": [ 304, 192.33333333333334, 510, 205.66666666666669 ], "spans": [], "index": 20 }, { "bbox": [ 304, 205.66666666666669, 510, 219.00000000000003 ], "spans": [], "index": 21 }, { "bbox": [ 304, 219.00000000000003, 510, 232.33333333333337 ], "spans": [], "index": 22 }, { "bbox": [ 304, 232.33333333333337, 510, 245.6666666666667 ], "spans": [], "index": 23 }, { "bbox": [ 304, 245.6666666666667, 510, 259.00000000000006 ], "spans": [], "index": 24 } ] }, { "type": "table_footnote", "bbox": [ 102, 275, 475, 286 ], "group_id": 0, "lines": [ { "bbox": [ 106, 273, 472, 288 ], "spans": [ { "bbox": [ 106, 273, 411, 288 ], "score": 1.0, "content": "learnable parameters than PEG. However, it only boosts the performance by", "type": "text" }, { "bbox": [ 412, 275, 434, 286 ], "score": 0.86, "content": "0 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 434, 273, 445, 288 ], "score": 1.0, "content": "to", "type": "text" }, { "bbox": [ 445, 275, 472, 285 ], "score": 0.88, "content": "6 8 . 6 \\%", "type": "inline_equation" } ], "index": 25 } ], "index": 25 } ], "index": 21.5 }, { "type": "text", "bbox": [ 107, 291, 505, 336 ], "lines": [ { "bbox": [ 105, 291, 505, 304 ], "spans": [ { "bbox": [ 105, 291, 505, 304 ], "score": 1.0, "content": "Another interesting finding is that fixing a learned PEG also helps training. When we initialize with", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 303, 505, 315 ], "spans": [ { "bbox": [ 105, 303, 505, 315 ], "score": 1.0, "content": "a learned PEG instead of the random values and train the tiny version of the model from scratch", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 313, 505, 326 ], "spans": [ { "bbox": [ 105, 313, 336, 326 ], "score": 1.0, "content": "while keeping the PEG fixed, the model can also achieve", "type": "text" }, { "bbox": [ 337, 314, 364, 324 ], "score": 0.87, "content": "7 2 . 3 \\%", "type": "inline_equation" }, { "bbox": [ 364, 313, 505, 326 ], "score": 1.0, "content": "top-1 accuracy on ImageNet. This", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 325, 281, 337 ], "spans": [ { "bbox": [ 105, 325, 244, 337 ], "score": 1.0, "content": "is very close to the learnable PEG", "type": "text" }, { "bbox": [ 245, 325, 277, 336 ], "score": 0.86, "content": "( 7 2 . 4 \\% )", "type": "inline_equation" }, { "bbox": [ 277, 325, 281, 337 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 29 } ], "index": 27.5, "bbox_fs": [ 105, 291, 505, 337 ] }, { "type": "title", "bbox": [ 107, 352, 237, 363 ], "lines": [ { "bbox": [ 105, 351, 239, 365 ], "spans": [ { "bbox": [ 105, 351, 239, 365 ], "score": 1.0, "content": "5.2 PEG POSITION IN CPVT", "type": "text" } ], "index": 30 } ], "index": 30 }, { "type": "text", "bbox": [ 107, 374, 505, 418 ], "lines": [ { "bbox": [ 105, 374, 505, 387 ], "spans": [ { "bbox": [ 105, 374, 505, 387 ], "score": 1.0, "content": "We also experiment by varying the position of the PEG in the model. Table 6 (left) presents the", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 384, 505, 397 ], "spans": [ { "bbox": [ 105, 384, 505, 397 ], "score": 1.0, "content": "ablations for variable positions (denoted as PosIdx) based on the tiny model. We consider the input", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 396, 505, 408 ], "spans": [ { "bbox": [ 105, 396, 505, 408 ], "score": 1.0, "content": "of the first encoder by index -1. Therefore, position 0 is the output of the first encoder block. PEG", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 407, 363, 420 ], "spans": [ { "bbox": [ 105, 407, 215, 420 ], "score": 1.0, "content": "shows strong performance", "type": "text" }, { "bbox": [ 216, 407, 255, 418 ], "score": 0.85, "content": "( \\sim 7 2 . 4 \\% )", "type": "inline_equation" }, { "bbox": [ 255, 407, 363, 420 ], "score": 1.0, "content": "when it is placed at [0, 3].", "type": "text" } ], "index": 34 } ], "index": 32.5, "bbox_fs": [ 105, 374, 505, 420 ] }, { "type": "text", "bbox": [ 106, 423, 505, 501 ], "lines": [ { "bbox": [ 105, 423, 505, 437 ], "spans": [ { "bbox": [ 105, 423, 505, 437 ], "score": 1.0, "content": "Note that positioning the PEG at 0 can have much better performance than positioning it at -1 (i.e.,", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 435, 505, 446 ], "spans": [ { "bbox": [ 106, 435, 505, 446 ], "score": 1.0, "content": "before the first encoder), as shown in Table 6 (left). We observe that the difference between the two", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 446, 505, 458 ], "spans": [ { "bbox": [ 106, 446, 505, 458 ], "score": 1.0, "content": "situations is they have different receptive fields. Specifically, the former has a global field while the", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 457, 505, 470 ], "spans": [ { "bbox": [ 105, 457, 505, 470 ], "score": 1.0, "content": "latter can only see a local area. Hence, they are supposed to work similarly well if we enlarge the", "type": "text" } ], "index": 38 }, { "bbox": [ 104, 466, 505, 482 ], "spans": [ { "bbox": [ 104, 466, 505, 482 ], "score": 1.0, "content": "convolution’s kernel size. To verify our hypothesis, we use a quite large kernel size 27 with a padding", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 479, 506, 491 ], "spans": [ { "bbox": [ 105, 479, 506, 491 ], "score": 1.0, "content": "size 13 at position -1, whose result is reported in Table 6 (right). It achieves similar performance to", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 490, 400, 502 ], "spans": [ { "bbox": [ 106, 490, 239, 502 ], "score": 1.0, "content": "the one positioning the PEG at 0", "type": "text" }, { "bbox": [ 240, 490, 271, 501 ], "score": 0.8, "content": "( 7 2 . 5 \\% )", "type": "inline_equation" }, { "bbox": [ 272, 490, 400, 502 ], "score": 1.0, "content": ", which verifies our assumption.", "type": "text" } ], "index": 41 } ], "index": 38, "bbox_fs": [ 104, 423, 506, 502 ] }, { "type": "table", "bbox": [ 115, 540, 262, 619 ], "blocks": [ { "type": "table_caption", "bbox": [ 123, 513, 487, 526 ], "group_id": 1, "lines": [ { "bbox": [ 122, 512, 487, 528 ], "spans": [ { "bbox": [ 122, 512, 487, 528 ], "score": 1.0, "content": "Table 6. Comparison of different plugin positions (left) and kernels (right) using DeiT-tiny", "type": "text" } ], "index": 42 } ], "index": 42 }, { "type": "table_body", "bbox": [ 115, 540, 262, 619 ], "group_id": 1, "lines": [ { "bbox": [ 115, 540, 262, 619 ], "spans": [ { "bbox": [ 115, 540, 262, 619 ], "score": 0.958, "html": "
PosIdxTop-1 (%)Top-5 (%)
none -168.288.7
0 370.690.2
72.491.2
72.391.1
71.790.8
6 1069.089.1
", "type": "table", "image_path": "0481d99dee42f3e65399bd2ca0e28796db6647911c0f5a0ee152c022cd181da9.jpg" } ] } ], "index": 44.0, "virtual_lines": [ { "bbox": [ 115, 540, 262, 579.5 ], "spans": [], "index": 43 }, { "bbox": [ 115, 579.5, 262, 619.0 ], "spans": [], "index": 45 } ] } ], "index": 43.0 }, { "type": "table", "bbox": [ 267, 562, 496, 597 ], "blocks": [ { "type": "table_body", "bbox": [ 267, 562, 496, 597 ], "group_id": 2, "lines": [ { "bbox": [ 267, 562, 496, 597 ], "spans": [ { "bbox": [ 267, 562, 496, 597 ], "score": 0.96, "html": "
PosIdxkernelParamsTop-1 (%)Top-5 (%)
-13×35.7M70.690.2
-127×275.8M72.591.3
", "type": "table", "image_path": "8f8db5718255214be93af4acfd1c2a152c0682df2fbe744bcae9d8859becad24.jpg" } ] } ], "index": 45.0, "virtual_lines": [ { "bbox": [ 267, 562, 496, 579.5 ], "spans": [], "index": 44 }, { "bbox": [ 267, 579.5, 496, 597.0 ], "spans": [], "index": 46 } ] } ], "index": 45.0 }, { "type": "title", "bbox": [ 106, 649, 358, 660 ], "lines": [ { "bbox": [ 105, 649, 359, 662 ], "spans": [ { "bbox": [ 105, 649, 359, 662 ], "score": 1.0, "content": "5.3 COMPARISONS WITH OTHER POSITIONAL ENCODINGS", "type": "text" } ], "index": 47 } ], "index": 47 }, { "type": "text", "bbox": [ 107, 670, 505, 704 ], "lines": [ { "bbox": [ 106, 670, 505, 684 ], "spans": [ { "bbox": [ 106, 670, 505, 684 ], "score": 1.0, "content": "We compare PEG with other commonly used encodings: absolute positional encoding (e.g. sinu-", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 681, 505, 695 ], "spans": [ { "bbox": [ 105, 681, 505, 695 ], "score": 1.0, "content": "soidal (Vaswani et al., 2017)), relative positional encoding (RPE) (Shaw et al., 2018) and learnable", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 693, 421, 704 ], "spans": [ { "bbox": [ 105, 693, 421, 704 ], "score": 1.0, "content": "encoding (LE) (Devlin et al., 2019; Radford et al., 2018), as shown in Table 7.", "type": "text" } ], "index": 50 } ], "index": 49, "bbox_fs": [ 105, 670, 505, 704 ] }, { "type": "text", "bbox": [ 107, 709, 504, 732 ], "lines": [ { "bbox": [ 106, 708, 505, 721 ], "spans": [ { "bbox": [ 106, 708, 180, 721 ], "score": 1.0, "content": "DeiT-tiny obtains", "type": "text" }, { "bbox": [ 180, 709, 208, 720 ], "score": 0.88, "content": "7 2 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 208, 708, 505, 721 ], "score": 1.0, "content": "with the learnable absolute PE. We experiment with the 2-D sinusoidal", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 720, 505, 733 ], "spans": [ { "bbox": [ 105, 720, 505, 733 ], "score": 1.0, "content": "encodings and it achieves on-par performance. For RPE, we follow (Shaw et al., 2018) and set the", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 232, 505, 248 ], "spans": [ { "bbox": [ 105, 232, 222, 248 ], "score": 1.0, "content": "local range hyper-parameter", "type": "text", "cross_page": true }, { "bbox": [ 222, 235, 233, 244 ], "score": 0.82, "content": "K", "type": "inline_equation", "cross_page": true }, { "bbox": [ 233, 232, 346, 248 ], "score": 1.0, "content": "as 8, with which we obtain", "type": "text", "cross_page": true }, { "bbox": [ 346, 235, 373, 245 ], "score": 0.87, "content": "70 . 5 \\%", "type": "inline_equation", "cross_page": true }, { "bbox": [ 374, 232, 505, 248 ], "score": 1.0, "content": ". RPE here does not encode any", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 106, 245, 352, 257 ], "spans": [ { "bbox": [ 106, 245, 352, 257 ], "score": 1.0, "content": "absolute position information, see discussion in D.1 and B.3.", "type": "text", "cross_page": true } ], "index": 10 } ], "index": 51.5, "bbox_fs": [ 105, 708, 505, 733 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 192, 107, 419, 209 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 70, 502, 93 ], "group_id": 0, "lines": [ { "bbox": [ 106, 68, 504, 84 ], "spans": [ { "bbox": [ 106, 68, 504, 84 ], "score": 1.0, "content": "Table 7. Comparison of various positional encoding strategies. LE: learnable positional encoding.", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 79, 244, 95 ], "spans": [ { "bbox": [ 106, 79, 244, 95 ], "score": 1.0, "content": "RPE: relative positional encoding", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table_body", "bbox": [ 192, 107, 419, 209 ], "group_id": 0, "lines": [ { "bbox": [ 192, 107, 419, 209 ], "spans": [ { "bbox": [ 192, 107, 419, 209 ], "score": 0.98, "html": "
ModelPEG PosEncodingTop-1 (%)Top-5 (%）
DeiT-tiny (2020)LE72.2 72.391.0 91.0
DeiT-tiny DeiT-tiny2D sin-cos 2DRPE70.590.0
CPVT-Ti= 0-1PEG72.491.2
CPVT-Ti0-1PEG+LE72.991.4
CPVT-Ti0-14×PEG+LE72.9
0-591.4
CPVT-TiPEG73.491.8
", "type": "table", "image_path": "95fd6a569fa1244f963febaf9a0efdc3cfabf31651fb6fcb321724eff294eb94.jpg" } ] } ], "index": 5, "virtual_lines": [ { "bbox": [ 192, 107, 419, 121.57142857142857 ], "spans": [], "index": 2 }, { "bbox": [ 192, 121.57142857142857, 419, 136.14285714285714 ], "spans": [], "index": 3 }, { "bbox": [ 192, 136.14285714285714, 419, 150.71428571428572 ], "spans": [], "index": 4 }, { "bbox": [ 192, 150.71428571428572, 419, 165.2857142857143 ], "spans": [], "index": 5 }, { "bbox": [ 192, 165.2857142857143, 419, 179.8571428571429 ], "spans": [], "index": 6 }, { "bbox": [ 192, 179.8571428571429, 419, 194.42857142857147 ], "spans": [], "index": 7 }, { "bbox": [ 192, 194.42857142857147, 419, 209.00000000000006 ], "spans": [], "index": 8 } ] } ], "index": 2.75 }, { "type": "text", "bbox": [ 107, 234, 503, 256 ], "lines": [ { "bbox": [ 105, 232, 505, 248 ], "spans": [ { "bbox": [ 105, 232, 222, 248 ], "score": 1.0, "content": "local range hyper-parameter", "type": "text" }, { "bbox": [ 222, 235, 233, 244 ], "score": 0.82, "content": "K", "type": "inline_equation" }, { "bbox": [ 233, 232, 346, 248 ], "score": 1.0, "content": "as 8, with which we obtain", "type": "text" }, { "bbox": [ 346, 235, 373, 245 ], "score": 0.87, "content": "70 . 5 \\%", "type": "inline_equation" }, { "bbox": [ 374, 232, 505, 248 ], "score": 1.0, "content": ". RPE here does not encode any", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 245, 352, 257 ], "spans": [ { "bbox": [ 106, 245, 352, 257 ], "score": 1.0, "content": "absolute position information, see discussion in D.1 and B.3.", "type": "text" } ], "index": 10 } ], "index": 9.5 }, { "type": "text", "bbox": [ 107, 262, 504, 295 ], "lines": [ { "bbox": [ 106, 263, 505, 273 ], "spans": [ { "bbox": [ 106, 263, 505, 273 ], "score": 1.0, "content": "Moreover, we combine the learnable absolute PE with a single-layer PEG. This boosts the baseline", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 273, 506, 285 ], "spans": [ { "bbox": [ 106, 273, 181, 285 ], "score": 1.0, "content": "CPVT-Ti (0-1) by", "type": "text" }, { "bbox": [ 181, 273, 203, 284 ], "score": 0.87, "content": "0 . 5 \\%", "type": "inline_equation" }, { "bbox": [ 203, 273, 360, 285 ], "score": 1.0, "content": ". If we use 4-layer PEG, it can achieve", "type": "text" }, { "bbox": [ 361, 273, 388, 284 ], "score": 0.88, "content": "7 2 . 9 \\%", "type": "inline_equation" }, { "bbox": [ 388, 273, 506, 285 ], "score": 1.0, "content": ". If we add a PEG to each of", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 284, 486, 297 ], "spans": [ { "bbox": [ 105, 284, 245, 297 ], "score": 1.0, "content": "the first five blocks, we can obtain", "type": "text" }, { "bbox": [ 245, 284, 272, 295 ], "score": 0.86, "content": "7 3 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 272, 284, 486, 297 ], "score": 1.0, "content": ", which is better than stacking them within one block.", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "text", "bbox": [ 106, 300, 505, 389 ], "lines": [ { "bbox": [ 106, 301, 506, 313 ], "spans": [ { "bbox": [ 106, 301, 506, 313 ], "score": 1.0, "content": "CPE is not a simple combination of APE and RPE. We further compare our method with a", "type": "text" } ], "index": 14 }, { "bbox": [ 104, 310, 506, 326 ], "spans": [ { "bbox": [ 104, 310, 506, 326 ], "score": 1.0, "content": "baseline with combination of APE and RPE. Specifically, we use learnable positional encoding", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 323, 505, 336 ], "spans": [ { "bbox": [ 105, 323, 505, 336 ], "score": 1.0, "content": "(LE) as DeiT at the beginning of the model and supply 2D RPE for every transformer block. This", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 334, 505, 347 ], "spans": [ { "bbox": [ 105, 334, 172, 347 ], "score": 1.0, "content": "setting achieves", "type": "text" }, { "bbox": [ 172, 334, 199, 345 ], "score": 0.87, "content": "7 2 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 199, 334, 468, 347 ], "score": 1.0, "content": "top-1 accuracy on ImageNet, which is comparable to a single PEG", "type": "text" }, { "bbox": [ 469, 334, 501, 345 ], "score": 0.87, "content": "( 7 2 . 4 \\% )", "type": "inline_equation" }, { "bbox": [ 501, 334, 505, 347 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 345, 505, 357 ], "spans": [ { "bbox": [ 105, 345, 505, 357 ], "score": 1.0, "content": "Nevertheless, this experiment does not necessarily indicate that our CPE is a simple combination", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 356, 505, 368 ], "spans": [ { "bbox": [ 105, 356, 505, 368 ], "score": 1.0, "content": "of APE and RPE. When tested on different resolutions, this baseline cannot scale well compared to", "type": "text" } ], "index": 19 }, { "bbox": [ 104, 367, 505, 379 ], "spans": [ { "bbox": [ 104, 367, 505, 379 ], "score": 1.0, "content": "ours (Table 8). RPE is not able to adequately mitigate the performance degradation on top of LE.", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 377, 269, 390 ], "spans": [ { "bbox": [ 105, 377, 269, 390 ], "score": 1.0, "content": "This shall be seen as a major difference.", "type": "text" } ], "index": 21 } ], "index": 17.5 }, { "type": "table", "bbox": [ 123, 437, 488, 473 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 399, 504, 423 ], "group_id": 1, "lines": [ { "bbox": [ 106, 398, 504, 412 ], "spans": [ { "bbox": [ 106, 398, 504, 412 ], "score": 1.0, "content": "Table 8. Direct evaluation on other resolutions without fine-tuning. The models are trained on", "type": "text" } ], "index": 22 }, { "bbox": [ 108, 411, 414, 423 ], "spans": [ { "bbox": [ 108, 411, 146, 421 ], "score": 0.87, "content": "2 2 4 \\times 2 2 4", "type": "inline_equation" }, { "bbox": [ 147, 411, 222, 423 ], "score": 1.0, "content": ". CPE outperforms", "type": "text" }, { "bbox": [ 223, 411, 260, 421 ], "score": 0.28, "content": "\\mathrm { L E + R P E }", "type": "inline_equation" }, { "bbox": [ 261, 411, 414, 423 ], "score": 1.0, "content": "combination on untrained resolutions.", "type": "text" } ], "index": 23 } ], "index": 22.5 }, { "type": "table_body", "bbox": [ 123, 437, 488, 473 ], "group_id": 1, "lines": [ { "bbox": [ 123, 437, 488, 473 ], "spans": [ { "bbox": [ 123, 437, 488, 473 ], "score": 0.971, "html": "
ModelPositionalParams160(%)224(%)384(%)448(%)512(%)
DeiT-tiny (LE+RPE)DeiT-tiny (PEG at Pos 0)40011192065.666.872.472.470.873.268.471.865.670.3
", "type": "table", "image_path": "6b8d3b6d4e5b316d0cd1fa2df431aa65ca8ce37bb57e56b851d5cb5d959fbc3f.jpg" } ] } ], "index": 25, "virtual_lines": [ { "bbox": [ 123, 437, 488, 449.0 ], "spans": [], "index": 24 }, { "bbox": [ 123, 449.0, 488, 461.0 ], "spans": [], "index": 25 }, { "bbox": [ 123, 461.0, 488, 473.0 ], "spans": [], "index": 26 } ] } ], "index": 23.75 }, { "type": "text", "bbox": [ 107, 489, 505, 544 ], "lines": [ { "bbox": [ 105, 488, 505, 501 ], "spans": [ { "bbox": [ 105, 488, 505, 501 ], "score": 1.0, "content": "PEG can continuously improve the performance if stacked more. We use LE not only at the", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 499, 505, 514 ], "spans": [ { "bbox": [ 105, 499, 505, 514 ], "score": 1.0, "content": "beginning but also in the next 5 layers to have a similar thing as 0-5 PEG configuration.This setting", "type": "text" } ], "index": 28 }, { "bbox": [ 104, 508, 506, 525 ], "spans": [ { "bbox": [ 104, 508, 143, 525 ], "score": 1.0, "content": "achieves", "type": "text" }, { "bbox": [ 144, 511, 171, 522 ], "score": 0.87, "content": "7 2 . 7 \\%", "type": "inline_equation" }, { "bbox": [ 171, 508, 333, 525 ], "score": 1.0, "content": "top-1 accuracy on ImageNet, which is", "type": "text" }, { "bbox": [ 333, 511, 356, 522 ], "score": 0.87, "content": "0 . 7 \\%", "type": "inline_equation" }, { "bbox": [ 356, 508, 506, 525 ], "score": 1.0, "content": "lower than PEG (0-5). This setting", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 522, 506, 536 ], "spans": [ { "bbox": [ 105, 522, 506, 536 ], "score": 1.0, "content": "suggests that it is also beneficial to have more of LEs, but not as good as ours. It is expected since", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 533, 342, 545 ], "spans": [ { "bbox": [ 106, 533, 342, 545 ], "score": 1.0, "content": "we exploit relative information via PEGs at the same time.", "type": "text" } ], "index": 31 } ], "index": 29 }, { "type": "title", "bbox": [ 108, 560, 195, 573 ], "lines": [ { "bbox": [ 105, 559, 197, 576 ], "spans": [ { "bbox": [ 105, 559, 197, 576 ], "score": 1.0, "content": "6 CONCLUSION", "type": "text" } ], "index": 32 } ], "index": 32 }, { "type": "text", "bbox": [ 107, 585, 505, 674 ], "lines": [ { "bbox": [ 105, 585, 505, 599 ], "spans": [ { "bbox": [ 105, 585, 505, 599 ], "score": 1.0, "content": "We introduced CPVT, a novel method to provide the position information in vision transformers,", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 597, 505, 609 ], "spans": [ { "bbox": [ 106, 597, 505, 609 ], "score": 1.0, "content": "which dynamically generates the position encodings based on the local neighbors of each input", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 608, 505, 621 ], "spans": [ { "bbox": [ 105, 608, 505, 621 ], "score": 1.0, "content": "token. Through extensive experimental studies, we demonstrate that our proposed positional en-", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 619, 505, 631 ], "spans": [ { "bbox": [ 105, 619, 505, 631 ], "score": 1.0, "content": "codings can achieve stronger performance than the previous positional encodings. The transformer", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 630, 505, 643 ], "spans": [ { "bbox": [ 105, 630, 505, 643 ], "score": 1.0, "content": "models with our positional encodings can naturally process longer input sequences and keep the", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 641, 505, 654 ], "spans": [ { "bbox": [ 105, 641, 505, 654 ], "score": 1.0, "content": "desired translation equivalence in vision tasks. Moreover, our positional encodings are easy to im-", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 652, 505, 664 ], "spans": [ { "bbox": [ 106, 652, 505, 664 ], "score": 1.0, "content": "plement and come with negligible cost. We look forward to a broader application of our method in", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 662, 395, 676 ], "spans": [ { "bbox": [ 105, 662, 395, 676 ], "score": 1.0, "content": "transformer-driven vision tasks like segmentation and video processing.", "type": "text" } ], "index": 40 } ], "index": 36.5 }, { "type": "title", "bbox": [ 108, 691, 175, 703 ], "lines": [ { "bbox": [ 106, 690, 176, 704 ], "spans": [ { "bbox": [ 106, 690, 176, 704 ], "score": 1.0, "content": "REFERENCES", "type": "text" } ], "index": 41 } ], "index": 41 }, { "type": "text", "bbox": [ 109, 709, 504, 732 ], "lines": [ { "bbox": [ 106, 709, 505, 722 ], "spans": [ { "bbox": [ 106, 709, 505, 722 ], "score": 1.0, "content": "Mart´ın Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu", "type": "text" } ], "index": 42 }, { "bbox": [ 115, 719, 505, 734 ], "spans": [ { "bbox": [ 115, 719, 505, 734 ], "score": 1.0, "content": "Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al. Tensorflow: A system for large-", "type": "text" } ], "index": 43 } ], "index": 42.5 } ], "page_idx": 8, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 308, 759 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "9", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 192, 107, 419, 209 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 70, 502, 93 ], "group_id": 0, "lines": [ { "bbox": [ 106, 68, 504, 84 ], "spans": [ { "bbox": [ 106, 68, 504, 84 ], "score": 1.0, "content": "Table 7. Comparison of various positional encoding strategies. LE: learnable positional encoding.", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 79, 244, 95 ], "spans": [ { "bbox": [ 106, 79, 244, 95 ], "score": 1.0, "content": "RPE: relative positional encoding", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table_body", "bbox": [ 192, 107, 419, 209 ], "group_id": 0, "lines": [ { "bbox": [ 192, 107, 419, 209 ], "spans": [ { "bbox": [ 192, 107, 419, 209 ], "score": 0.98, "html": "
ModelPEG PosEncodingTop-1 (%)Top-5 (%）
DeiT-tiny (2020)LE72.2 72.391.0 91.0
DeiT-tiny DeiT-tiny2D sin-cos 2DRPE70.590.0
CPVT-Ti= 0-1PEG72.491.2
CPVT-Ti0-1PEG+LE72.991.4
CPVT-Ti0-14×PEG+LE72.9
0-591.4
CPVT-TiPEG73.491.8
", "type": "table", "image_path": "95fd6a569fa1244f963febaf9a0efdc3cfabf31651fb6fcb321724eff294eb94.jpg" } ] } ], "index": 5, "virtual_lines": [ { "bbox": [ 192, 107, 419, 121.57142857142857 ], "spans": [], "index": 2 }, { "bbox": [ 192, 121.57142857142857, 419, 136.14285714285714 ], "spans": [], "index": 3 }, { "bbox": [ 192, 136.14285714285714, 419, 150.71428571428572 ], "spans": [], "index": 4 }, { "bbox": [ 192, 150.71428571428572, 419, 165.2857142857143 ], "spans": [], "index": 5 }, { "bbox": [ 192, 165.2857142857143, 419, 179.8571428571429 ], "spans": [], "index": 6 }, { "bbox": [ 192, 179.8571428571429, 419, 194.42857142857147 ], "spans": [], "index": 7 }, { "bbox": [ 192, 194.42857142857147, 419, 209.00000000000006 ], "spans": [], "index": 8 } ] } ], "index": 2.75 }, { "type": "text", "bbox": [ 107, 234, 503, 256 ], "lines": [], "index": 9.5, "bbox_fs": [ 105, 232, 505, 257 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 262, 504, 295 ], "lines": [ { "bbox": [ 106, 263, 505, 273 ], "spans": [ { "bbox": [ 106, 263, 505, 273 ], "score": 1.0, "content": "Moreover, we combine the learnable absolute PE with a single-layer PEG. This boosts the baseline", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 273, 506, 285 ], "spans": [ { "bbox": [ 106, 273, 181, 285 ], "score": 1.0, "content": "CPVT-Ti (0-1) by", "type": "text" }, { "bbox": [ 181, 273, 203, 284 ], "score": 0.87, "content": "0 . 5 \\%", "type": "inline_equation" }, { "bbox": [ 203, 273, 360, 285 ], "score": 1.0, "content": ". If we use 4-layer PEG, it can achieve", "type": "text" }, { "bbox": [ 361, 273, 388, 284 ], "score": 0.88, "content": "7 2 . 9 \\%", "type": "inline_equation" }, { "bbox": [ 388, 273, 506, 285 ], "score": 1.0, "content": ". If we add a PEG to each of", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 284, 486, 297 ], "spans": [ { "bbox": [ 105, 284, 245, 297 ], "score": 1.0, "content": "the first five blocks, we can obtain", "type": "text" }, { "bbox": [ 245, 284, 272, 295 ], "score": 0.86, "content": "7 3 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 272, 284, 486, 297 ], "score": 1.0, "content": ", which is better than stacking them within one block.", "type": "text" } ], "index": 13 } ], "index": 12, "bbox_fs": [ 105, 263, 506, 297 ] }, { "type": "text", "bbox": [ 106, 300, 505, 389 ], "lines": [ { "bbox": [ 106, 301, 506, 313 ], "spans": [ { "bbox": [ 106, 301, 506, 313 ], "score": 1.0, "content": "CPE is not a simple combination of APE and RPE. We further compare our method with a", "type": "text" } ], "index": 14 }, { "bbox": [ 104, 310, 506, 326 ], "spans": [ { "bbox": [ 104, 310, 506, 326 ], "score": 1.0, "content": "baseline with combination of APE and RPE. Specifically, we use learnable positional encoding", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 323, 505, 336 ], "spans": [ { "bbox": [ 105, 323, 505, 336 ], "score": 1.0, "content": "(LE) as DeiT at the beginning of the model and supply 2D RPE for every transformer block. This", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 334, 505, 347 ], "spans": [ { "bbox": [ 105, 334, 172, 347 ], "score": 1.0, "content": "setting achieves", "type": "text" }, { "bbox": [ 172, 334, 199, 345 ], "score": 0.87, "content": "7 2 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 199, 334, 468, 347 ], "score": 1.0, "content": "top-1 accuracy on ImageNet, which is comparable to a single PEG", "type": "text" }, { "bbox": [ 469, 334, 501, 345 ], "score": 0.87, "content": "( 7 2 . 4 \\% )", "type": "inline_equation" }, { "bbox": [ 501, 334, 505, 347 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 345, 505, 357 ], "spans": [ { "bbox": [ 105, 345, 505, 357 ], "score": 1.0, "content": "Nevertheless, this experiment does not necessarily indicate that our CPE is a simple combination", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 356, 505, 368 ], "spans": [ { "bbox": [ 105, 356, 505, 368 ], "score": 1.0, "content": "of APE and RPE. When tested on different resolutions, this baseline cannot scale well compared to", "type": "text" } ], "index": 19 }, { "bbox": [ 104, 367, 505, 379 ], "spans": [ { "bbox": [ 104, 367, 505, 379 ], "score": 1.0, "content": "ours (Table 8). RPE is not able to adequately mitigate the performance degradation on top of LE.", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 377, 269, 390 ], "spans": [ { "bbox": [ 105, 377, 269, 390 ], "score": 1.0, "content": "This shall be seen as a major difference.", "type": "text" } ], "index": 21 } ], "index": 17.5, "bbox_fs": [ 104, 301, 506, 390 ] }, { "type": "table", "bbox": [ 123, 437, 488, 473 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 399, 504, 423 ], "group_id": 1, "lines": [ { "bbox": [ 106, 398, 504, 412 ], "spans": [ { "bbox": [ 106, 398, 504, 412 ], "score": 1.0, "content": "Table 8. Direct evaluation on other resolutions without fine-tuning. The models are trained on", "type": "text" } ], "index": 22 }, { "bbox": [ 108, 411, 414, 423 ], "spans": [ { "bbox": [ 108, 411, 146, 421 ], "score": 0.87, "content": "2 2 4 \\times 2 2 4", "type": "inline_equation" }, { "bbox": [ 147, 411, 222, 423 ], "score": 1.0, "content": ". CPE outperforms", "type": "text" }, { "bbox": [ 223, 411, 260, 421 ], "score": 0.28, "content": "\\mathrm { L E + R P E }", "type": "inline_equation" }, { "bbox": [ 261, 411, 414, 423 ], "score": 1.0, "content": "combination on untrained resolutions.", "type": "text" } ], "index": 23 } ], "index": 22.5 }, { "type": "table_body", "bbox": [ 123, 437, 488, 473 ], "group_id": 1, "lines": [ { "bbox": [ 123, 437, 488, 473 ], "spans": [ { "bbox": [ 123, 437, 488, 473 ], "score": 0.971, "html": "
ModelPositionalParams160(%)224(%)384(%)448(%)512(%)
DeiT-tiny (LE+RPE)DeiT-tiny (PEG at Pos 0)40011192065.666.872.472.470.873.268.471.865.670.3
", "type": "table", "image_path": "6b8d3b6d4e5b316d0cd1fa2df431aa65ca8ce37bb57e56b851d5cb5d959fbc3f.jpg" } ] } ], "index": 25, "virtual_lines": [ { "bbox": [ 123, 437, 488, 449.0 ], "spans": [], "index": 24 }, { "bbox": [ 123, 449.0, 488, 461.0 ], "spans": [], "index": 25 }, { "bbox": [ 123, 461.0, 488, 473.0 ], "spans": [], "index": 26 } ] } ], "index": 23.75 }, { "type": "text", "bbox": [ 107, 489, 505, 544 ], "lines": [ { "bbox": [ 105, 488, 505, 501 ], "spans": [ { "bbox": [ 105, 488, 505, 501 ], "score": 1.0, "content": "PEG can continuously improve the performance if stacked more. We use LE not only at the", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 499, 505, 514 ], "spans": [ { "bbox": [ 105, 499, 505, 514 ], "score": 1.0, "content": "beginning but also in the next 5 layers to have a similar thing as 0-5 PEG configuration.This setting", "type": "text" } ], "index": 28 }, { "bbox": [ 104, 508, 506, 525 ], "spans": [ { "bbox": [ 104, 508, 143, 525 ], "score": 1.0, "content": "achieves", "type": "text" }, { "bbox": [ 144, 511, 171, 522 ], "score": 0.87, "content": "7 2 . 7 \\%", "type": "inline_equation" }, { "bbox": [ 171, 508, 333, 525 ], "score": 1.0, "content": "top-1 accuracy on ImageNet, which is", "type": "text" }, { "bbox": [ 333, 511, 356, 522 ], "score": 0.87, "content": "0 . 7 \\%", "type": "inline_equation" }, { "bbox": [ 356, 508, 506, 525 ], "score": 1.0, "content": "lower than PEG (0-5). This setting", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 522, 506, 536 ], "spans": [ { "bbox": [ 105, 522, 506, 536 ], "score": 1.0, "content": "suggests that it is also beneficial to have more of LEs, but not as good as ours. It is expected since", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 533, 342, 545 ], "spans": [ { "bbox": [ 106, 533, 342, 545 ], "score": 1.0, "content": "we exploit relative information via PEGs at the same time.", "type": "text" } ], "index": 31 } ], "index": 29, "bbox_fs": [ 104, 488, 506, 545 ] }, { "type": "title", "bbox": [ 108, 560, 195, 573 ], "lines": [ { "bbox": [ 105, 559, 197, 576 ], "spans": [ { "bbox": [ 105, 559, 197, 576 ], "score": 1.0, "content": "6 CONCLUSION", "type": "text" } ], "index": 32 } ], "index": 32 }, { "type": "text", "bbox": [ 107, 585, 505, 674 ], "lines": [ { "bbox": [ 105, 585, 505, 599 ], "spans": [ { "bbox": [ 105, 585, 505, 599 ], "score": 1.0, "content": "We introduced CPVT, a novel method to provide the position information in vision transformers,", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 597, 505, 609 ], "spans": [ { "bbox": [ 106, 597, 505, 609 ], "score": 1.0, "content": "which dynamically generates the position encodings based on the local neighbors of each input", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 608, 505, 621 ], "spans": [ { "bbox": [ 105, 608, 505, 621 ], "score": 1.0, "content": "token. Through extensive experimental studies, we demonstrate that our proposed positional en-", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 619, 505, 631 ], "spans": [ { "bbox": [ 105, 619, 505, 631 ], "score": 1.0, "content": "codings can achieve stronger performance than the previous positional encodings. The transformer", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 630, 505, 643 ], "spans": [ { "bbox": [ 105, 630, 505, 643 ], "score": 1.0, "content": "models with our positional encodings can naturally process longer input sequences and keep the", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 641, 505, 654 ], "spans": [ { "bbox": [ 105, 641, 505, 654 ], "score": 1.0, "content": "desired translation equivalence in vision tasks. Moreover, our positional encodings are easy to im-", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 652, 505, 664 ], "spans": [ { "bbox": [ 106, 652, 505, 664 ], "score": 1.0, "content": "plement and come with negligible cost. We look forward to a broader application of our method in", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 662, 395, 676 ], "spans": [ { "bbox": [ 105, 662, 395, 676 ], "score": 1.0, "content": "transformer-driven vision tasks like segmentation and video processing.", "type": "text" } ], "index": 40 } ], "index": 36.5, "bbox_fs": [ 105, 585, 505, 676 ] }, { "type": "title", "bbox": [ 108, 691, 175, 703 ], "lines": [ { "bbox": [ 106, 690, 176, 704 ], "spans": [ { "bbox": [ 106, 690, 176, 704 ], "score": 1.0, "content": "REFERENCES", "type": "text" } ], "index": 41 } ], "index": 41 }, { "type": "text", "bbox": [ 109, 709, 504, 732 ], "lines": [ { "bbox": [ 106, 709, 505, 722 ], "spans": [ { "bbox": [ 106, 709, 505, 722 ], "score": 1.0, "content": "Mart´ın Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu", "type": "text" } ], "index": 42 }, { "bbox": [ 115, 719, 505, 734 ], "spans": [ { "bbox": [ 115, 719, 505, 734 ], "score": 1.0, "content": "Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al. Tensorflow: A system for large-", "type": "text" } ], "index": 43 }, { "bbox": [ 116, 81, 504, 95 ], "spans": [ { "bbox": [ 116, 81, 504, 95 ], "score": 1.0, "content": "scale machine learning. In 12th {USENIX} symposium on operating systems design and imple-", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 115, 93, 295, 106 ], "spans": [ { "bbox": [ 115, 93, 295, 106 ], "score": 1.0, "content": "mentation ({OSDI} 16), pp. 265–283, 2016.", "type": "text", "cross_page": true } ], "index": 1 } ], "index": 42.5, "bbox_fs": [ 106, 709, 505, 734 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 111, 82, 504, 105 ], "lines": [ { "bbox": [ 116, 81, 504, 95 ], "spans": [ { "bbox": [ 116, 81, 504, 95 ], "score": 1.0, "content": "scale machine learning. In 12th {USENIX} symposium on operating systems design and imple-", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 295, 106 ], "spans": [ { "bbox": [ 115, 93, 295, 106 ], "score": 1.0, "content": "mentation ({OSDI} 16), pp. 265–283, 2016.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 108, 112, 504, 146 ], "lines": [ { "bbox": [ 105, 111, 506, 126 ], "spans": [ { "bbox": [ 105, 111, 506, 126 ], "score": 1.0, "content": "Irwan Bello. Lambdanetworks: Modeling long-range interactions without attention. In International", "type": "text" } ], "index": 2 }, { "bbox": [ 117, 124, 505, 136 ], "spans": [ { "bbox": [ 117, 124, 505, 136 ], "score": 1.0, "content": "Conference on Learning Representations, 2021. URL https://openreview.net/forum?", "type": "text" } ], "index": 3 }, { "bbox": [ 116, 134, 206, 148 ], "spans": [ { "bbox": [ 116, 134, 206, 148 ], "score": 1.0, "content": "id=xTJEN-ggl1b.", "type": "text" } ], "index": 4 } ], "index": 3 }, { "type": "text", "bbox": [ 106, 153, 505, 187 ], "lines": [ { "bbox": [ 105, 153, 505, 167 ], "spans": [ { "bbox": [ 105, 153, 505, 167 ], "score": 1.0, "content": "Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented", "type": "text" } ], "index": 5 }, { "bbox": [ 116, 165, 505, 177 ], "spans": [ { "bbox": [ 116, 165, 505, 177 ], "score": 1.0, "content": "convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer", "type": "text" } ], "index": 6 }, { "bbox": [ 117, 175, 236, 187 ], "spans": [ { "bbox": [ 117, 175, 236, 187 ], "score": 1.0, "content": "Vision, pp. 3286–3295, 2019.", "type": "text" } ], "index": 7 } ], "index": 6 }, { "type": "text", "bbox": [ 105, 194, 504, 217 ], "lines": [ { "bbox": [ 106, 195, 505, 207 ], "spans": [ { "bbox": [ 106, 195, 505, 207 ], "score": 1.0, "content": "Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and A ´ aron van den Oord. Are ¨", "type": "text" } ], "index": 8 }, { "bbox": [ 115, 206, 380, 218 ], "spans": [ { "bbox": [ 115, 206, 380, 218 ], "score": 1.0, "content": "we done with imagenet? arXiv preprint arXiv:2006.07159, 2020.", "type": "text" } ], "index": 9 } ], "index": 8.5 }, { "type": "text", "bbox": [ 106, 224, 505, 258 ], "lines": [ { "bbox": [ 105, 224, 505, 238 ], "spans": [ { "bbox": [ 105, 224, 505, 238 ], "score": 1.0, "content": "Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal,", "type": "text" } ], "index": 10 }, { "bbox": [ 116, 236, 505, 248 ], "spans": [ { "bbox": [ 116, 236, 505, 248 ], "score": 1.0, "content": "Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are", "type": "text" } ], "index": 11 }, { "bbox": [ 116, 246, 353, 259 ], "spans": [ { "bbox": [ 116, 246, 353, 259 ], "score": 1.0, "content": "few-shot learners. arXiv preprint arXiv:2005.14165, 2020.", "type": "text" } ], "index": 12 } ], "index": 11 }, { "type": "text", "bbox": [ 107, 265, 505, 300 ], "lines": [ { "bbox": [ 106, 266, 505, 278 ], "spans": [ { "bbox": [ 106, 266, 505, 278 ], "score": 1.0, "content": "Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and", "type": "text" } ], "index": 13 }, { "bbox": [ 116, 276, 505, 290 ], "spans": [ { "bbox": [ 116, 276, 505, 290 ], "score": 1.0, "content": "Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference", "type": "text" } ], "index": 14 }, { "bbox": [ 116, 288, 319, 300 ], "spans": [ { "bbox": [ 116, 288, 319, 300 ], "score": 1.0, "content": "on Computer Vision, pp. 213–229. Springer, 2020.", "type": "text" } ], "index": 15 } ], "index": 14 }, { "type": "text", "bbox": [ 106, 306, 505, 340 ], "lines": [ { "bbox": [ 106, 306, 505, 320 ], "spans": [ { "bbox": [ 106, 306, 505, 320 ], "score": 1.0, "content": "Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chun-", "type": "text" } ], "index": 16 }, { "bbox": [ 114, 318, 506, 331 ], "spans": [ { "bbox": [ 114, 318, 506, 331 ], "score": 1.0, "content": "jing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. arXiv preprint", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 329, 219, 340 ], "spans": [ { "bbox": [ 115, 329, 219, 340 ], "score": 1.0, "content": "arXiv:2012.00364, 2020.", "type": "text" } ], "index": 18 } ], "index": 17 }, { "type": "text", "bbox": [ 106, 347, 504, 382 ], "lines": [ { "bbox": [ 106, 347, 505, 361 ], "spans": [ { "bbox": [ 106, 347, 505, 361 ], "score": 1.0, "content": "Tianqi Chen, Mu Li, Yutian Li, Min Lin, Naiyan Wang, Minjie Wang, Tianjun Xiao, Bing Xu,", "type": "text" } ], "index": 19 }, { "bbox": [ 116, 359, 505, 371 ], "spans": [ { "bbox": [ 116, 359, 505, 371 ], "score": 1.0, "content": "Chiyuan Zhang, and Zheng Zhang. Mxnet: A flexible and efficient machine learning library for", "type": "text" } ], "index": 20 }, { "bbox": [ 115, 370, 421, 383 ], "spans": [ { "bbox": [ 115, 370, 421, 383 ], "score": 1.0, "content": "heterogeneous distributed systems. arXiv preprint arXiv:1512.01274, 2015.", "type": "text" } ], "index": 21 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 388, 504, 423 ], "lines": [ { "bbox": [ 105, 388, 506, 402 ], "spans": [ { "bbox": [ 105, 388, 506, 402 ], "score": 1.0, "content": "Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 399, 506, 413 ], "spans": [ { "bbox": [ 115, 399, 506, 413 ], "score": 1.0, "content": "Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 411, 204, 423 ], "spans": [ { "bbox": [ 115, 411, 204, 423 ], "score": 1.0, "content": "NeurIPS 2021, 2021.", "type": "text" } ], "index": 24 } ], "index": 23 }, { "type": "text", "bbox": [ 107, 429, 505, 464 ], "lines": [ { "bbox": [ 105, 429, 505, 443 ], "spans": [ { "bbox": [ 105, 429, 505, 443 ], "score": 1.0, "content": "Jean-Baptiste Cordonnier, Andreas Loukas, and Martin Jaggi. On the relationship between self-", "type": "text" } ], "index": 25 }, { "bbox": [ 115, 440, 505, 454 ], "spans": [ { "bbox": [ 115, 440, 505, 454 ], "score": 1.0, "content": "attention and convolutional layers. In International Conference on Learning Representations,", "type": "text" } ], "index": 26 }, { "bbox": [ 115, 451, 142, 464 ], "spans": [ { "bbox": [ 115, 451, 142, 464 ], "score": 1.0, "content": "2020.", "type": "text" } ], "index": 27 } ], "index": 26 }, { "type": "text", "bbox": [ 106, 470, 505, 505 ], "lines": [ { "bbox": [ 106, 471, 505, 484 ], "spans": [ { "bbox": [ 106, 471, 505, 484 ], "score": 1.0, "content": "Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data", "type": "text" } ], "index": 28 }, { "bbox": [ 115, 481, 506, 496 ], "spans": [ { "bbox": [ 115, 481, 506, 496 ], "score": 1.0, "content": "augmentation with a reduced search space. Advances in Neural Information Processing Systems,", "type": "text" } ], "index": 29 }, { "bbox": [ 116, 493, 156, 505 ], "spans": [ { "bbox": [ 116, 493, 156, 505 ], "score": 1.0, "content": "33, 2020.", "type": "text" } ], "index": 30 } ], "index": 29 }, { "type": "text", "bbox": [ 105, 511, 503, 535 ], "lines": [ { "bbox": [ 106, 512, 505, 525 ], "spans": [ { "bbox": [ 106, 512, 505, 525 ], "score": 1.0, "content": "Zihang Dai, Hanxiao Liu, Quoc Le, and Mingxing Tan. Coatnet: Marrying convolution and attention", "type": "text" } ], "index": 31 }, { "bbox": [ 116, 523, 441, 536 ], "spans": [ { "bbox": [ 116, 523, 441, 536 ], "score": 1.0, "content": "for all data sizes. Advances in Neural Information Processing Systems, 34, 2021.", "type": "text" } ], "index": 32 } ], "index": 31.5 }, { "type": "text", "bbox": [ 106, 542, 505, 576 ], "lines": [ { "bbox": [ 105, 542, 504, 555 ], "spans": [ { "bbox": [ 105, 542, 504, 555 ], "score": 1.0, "content": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hi-", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 553, 505, 567 ], "spans": [ { "bbox": [ 115, 553, 505, 567 ], "score": 1.0, "content": "erarchical image database. In 2009 IEEE conference on computer vision and pattern recognition,", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 565, 218, 576 ], "spans": [ { "bbox": [ 115, 565, 218, 576 ], "score": 1.0, "content": "pp. 248–255. Ieee, 2009.", "type": "text" } ], "index": 35 } ], "index": 34 }, { "type": "text", "bbox": [ 106, 583, 506, 628 ], "lines": [ { "bbox": [ 105, 583, 506, 596 ], "spans": [ { "bbox": [ 105, 583, 506, 596 ], "score": 1.0, "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 594, 506, 608 ], "spans": [ { "bbox": [ 115, 594, 506, 608 ], "score": 1.0, "content": "bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 605, 505, 619 ], "spans": [ { "bbox": [ 115, 605, 505, 619 ], "score": 1.0, "content": "the North American Chapter of the Association for Computational Linguistics: Human Language", "type": "text" } ], "index": 38 }, { "bbox": [ 115, 616, 406, 629 ], "spans": [ { "bbox": [ 115, 616, 406, 629 ], "score": 1.0, "content": "Technologies, Volume 1 (Long and Short Papers), pp. 4171–4186, 2019.", "type": "text" } ], "index": 39 } ], "index": 37.5 }, { "type": "text", "bbox": [ 107, 635, 505, 691 ], "lines": [ { "bbox": [ 106, 635, 505, 647 ], "spans": [ { "bbox": [ 106, 635, 505, 647 ], "score": 1.0, "content": "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas", "type": "text" } ], "index": 40 }, { "bbox": [ 117, 647, 504, 659 ], "spans": [ { "bbox": [ 117, 647, 504, 659 ], "score": 1.0, "content": "Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszko-", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 656, 505, 671 ], "spans": [ { "bbox": [ 115, 656, 505, 671 ], "score": 1.0, "content": "reit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recogni-", "type": "text" } ], "index": 42 }, { "bbox": [ 115, 668, 505, 682 ], "spans": [ { "bbox": [ 115, 668, 505, 682 ], "score": 1.0, "content": "tion at scale. In International Conference on Learning Representations, 2021. URL https:", "type": "text" } ], "index": 43 }, { "bbox": [ 116, 679, 336, 693 ], "spans": [ { "bbox": [ 116, 679, 266, 693 ], "score": 1.0, "content": "//openreview.net/forum?id", "type": "text" }, { "bbox": [ 267, 681, 273, 689 ], "score": 0.38, "content": "=", "type": "inline_equation" }, { "bbox": [ 273, 679, 336, 693 ], "score": 1.0, "content": "YicbFdNTTy.", "type": "text" } ], "index": 44 } ], "index": 42 }, { "type": "text", "bbox": [ 108, 699, 504, 732 ], "lines": [ { "bbox": [ 106, 698, 506, 712 ], "spans": [ { "bbox": [ 106, 698, 506, 712 ], "score": 1.0, "content": "Stephane d’Ascoli, Hugo Touvron, Matthew L Leavitt, Ari S Morcos, Giulio Biroli, and Levent ´", "type": "text" } ], "index": 45 }, { "bbox": [ 116, 709, 506, 723 ], "spans": [ { "bbox": [ 116, 709, 506, 723 ], "score": 1.0, "content": "Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In Inter-", "type": "text" } ], "index": 46 }, { "bbox": [ 115, 721, 413, 732 ], "spans": [ { "bbox": [ 115, 721, 413, 732 ], "score": 1.0, "content": "national Conference on Machine Learning, pp. 2286–2296. PMLR, 2021.", "type": "text" } ], "index": 47 } ], "index": 46 } ], "page_idx": 9, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "10", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 111, 82, 504, 105 ], "lines": [], "index": 0.5, "bbox_fs": [ 115, 81, 504, 106 ], "lines_deleted": true }, { "type": "text", "bbox": [ 108, 112, 504, 146 ], "lines": [ { "bbox": [ 105, 111, 506, 126 ], "spans": [ { "bbox": [ 105, 111, 506, 126 ], "score": 1.0, "content": "Irwan Bello. Lambdanetworks: Modeling long-range interactions without attention. In International", "type": "text" } ], "index": 2 }, { "bbox": [ 117, 124, 505, 136 ], "spans": [ { "bbox": [ 117, 124, 505, 136 ], "score": 1.0, "content": "Conference on Learning Representations, 2021. URL https://openreview.net/forum?", "type": "text" } ], "index": 3 }, { "bbox": [ 116, 134, 206, 148 ], "spans": [ { "bbox": [ 116, 134, 206, 148 ], "score": 1.0, "content": "id=xTJEN-ggl1b.", "type": "text" } ], "index": 4 } ], "index": 3, "bbox_fs": [ 105, 111, 506, 148 ] }, { "type": "text", "bbox": [ 106, 153, 505, 187 ], "lines": [ { "bbox": [ 105, 153, 505, 167 ], "spans": [ { "bbox": [ 105, 153, 505, 167 ], "score": 1.0, "content": "Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented", "type": "text" } ], "index": 5 }, { "bbox": [ 116, 165, 505, 177 ], "spans": [ { "bbox": [ 116, 165, 505, 177 ], "score": 1.0, "content": "convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer", "type": "text" } ], "index": 6 }, { "bbox": [ 117, 175, 236, 187 ], "spans": [ { "bbox": [ 117, 175, 236, 187 ], "score": 1.0, "content": "Vision, pp. 3286–3295, 2019.", "type": "text" } ], "index": 7 } ], "index": 6, "bbox_fs": [ 105, 153, 505, 187 ] }, { "type": "text", "bbox": [ 105, 194, 504, 217 ], "lines": [ { "bbox": [ 106, 195, 505, 207 ], "spans": [ { "bbox": [ 106, 195, 505, 207 ], "score": 1.0, "content": "Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and A ´ aron van den Oord. Are ¨", "type": "text" } ], "index": 8 }, { "bbox": [ 115, 206, 380, 218 ], "spans": [ { "bbox": [ 115, 206, 380, 218 ], "score": 1.0, "content": "we done with imagenet? arXiv preprint arXiv:2006.07159, 2020.", "type": "text" } ], "index": 9 } ], "index": 8.5, "bbox_fs": [ 106, 195, 505, 218 ] }, { "type": "text", "bbox": [ 106, 224, 505, 258 ], "lines": [ { "bbox": [ 105, 224, 505, 238 ], "spans": [ { "bbox": [ 105, 224, 505, 238 ], "score": 1.0, "content": "Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal,", "type": "text" } ], "index": 10 }, { "bbox": [ 116, 236, 505, 248 ], "spans": [ { "bbox": [ 116, 236, 505, 248 ], "score": 1.0, "content": "Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are", "type": "text" } ], "index": 11 }, { "bbox": [ 116, 246, 353, 259 ], "spans": [ { "bbox": [ 116, 246, 353, 259 ], "score": 1.0, "content": "few-shot learners. arXiv preprint arXiv:2005.14165, 2020.", "type": "text" } ], "index": 12 } ], "index": 11, "bbox_fs": [ 105, 224, 505, 259 ] }, { "type": "text", "bbox": [ 107, 265, 505, 300 ], "lines": [ { "bbox": [ 106, 266, 505, 278 ], "spans": [ { "bbox": [ 106, 266, 505, 278 ], "score": 1.0, "content": "Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and", "type": "text" } ], "index": 13 }, { "bbox": [ 116, 276, 505, 290 ], "spans": [ { "bbox": [ 116, 276, 505, 290 ], "score": 1.0, "content": "Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference", "type": "text" } ], "index": 14 }, { "bbox": [ 116, 288, 319, 300 ], "spans": [ { "bbox": [ 116, 288, 319, 300 ], "score": 1.0, "content": "on Computer Vision, pp. 213–229. Springer, 2020.", "type": "text" } ], "index": 15 } ], "index": 14, "bbox_fs": [ 106, 266, 505, 300 ] }, { "type": "text", "bbox": [ 106, 306, 505, 340 ], "lines": [ { "bbox": [ 106, 306, 505, 320 ], "spans": [ { "bbox": [ 106, 306, 505, 320 ], "score": 1.0, "content": "Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chun-", "type": "text" } ], "index": 16 }, { "bbox": [ 114, 318, 506, 331 ], "spans": [ { "bbox": [ 114, 318, 506, 331 ], "score": 1.0, "content": "jing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. arXiv preprint", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 329, 219, 340 ], "spans": [ { "bbox": [ 115, 329, 219, 340 ], "score": 1.0, "content": "arXiv:2012.00364, 2020.", "type": "text" } ], "index": 18 } ], "index": 17, "bbox_fs": [ 106, 306, 506, 340 ] }, { "type": "text", "bbox": [ 106, 347, 504, 382 ], "lines": [ { "bbox": [ 106, 347, 505, 361 ], "spans": [ { "bbox": [ 106, 347, 505, 361 ], "score": 1.0, "content": "Tianqi Chen, Mu Li, Yutian Li, Min Lin, Naiyan Wang, Minjie Wang, Tianjun Xiao, Bing Xu,", "type": "text" } ], "index": 19 }, { "bbox": [ 116, 359, 505, 371 ], "spans": [ { "bbox": [ 116, 359, 505, 371 ], "score": 1.0, "content": "Chiyuan Zhang, and Zheng Zhang. Mxnet: A flexible and efficient machine learning library for", "type": "text" } ], "index": 20 }, { "bbox": [ 115, 370, 421, 383 ], "spans": [ { "bbox": [ 115, 370, 421, 383 ], "score": 1.0, "content": "heterogeneous distributed systems. arXiv preprint arXiv:1512.01274, 2015.", "type": "text" } ], "index": 21 } ], "index": 20, "bbox_fs": [ 106, 347, 505, 383 ] }, { "type": "text", "bbox": [ 107, 388, 504, 423 ], "lines": [ { "bbox": [ 105, 388, 506, 402 ], "spans": [ { "bbox": [ 105, 388, 506, 402 ], "score": 1.0, "content": "Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 399, 506, 413 ], "spans": [ { "bbox": [ 115, 399, 506, 413 ], "score": 1.0, "content": "Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 411, 204, 423 ], "spans": [ { "bbox": [ 115, 411, 204, 423 ], "score": 1.0, "content": "NeurIPS 2021, 2021.", "type": "text" } ], "index": 24 } ], "index": 23, "bbox_fs": [ 105, 388, 506, 423 ] }, { "type": "text", "bbox": [ 107, 429, 505, 464 ], "lines": [ { "bbox": [ 105, 429, 505, 443 ], "spans": [ { "bbox": [ 105, 429, 505, 443 ], "score": 1.0, "content": "Jean-Baptiste Cordonnier, Andreas Loukas, and Martin Jaggi. On the relationship between self-", "type": "text" } ], "index": 25 }, { "bbox": [ 115, 440, 505, 454 ], "spans": [ { "bbox": [ 115, 440, 505, 454 ], "score": 1.0, "content": "attention and convolutional layers. In International Conference on Learning Representations,", "type": "text" } ], "index": 26 }, { "bbox": [ 115, 451, 142, 464 ], "spans": [ { "bbox": [ 115, 451, 142, 464 ], "score": 1.0, "content": "2020.", "type": "text" } ], "index": 27 } ], "index": 26, "bbox_fs": [ 105, 429, 505, 464 ] }, { "type": "text", "bbox": [ 106, 470, 505, 505 ], "lines": [ { "bbox": [ 106, 471, 505, 484 ], "spans": [ { "bbox": [ 106, 471, 505, 484 ], "score": 1.0, "content": "Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data", "type": "text" } ], "index": 28 }, { "bbox": [ 115, 481, 506, 496 ], "spans": [ { "bbox": [ 115, 481, 506, 496 ], "score": 1.0, "content": "augmentation with a reduced search space. Advances in Neural Information Processing Systems,", "type": "text" } ], "index": 29 }, { "bbox": [ 116, 493, 156, 505 ], "spans": [ { "bbox": [ 116, 493, 156, 505 ], "score": 1.0, "content": "33, 2020.", "type": "text" } ], "index": 30 } ], "index": 29, "bbox_fs": [ 106, 471, 506, 505 ] }, { "type": "text", "bbox": [ 105, 511, 503, 535 ], "lines": [ { "bbox": [ 106, 512, 505, 525 ], "spans": [ { "bbox": [ 106, 512, 505, 525 ], "score": 1.0, "content": "Zihang Dai, Hanxiao Liu, Quoc Le, and Mingxing Tan. Coatnet: Marrying convolution and attention", "type": "text" } ], "index": 31 }, { "bbox": [ 116, 523, 441, 536 ], "spans": [ { "bbox": [ 116, 523, 441, 536 ], "score": 1.0, "content": "for all data sizes. Advances in Neural Information Processing Systems, 34, 2021.", "type": "text" } ], "index": 32 } ], "index": 31.5, "bbox_fs": [ 106, 512, 505, 536 ] }, { "type": "text", "bbox": [ 106, 542, 505, 576 ], "lines": [ { "bbox": [ 105, 542, 504, 555 ], "spans": [ { "bbox": [ 105, 542, 504, 555 ], "score": 1.0, "content": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hi-", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 553, 505, 567 ], "spans": [ { "bbox": [ 115, 553, 505, 567 ], "score": 1.0, "content": "erarchical image database. In 2009 IEEE conference on computer vision and pattern recognition,", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 565, 218, 576 ], "spans": [ { "bbox": [ 115, 565, 218, 576 ], "score": 1.0, "content": "pp. 248–255. Ieee, 2009.", "type": "text" } ], "index": 35 } ], "index": 34, "bbox_fs": [ 105, 542, 505, 576 ] }, { "type": "text", "bbox": [ 106, 583, 506, 628 ], "lines": [ { "bbox": [ 105, 583, 506, 596 ], "spans": [ { "bbox": [ 105, 583, 506, 596 ], "score": 1.0, "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 594, 506, 608 ], "spans": [ { "bbox": [ 115, 594, 506, 608 ], "score": 1.0, "content": "bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 605, 505, 619 ], "spans": [ { "bbox": [ 115, 605, 505, 619 ], "score": 1.0, "content": "the North American Chapter of the Association for Computational Linguistics: Human Language", "type": "text" } ], "index": 38 }, { "bbox": [ 115, 616, 406, 629 ], "spans": [ { "bbox": [ 115, 616, 406, 629 ], "score": 1.0, "content": "Technologies, Volume 1 (Long and Short Papers), pp. 4171–4186, 2019.", "type": "text" } ], "index": 39 } ], "index": 37.5, "bbox_fs": [ 105, 583, 506, 629 ] }, { "type": "text", "bbox": [ 107, 635, 505, 691 ], "lines": [ { "bbox": [ 106, 635, 505, 647 ], "spans": [ { "bbox": [ 106, 635, 505, 647 ], "score": 1.0, "content": "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas", "type": "text" } ], "index": 40 }, { "bbox": [ 117, 647, 504, 659 ], "spans": [ { "bbox": [ 117, 647, 504, 659 ], "score": 1.0, "content": "Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszko-", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 656, 505, 671 ], "spans": [ { "bbox": [ 115, 656, 505, 671 ], "score": 1.0, "content": "reit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recogni-", "type": "text" } ], "index": 42 }, { "bbox": [ 115, 668, 505, 682 ], "spans": [ { "bbox": [ 115, 668, 505, 682 ], "score": 1.0, "content": "tion at scale. In International Conference on Learning Representations, 2021. URL https:", "type": "text" } ], "index": 43 }, { "bbox": [ 116, 679, 336, 693 ], "spans": [ { "bbox": [ 116, 679, 266, 693 ], "score": 1.0, "content": "//openreview.net/forum?id", "type": "text" }, { "bbox": [ 267, 681, 273, 689 ], "score": 0.38, "content": "=", "type": "inline_equation" }, { "bbox": [ 273, 679, 336, 693 ], "score": 1.0, "content": "YicbFdNTTy.", "type": "text" } ], "index": 44 } ], "index": 42, "bbox_fs": [ 106, 635, 505, 693 ] }, { "type": "text", "bbox": [ 108, 699, 504, 732 ], "lines": [ { "bbox": [ 106, 698, 506, 712 ], "spans": [ { "bbox": [ 106, 698, 506, 712 ], "score": 1.0, "content": "Stephane d’Ascoli, Hugo Touvron, Matthew L Leavitt, Ari S Morcos, Giulio Biroli, and Levent ´", "type": "text" } ], "index": 45 }, { "bbox": [ 116, 709, 506, 723 ], "spans": [ { "bbox": [ 116, 709, 506, 723 ], "score": 1.0, "content": "Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In Inter-", "type": "text" } ], "index": 46 }, { "bbox": [ 115, 721, 413, 732 ], "spans": [ { "bbox": [ 115, 721, 413, 732 ], "score": 1.0, "content": "national Conference on Machine Learning, pp. 2286–2296. PMLR, 2021.", "type": "text" } ], "index": 47 } ], "index": 46, "bbox_fs": [ 106, 698, 506, 732 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 108, 82, 504, 116 ], "lines": [ { "bbox": [ 106, 81, 506, 96 ], "spans": [ { "bbox": [ 106, 81, 506, 96 ], "score": 1.0, "content": "Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo", "type": "text" } ], "index": 0 }, { "bbox": [ 117, 94, 505, 106 ], "spans": [ { "bbox": [ 117, 94, 505, 106 ], "score": 1.0, "content": "Wang, Zhengdong Zhang, Yonghui Wu, et al. Conformer: Convolution-augmented transformer", "type": "text" } ], "index": 1 }, { "bbox": [ 115, 104, 374, 118 ], "spans": [ { "bbox": [ 115, 104, 374, 118 ], "score": 1.0, "content": "for speech recognition. arXiv preprint arXiv:2005.08100, 2020.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "text", "bbox": [ 108, 122, 503, 156 ], "lines": [ { "bbox": [ 105, 121, 505, 138 ], "spans": [ { "bbox": [ 105, 121, 505, 138 ], "score": 1.0, "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 132, 505, 148 ], "spans": [ { "bbox": [ 115, 132, 505, 148 ], "score": 1.0, "content": "nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.", "type": "text" } ], "index": 4 }, { "bbox": [ 116, 145, 181, 156 ], "spans": [ { "bbox": [ 116, 145, 181, 156 ], "score": 1.0, "content": "770–778, 2016.", "type": "text" } ], "index": 5 } ], "index": 4 }, { "type": "text", "bbox": [ 106, 163, 504, 186 ], "lines": [ { "bbox": [ 105, 162, 505, 177 ], "spans": [ { "bbox": [ 105, 162, 505, 177 ], "score": 1.0, "content": "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: Decoding-enhanced bert", "type": "text" } ], "index": 6 }, { "bbox": [ 116, 174, 393, 187 ], "spans": [ { "bbox": [ 116, 174, 393, 187 ], "score": 1.0, "content": "with disentangled attention. arXiv preprint arXiv:2006.03654, 2020.", "type": "text" } ], "index": 7 } ], "index": 6.5 }, { "type": "text", "bbox": [ 108, 192, 504, 227 ], "lines": [ { "bbox": [ 105, 192, 505, 207 ], "spans": [ { "bbox": [ 105, 192, 505, 207 ], "score": 1.0, "content": "Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Aug-", "type": "text" } ], "index": 8 }, { "bbox": [ 115, 204, 505, 217 ], "spans": [ { "bbox": [ 115, 204, 505, 217 ], "score": 1.0, "content": "ment your batch: Improving generalization through instance repetition. In Proceedings of the", "type": "text" } ], "index": 9 }, { "bbox": [ 115, 215, 484, 229 ], "spans": [ { "bbox": [ 115, 215, 484, 229 ], "score": 1.0, "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8129–8138, 2020.", "type": "text" } ], "index": 10 } ], "index": 9 }, { "type": "text", "bbox": [ 106, 233, 506, 267 ], "lines": [ { "bbox": [ 106, 233, 505, 246 ], "spans": [ { "bbox": [ 106, 233, 505, 246 ], "score": 1.0, "content": "Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Noam Shazeer, Ian Simon, Curtis", "type": "text" } ], "index": 11 }, { "bbox": [ 115, 244, 506, 258 ], "spans": [ { "bbox": [ 115, 244, 506, 258 ], "score": 1.0, "content": "Hawthorne, Andrew M Dai, Matthew D Hoffman, Monica Dinculescu, and Douglas Eck. Music", "type": "text" } ], "index": 12 }, { "bbox": [ 116, 256, 368, 268 ], "spans": [ { "bbox": [ 116, 256, 368, 268 ], "score": 1.0, "content": "transformer. In Advances in Neural Processing Systems, 2018.", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "text", "bbox": [ 106, 273, 505, 297 ], "lines": [ { "bbox": [ 107, 274, 505, 286 ], "spans": [ { "bbox": [ 107, 274, 505, 286 ], "score": 1.0, "content": "Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 285, 484, 297 ], "spans": [ { "bbox": [ 115, 285, 484, 297 ], "score": 1.0, "content": "stochastic depth. In European conference on computer vision, pp. 646–661. Springer, 2016.", "type": "text" } ], "index": 15 } ], "index": 14.5 }, { "type": "text", "bbox": [ 107, 303, 504, 327 ], "lines": [ { "bbox": [ 106, 304, 504, 316 ], "spans": [ { "bbox": [ 106, 304, 504, 316 ], "score": 1.0, "content": "Md Amirul Islam, Sen Jia, and Neil DB Bruce. How much position information do convolutional", "type": "text" } ], "index": 16 }, { "bbox": [ 116, 315, 478, 327 ], "spans": [ { "bbox": [ 116, 315, 478, 327 ], "score": 1.0, "content": "neural networks encode? In International Conference on Learning Representations, 2020.", "type": "text" } ], "index": 17 } ], "index": 16.5 }, { "type": "text", "bbox": [ 107, 332, 505, 367 ], "lines": [ { "bbox": [ 105, 332, 505, 346 ], "spans": [ { "bbox": [ 105, 332, 505, 346 ], "score": 1.0, "content": "Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollar. Panoptic feature pyramid net- ´", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 344, 505, 357 ], "spans": [ { "bbox": [ 115, 344, 505, 357 ], "score": 1.0, "content": "works. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,", "type": "text" } ], "index": 19 }, { "bbox": [ 115, 356, 206, 367 ], "spans": [ { "bbox": [ 115, 356, 206, 367 ], "score": 1.0, "content": "pp. 6399–6408, 2019.", "type": "text" } ], "index": 20 } ], "index": 19 }, { "type": "text", "bbox": [ 107, 373, 505, 407 ], "lines": [ { "bbox": [ 106, 374, 504, 386 ], "spans": [ { "bbox": [ 106, 374, 504, 386 ], "score": 1.0, "content": "Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense´", "type": "text" } ], "index": 21 }, { "bbox": [ 115, 384, 505, 399 ], "spans": [ { "bbox": [ 115, 384, 505, 399 ], "score": 1.0, "content": "object detection. In Proceedings of the IEEE international conference on computer vision, pp.", "type": "text" } ], "index": 22 }, { "bbox": [ 116, 396, 191, 407 ], "spans": [ { "bbox": [ 116, 396, 191, 407 ], "score": 1.0, "content": "2980–2988, 2017.", "type": "text" } ], "index": 23 } ], "index": 22 }, { "type": "text", "bbox": [ 107, 414, 505, 448 ], "lines": [ { "bbox": [ 105, 414, 505, 428 ], "spans": [ { "bbox": [ 105, 414, 505, 428 ], "score": 1.0, "content": "Xuanqing Liu, Hsiang-Fu Yu, Inderjit Dhillon, and Cho-Jui Hsieh. Learning to encode position for", "type": "text" } ], "index": 24 }, { "bbox": [ 116, 426, 505, 438 ], "spans": [ { "bbox": [ 116, 426, 505, 438 ], "score": 1.0, "content": "transformer with continuous dynamical model. In International Conference on Machine Learn-", "type": "text" } ], "index": 25 }, { "bbox": [ 116, 437, 255, 448 ], "spans": [ { "bbox": [ 116, 437, 255, 448 ], "score": 1.0, "content": "ing, pp. 6327–6335. PMLR, 2020.", "type": "text" } ], "index": 26 } ], "index": 25 }, { "type": "text", "bbox": [ 106, 454, 505, 488 ], "lines": [ { "bbox": [ 105, 453, 505, 468 ], "spans": [ { "bbox": [ 105, 453, 505, 468 ], "score": 1.0, "content": "Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 466, 505, 479 ], "spans": [ { "bbox": [ 115, 466, 505, 479 ], "score": 1.0, "content": "Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint", "type": "text" } ], "index": 28 }, { "bbox": [ 116, 478, 219, 487 ], "spans": [ { "bbox": [ 116, 478, 219, 487 ], "score": 1.0, "content": "arXiv:2103.14030, 2021.", "type": "text" } ], "index": 29 } ], "index": 28 }, { "type": "text", "bbox": [ 106, 495, 503, 529 ], "lines": [ { "bbox": [ 106, 496, 505, 508 ], "spans": [ { "bbox": [ 106, 496, 505, 508 ], "score": 1.0, "content": "Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Confer-", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 507, 504, 519 ], "spans": [ { "bbox": [ 115, 507, 497, 519 ], "score": 1.0, "content": "ence on Learning Representations, 2019. URL https://openreview.net/forum?id", "type": "text" }, { "bbox": [ 497, 509, 504, 516 ], "score": 0.32, "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", "type": "inline_equation" } ], "index": 31 }, { "bbox": [ 115, 518, 181, 530 ], "spans": [ { "bbox": [ 115, 518, 181, 530 ], "score": 1.0, "content": "Bkg6RiCqY7.", "type": "text" } ], "index": 32 } ], "index": 31 }, { "type": "text", "bbox": [ 104, 536, 504, 559 ], "lines": [ { "bbox": [ 106, 537, 504, 548 ], "spans": [ { "bbox": [ 106, 537, 504, 548 ], "score": 1.0, "content": "Abdelrahman Mohamed, Dmytro Okhonko, and Luke Zettlemoyer. Transformers with convolutional", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 547, 342, 559 ], "spans": [ { "bbox": [ 115, 547, 342, 559 ], "score": 1.0, "content": "context for asr. arXiv preprint arXiv:1904.11660, 2019.", "type": "text" } ], "index": 34 } ], "index": 33.5 }, { "type": "text", "bbox": [ 107, 565, 506, 621 ], "lines": [ { "bbox": [ 106, 566, 505, 578 ], "spans": [ { "bbox": [ 106, 566, 505, 578 ], "score": 1.0, "content": "Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and", "type": "text" } ], "index": 35 }, { "bbox": [ 115, 576, 506, 590 ], "spans": [ { "bbox": [ 115, 576, 506, 590 ], "score": 1.0, "content": "Dustin Tran. Image transformer. In Jennifer Dy and Andreas Krause (eds.), Proceedings of", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 587, 506, 601 ], "spans": [ { "bbox": [ 115, 587, 506, 601 ], "score": 1.0, "content": "the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 597, 505, 612 ], "spans": [ { "bbox": [ 115, 597, 505, 612 ], "score": 1.0, "content": "Learning Research, pp. 4055–4064, Stockholmsmassan, Stockholm Sweden, 10–15 Jul 2018. ¨", "type": "text" } ], "index": 38 }, { "bbox": [ 115, 609, 149, 622 ], "spans": [ { "bbox": [ 115, 609, 149, 622 ], "score": 1.0, "content": "PMLR.", "type": "text" } ], "index": 39 } ], "index": 37 }, { "type": "text", "bbox": [ 106, 628, 505, 673 ], "lines": [ { "bbox": [ 105, 627, 506, 641 ], "spans": [ { "bbox": [ 105, 627, 506, 641 ], "score": 1.0, "content": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor", "type": "text" } ], "index": 40 }, { "bbox": [ 116, 639, 505, 652 ], "spans": [ { "bbox": [ 116, 639, 505, 652 ], "score": 1.0, "content": "Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style,", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 650, 505, 663 ], "spans": [ { "bbox": [ 115, 650, 505, 663 ], "score": 1.0, "content": "high-performance deep learning library. Advances in Neural Information Processing Systems, 32:", "type": "text" } ], "index": 42 }, { "bbox": [ 116, 662, 191, 673 ], "spans": [ { "bbox": [ 116, 662, 191, 673 ], "score": 1.0, "content": "8026–8037, 2019.", "type": "text" } ], "index": 43 } ], "index": 41.5 }, { "type": "text", "bbox": [ 106, 680, 502, 703 ], "lines": [ { "bbox": [ 106, 678, 504, 693 ], "spans": [ { "bbox": [ 106, 678, 504, 693 ], "score": 1.0, "content": "Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language under-", "type": "text" } ], "index": 44 }, { "bbox": [ 115, 691, 287, 703 ], "spans": [ { "bbox": [ 115, 691, 287, 703 ], "score": 1.0, "content": "standing by generative pre-training. 2018.", "type": "text" } ], "index": 45 } ], "index": 44.5 }, { "type": "text", "bbox": [ 107, 709, 503, 732 ], "lines": [ { "bbox": [ 105, 707, 505, 724 ], "spans": [ { "bbox": [ 105, 707, 505, 724 ], "score": 1.0, "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language", "type": "text" } ], "index": 46 }, { "bbox": [ 116, 720, 405, 733 ], "spans": [ { "bbox": [ 116, 720, 405, 733 ], "score": 1.0, "content": "models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.", "type": "text" } ], "index": 47 } ], "index": 46.5 } ], "page_idx": 10, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 310, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 765 ], "spans": [ { "bbox": [ 299, 750, 312, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 13 } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 108, 82, 504, 116 ], "lines": [ { "bbox": [ 106, 81, 506, 96 ], "spans": [ { "bbox": [ 106, 81, 506, 96 ], "score": 1.0, "content": "Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo", "type": "text" } ], "index": 0 }, { "bbox": [ 117, 94, 505, 106 ], "spans": [ { "bbox": [ 117, 94, 505, 106 ], "score": 1.0, "content": "Wang, Zhengdong Zhang, Yonghui Wu, et al. Conformer: Convolution-augmented transformer", "type": "text" } ], "index": 1 }, { "bbox": [ 115, 104, 374, 118 ], "spans": [ { "bbox": [ 115, 104, 374, 118 ], "score": 1.0, "content": "for speech recognition. arXiv preprint arXiv:2005.08100, 2020.", "type": "text" } ], "index": 2 } ], "index": 1, "bbox_fs": [ 106, 81, 506, 118 ] }, { "type": "text", "bbox": [ 108, 122, 503, 156 ], "lines": [ { "bbox": [ 105, 121, 505, 138 ], "spans": [ { "bbox": [ 105, 121, 505, 138 ], "score": 1.0, "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 132, 505, 148 ], "spans": [ { "bbox": [ 115, 132, 505, 148 ], "score": 1.0, "content": "nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.", "type": "text" } ], "index": 4 }, { "bbox": [ 116, 145, 181, 156 ], "spans": [ { "bbox": [ 116, 145, 181, 156 ], "score": 1.0, "content": "770–778, 2016.", "type": "text" } ], "index": 5 } ], "index": 4, "bbox_fs": [ 105, 121, 505, 156 ] }, { "type": "text", "bbox": [ 106, 163, 504, 186 ], "lines": [ { "bbox": [ 105, 162, 505, 177 ], "spans": [ { "bbox": [ 105, 162, 505, 177 ], "score": 1.0, "content": "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: Decoding-enhanced bert", "type": "text" } ], "index": 6 }, { "bbox": [ 116, 174, 393, 187 ], "spans": [ { "bbox": [ 116, 174, 393, 187 ], "score": 1.0, "content": "with disentangled attention. arXiv preprint arXiv:2006.03654, 2020.", "type": "text" } ], "index": 7 } ], "index": 6.5, "bbox_fs": [ 105, 162, 505, 187 ] }, { "type": "text", "bbox": [ 108, 192, 504, 227 ], "lines": [ { "bbox": [ 105, 192, 505, 207 ], "spans": [ { "bbox": [ 105, 192, 505, 207 ], "score": 1.0, "content": "Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Aug-", "type": "text" } ], "index": 8 }, { "bbox": [ 115, 204, 505, 217 ], "spans": [ { "bbox": [ 115, 204, 505, 217 ], "score": 1.0, "content": "ment your batch: Improving generalization through instance repetition. In Proceedings of the", "type": "text" } ], "index": 9 }, { "bbox": [ 115, 215, 484, 229 ], "spans": [ { "bbox": [ 115, 215, 484, 229 ], "score": 1.0, "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8129–8138, 2020.", "type": "text" } ], "index": 10 } ], "index": 9, "bbox_fs": [ 105, 192, 505, 229 ] }, { "type": "text", "bbox": [ 106, 233, 506, 267 ], "lines": [ { "bbox": [ 106, 233, 505, 246 ], "spans": [ { "bbox": [ 106, 233, 505, 246 ], "score": 1.0, "content": "Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Noam Shazeer, Ian Simon, Curtis", "type": "text" } ], "index": 11 }, { "bbox": [ 115, 244, 506, 258 ], "spans": [ { "bbox": [ 115, 244, 506, 258 ], "score": 1.0, "content": "Hawthorne, Andrew M Dai, Matthew D Hoffman, Monica Dinculescu, and Douglas Eck. Music", "type": "text" } ], "index": 12 }, { "bbox": [ 116, 256, 368, 268 ], "spans": [ { "bbox": [ 116, 256, 368, 268 ], "score": 1.0, "content": "transformer. In Advances in Neural Processing Systems, 2018.", "type": "text" } ], "index": 13 } ], "index": 12, "bbox_fs": [ 106, 233, 506, 268 ] }, { "type": "text", "bbox": [ 106, 273, 505, 297 ], "lines": [ { "bbox": [ 107, 274, 505, 286 ], "spans": [ { "bbox": [ 107, 274, 505, 286 ], "score": 1.0, "content": "Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 285, 484, 297 ], "spans": [ { "bbox": [ 115, 285, 484, 297 ], "score": 1.0, "content": "stochastic depth. In European conference on computer vision, pp. 646–661. Springer, 2016.", "type": "text" } ], "index": 15 } ], "index": 14.5, "bbox_fs": [ 107, 274, 505, 297 ] }, { "type": "text", "bbox": [ 107, 303, 504, 327 ], "lines": [ { "bbox": [ 106, 304, 504, 316 ], "spans": [ { "bbox": [ 106, 304, 504, 316 ], "score": 1.0, "content": "Md Amirul Islam, Sen Jia, and Neil DB Bruce. How much position information do convolutional", "type": "text" } ], "index": 16 }, { "bbox": [ 116, 315, 478, 327 ], "spans": [ { "bbox": [ 116, 315, 478, 327 ], "score": 1.0, "content": "neural networks encode? In International Conference on Learning Representations, 2020.", "type": "text" } ], "index": 17 } ], "index": 16.5, "bbox_fs": [ 106, 304, 504, 327 ] }, { "type": "text", "bbox": [ 107, 332, 505, 367 ], "lines": [ { "bbox": [ 105, 332, 505, 346 ], "spans": [ { "bbox": [ 105, 332, 505, 346 ], "score": 1.0, "content": "Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollar. Panoptic feature pyramid net- ´", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 344, 505, 357 ], "spans": [ { "bbox": [ 115, 344, 505, 357 ], "score": 1.0, "content": "works. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,", "type": "text" } ], "index": 19 }, { "bbox": [ 115, 356, 206, 367 ], "spans": [ { "bbox": [ 115, 356, 206, 367 ], "score": 1.0, "content": "pp. 6399–6408, 2019.", "type": "text" } ], "index": 20 } ], "index": 19, "bbox_fs": [ 105, 332, 505, 367 ] }, { "type": "text", "bbox": [ 107, 373, 505, 407 ], "lines": [ { "bbox": [ 106, 374, 504, 386 ], "spans": [ { "bbox": [ 106, 374, 504, 386 ], "score": 1.0, "content": "Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense´", "type": "text" } ], "index": 21 }, { "bbox": [ 115, 384, 505, 399 ], "spans": [ { "bbox": [ 115, 384, 505, 399 ], "score": 1.0, "content": "object detection. In Proceedings of the IEEE international conference on computer vision, pp.", "type": "text" } ], "index": 22 }, { "bbox": [ 116, 396, 191, 407 ], "spans": [ { "bbox": [ 116, 396, 191, 407 ], "score": 1.0, "content": "2980–2988, 2017.", "type": "text" } ], "index": 23 } ], "index": 22, "bbox_fs": [ 106, 374, 505, 407 ] }, { "type": "text", "bbox": [ 107, 414, 505, 448 ], "lines": [ { "bbox": [ 105, 414, 505, 428 ], "spans": [ { "bbox": [ 105, 414, 505, 428 ], "score": 1.0, "content": "Xuanqing Liu, Hsiang-Fu Yu, Inderjit Dhillon, and Cho-Jui Hsieh. Learning to encode position for", "type": "text" } ], "index": 24 }, { "bbox": [ 116, 426, 505, 438 ], "spans": [ { "bbox": [ 116, 426, 505, 438 ], "score": 1.0, "content": "transformer with continuous dynamical model. In International Conference on Machine Learn-", "type": "text" } ], "index": 25 }, { "bbox": [ 116, 437, 255, 448 ], "spans": [ { "bbox": [ 116, 437, 255, 448 ], "score": 1.0, "content": "ing, pp. 6327–6335. PMLR, 2020.", "type": "text" } ], "index": 26 } ], "index": 25, "bbox_fs": [ 105, 414, 505, 448 ] }, { "type": "text", "bbox": [ 106, 454, 505, 488 ], "lines": [ { "bbox": [ 105, 453, 505, 468 ], "spans": [ { "bbox": [ 105, 453, 505, 468 ], "score": 1.0, "content": "Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 466, 505, 479 ], "spans": [ { "bbox": [ 115, 466, 505, 479 ], "score": 1.0, "content": "Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint", "type": "text" } ], "index": 28 }, { "bbox": [ 116, 478, 219, 487 ], "spans": [ { "bbox": [ 116, 478, 219, 487 ], "score": 1.0, "content": "arXiv:2103.14030, 2021.", "type": "text" } ], "index": 29 } ], "index": 28, "bbox_fs": [ 105, 453, 505, 487 ] }, { "type": "text", "bbox": [ 106, 495, 503, 529 ], "lines": [ { "bbox": [ 106, 496, 505, 508 ], "spans": [ { "bbox": [ 106, 496, 505, 508 ], "score": 1.0, "content": "Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Confer-", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 507, 504, 519 ], "spans": [ { "bbox": [ 115, 507, 497, 519 ], "score": 1.0, "content": "ence on Learning Representations, 2019. URL https://openreview.net/forum?id", "type": "text" }, { "bbox": [ 497, 509, 504, 516 ], "score": 0.32, "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", "type": "inline_equation" } ], "index": 31 }, { "bbox": [ 115, 518, 181, 530 ], "spans": [ { "bbox": [ 115, 518, 181, 530 ], "score": 1.0, "content": "Bkg6RiCqY7.", "type": "text" } ], "index": 32 } ], "index": 31, "bbox_fs": [ 106, 496, 505, 530 ] }, { "type": "text", "bbox": [ 104, 536, 504, 559 ], "lines": [ { "bbox": [ 106, 537, 504, 548 ], "spans": [ { "bbox": [ 106, 537, 504, 548 ], "score": 1.0, "content": "Abdelrahman Mohamed, Dmytro Okhonko, and Luke Zettlemoyer. Transformers with convolutional", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 547, 342, 559 ], "spans": [ { "bbox": [ 115, 547, 342, 559 ], "score": 1.0, "content": "context for asr. arXiv preprint arXiv:1904.11660, 2019.", "type": "text" } ], "index": 34 } ], "index": 33.5, "bbox_fs": [ 106, 537, 504, 559 ] }, { "type": "text", "bbox": [ 107, 565, 506, 621 ], "lines": [ { "bbox": [ 106, 566, 505, 578 ], "spans": [ { "bbox": [ 106, 566, 505, 578 ], "score": 1.0, "content": "Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and", "type": "text" } ], "index": 35 }, { "bbox": [ 115, 576, 506, 590 ], "spans": [ { "bbox": [ 115, 576, 506, 590 ], "score": 1.0, "content": "Dustin Tran. Image transformer. In Jennifer Dy and Andreas Krause (eds.), Proceedings of", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 587, 506, 601 ], "spans": [ { "bbox": [ 115, 587, 506, 601 ], "score": 1.0, "content": "the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 597, 505, 612 ], "spans": [ { "bbox": [ 115, 597, 505, 612 ], "score": 1.0, "content": "Learning Research, pp. 4055–4064, Stockholmsmassan, Stockholm Sweden, 10–15 Jul 2018. ¨", "type": "text" } ], "index": 38 }, { "bbox": [ 115, 609, 149, 622 ], "spans": [ { "bbox": [ 115, 609, 149, 622 ], "score": 1.0, "content": "PMLR.", "type": "text" } ], "index": 39 } ], "index": 37, "bbox_fs": [ 106, 566, 506, 622 ] }, { "type": "text", "bbox": [ 106, 628, 505, 673 ], "lines": [ { "bbox": [ 105, 627, 506, 641 ], "spans": [ { "bbox": [ 105, 627, 506, 641 ], "score": 1.0, "content": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor", "type": "text" } ], "index": 40 }, { "bbox": [ 116, 639, 505, 652 ], "spans": [ { "bbox": [ 116, 639, 505, 652 ], "score": 1.0, "content": "Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style,", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 650, 505, 663 ], "spans": [ { "bbox": [ 115, 650, 505, 663 ], "score": 1.0, "content": "high-performance deep learning library. Advances in Neural Information Processing Systems, 32:", "type": "text" } ], "index": 42 }, { "bbox": [ 116, 662, 191, 673 ], "spans": [ { "bbox": [ 116, 662, 191, 673 ], "score": 1.0, "content": "8026–8037, 2019.", "type": "text" } ], "index": 43 } ], "index": 41.5, "bbox_fs": [ 105, 627, 506, 673 ] }, { "type": "text", "bbox": [ 106, 680, 502, 703 ], "lines": [ { "bbox": [ 106, 678, 504, 693 ], "spans": [ { "bbox": [ 106, 678, 504, 693 ], "score": 1.0, "content": "Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language under-", "type": "text" } ], "index": 44 }, { "bbox": [ 115, 691, 287, 703 ], "spans": [ { "bbox": [ 115, 691, 287, 703 ], "score": 1.0, "content": "standing by generative pre-training. 2018.", "type": "text" } ], "index": 45 } ], "index": 44.5, "bbox_fs": [ 106, 678, 504, 703 ] }, { "type": "text", "bbox": [ 107, 709, 503, 732 ], "lines": [ { "bbox": [ 105, 707, 505, 724 ], "spans": [ { "bbox": [ 105, 707, 505, 724 ], "score": 1.0, "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language", "type": "text" } ], "index": 46 }, { "bbox": [ 116, 720, 405, 733 ], "spans": [ { "bbox": [ 116, 720, 405, 733 ], "score": 1.0, "content": "models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.", "type": "text" } ], "index": 47 } ], "index": 46.5, "bbox_fs": [ 105, 707, 505, 733 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 108, 82, 503, 115 ], "lines": [ { "bbox": [ 106, 81, 505, 96 ], "spans": [ { "bbox": [ 106, 81, 505, 96 ], "score": 1.0, "content": "Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollar. Designing ´", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 505, 107 ], "spans": [ { "bbox": [ 115, 93, 505, 107 ], "score": 1.0, "content": "network design spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and", "type": "text" } ], "index": 1 }, { "bbox": [ 115, 104, 300, 117 ], "spans": [ { "bbox": [ 115, 104, 300, 117 ], "score": 1.0, "content": "Pattern Recognition, pp. 10428–10436, 2020.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "text", "bbox": [ 106, 125, 505, 191 ], "lines": [ { "bbox": [ 105, 123, 506, 138 ], "spans": [ { "bbox": [ 105, 123, 506, 138 ], "score": 1.0, "content": "Prajit Ramachandran, Niki Parmar, Ashish Vaswani, Irwan Bello, Anselm Levskaya,", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 136, 505, 147 ], "spans": [ { "bbox": [ 115, 136, 194, 147 ], "score": 1.0, "content": "and Jon Shlens.", "type": "text" }, { "bbox": [ 212, 136, 415, 147 ], "score": 1.0, "content": "Stand-alone self-attention in vision models.", "type": "text" }, { "bbox": [ 431, 136, 505, 147 ], "score": 1.0, "content": "In H. Wallach,", "type": "text" } ], "index": 4 }, { "bbox": [ 115, 146, 505, 159 ], "spans": [ { "bbox": [ 115, 146, 505, 159 ], "score": 1.0, "content": "H. Larochelle, A. Beygelzimer, F. d'Alche-Buc, E. Fox, and R. Garnett (eds.), ´ Ad-", "type": "text" } ], "index": 5 }, { "bbox": [ 115, 158, 505, 170 ], "spans": [ { "bbox": [ 115, 158, 505, 170 ], "score": 1.0, "content": "vances in Neural Information Processing Systems, volume 32, pp. 68–80. Curran Asso-", "type": "text" } ], "index": 6 }, { "bbox": [ 115, 168, 505, 182 ], "spans": [ { "bbox": [ 115, 168, 505, 182 ], "score": 1.0, "content": "ciates, Inc., 2019. URL https://proceedings.neurips.cc/paper/2019/file/", "type": "text" } ], "index": 7 }, { "bbox": [ 116, 178, 373, 192 ], "spans": [ { "bbox": [ 116, 178, 373, 192 ], "score": 1.0, "content": "3416a75f4cea9109507cacd8e2f2aefc-Paper.pdf.", "type": "text" } ], "index": 8 } ], "index": 5.5 }, { "type": "text", "bbox": [ 108, 200, 504, 234 ], "lines": [ { "bbox": [ 105, 199, 505, 214 ], "spans": [ { "bbox": [ 105, 199, 505, 214 ], "score": 1.0, "content": "Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representa-", "type": "text" } ], "index": 9 }, { "bbox": [ 115, 211, 505, 224 ], "spans": [ { "bbox": [ 115, 211, 505, 224 ], "score": 1.0, "content": "tions. In Proceedings of the 2018 Conference of the North American Chapter of the Association", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 223, 496, 235 ], "spans": [ { "bbox": [ 115, 223, 496, 235 ], "score": 1.0, "content": "for Computational Linguistics: Human Language Technologies, Volume 2, pp. 464–468, 2018.", "type": "text" } ], "index": 11 } ], "index": 10 }, { "type": "text", "bbox": [ 108, 243, 503, 276 ], "lines": [ { "bbox": [ 106, 243, 505, 256 ], "spans": [ { "bbox": [ 106, 243, 505, 256 ], "score": 1.0, "content": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.", "type": "text" } ], "index": 12 }, { "bbox": [ 116, 254, 505, 266 ], "spans": [ { "bbox": [ 116, 254, 505, 266 ], "score": 1.0, "content": "Dropout: a simple way to prevent neural networks from overfitting. The journal of machine", "type": "text" } ], "index": 13 }, { "bbox": [ 116, 265, 291, 277 ], "spans": [ { "bbox": [ 116, 265, 291, 277 ], "score": 1.0, "content": "learning research, 15(1):1929–1958, 2014.", "type": "text" } ], "index": 14 } ], "index": 13 }, { "type": "text", "bbox": [ 106, 285, 504, 308 ], "lines": [ { "bbox": [ 106, 285, 505, 298 ], "spans": [ { "bbox": [ 106, 285, 505, 298 ], "score": 1.0, "content": "Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer", "type": "text" } ], "index": 15 }, { "bbox": [ 116, 296, 411, 309 ], "spans": [ { "bbox": [ 116, 296, 411, 309 ], "score": 1.0, "content": "with rotary position embedding. arXiv preprint arXiv:2104.09864, 2021.", "type": "text" } ], "index": 16 } ], "index": 15.5 }, { "type": "text", "bbox": [ 106, 316, 505, 351 ], "lines": [ { "bbox": [ 106, 316, 505, 330 ], "spans": [ { "bbox": [ 106, 316, 505, 330 ], "score": 1.0, "content": "Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable ef-", "type": "text" } ], "index": 17 }, { "bbox": [ 116, 328, 505, 340 ], "spans": [ { "bbox": [ 116, 328, 505, 340 ], "score": 1.0, "content": "fectiveness of data in deep learning era. In Proceedings of the IEEE international conference on", "type": "text" } ], "index": 18 }, { "bbox": [ 116, 340, 265, 351 ], "spans": [ { "bbox": [ 116, 340, 265, 351 ], "score": 1.0, "content": "computer vision, pp. 843–852, 2017.", "type": "text" } ], "index": 19 } ], "index": 18 }, { "type": "text", "bbox": [ 107, 359, 504, 393 ], "lines": [ { "bbox": [ 106, 359, 504, 373 ], "spans": [ { "bbox": [ 106, 359, 504, 373 ], "score": 1.0, "content": "Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethink-", "type": "text" } ], "index": 20 }, { "bbox": [ 115, 370, 505, 384 ], "spans": [ { "bbox": [ 115, 370, 505, 384 ], "score": 1.0, "content": "ing the inception architecture for computer vision. In Proceedings of the IEEE conference on", "type": "text" } ], "index": 21 }, { "bbox": [ 116, 381, 371, 394 ], "spans": [ { "bbox": [ 116, 381, 371, 394 ], "score": 1.0, "content": "computer vision and pattern recognition, pp. 2818–2826, 2016.", "type": "text" } ], "index": 22 } ], "index": 21 }, { "type": "text", "bbox": [ 106, 402, 503, 425 ], "lines": [ { "bbox": [ 106, 402, 505, 415 ], "spans": [ { "bbox": [ 106, 402, 505, 415 ], "score": 1.0, "content": "Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural net-", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 413, 474, 425 ], "spans": [ { "bbox": [ 115, 413, 474, 425 ], "score": 1.0, "content": "works. In International Conference on Machine Learning, pp. 6105–6114. PMLR, 2019.", "type": "text" } ], "index": 24 } ], "index": 23.5 }, { "type": "text", "bbox": [ 106, 433, 505, 467 ], "lines": [ { "bbox": [ 106, 433, 504, 446 ], "spans": [ { "bbox": [ 106, 433, 504, 446 ], "score": 1.0, "content": "Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and", "type": "text" } ], "index": 25 }, { "bbox": [ 115, 443, 506, 457 ], "spans": [ { "bbox": [ 115, 443, 506, 457 ], "score": 1.0, "content": "Herve J ´ egou. Training data-efficient image transformers & distillation through attention. ´ arXiv", "type": "text" } ], "index": 26 }, { "bbox": [ 115, 456, 254, 468 ], "spans": [ { "bbox": [ 115, 456, 254, 468 ], "score": 1.0, "content": "preprint arXiv:2012.12877, 2020.", "type": "text" } ], "index": 27 } ], "index": 26 }, { "type": "text", "bbox": [ 106, 475, 504, 510 ], "lines": [ { "bbox": [ 105, 475, 506, 489 ], "spans": [ { "bbox": [ 105, 475, 506, 489 ], "score": 1.0, "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", "type": "text" } ], "index": 28 }, { "bbox": [ 116, 487, 505, 500 ], "spans": [ { "bbox": [ 116, 487, 505, 500 ], "score": 1.0, "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st In-", "type": "text" } ], "index": 29 }, { "bbox": [ 116, 499, 476, 510 ], "spans": [ { "bbox": [ 116, 499, 476, 510 ], "score": 1.0, "content": "ternational Conference on Neural Information Processing Systems, pp. 6000–6010, 2017.", "type": "text" } ], "index": 30 } ], "index": 29 }, { "type": "text", "bbox": [ 105, 518, 503, 541 ], "lines": [ { "bbox": [ 106, 519, 504, 532 ], "spans": [ { "bbox": [ 106, 519, 504, 532 ], "score": 1.0, "content": "Benyou Wang, Donghao Zhao, Christina Lioma, Qiuchi Li, Peng Zhang, and Jakob Grue Simonsen.", "type": "text" } ], "index": 31 }, { "bbox": [ 115, 529, 466, 542 ], "spans": [ { "bbox": [ 115, 529, 466, 542 ], "score": 1.0, "content": "Encoding word order in complex embeddings. arXiv preprint arXiv:1912.12333, 2019.", "type": "text" } ], "index": 32 } ], "index": 31.5 }, { "type": "text", "bbox": [ 106, 550, 505, 584 ], "lines": [ { "bbox": [ 106, 549, 505, 564 ], "spans": [ { "bbox": [ 106, 549, 505, 564 ], "score": 1.0, "content": "Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo,", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 561, 506, 574 ], "spans": [ { "bbox": [ 115, 561, 506, 574 ], "score": 1.0, "content": "and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 572, 336, 585 ], "spans": [ { "bbox": [ 115, 572, 336, 585 ], "score": 1.0, "content": "convolutions. arXiv preprint arXiv:2102.12122, 2021.", "type": "text" } ], "index": 35 } ], "index": 34 }, { "type": "text", "bbox": [ 107, 592, 505, 626 ], "lines": [ { "bbox": [ 106, 592, 506, 606 ], "spans": [ { "bbox": [ 106, 592, 506, 606 ], "score": 1.0, "content": "Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 603, 505, 616 ], "spans": [ { "bbox": [ 115, 603, 505, 616 ], "score": 1.0, "content": "Huaxia Xia. End-to-end video instance segmentation with transformers. arXiv preprint", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 614, 219, 626 ], "spans": [ { "bbox": [ 115, 614, 219, 626 ], "score": 1.0, "content": "arXiv:2011.14503, 2020.", "type": "text" } ], "index": 38 } ], "index": 37 }, { "type": "text", "bbox": [ 104, 635, 504, 658 ], "lines": [ { "bbox": [ 106, 635, 505, 648 ], "spans": [ { "bbox": [ 106, 635, 505, 648 ], "score": 1.0, "content": "Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt:", "type": "text" } ], "index": 39 }, { "bbox": [ 116, 646, 475, 659 ], "spans": [ { "bbox": [ 116, 646, 475, 659 ], "score": 1.0, "content": "Introducing convolutions to vision transformers. arXiv preprint arXiv:2103.15808, 2021.", "type": "text" } ], "index": 40 } ], "index": 39.5 }, { "type": "text", "bbox": [ 104, 667, 504, 690 ], "lines": [ { "bbox": [ 106, 667, 504, 680 ], "spans": [ { "bbox": [ 106, 667, 504, 680 ], "score": 1.0, "content": "Baosong Yang, Longyue Wang, Derek Wong, Lidia S Chao, and Zhaopeng Tu. Convolutional self-", "type": "text" } ], "index": 41 }, { "bbox": [ 116, 678, 363, 690 ], "spans": [ { "bbox": [ 116, 678, 363, 690 ], "score": 1.0, "content": "attention networks. arXiv preprint arXiv:1904.03107, 2019a.", "type": "text" } ], "index": 42 } ], "index": 41.5 }, { "type": "text", "bbox": [ 107, 699, 505, 731 ], "lines": [ { "bbox": [ 105, 699, 505, 711 ], "spans": [ { "bbox": [ 105, 699, 505, 711 ], "score": 1.0, "content": "Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V", "type": "text" } ], "index": 43 }, { "bbox": [ 115, 709, 505, 722 ], "spans": [ { "bbox": [ 115, 709, 505, 722 ], "score": 1.0, "content": "Le. Xlnet: Generalized autoregressive pretraining for language understanding. arXiv preprint", "type": "text" } ], "index": 44 }, { "bbox": [ 115, 720, 224, 732 ], "spans": [ { "bbox": [ 115, 720, 224, 732 ], "score": 1.0, "content": "arXiv:1906.08237, 2019b.", "type": "text" } ], "index": 45 } ], "index": 44 } ], "page_idx": 11, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "12", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 108, 82, 503, 115 ], "lines": [ { "bbox": [ 106, 81, 505, 96 ], "spans": [ { "bbox": [ 106, 81, 505, 96 ], "score": 1.0, "content": "Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollar. Designing ´", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 505, 107 ], "spans": [ { "bbox": [ 115, 93, 505, 107 ], "score": 1.0, "content": "network design spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and", "type": "text" } ], "index": 1 }, { "bbox": [ 115, 104, 300, 117 ], "spans": [ { "bbox": [ 115, 104, 300, 117 ], "score": 1.0, "content": "Pattern Recognition, pp. 10428–10436, 2020.", "type": "text" } ], "index": 2 } ], "index": 1, "bbox_fs": [ 106, 81, 505, 117 ] }, { "type": "text", "bbox": [ 106, 125, 505, 191 ], "lines": [ { "bbox": [ 105, 123, 506, 138 ], "spans": [ { "bbox": [ 105, 123, 506, 138 ], "score": 1.0, "content": "Prajit Ramachandran, Niki Parmar, Ashish Vaswani, Irwan Bello, Anselm Levskaya,", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 136, 505, 147 ], "spans": [ { "bbox": [ 115, 136, 194, 147 ], "score": 1.0, "content": "and Jon Shlens.", "type": "text" }, { "bbox": [ 212, 136, 415, 147 ], "score": 1.0, "content": "Stand-alone self-attention in vision models.", "type": "text" }, { "bbox": [ 431, 136, 505, 147 ], "score": 1.0, "content": "In H. Wallach,", "type": "text" } ], "index": 4 }, { "bbox": [ 115, 146, 505, 159 ], "spans": [ { "bbox": [ 115, 146, 505, 159 ], "score": 1.0, "content": "H. Larochelle, A. Beygelzimer, F. d'Alche-Buc, E. Fox, and R. Garnett (eds.), ´ Ad-", "type": "text" } ], "index": 5 }, { "bbox": [ 115, 158, 505, 170 ], "spans": [ { "bbox": [ 115, 158, 505, 170 ], "score": 1.0, "content": "vances in Neural Information Processing Systems, volume 32, pp. 68–80. Curran Asso-", "type": "text" } ], "index": 6 }, { "bbox": [ 115, 168, 505, 182 ], "spans": [ { "bbox": [ 115, 168, 505, 182 ], "score": 1.0, "content": "ciates, Inc., 2019. URL https://proceedings.neurips.cc/paper/2019/file/", "type": "text" } ], "index": 7 }, { "bbox": [ 116, 178, 373, 192 ], "spans": [ { "bbox": [ 116, 178, 373, 192 ], "score": 1.0, "content": "3416a75f4cea9109507cacd8e2f2aefc-Paper.pdf.", "type": "text" } ], "index": 8 } ], "index": 5.5, "bbox_fs": [ 105, 123, 506, 192 ] }, { "type": "text", "bbox": [ 108, 200, 504, 234 ], "lines": [ { "bbox": [ 105, 199, 505, 214 ], "spans": [ { "bbox": [ 105, 199, 505, 214 ], "score": 1.0, "content": "Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representa-", "type": "text" } ], "index": 9 }, { "bbox": [ 115, 211, 505, 224 ], "spans": [ { "bbox": [ 115, 211, 505, 224 ], "score": 1.0, "content": "tions. In Proceedings of the 2018 Conference of the North American Chapter of the Association", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 223, 496, 235 ], "spans": [ { "bbox": [ 115, 223, 496, 235 ], "score": 1.0, "content": "for Computational Linguistics: Human Language Technologies, Volume 2, pp. 464–468, 2018.", "type": "text" } ], "index": 11 } ], "index": 10, "bbox_fs": [ 105, 199, 505, 235 ] }, { "type": "text", "bbox": [ 108, 243, 503, 276 ], "lines": [ { "bbox": [ 106, 243, 505, 256 ], "spans": [ { "bbox": [ 106, 243, 505, 256 ], "score": 1.0, "content": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.", "type": "text" } ], "index": 12 }, { "bbox": [ 116, 254, 505, 266 ], "spans": [ { "bbox": [ 116, 254, 505, 266 ], "score": 1.0, "content": "Dropout: a simple way to prevent neural networks from overfitting. The journal of machine", "type": "text" } ], "index": 13 }, { "bbox": [ 116, 265, 291, 277 ], "spans": [ { "bbox": [ 116, 265, 291, 277 ], "score": 1.0, "content": "learning research, 15(1):1929–1958, 2014.", "type": "text" } ], "index": 14 } ], "index": 13, "bbox_fs": [ 106, 243, 505, 277 ] }, { "type": "text", "bbox": [ 106, 285, 504, 308 ], "lines": [ { "bbox": [ 106, 285, 505, 298 ], "spans": [ { "bbox": [ 106, 285, 505, 298 ], "score": 1.0, "content": "Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer", "type": "text" } ], "index": 15 }, { "bbox": [ 116, 296, 411, 309 ], "spans": [ { "bbox": [ 116, 296, 411, 309 ], "score": 1.0, "content": "with rotary position embedding. arXiv preprint arXiv:2104.09864, 2021.", "type": "text" } ], "index": 16 } ], "index": 15.5, "bbox_fs": [ 106, 285, 505, 309 ] }, { "type": "text", "bbox": [ 106, 316, 505, 351 ], "lines": [ { "bbox": [ 106, 316, 505, 330 ], "spans": [ { "bbox": [ 106, 316, 505, 330 ], "score": 1.0, "content": "Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable ef-", "type": "text" } ], "index": 17 }, { "bbox": [ 116, 328, 505, 340 ], "spans": [ { "bbox": [ 116, 328, 505, 340 ], "score": 1.0, "content": "fectiveness of data in deep learning era. In Proceedings of the IEEE international conference on", "type": "text" } ], "index": 18 }, { "bbox": [ 116, 340, 265, 351 ], "spans": [ { "bbox": [ 116, 340, 265, 351 ], "score": 1.0, "content": "computer vision, pp. 843–852, 2017.", "type": "text" } ], "index": 19 } ], "index": 18, "bbox_fs": [ 106, 316, 505, 351 ] }, { "type": "text", "bbox": [ 107, 359, 504, 393 ], "lines": [ { "bbox": [ 106, 359, 504, 373 ], "spans": [ { "bbox": [ 106, 359, 504, 373 ], "score": 1.0, "content": "Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethink-", "type": "text" } ], "index": 20 }, { "bbox": [ 115, 370, 505, 384 ], "spans": [ { "bbox": [ 115, 370, 505, 384 ], "score": 1.0, "content": "ing the inception architecture for computer vision. In Proceedings of the IEEE conference on", "type": "text" } ], "index": 21 }, { "bbox": [ 116, 381, 371, 394 ], "spans": [ { "bbox": [ 116, 381, 371, 394 ], "score": 1.0, "content": "computer vision and pattern recognition, pp. 2818–2826, 2016.", "type": "text" } ], "index": 22 } ], "index": 21, "bbox_fs": [ 106, 359, 505, 394 ] }, { "type": "text", "bbox": [ 106, 402, 503, 425 ], "lines": [ { "bbox": [ 106, 402, 505, 415 ], "spans": [ { "bbox": [ 106, 402, 505, 415 ], "score": 1.0, "content": "Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural net-", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 413, 474, 425 ], "spans": [ { "bbox": [ 115, 413, 474, 425 ], "score": 1.0, "content": "works. In International Conference on Machine Learning, pp. 6105–6114. PMLR, 2019.", "type": "text" } ], "index": 24 } ], "index": 23.5, "bbox_fs": [ 106, 402, 505, 425 ] }, { "type": "text", "bbox": [ 106, 433, 505, 467 ], "lines": [ { "bbox": [ 106, 433, 504, 446 ], "spans": [ { "bbox": [ 106, 433, 504, 446 ], "score": 1.0, "content": "Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and", "type": "text" } ], "index": 25 }, { "bbox": [ 115, 443, 506, 457 ], "spans": [ { "bbox": [ 115, 443, 506, 457 ], "score": 1.0, "content": "Herve J ´ egou. Training data-efficient image transformers & distillation through attention. ´ arXiv", "type": "text" } ], "index": 26 }, { "bbox": [ 115, 456, 254, 468 ], "spans": [ { "bbox": [ 115, 456, 254, 468 ], "score": 1.0, "content": "preprint arXiv:2012.12877, 2020.", "type": "text" } ], "index": 27 } ], "index": 26, "bbox_fs": [ 106, 433, 506, 468 ] }, { "type": "text", "bbox": [ 106, 475, 504, 510 ], "lines": [ { "bbox": [ 105, 475, 506, 489 ], "spans": [ { "bbox": [ 105, 475, 506, 489 ], "score": 1.0, "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", "type": "text" } ], "index": 28 }, { "bbox": [ 116, 487, 505, 500 ], "spans": [ { "bbox": [ 116, 487, 505, 500 ], "score": 1.0, "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st In-", "type": "text" } ], "index": 29 }, { "bbox": [ 116, 499, 476, 510 ], "spans": [ { "bbox": [ 116, 499, 476, 510 ], "score": 1.0, "content": "ternational Conference on Neural Information Processing Systems, pp. 6000–6010, 2017.", "type": "text" } ], "index": 30 } ], "index": 29, "bbox_fs": [ 105, 475, 506, 510 ] }, { "type": "text", "bbox": [ 105, 518, 503, 541 ], "lines": [ { "bbox": [ 106, 519, 504, 532 ], "spans": [ { "bbox": [ 106, 519, 504, 532 ], "score": 1.0, "content": "Benyou Wang, Donghao Zhao, Christina Lioma, Qiuchi Li, Peng Zhang, and Jakob Grue Simonsen.", "type": "text" } ], "index": 31 }, { "bbox": [ 115, 529, 466, 542 ], "spans": [ { "bbox": [ 115, 529, 466, 542 ], "score": 1.0, "content": "Encoding word order in complex embeddings. arXiv preprint arXiv:1912.12333, 2019.", "type": "text" } ], "index": 32 } ], "index": 31.5, "bbox_fs": [ 106, 519, 504, 542 ] }, { "type": "text", "bbox": [ 106, 550, 505, 584 ], "lines": [ { "bbox": [ 106, 549, 505, 564 ], "spans": [ { "bbox": [ 106, 549, 505, 564 ], "score": 1.0, "content": "Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo,", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 561, 506, 574 ], "spans": [ { "bbox": [ 115, 561, 506, 574 ], "score": 1.0, "content": "and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 572, 336, 585 ], "spans": [ { "bbox": [ 115, 572, 336, 585 ], "score": 1.0, "content": "convolutions. arXiv preprint arXiv:2102.12122, 2021.", "type": "text" } ], "index": 35 } ], "index": 34, "bbox_fs": [ 106, 549, 506, 585 ] }, { "type": "text", "bbox": [ 107, 592, 505, 626 ], "lines": [ { "bbox": [ 106, 592, 506, 606 ], "spans": [ { "bbox": [ 106, 592, 506, 606 ], "score": 1.0, "content": "Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 603, 505, 616 ], "spans": [ { "bbox": [ 115, 603, 505, 616 ], "score": 1.0, "content": "Huaxia Xia. End-to-end video instance segmentation with transformers. arXiv preprint", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 614, 219, 626 ], "spans": [ { "bbox": [ 115, 614, 219, 626 ], "score": 1.0, "content": "arXiv:2011.14503, 2020.", "type": "text" } ], "index": 38 } ], "index": 37, "bbox_fs": [ 106, 592, 506, 626 ] }, { "type": "text", "bbox": [ 104, 635, 504, 658 ], "lines": [ { "bbox": [ 106, 635, 505, 648 ], "spans": [ { "bbox": [ 106, 635, 505, 648 ], "score": 1.0, "content": "Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt:", "type": "text" } ], "index": 39 }, { "bbox": [ 116, 646, 475, 659 ], "spans": [ { "bbox": [ 116, 646, 475, 659 ], "score": 1.0, "content": "Introducing convolutions to vision transformers. arXiv preprint arXiv:2103.15808, 2021.", "type": "text" } ], "index": 40 } ], "index": 39.5, "bbox_fs": [ 106, 635, 505, 659 ] }, { "type": "text", "bbox": [ 104, 667, 504, 690 ], "lines": [ { "bbox": [ 106, 667, 504, 680 ], "spans": [ { "bbox": [ 106, 667, 504, 680 ], "score": 1.0, "content": "Baosong Yang, Longyue Wang, Derek Wong, Lidia S Chao, and Zhaopeng Tu. Convolutional self-", "type": "text" } ], "index": 41 }, { "bbox": [ 116, 678, 363, 690 ], "spans": [ { "bbox": [ 116, 678, 363, 690 ], "score": 1.0, "content": "attention networks. arXiv preprint arXiv:1904.03107, 2019a.", "type": "text" } ], "index": 42 } ], "index": 41.5, "bbox_fs": [ 106, 667, 504, 690 ] }, { "type": "text", "bbox": [ 107, 699, 505, 731 ], "lines": [ { "bbox": [ 105, 699, 505, 711 ], "spans": [ { "bbox": [ 105, 699, 505, 711 ], "score": 1.0, "content": "Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V", "type": "text" } ], "index": 43 }, { "bbox": [ 115, 709, 505, 722 ], "spans": [ { "bbox": [ 115, 709, 505, 722 ], "score": 1.0, "content": "Le. Xlnet: Generalized autoregressive pretraining for language understanding. arXiv preprint", "type": "text" } ], "index": 44 }, { "bbox": [ 115, 720, 224, 732 ], "spans": [ { "bbox": [ 115, 720, 224, 732 ], "score": 1.0, "content": "arXiv:1906.08237, 2019b.", "type": "text" } ], "index": 45 } ], "index": 44, "bbox_fs": [ 105, 699, 505, 732 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 108, 82, 504, 116 ], "lines": [ { "bbox": [ 106, 82, 505, 95 ], "spans": [ { "bbox": [ 106, 82, 505, 95 ], "score": 1.0, "content": "Adams Wei Yu, David Dohan, Minh-Thang Luong, Rui Zhao, Kai Chen, Mohammad Norouzi, and", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 506, 107 ], "spans": [ { "bbox": [ 115, 93, 506, 107 ], "score": 1.0, "content": "Quoc V Le. Qanet: Combining local convolution with global self-attention for reading compre-", "type": "text" } ], "index": 1 }, { "bbox": [ 115, 105, 316, 117 ], "spans": [ { "bbox": [ 115, 105, 316, 117 ], "score": 1.0, "content": "hension. arXiv preprint arXiv:1804.09541, 2018.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "text", "bbox": [ 109, 123, 503, 157 ], "lines": [ { "bbox": [ 106, 123, 504, 136 ], "spans": [ { "bbox": [ 106, 123, 504, 136 ], "score": 1.0, "content": "Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo.", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 134, 505, 147 ], "spans": [ { "bbox": [ 115, 134, 505, 147 ], "score": 1.0, "content": "Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceed-", "type": "text" } ], "index": 4 }, { "bbox": [ 115, 145, 483, 159 ], "spans": [ { "bbox": [ 115, 145, 483, 159 ], "score": 1.0, "content": "ings of the IEEE/CVF International Conference on Computer Vision, pp. 6023–6032, 2019.", "type": "text" } ], "index": 5 } ], "index": 4 }, { "type": "text", "bbox": [ 108, 163, 503, 198 ], "lines": [ { "bbox": [ 106, 163, 504, 177 ], "spans": [ { "bbox": [ 106, 163, 504, 177 ], "score": 1.0, "content": "Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond em-", "type": "text" } ], "index": 6 }, { "bbox": [ 115, 174, 505, 188 ], "spans": [ { "bbox": [ 115, 174, 505, 188 ], "score": 1.0, "content": "pirical risk minimization. In International Conference on Learning Representations, 2018. URL", "type": "text" } ], "index": 7 }, { "bbox": [ 115, 186, 367, 200 ], "spans": [ { "bbox": [ 115, 186, 302, 200 ], "score": 1.0, "content": "https://openreview.net/forum?id", "type": "text" }, { "bbox": [ 302, 187, 309, 196 ], "score": 0.5, "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", "type": "inline_equation" }, { "bbox": [ 309, 186, 367, 200 ], "score": 1.0, "content": "r1Ddp1-Rb.", "type": "text" } ], "index": 8 } ], "index": 7 }, { "type": "text", "bbox": [ 107, 204, 505, 239 ], "lines": [ { "bbox": [ 105, 204, 505, 218 ], "spans": [ { "bbox": [ 105, 204, 505, 218 ], "score": 1.0, "content": "Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei", "type": "text" } ], "index": 9 }, { "bbox": [ 116, 217, 505, 229 ], "spans": [ { "bbox": [ 116, 217, 505, 229 ], "score": 1.0, "content": "Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 227, 500, 239 ], "spans": [ { "bbox": [ 115, 227, 500, 239 ], "score": 1.0, "content": "a sequence-to-sequence perspective with transformers. arXiv preprint arXiv:2012.15840, 2020.", "type": "text" } ], "index": 11 } ], "index": 10 }, { "type": "text", "bbox": [ 106, 245, 504, 279 ], "lines": [ { "bbox": [ 104, 244, 505, 260 ], "spans": [ { "bbox": [ 104, 244, 505, 260 ], "score": 1.0, "content": "Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data aug-", "type": "text" } ], "index": 12 }, { "bbox": [ 115, 256, 506, 270 ], "spans": [ { "bbox": [ 115, 256, 506, 270 ], "score": 1.0, "content": "mentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pp.", "type": "text" } ], "index": 13 }, { "bbox": [ 117, 268, 201, 279 ], "spans": [ { "bbox": [ 117, 268, 201, 279 ], "score": 1.0, "content": "13001–13008, 2020.", "type": "text" } ], "index": 14 } ], "index": 13 }, { "type": "text", "bbox": [ 106, 286, 504, 320 ], "lines": [ { "bbox": [ 105, 286, 506, 299 ], "spans": [ { "bbox": [ 105, 286, 506, 299 ], "score": 1.0, "content": "Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene", "type": "text" } ], "index": 15 }, { "bbox": [ 115, 298, 506, 311 ], "spans": [ { "bbox": [ 115, 298, 506, 311 ], "score": 1.0, "content": "parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 309, 278, 321 ], "spans": [ { "bbox": [ 115, 309, 278, 321 ], "score": 1.0, "content": "pattern recognition, pp. 633–641, 2017.", "type": "text" } ], "index": 17 } ], "index": 16 }, { "type": "text", "bbox": [ 106, 327, 504, 361 ], "lines": [ { "bbox": [ 106, 327, 505, 341 ], "spans": [ { "bbox": [ 106, 327, 505, 341 ], "score": 1.0, "content": "Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: De-", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 337, 505, 353 ], "spans": [ { "bbox": [ 115, 337, 505, 353 ], "score": 1.0, "content": "formable transformers for end-to-end object detection. In International Conference on Learning", "type": "text" } ], "index": 19 }, { "bbox": [ 116, 349, 496, 363 ], "spans": [ { "bbox": [ 116, 349, 419, 363 ], "score": 1.0, "content": "Representations, 2021. URL https://openreview.net/forum?id", "type": "text" }, { "bbox": [ 419, 351, 426, 359 ], "score": 0.29, "content": "{ . } =", "type": "inline_equation" }, { "bbox": [ 426, 349, 496, 363 ], "score": 1.0, "content": "gZ9hCDWe6ke.", "type": "text" } ], "index": 20 } ], "index": 19 } ], "page_idx": 12, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "13", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 108, 82, 504, 116 ], "lines": [ { "bbox": [ 106, 82, 505, 95 ], "spans": [ { "bbox": [ 106, 82, 505, 95 ], "score": 1.0, "content": "Adams Wei Yu, David Dohan, Minh-Thang Luong, Rui Zhao, Kai Chen, Mohammad Norouzi, and", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 506, 107 ], "spans": [ { "bbox": [ 115, 93, 506, 107 ], "score": 1.0, "content": "Quoc V Le. Qanet: Combining local convolution with global self-attention for reading compre-", "type": "text" } ], "index": 1 }, { "bbox": [ 115, 105, 316, 117 ], "spans": [ { "bbox": [ 115, 105, 316, 117 ], "score": 1.0, "content": "hension. arXiv preprint arXiv:1804.09541, 2018.", "type": "text" } ], "index": 2 } ], "index": 1, "bbox_fs": [ 106, 82, 506, 117 ] }, { "type": "text", "bbox": [ 109, 123, 503, 157 ], "lines": [ { "bbox": [ 106, 123, 504, 136 ], "spans": [ { "bbox": [ 106, 123, 504, 136 ], "score": 1.0, "content": "Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo.", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 134, 505, 147 ], "spans": [ { "bbox": [ 115, 134, 505, 147 ], "score": 1.0, "content": "Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceed-", "type": "text" } ], "index": 4 }, { "bbox": [ 115, 145, 483, 159 ], "spans": [ { "bbox": [ 115, 145, 483, 159 ], "score": 1.0, "content": "ings of the IEEE/CVF International Conference on Computer Vision, pp. 6023–6032, 2019.", "type": "text" } ], "index": 5 } ], "index": 4, "bbox_fs": [ 106, 123, 505, 159 ] }, { "type": "text", "bbox": [ 108, 163, 503, 198 ], "lines": [ { "bbox": [ 106, 163, 504, 177 ], "spans": [ { "bbox": [ 106, 163, 504, 177 ], "score": 1.0, "content": "Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond em-", "type": "text" } ], "index": 6 }, { "bbox": [ 115, 174, 505, 188 ], "spans": [ { "bbox": [ 115, 174, 505, 188 ], "score": 1.0, "content": "pirical risk minimization. In International Conference on Learning Representations, 2018. URL", "type": "text" } ], "index": 7 }, { "bbox": [ 115, 186, 367, 200 ], "spans": [ { "bbox": [ 115, 186, 302, 200 ], "score": 1.0, "content": "https://openreview.net/forum?id", "type": "text" }, { "bbox": [ 302, 187, 309, 196 ], "score": 0.5, "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", "type": "inline_equation" }, { "bbox": [ 309, 186, 367, 200 ], "score": 1.0, "content": "r1Ddp1-Rb.", "type": "text" } ], "index": 8 } ], "index": 7, "bbox_fs": [ 106, 163, 505, 200 ] }, { "type": "text", "bbox": [ 107, 204, 505, 239 ], "lines": [ { "bbox": [ 105, 204, 505, 218 ], "spans": [ { "bbox": [ 105, 204, 505, 218 ], "score": 1.0, "content": "Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei", "type": "text" } ], "index": 9 }, { "bbox": [ 116, 217, 505, 229 ], "spans": [ { "bbox": [ 116, 217, 505, 229 ], "score": 1.0, "content": "Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 227, 500, 239 ], "spans": [ { "bbox": [ 115, 227, 500, 239 ], "score": 1.0, "content": "a sequence-to-sequence perspective with transformers. arXiv preprint arXiv:2012.15840, 2020.", "type": "text" } ], "index": 11 } ], "index": 10, "bbox_fs": [ 105, 204, 505, 239 ] }, { "type": "text", "bbox": [ 106, 245, 504, 279 ], "lines": [ { "bbox": [ 104, 244, 505, 260 ], "spans": [ { "bbox": [ 104, 244, 505, 260 ], "score": 1.0, "content": "Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data aug-", "type": "text" } ], "index": 12 }, { "bbox": [ 115, 256, 506, 270 ], "spans": [ { "bbox": [ 115, 256, 506, 270 ], "score": 1.0, "content": "mentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pp.", "type": "text" } ], "index": 13 }, { "bbox": [ 117, 268, 201, 279 ], "spans": [ { "bbox": [ 117, 268, 201, 279 ], "score": 1.0, "content": "13001–13008, 2020.", "type": "text" } ], "index": 14 } ], "index": 13, "bbox_fs": [ 104, 244, 506, 279 ] }, { "type": "text", "bbox": [ 106, 286, 504, 320 ], "lines": [ { "bbox": [ 105, 286, 506, 299 ], "spans": [ { "bbox": [ 105, 286, 506, 299 ], "score": 1.0, "content": "Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene", "type": "text" } ], "index": 15 }, { "bbox": [ 115, 298, 506, 311 ], "spans": [ { "bbox": [ 115, 298, 506, 311 ], "score": 1.0, "content": "parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 309, 278, 321 ], "spans": [ { "bbox": [ 115, 309, 278, 321 ], "score": 1.0, "content": "pattern recognition, pp. 633–641, 2017.", "type": "text" } ], "index": 17 } ], "index": 16, "bbox_fs": [ 105, 286, 506, 321 ] }, { "type": "text", "bbox": [ 106, 327, 504, 361 ], "lines": [ { "bbox": [ 106, 327, 505, 341 ], "spans": [ { "bbox": [ 106, 327, 505, 341 ], "score": 1.0, "content": "Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: De-", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 337, 505, 353 ], "spans": [ { "bbox": [ 115, 337, 505, 353 ], "score": 1.0, "content": "formable transformers for end-to-end object detection. In International Conference on Learning", "type": "text" } ], "index": 19 }, { "bbox": [ 116, 349, 496, 363 ], "spans": [ { "bbox": [ 116, 349, 419, 363 ], "score": 1.0, "content": "Representations, 2021. URL https://openreview.net/forum?id", "type": "text" }, { "bbox": [ 419, 351, 426, 359 ], "score": 0.29, "content": "{ . } =", "type": "inline_equation" }, { "bbox": [ 426, 349, 496, 363 ], "score": 1.0, "content": "gZ9hCDWe6ke.", "type": "text" } ], "index": 20 } ], "index": 19, "bbox_fs": [ 106, 327, 505, 363 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 81, 285, 94 ], "lines": [ { "bbox": [ 105, 79, 287, 96 ], "spans": [ { "bbox": [ 105, 79, 287, 96 ], "score": 1.0, "content": "A TRANSLATION EQUIVARIANCE", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 106, 505, 206 ], "lines": [ { "bbox": [ 106, 106, 505, 119 ], "spans": [ { "bbox": [ 106, 106, 505, 119 ], "score": 1.0, "content": "The term translation-equivariance means the output feature maps can be equally translated with", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 118, 505, 130 ], "spans": [ { "bbox": [ 106, 118, 505, 130 ], "score": 1.0, "content": "the input signal. Imagine there is a person in the left-top of an image, if the person is moved to", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 128, 505, 142 ], "spans": [ { "bbox": [ 105, 128, 505, 142 ], "score": 1.0, "content": "the right-bottom, the output feature maps will change accordingly. This property is very important", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 139, 505, 153 ], "spans": [ { "bbox": [ 106, 139, 505, 153 ], "score": 1.0, "content": "to the success of convolution network. Convolution (ignoring paddings), RPE, and self-attention", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 150, 505, 163 ], "spans": [ { "bbox": [ 105, 150, 505, 163 ], "score": 1.0, "content": "are all translation-equivariant operations (regardless of their receptive field). It’s nontrivial to make", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 161, 504, 173 ], "spans": [ { "bbox": [ 106, 161, 504, 173 ], "score": 1.0, "content": "absolute positional encodings like DeiT (using learnable positional encoding) translation-equivariant", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 172, 506, 185 ], "spans": [ { "bbox": [ 105, 172, 506, 185 ], "score": 1.0, "content": "since different absolute positions will be added if the input signal is translated. Note that our method", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 183, 506, 195 ], "spans": [ { "bbox": [ 105, 183, 506, 195 ], "score": 1.0, "content": "is not strictly translation-equivariant because of the zero padding. Instead, it provides a kind of", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 192, 371, 209 ], "spans": [ { "bbox": [ 105, 192, 371, 209 ], "score": 1.0, "content": "stronger explicit bias towards the translation-equivariant property.", "type": "text" } ], "index": 9 } ], "index": 5 }, { "type": "title", "bbox": [ 108, 222, 244, 234 ], "lines": [ { "bbox": [ 105, 221, 245, 236 ], "spans": [ { "bbox": [ 105, 221, 245, 236 ], "score": 1.0, "content": "B EXPERIMENT DETAILS", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "title", "bbox": [ 108, 246, 290, 258 ], "lines": [ { "bbox": [ 105, 246, 291, 259 ], "spans": [ { "bbox": [ 105, 246, 291, 259 ], "score": 1.0, "content": "B.1 ARCHITECTURE VARIANTS OF CPVT", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "table", "bbox": [ 214, 318, 396, 361 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 270, 503, 305 ], "group_id": 0, "lines": [ { "bbox": [ 106, 270, 505, 284 ], "spans": [ { "bbox": [ 106, 270, 505, 284 ], "score": 1.0, "content": "Table 9. CPVT architecture variants. The larger model, CPVT-B, has the same architecture as ViT-", "type": "text" } ], "index": 12 }, { "bbox": [ 107, 281, 504, 294 ], "spans": [ { "bbox": [ 107, 281, 504, 294 ], "score": 1.0, "content": "B (Dosovitskiy et al., 2021) and DeiT-B (Touvron et al., 2020). CPVT-S and CPVT-Ti have the", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 292, 345, 306 ], "spans": [ { "bbox": [ 106, 292, 345, 306 ], "score": 1.0, "content": "same architecture as DeiT-small and DeiT-tiny respectively", "type": "text" } ], "index": 14 } ], "index": 13 }, { "type": "table_body", "bbox": [ 214, 318, 396, 361 ], "group_id": 0, "lines": [ { "bbox": [ 214, 318, 396, 361 ], "spans": [ { "bbox": [ 214, 318, 396, 361 ], "score": 0.974, "html": "
Model#channels#heads#layers#params
CPVT-Ti1923126M
CPVT-S38461222M
CPVT-B768121286M
", "type": "table", "image_path": "2d0f7ad56c59c2d4c1f15b5e9b951e289749c3731062b05327bb45b3c7e140dd.jpg" } ] } ], "index": 16, "virtual_lines": [ { "bbox": [ 214, 318, 396, 332.3333333333333 ], "spans": [], "index": 15 }, { "bbox": [ 214, 332.3333333333333, 396, 346.66666666666663 ], "spans": [], "index": 16 }, { "bbox": [ 214, 346.66666666666663, 396, 360.99999999999994 ], "spans": [], "index": 17 } ] } ], "index": 14.5 }, { "type": "title", "bbox": [ 108, 388, 284, 399 ], "lines": [ { "bbox": [ 105, 388, 286, 400 ], "spans": [ { "bbox": [ 105, 388, 286, 400 ], "score": 1.0, "content": "B.2 THE HYPERPARAMETERS OF CPVT", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 108, 408, 506, 442 ], "lines": [ { "bbox": [ 105, 408, 505, 421 ], "spans": [ { "bbox": [ 105, 408, 505, 421 ], "score": 1.0, "content": "As for the ImageNet classification task, we use exactly the same hyperparameters as DeiT except for", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 420, 505, 432 ], "spans": [ { "bbox": [ 105, 420, 505, 432 ], "score": 1.0, "content": "the base model because it is not always stably trained using AdamW. The detailed setting is shown", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 430, 156, 443 ], "spans": [ { "bbox": [ 105, 430, 156, 443 ], "score": 1.0, "content": "in Table 10.", "type": "text" } ], "index": 21 } ], "index": 20 }, { "type": "table", "bbox": [ 164, 471, 446, 634 ], "blocks": [ { "type": "table_caption", "bbox": [ 198, 450, 413, 463 ], "group_id": 1, "lines": [ { "bbox": [ 197, 450, 414, 463 ], "spans": [ { "bbox": [ 197, 450, 414, 463 ], "score": 1.0, "content": "Table 10. Hyper-parameters for ViT, DeiT and CPVT", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "table_body", "bbox": [ 164, 471, 446, 634 ], "group_id": 1, "lines": [ { "bbox": [ 164, 471, 446, 634 ], "spans": [ { "bbox": [ 164, 471, 446, 634 ], "score": 0.983, "html": "
MethodsViTDeiTCPVT
Epochs Batch size300 4096300 1024300 1024
OptimizerAdamWAdamWLAMB
Learning rate decaycosinecosinecosine
Weight decay0.30.050.05
Warmup epochs3.455
Label smoothing ε (Szegedy et al., 2016)X0.1 X0.1
Dropout (Srivastava et al.,2014)0.1X
Stoch.Depth (Huang et al., 2016)X0.1 √0.1
Repeated Aug (Hoffer et al., 2020)XX√
Gradient Clip.√9/0.5X
Rand Augment (Cubuk et al., 2020)X9/0.5
Mixup prob. (Zhang et al.,2018)X0.80.8
Cutmix prob. (Yun et al., 2019)X1.01.0
Erasing prob. (Zhong et al.,2020)X0.250.25
", "type": "table", "image_path": "164ce8cb1f4d7b162d7d0eb721d68b501123589b84ae390079aa08d3889b476d.jpg" } ] } ], "index": 24, "virtual_lines": [ { "bbox": [ 164, 471, 446, 525.3333333333334 ], "spans": [], "index": 23 }, { "bbox": [ 164, 525.3333333333334, 446, 579.6666666666667 ], "spans": [], "index": 24 }, { "bbox": [ 164, 579.6666666666667, 446, 634.0000000000001 ], "spans": [], "index": 25 } ] } ], "index": 23.0 }, { "type": "title", "bbox": [ 108, 655, 276, 667 ], "lines": [ { "bbox": [ 105, 655, 278, 668 ], "spans": [ { "bbox": [ 105, 655, 278, 668 ], "score": 1.0, "content": "B.3 IMPORTANCE OF ZERO PADDINGS", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 107, 676, 505, 732 ], "lines": [ { "bbox": [ 107, 677, 505, 689 ], "spans": [ { "bbox": [ 107, 677, 505, 689 ], "score": 1.0, "content": "We design an experiment to verify the importance of the zero paddings, which can help the model", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 687, 505, 700 ], "spans": [ { "bbox": [ 106, 687, 505, 700 ], "score": 1.0, "content": "infer the absolute positional information. Specifically, we use CPVT-S and simply remove the zero", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 699, 506, 711 ], "spans": [ { "bbox": [ 105, 699, 506, 711 ], "score": 1.0, "content": "paddings from CPVT while keeping all other settings unchanged. Table 11 shows that this can", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 708, 505, 724 ], "spans": [ { "bbox": [ 105, 708, 154, 724 ], "score": 1.0, "content": "only obtain", "type": "text" }, { "bbox": [ 154, 710, 181, 720 ], "score": 0.88, "content": "7 0 . 5 \\%", "type": "inline_equation" }, { "bbox": [ 182, 708, 505, 724 ], "score": 1.0, "content": ", which indicates that the zero paddings and absolute positional information play", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 721, 258, 733 ], "spans": [ { "bbox": [ 106, 721, 258, 733 ], "score": 1.0, "content": "important roles in classifying objects.", "type": "text" } ], "index": 31 } ], "index": 29 } ], "page_idx": 13, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "14", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 81, 285, 94 ], "lines": [ { "bbox": [ 105, 79, 287, 96 ], "spans": [ { "bbox": [ 105, 79, 287, 96 ], "score": 1.0, "content": "A TRANSLATION EQUIVARIANCE", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 106, 505, 206 ], "lines": [ { "bbox": [ 106, 106, 505, 119 ], "spans": [ { "bbox": [ 106, 106, 505, 119 ], "score": 1.0, "content": "The term translation-equivariance means the output feature maps can be equally translated with", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 118, 505, 130 ], "spans": [ { "bbox": [ 106, 118, 505, 130 ], "score": 1.0, "content": "the input signal. Imagine there is a person in the left-top of an image, if the person is moved to", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 128, 505, 142 ], "spans": [ { "bbox": [ 105, 128, 505, 142 ], "score": 1.0, "content": "the right-bottom, the output feature maps will change accordingly. This property is very important", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 139, 505, 153 ], "spans": [ { "bbox": [ 106, 139, 505, 153 ], "score": 1.0, "content": "to the success of convolution network. Convolution (ignoring paddings), RPE, and self-attention", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 150, 505, 163 ], "spans": [ { "bbox": [ 105, 150, 505, 163 ], "score": 1.0, "content": "are all translation-equivariant operations (regardless of their receptive field). It’s nontrivial to make", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 161, 504, 173 ], "spans": [ { "bbox": [ 106, 161, 504, 173 ], "score": 1.0, "content": "absolute positional encodings like DeiT (using learnable positional encoding) translation-equivariant", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 172, 506, 185 ], "spans": [ { "bbox": [ 105, 172, 506, 185 ], "score": 1.0, "content": "since different absolute positions will be added if the input signal is translated. Note that our method", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 183, 506, 195 ], "spans": [ { "bbox": [ 105, 183, 506, 195 ], "score": 1.0, "content": "is not strictly translation-equivariant because of the zero padding. Instead, it provides a kind of", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 192, 371, 209 ], "spans": [ { "bbox": [ 105, 192, 371, 209 ], "score": 1.0, "content": "stronger explicit bias towards the translation-equivariant property.", "type": "text" } ], "index": 9 } ], "index": 5, "bbox_fs": [ 105, 106, 506, 209 ] }, { "type": "title", "bbox": [ 108, 222, 244, 234 ], "lines": [ { "bbox": [ 105, 221, 245, 236 ], "spans": [ { "bbox": [ 105, 221, 245, 236 ], "score": 1.0, "content": "B EXPERIMENT DETAILS", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "title", "bbox": [ 108, 246, 290, 258 ], "lines": [ { "bbox": [ 105, 246, 291, 259 ], "spans": [ { "bbox": [ 105, 246, 291, 259 ], "score": 1.0, "content": "B.1 ARCHITECTURE VARIANTS OF CPVT", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "table", "bbox": [ 214, 318, 396, 361 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 270, 503, 305 ], "group_id": 0, "lines": [ { "bbox": [ 106, 270, 505, 284 ], "spans": [ { "bbox": [ 106, 270, 505, 284 ], "score": 1.0, "content": "Table 9. CPVT architecture variants. The larger model, CPVT-B, has the same architecture as ViT-", "type": "text" } ], "index": 12 }, { "bbox": [ 107, 281, 504, 294 ], "spans": [ { "bbox": [ 107, 281, 504, 294 ], "score": 1.0, "content": "B (Dosovitskiy et al., 2021) and DeiT-B (Touvron et al., 2020). CPVT-S and CPVT-Ti have the", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 292, 345, 306 ], "spans": [ { "bbox": [ 106, 292, 345, 306 ], "score": 1.0, "content": "same architecture as DeiT-small and DeiT-tiny respectively", "type": "text" } ], "index": 14 } ], "index": 13 }, { "type": "table_body", "bbox": [ 214, 318, 396, 361 ], "group_id": 0, "lines": [ { "bbox": [ 214, 318, 396, 361 ], "spans": [ { "bbox": [ 214, 318, 396, 361 ], "score": 0.974, "html": "
Model#channels#heads#layers#params
CPVT-Ti1923126M
CPVT-S38461222M
CPVT-B768121286M
", "type": "table", "image_path": "2d0f7ad56c59c2d4c1f15b5e9b951e289749c3731062b05327bb45b3c7e140dd.jpg" } ] } ], "index": 16, "virtual_lines": [ { "bbox": [ 214, 318, 396, 332.3333333333333 ], "spans": [], "index": 15 }, { "bbox": [ 214, 332.3333333333333, 396, 346.66666666666663 ], "spans": [], "index": 16 }, { "bbox": [ 214, 346.66666666666663, 396, 360.99999999999994 ], "spans": [], "index": 17 } ] } ], "index": 14.5 }, { "type": "title", "bbox": [ 108, 388, 284, 399 ], "lines": [ { "bbox": [ 105, 388, 286, 400 ], "spans": [ { "bbox": [ 105, 388, 286, 400 ], "score": 1.0, "content": "B.2 THE HYPERPARAMETERS OF CPVT", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 108, 408, 506, 442 ], "lines": [ { "bbox": [ 105, 408, 505, 421 ], "spans": [ { "bbox": [ 105, 408, 505, 421 ], "score": 1.0, "content": "As for the ImageNet classification task, we use exactly the same hyperparameters as DeiT except for", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 420, 505, 432 ], "spans": [ { "bbox": [ 105, 420, 505, 432 ], "score": 1.0, "content": "the base model because it is not always stably trained using AdamW. The detailed setting is shown", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 430, 156, 443 ], "spans": [ { "bbox": [ 105, 430, 156, 443 ], "score": 1.0, "content": "in Table 10.", "type": "text" } ], "index": 21 } ], "index": 20, "bbox_fs": [ 105, 408, 505, 443 ] }, { "type": "table", "bbox": [ 164, 471, 446, 634 ], "blocks": [ { "type": "table_caption", "bbox": [ 198, 450, 413, 463 ], "group_id": 1, "lines": [ { "bbox": [ 197, 450, 414, 463 ], "spans": [ { "bbox": [ 197, 450, 414, 463 ], "score": 1.0, "content": "Table 10. Hyper-parameters for ViT, DeiT and CPVT", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "table_body", "bbox": [ 164, 471, 446, 634 ], "group_id": 1, "lines": [ { "bbox": [ 164, 471, 446, 634 ], "spans": [ { "bbox": [ 164, 471, 446, 634 ], "score": 0.983, "html": "
MethodsViTDeiTCPVT
Epochs Batch size300 4096300 1024300 1024
OptimizerAdamWAdamWLAMB
Learning rate decaycosinecosinecosine
Weight decay0.30.050.05
Warmup epochs3.455
Label smoothing ε (Szegedy et al., 2016)X0.1 X0.1
Dropout (Srivastava et al.,2014)0.1X
Stoch.Depth (Huang et al., 2016)X0.1 √0.1
Repeated Aug (Hoffer et al., 2020)XX√
Gradient Clip.√9/0.5X
Rand Augment (Cubuk et al., 2020)X9/0.5
Mixup prob. (Zhang et al.,2018)X0.80.8
Cutmix prob. (Yun et al., 2019)X1.01.0
Erasing prob. (Zhong et al.,2020)X0.250.25
", "type": "table", "image_path": "164ce8cb1f4d7b162d7d0eb721d68b501123589b84ae390079aa08d3889b476d.jpg" } ] } ], "index": 24, "virtual_lines": [ { "bbox": [ 164, 471, 446, 525.3333333333334 ], "spans": [], "index": 23 }, { "bbox": [ 164, 525.3333333333334, 446, 579.6666666666667 ], "spans": [], "index": 24 }, { "bbox": [ 164, 579.6666666666667, 446, 634.0000000000001 ], "spans": [], "index": 25 } ] } ], "index": 23.0 }, { "type": "title", "bbox": [ 108, 655, 276, 667 ], "lines": [ { "bbox": [ 105, 655, 278, 668 ], "spans": [ { "bbox": [ 105, 655, 278, 668 ], "score": 1.0, "content": "B.3 IMPORTANCE OF ZERO PADDINGS", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 107, 676, 505, 732 ], "lines": [ { "bbox": [ 107, 677, 505, 689 ], "spans": [ { "bbox": [ 107, 677, 505, 689 ], "score": 1.0, "content": "We design an experiment to verify the importance of the zero paddings, which can help the model", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 687, 505, 700 ], "spans": [ { "bbox": [ 106, 687, 505, 700 ], "score": 1.0, "content": "infer the absolute positional information. Specifically, we use CPVT-S and simply remove the zero", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 699, 506, 711 ], "spans": [ { "bbox": [ 105, 699, 506, 711 ], "score": 1.0, "content": "paddings from CPVT while keeping all other settings unchanged. Table 11 shows that this can", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 708, 505, 724 ], "spans": [ { "bbox": [ 105, 708, 154, 724 ], "score": 1.0, "content": "only obtain", "type": "text" }, { "bbox": [ 154, 710, 181, 720 ], "score": 0.88, "content": "7 0 . 5 \\%", "type": "inline_equation" }, { "bbox": [ 182, 708, 505, 724 ], "score": 1.0, "content": ", which indicates that the zero paddings and absolute positional information play", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 721, 258, 733 ], "spans": [ { "bbox": [ 106, 721, 258, 733 ], "score": 1.0, "content": "important roles in classifying objects.", "type": "text" } ], "index": 31 } ], "index": 29, "bbox_fs": [ 105, 677, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 196, 106, 414, 139 ], "blocks": [ { "type": "table_caption", "bbox": [ 152, 80, 459, 92 ], "group_id": 0, "lines": [ { "bbox": [ 151, 78, 459, 95 ], "spans": [ { "bbox": [ 151, 78, 459, 95 ], "score": 1.0, "content": "Table 11. Ablation study on ImageNet performance w/ or w/o zero paddings", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "table_body", "bbox": [ 196, 106, 414, 139 ], "group_id": 0, "lines": [ { "bbox": [ 196, 106, 414, 139 ], "spans": [ { "bbox": [ 196, 106, 414, 139 ], "score": 0.969, "html": "
ModelPaddingTop-1 Acc(%)Top-5 Acc(%)
CPVT-Ti√72.491.2
X70.589.8
", "type": "table", "image_path": "88c034eed6fc8b897c16fd609b0747d75c5d2d484b983f1b2b68b89b5089c291.jpg" } ] } ], "index": 1.5, "virtual_lines": [ { "bbox": [ 196, 106, 414, 122.5 ], "spans": [], "index": 1 }, { "bbox": [ 196, 122.5, 414, 139.0 ], "spans": [], "index": 2 } ] } ], "index": 0.75 }, { "type": "title", "bbox": [ 108, 167, 280, 178 ], "lines": [ { "bbox": [ 105, 166, 281, 180 ], "spans": [ { "bbox": [ 105, 166, 281, 180 ], "score": 1.0, "content": "B.4 SINGLE PEG VS. MULTIPLE PEGS", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 106, 189, 506, 255 ], "lines": [ { "bbox": [ 106, 188, 506, 201 ], "spans": [ { "bbox": [ 106, 188, 506, 201 ], "score": 1.0, "content": "We further evaluate whether or not using multi-position encodings can benefit the performance in", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 199, 506, 212 ], "spans": [ { "bbox": [ 106, 199, 233, 212 ], "score": 1.0, "content": "Table 12. Notice we denote by", "type": "text" }, { "bbox": [ 234, 200, 247, 211 ], "score": 0.46, "content": "i \\mathrm { - } j", "type": "inline_equation" }, { "bbox": [ 247, 199, 453, 212 ], "score": 1.0, "content": "the inserted positions of PEG which start from the", "type": "text" }, { "bbox": [ 453, 200, 459, 209 ], "score": 0.62, "content": "i", "type": "inline_equation" }, { "bbox": [ 459, 199, 506, 212 ], "score": 1.0, "content": "-th encoder", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 210, 506, 223 ], "spans": [ { "bbox": [ 105, 210, 163, 223 ], "score": 1.0, "content": "and end at the", "type": "text" }, { "bbox": [ 163, 211, 184, 222 ], "score": 0.89, "content": "j - 1", "type": "inline_equation" }, { "bbox": [ 185, 210, 506, 223 ], "score": 1.0, "content": "-th one (inclusion). By inserting PEGs to five positions, the top-1 accuracy of the", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 221, 506, 234 ], "spans": [ { "bbox": [ 106, 221, 200, 234 ], "score": 1.0, "content": "tiny model can achieve", "type": "text" }, { "bbox": [ 201, 222, 228, 232 ], "score": 0.88, "content": "7 3 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 228, 221, 352, 234 ], "score": 1.0, "content": ", which surpasses DeiT-tiny by", "type": "text" }, { "bbox": [ 352, 222, 374, 232 ], "score": 0.86, "content": "1 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 374, 221, 506, 234 ], "score": 1.0, "content": ". Similarly, CPVT-S can achieve", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 232, 506, 246 ], "spans": [ { "bbox": [ 106, 233, 133, 243 ], "score": 0.86, "content": "8 0 . 5 \\%", "type": "inline_equation" }, { "bbox": [ 134, 232, 506, 246 ], "score": 1.0, "content": ". It turns out more PEGs do help, but up to a level where more PEGs become incremental (0-5", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 244, 148, 255 ], "spans": [ { "bbox": [ 106, 244, 148, 255 ], "score": 1.0, "content": "vs. 0-11).", "type": "text" } ], "index": 9 } ], "index": 6.5 }, { "type": "table", "bbox": [ 210, 293, 401, 376 ], "blocks": [ { "type": "table_caption", "bbox": [ 186, 267, 424, 279 ], "group_id": 1, "lines": [ { "bbox": [ 186, 265, 424, 281 ], "spans": [ { "bbox": [ 186, 265, 424, 281 ], "score": 1.0, "content": "Table 12. CPVT’s sensitivity to number of plugin positions", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "table_body", "bbox": [ 210, 293, 401, 376 ], "group_id": 1, "lines": [ { "bbox": [ 210, 293, 401, 376 ], "spans": [ { "bbox": [ 210, 293, 401, 376 ], "score": 0.978, "html": "
PositionsModelParams (M)Top-1 Acc (%）Top-5 Acc (%)
0-1tiny5.772.491.2
0-5tiny5.973.491.8
0-11tiny6.173.491.8
0-1 0-5small small22.0 22.979.9 80.595.0 95.2
", "type": "table", "image_path": "0b5b1050203055bb3eb14543fdc8cd11d9192d3c5b12399d875656aeaac37c17.jpg" } ] } ], "index": 13.5, "virtual_lines": [ { "bbox": [ 210, 293, 401, 306.8333333333333 ], "spans": [], "index": 11 }, { "bbox": [ 210, 306.8333333333333, 401, 320.66666666666663 ], "spans": [], "index": 12 }, { "bbox": [ 210, 320.66666666666663, 401, 334.49999999999994 ], "spans": [], "index": 13 }, { "bbox": [ 210, 334.49999999999994, 401, 348.33333333333326 ], "spans": [], "index": 14 }, { "bbox": [ 210, 348.33333333333326, 401, 362.1666666666666 ], "spans": [], "index": 15 }, { "bbox": [ 210, 362.1666666666666, 401, 375.9999999999999 ], "spans": [], "index": 16 } ] } ], "index": 11.75 }, { "type": "title", "bbox": [ 106, 406, 348, 417 ], "lines": [ { "bbox": [ 105, 406, 349, 419 ], "spans": [ { "bbox": [ 105, 406, 349, 419 ], "score": 1.0, "content": "B.5 CLASSFICATION EVALUATION OF SWIN WITH PEG", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 104, 428, 503, 451 ], "lines": [ { "bbox": [ 106, 427, 505, 441 ], "spans": [ { "bbox": [ 106, 427, 505, 441 ], "score": 1.0, "content": "We show the validation curves when training Swin (Liu et al., 2021) equipped with PEG in Figure 4.", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 438, 398, 452 ], "spans": [ { "bbox": [ 105, 438, 219, 452 ], "score": 1.0, "content": "It can boost Swin-tiny from", "type": "text" }, { "bbox": [ 219, 439, 251, 450 ], "score": 0.87, "content": "8 1 . 1 0 \\%", "type": "inline_equation" }, { "bbox": [ 252, 438, 262, 452 ], "score": 1.0, "content": "to", "type": "text" }, { "bbox": [ 263, 439, 295, 450 ], "score": 0.85, "content": "8 2 . 2 5 \\%", "type": "inline_equation" }, { "bbox": [ 296, 439, 339, 451 ], "score": 0.88, "content": "( + 1 . 1 5 \\% \\uparrow )", "type": "inline_equation" }, { "bbox": [ 339, 438, 398, 452 ], "score": 1.0, "content": "on ImageNet.", "type": "text" } ], "index": 19 } ], "index": 18.5 }, { "type": "image", "bbox": [ 191, 463, 420, 554 ], "blocks": [ { "type": "image_body", "bbox": [ 191, 463, 420, 554 ], "group_id": 0, "lines": [ { "bbox": [ 191, 463, 420, 554 ], "spans": [ { "bbox": [ 191, 463, 420, 554 ], "score": 0.968, "type": "image", "image_path": "37115738de8c41b823e0aa8bb65e1c60638e6d07fe274cb734b5b2f7bcc4fd09.jpg" } ] } ], "index": 23, "virtual_lines": [ { "bbox": [ 191, 463, 420, 476.0 ], "spans": [], "index": 20 }, { "bbox": [ 191, 476.0, 420, 489.0 ], "spans": [], "index": 21 }, { "bbox": [ 191, 489.0, 420, 502.0 ], "spans": [], "index": 22 }, { "bbox": [ 191, 502.0, 420, 515.0 ], "spans": [], "index": 23 }, { "bbox": [ 191, 515.0, 420, 528.0 ], "spans": [], "index": 24 }, { "bbox": [ 191, 528.0, 420, 541.0 ], "spans": [], "index": 25 }, { "bbox": [ 191, 541.0, 420, 554.0 ], "spans": [], "index": 26 } ] }, { "type": "image_caption", "bbox": [ 170, 569, 441, 581 ], "group_id": 0, "lines": [ { "bbox": [ 168, 567, 442, 584 ], "spans": [ { "bbox": [ 168, 567, 369, 584 ], "score": 1.0, "content": "Figure 4. CPE boosts Swin Tiny on ImageNet by", "type": "text" }, { "bbox": [ 369, 569, 396, 580 ], "score": 0.86, "content": "1 . 1 5 \\%", "type": "inline_equation" }, { "bbox": [ 396, 567, 442, 584 ], "score": 1.0, "content": "top-1 Acc.", "type": "text" } ], "index": 27 } ], "index": 27 } ], "index": 25.0 }, { "type": "title", "bbox": [ 106, 606, 349, 618 ], "lines": [ { "bbox": [ 105, 606, 350, 619 ], "spans": [ { "bbox": [ 105, 606, 350, 619 ], "score": 1.0, "content": "B.6 EVALUATION ON SEGMENTATION AND DETECTION", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "text", "bbox": [ 106, 628, 505, 672 ], "lines": [ { "bbox": [ 105, 628, 506, 641 ], "spans": [ { "bbox": [ 105, 628, 506, 641 ], "score": 1.0, "content": "Semantic segmentation on ADE20K. We evaluate the performance of PEG on the ADE20K", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 640, 505, 651 ], "spans": [ { "bbox": [ 106, 640, 505, 651 ], "score": 1.0, "content": "(Zhou et al., 2017) segmentation task. Based on the Semantic FPN framework (Kirillov et al.,", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 650, 505, 663 ], "spans": [ { "bbox": [ 105, 650, 505, 663 ], "score": 1.0, "content": "2019), PVT achieves much better results than ResNet (He et al., 2016) baselines. Under carefully", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 661, 367, 673 ], "spans": [ { "bbox": [ 105, 661, 315, 673 ], "score": 1.0, "content": "controlled settings, PEG further boosts PVT-tiny by", "type": "text" }, { "bbox": [ 315, 661, 337, 672 ], "score": 0.84, "content": "3 . 1 \\%", "type": "inline_equation" }, { "bbox": [ 338, 661, 367, 673 ], "score": 1.0, "content": "mIoU.", "type": "text" } ], "index": 32 } ], "index": 30.5 }, { "type": "text", "bbox": [ 107, 687, 505, 731 ], "lines": [ { "bbox": [ 106, 687, 505, 700 ], "spans": [ { "bbox": [ 106, 687, 505, 700 ], "score": 1.0, "content": "Object detection on COCO. We also perform controlled experiments with the RetinaNet (Lin", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 699, 505, 711 ], "spans": [ { "bbox": [ 105, 699, 505, 711 ], "score": 1.0, "content": "et al., 2017) framework on the COCO detection task. The results are shown in Table 13. In the", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 710, 505, 722 ], "spans": [ { "bbox": [ 105, 710, 142, 722 ], "score": 1.0, "content": "standard", "type": "text" }, { "bbox": [ 143, 710, 157, 720 ], "score": 0.86, "content": "1 \\times", "type": "inline_equation" }, { "bbox": [ 157, 710, 309, 722 ], "score": 1.0, "content": "schedule, PEG improves PVT-tiny by", "type": "text" }, { "bbox": [ 309, 710, 332, 721 ], "score": 0.9, "content": "2 . 0 \\%", "type": "inline_equation" }, { "bbox": [ 332, 710, 405, 722 ], "score": 1.0, "content": "mAP. PEG brings", "type": "text" }, { "bbox": [ 406, 710, 428, 720 ], "score": 0.84, "content": "2 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 428, 710, 505, 722 ], "score": 1.0, "content": "higher mAP under", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 721, 178, 731 ], "spans": [ { "bbox": [ 106, 721, 121, 731 ], "score": 1.0, "content": "the", "type": "text" }, { "bbox": [ 121, 721, 138, 731 ], "score": 0.84, "content": "3 \\times", "type": "inline_equation" }, { "bbox": [ 139, 721, 178, 731 ], "score": 1.0, "content": "schedule.", "type": "text" } ], "index": 36 } ], "index": 34.5 } ], "page_idx": 14, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "15", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 196, 106, 414, 139 ], "blocks": [ { "type": "table_caption", "bbox": [ 152, 80, 459, 92 ], "group_id": 0, "lines": [ { "bbox": [ 151, 78, 459, 95 ], "spans": [ { "bbox": [ 151, 78, 459, 95 ], "score": 1.0, "content": "Table 11. Ablation study on ImageNet performance w/ or w/o zero paddings", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "table_body", "bbox": [ 196, 106, 414, 139 ], "group_id": 0, "lines": [ { "bbox": [ 196, 106, 414, 139 ], "spans": [ { "bbox": [ 196, 106, 414, 139 ], "score": 0.969, "html": "
ModelPaddingTop-1 Acc(%)Top-5 Acc(%)
CPVT-Ti√72.491.2
X70.589.8
", "type": "table", "image_path": "88c034eed6fc8b897c16fd609b0747d75c5d2d484b983f1b2b68b89b5089c291.jpg" } ] } ], "index": 1.5, "virtual_lines": [ { "bbox": [ 196, 106, 414, 122.5 ], "spans": [], "index": 1 }, { "bbox": [ 196, 122.5, 414, 139.0 ], "spans": [], "index": 2 } ] } ], "index": 0.75 }, { "type": "title", "bbox": [ 108, 167, 280, 178 ], "lines": [ { "bbox": [ 105, 166, 281, 180 ], "spans": [ { "bbox": [ 105, 166, 281, 180 ], "score": 1.0, "content": "B.4 SINGLE PEG VS. MULTIPLE PEGS", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 106, 189, 506, 255 ], "lines": [ { "bbox": [ 106, 188, 506, 201 ], "spans": [ { "bbox": [ 106, 188, 506, 201 ], "score": 1.0, "content": "We further evaluate whether or not using multi-position encodings can benefit the performance in", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 199, 506, 212 ], "spans": [ { "bbox": [ 106, 199, 233, 212 ], "score": 1.0, "content": "Table 12. Notice we denote by", "type": "text" }, { "bbox": [ 234, 200, 247, 211 ], "score": 0.46, "content": "i \\mathrm { - } j", "type": "inline_equation" }, { "bbox": [ 247, 199, 453, 212 ], "score": 1.0, "content": "the inserted positions of PEG which start from the", "type": "text" }, { "bbox": [ 453, 200, 459, 209 ], "score": 0.62, "content": "i", "type": "inline_equation" }, { "bbox": [ 459, 199, 506, 212 ], "score": 1.0, "content": "-th encoder", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 210, 506, 223 ], "spans": [ { "bbox": [ 105, 210, 163, 223 ], "score": 1.0, "content": "and end at the", "type": "text" }, { "bbox": [ 163, 211, 184, 222 ], "score": 0.89, "content": "j - 1", "type": "inline_equation" }, { "bbox": [ 185, 210, 506, 223 ], "score": 1.0, "content": "-th one (inclusion). By inserting PEGs to five positions, the top-1 accuracy of the", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 221, 506, 234 ], "spans": [ { "bbox": [ 106, 221, 200, 234 ], "score": 1.0, "content": "tiny model can achieve", "type": "text" }, { "bbox": [ 201, 222, 228, 232 ], "score": 0.88, "content": "7 3 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 228, 221, 352, 234 ], "score": 1.0, "content": ", which surpasses DeiT-tiny by", "type": "text" }, { "bbox": [ 352, 222, 374, 232 ], "score": 0.86, "content": "1 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 374, 221, 506, 234 ], "score": 1.0, "content": ". Similarly, CPVT-S can achieve", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 232, 506, 246 ], "spans": [ { "bbox": [ 106, 233, 133, 243 ], "score": 0.86, "content": "8 0 . 5 \\%", "type": "inline_equation" }, { "bbox": [ 134, 232, 506, 246 ], "score": 1.0, "content": ". It turns out more PEGs do help, but up to a level where more PEGs become incremental (0-5", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 244, 148, 255 ], "spans": [ { "bbox": [ 106, 244, 148, 255 ], "score": 1.0, "content": "vs. 0-11).", "type": "text" } ], "index": 9 } ], "index": 6.5, "bbox_fs": [ 105, 188, 506, 255 ] }, { "type": "table", "bbox": [ 210, 293, 401, 376 ], "blocks": [ { "type": "table_caption", "bbox": [ 186, 267, 424, 279 ], "group_id": 1, "lines": [ { "bbox": [ 186, 265, 424, 281 ], "spans": [ { "bbox": [ 186, 265, 424, 281 ], "score": 1.0, "content": "Table 12. CPVT’s sensitivity to number of plugin positions", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "table_body", "bbox": [ 210, 293, 401, 376 ], "group_id": 1, "lines": [ { "bbox": [ 210, 293, 401, 376 ], "spans": [ { "bbox": [ 210, 293, 401, 376 ], "score": 0.978, "html": "
PositionsModelParams (M)Top-1 Acc (%）Top-5 Acc (%)
0-1tiny5.772.491.2
0-5tiny5.973.491.8
0-11tiny6.173.491.8
0-1 0-5small small22.0 22.979.9 80.595.0 95.2
", "type": "table", "image_path": "0b5b1050203055bb3eb14543fdc8cd11d9192d3c5b12399d875656aeaac37c17.jpg" } ] } ], "index": 13.5, "virtual_lines": [ { "bbox": [ 210, 293, 401, 306.8333333333333 ], "spans": [], "index": 11 }, { "bbox": [ 210, 306.8333333333333, 401, 320.66666666666663 ], "spans": [], "index": 12 }, { "bbox": [ 210, 320.66666666666663, 401, 334.49999999999994 ], "spans": [], "index": 13 }, { "bbox": [ 210, 334.49999999999994, 401, 348.33333333333326 ], "spans": [], "index": 14 }, { "bbox": [ 210, 348.33333333333326, 401, 362.1666666666666 ], "spans": [], "index": 15 }, { "bbox": [ 210, 362.1666666666666, 401, 375.9999999999999 ], "spans": [], "index": 16 } ] } ], "index": 11.75 }, { "type": "title", "bbox": [ 106, 406, 348, 417 ], "lines": [ { "bbox": [ 105, 406, 349, 419 ], "spans": [ { "bbox": [ 105, 406, 349, 419 ], "score": 1.0, "content": "B.5 CLASSFICATION EVALUATION OF SWIN WITH PEG", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "list", "bbox": [ 104, 428, 503, 451 ], "lines": [ { "bbox": [ 106, 427, 505, 441 ], "spans": [ { "bbox": [ 106, 427, 505, 441 ], "score": 1.0, "content": "We show the validation curves when training Swin (Liu et al., 2021) equipped with PEG in Figure 4.", "type": "text" } ], "index": 18, "is_list_end_line": true }, { "bbox": [ 105, 438, 398, 452 ], "spans": [ { "bbox": [ 105, 438, 219, 452 ], "score": 1.0, "content": "It can boost Swin-tiny from", "type": "text" }, { "bbox": [ 219, 439, 251, 450 ], "score": 0.87, "content": "8 1 . 1 0 \\%", "type": "inline_equation" }, { "bbox": [ 252, 438, 262, 452 ], "score": 1.0, "content": "to", "type": "text" }, { "bbox": [ 263, 439, 295, 450 ], "score": 0.85, "content": "8 2 . 2 5 \\%", "type": "inline_equation" }, { "bbox": [ 296, 439, 339, 451 ], "score": 0.88, "content": "( + 1 . 1 5 \\% \\uparrow )", "type": "inline_equation" }, { "bbox": [ 339, 438, 398, 452 ], "score": 1.0, "content": "on ImageNet.", "type": "text" } ], "index": 19, "is_list_start_line": true, "is_list_end_line": true } ], "index": 18.5, "bbox_fs": [ 105, 427, 505, 452 ] }, { "type": "image", "bbox": [ 191, 463, 420, 554 ], "blocks": [ { "type": "image_body", "bbox": [ 191, 463, 420, 554 ], "group_id": 0, "lines": [ { "bbox": [ 191, 463, 420, 554 ], "spans": [ { "bbox": [ 191, 463, 420, 554 ], "score": 0.968, "type": "image", "image_path": "37115738de8c41b823e0aa8bb65e1c60638e6d07fe274cb734b5b2f7bcc4fd09.jpg" } ] } ], "index": 23, "virtual_lines": [ { "bbox": [ 191, 463, 420, 476.0 ], "spans": [], "index": 20 }, { "bbox": [ 191, 476.0, 420, 489.0 ], "spans": [], "index": 21 }, { "bbox": [ 191, 489.0, 420, 502.0 ], "spans": [], "index": 22 }, { "bbox": [ 191, 502.0, 420, 515.0 ], "spans": [], "index": 23 }, { "bbox": [ 191, 515.0, 420, 528.0 ], "spans": [], "index": 24 }, { "bbox": [ 191, 528.0, 420, 541.0 ], "spans": [], "index": 25 }, { "bbox": [ 191, 541.0, 420, 554.0 ], "spans": [], "index": 26 } ] }, { "type": "image_caption", "bbox": [ 170, 569, 441, 581 ], "group_id": 0, "lines": [ { "bbox": [ 168, 567, 442, 584 ], "spans": [ { "bbox": [ 168, 567, 369, 584 ], "score": 1.0, "content": "Figure 4. CPE boosts Swin Tiny on ImageNet by", "type": "text" }, { "bbox": [ 369, 569, 396, 580 ], "score": 0.86, "content": "1 . 1 5 \\%", "type": "inline_equation" }, { "bbox": [ 396, 567, 442, 584 ], "score": 1.0, "content": "top-1 Acc.", "type": "text" } ], "index": 27 } ], "index": 27 } ], "index": 25.0 }, { "type": "title", "bbox": [ 106, 606, 349, 618 ], "lines": [ { "bbox": [ 105, 606, 350, 619 ], "spans": [ { "bbox": [ 105, 606, 350, 619 ], "score": 1.0, "content": "B.6 EVALUATION ON SEGMENTATION AND DETECTION", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "text", "bbox": [ 106, 628, 505, 672 ], "lines": [ { "bbox": [ 105, 628, 506, 641 ], "spans": [ { "bbox": [ 105, 628, 506, 641 ], "score": 1.0, "content": "Semantic segmentation on ADE20K. We evaluate the performance of PEG on the ADE20K", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 640, 505, 651 ], "spans": [ { "bbox": [ 106, 640, 505, 651 ], "score": 1.0, "content": "(Zhou et al., 2017) segmentation task. Based on the Semantic FPN framework (Kirillov et al.,", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 650, 505, 663 ], "spans": [ { "bbox": [ 105, 650, 505, 663 ], "score": 1.0, "content": "2019), PVT achieves much better results than ResNet (He et al., 2016) baselines. Under carefully", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 661, 367, 673 ], "spans": [ { "bbox": [ 105, 661, 315, 673 ], "score": 1.0, "content": "controlled settings, PEG further boosts PVT-tiny by", "type": "text" }, { "bbox": [ 315, 661, 337, 672 ], "score": 0.84, "content": "3 . 1 \\%", "type": "inline_equation" }, { "bbox": [ 338, 661, 367, 673 ], "score": 1.0, "content": "mIoU.", "type": "text" } ], "index": 32 } ], "index": 30.5, "bbox_fs": [ 105, 628, 506, 673 ] }, { "type": "text", "bbox": [ 107, 687, 505, 731 ], "lines": [ { "bbox": [ 106, 687, 505, 700 ], "spans": [ { "bbox": [ 106, 687, 505, 700 ], "score": 1.0, "content": "Object detection on COCO. We also perform controlled experiments with the RetinaNet (Lin", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 699, 505, 711 ], "spans": [ { "bbox": [ 105, 699, 505, 711 ], "score": 1.0, "content": "et al., 2017) framework on the COCO detection task. The results are shown in Table 13. In the", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 710, 505, 722 ], "spans": [ { "bbox": [ 105, 710, 142, 722 ], "score": 1.0, "content": "standard", "type": "text" }, { "bbox": [ 143, 710, 157, 720 ], "score": 0.86, "content": "1 \\times", "type": "inline_equation" }, { "bbox": [ 157, 710, 309, 722 ], "score": 1.0, "content": "schedule, PEG improves PVT-tiny by", "type": "text" }, { "bbox": [ 309, 710, 332, 721 ], "score": 0.9, "content": "2 . 0 \\%", "type": "inline_equation" }, { "bbox": [ 332, 710, 405, 722 ], "score": 1.0, "content": "mAP. PEG brings", "type": "text" }, { "bbox": [ 406, 710, 428, 720 ], "score": 0.84, "content": "2 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 428, 710, 505, 722 ], "score": 1.0, "content": "higher mAP under", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 721, 178, 731 ], "spans": [ { "bbox": [ 106, 721, 121, 731 ], "score": 1.0, "content": "the", "type": "text" }, { "bbox": [ 121, 721, 138, 731 ], "score": 0.84, "content": "3 \\times", "type": "inline_equation" }, { "bbox": [ 139, 721, 178, 731 ], "score": 1.0, "content": "schedule.", "type": "text" } ], "index": 36 } ], "index": 34.5, "bbox_fs": [ 105, 687, 505, 731 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 110, 117, 502, 241 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 80, 502, 102 ], "group_id": 0, "lines": [ { "bbox": [ 106, 77, 504, 95 ], "spans": [ { "bbox": [ 106, 77, 504, 95 ], "score": 1.0, "content": "Table 13. Our method boosts the performance of PVT on ImageNet classification, ADE20K seg-", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 91, 236, 104 ], "spans": [ { "bbox": [ 106, 91, 236, 104 ], "score": 1.0, "content": "mentation and COCO detection", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table_body", "bbox": [ 110, 117, 502, 241 ], "group_id": 0, "lines": [ { "bbox": [ 110, 117, 502, 241 ], "spans": [ { "bbox": [ 110, 117, 502, 241 ], "score": 0.983, "html": "
BackboneImageNetSemantic FPN on ADE20KRetinaNet on COCO
Params (M)Top-1 (%)Params (M)mIoU (%）Params (M)mAP (%,1x)mAP (%,3×,+MS)
ResNet-18 (He et al.,2016)1269.81632.92131.835.4
PVT-tiny (Wang et al., 2021)1375.01735.72336.739.4
PVT-tiny+PEG1377.31738.02338.041.8
PVT-tiny+GAP1375.91736.02336.939.7
PVT-tiny+PEG+GAP1378.11738.82338.741.8
PVT-small (Wang et al., 2021)2579.82839.83440.442.2
PVT-small+PEG+GAP2581.22844.33443.045.2
PVT-Medium (Wang et al.,2021)4481.24841.65441.943.2
PVT-Medium+PEG+GAP4482.74844.95444.346.4
", "type": "table", "image_path": "6a5c348be4de1f0ebc180f4d83f11cd2431597c4136565afdef8f00598fb7f96.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 110, 117, 502, 158.33333333333334 ], "spans": [], "index": 2 }, { "bbox": [ 110, 158.33333333333334, 502, 199.66666666666669 ], "spans": [], "index": 3 }, { "bbox": [ 110, 199.66666666666669, 502, 241.00000000000003 ], "spans": [], "index": 4 } ] } ], "index": 1.75 }, { "type": "title", "bbox": [ 108, 264, 296, 276 ], "lines": [ { "bbox": [ 106, 264, 296, 277 ], "spans": [ { "bbox": [ 106, 264, 296, 277 ], "score": 1.0, "content": "B.7 ABLATION ON OTHER FORMS OF PEG", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 107, 285, 505, 351 ], "lines": [ { "bbox": [ 106, 284, 505, 298 ], "spans": [ { "bbox": [ 106, 284, 505, 298 ], "score": 1.0, "content": "We explore several forms of PEG based on the tiny model, which change the type of convolution,", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 294, 505, 310 ], "spans": [ { "bbox": [ 105, 294, 505, 310 ], "score": 1.0, "content": "kernel size and layers. The inserted position is 0. The result is shown in Table 14. When we use large", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 307, 505, 320 ], "spans": [ { "bbox": [ 106, 307, 144, 320 ], "score": 1.0, "content": "kernel of", "type": "text" }, { "bbox": [ 145, 307, 164, 318 ], "score": 0.89, "content": "7 \\times 7", "type": "inline_equation" }, { "bbox": [ 164, 307, 505, 320 ], "score": 1.0, "content": "or dense convolution, the performance improvement is limited. Stacking more layers", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 318, 505, 331 ], "spans": [ { "bbox": [ 106, 318, 505, 331 ], "score": 1.0, "content": "of depth-wise convolution doesn’t bring significant improvement. Therefore, we use the simplest", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 328, 505, 342 ], "spans": [ { "bbox": [ 105, 328, 505, 342 ], "score": 1.0, "content": "form as our default implementation. It indicates that this design is enough to provide good position", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 339, 159, 352 ], "spans": [ { "bbox": [ 106, 339, 159, 352 ], "score": 1.0, "content": "information.", "type": "text" } ], "index": 11 } ], "index": 8.5 }, { "type": "table", "bbox": [ 184, 384, 426, 448 ], "blocks": [ { "type": "table_caption", "bbox": [ 120, 359, 489, 371 ], "group_id": 1, "lines": [ { "bbox": [ 119, 357, 490, 373 ], "spans": [ { "bbox": [ 119, 357, 405, 373 ], "score": 1.0, "content": "Table 14. Other forms of PEG. The simple form of a single depth-wise", "type": "text" }, { "bbox": [ 405, 360, 424, 370 ], "score": 0.86, "content": "3 \\times 3", "type": "inline_equation" }, { "bbox": [ 425, 357, 490, 373 ], "score": 1.0, "content": "is good enough.", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "table_body", "bbox": [ 184, 384, 426, 448 ], "group_id": 1, "lines": [ { "bbox": [ 184, 384, 426, 448 ], "spans": [ { "bbox": [ 184, 384, 426, 448 ], "score": 0.98, "html": "
VariantsModelTop-1 Acc (%)
1 Depthwise Conv 3×3tiny72.4
1 Depthwise Conv 7×7tiny72.5
4 *(Depthwise Conv 3×3+BN+ReLU)tiny72.4
1 Dense Conv 3×3tiny72.3
4 * (Dense Conv 3×3+BN+ReLU)tiny72.5
", "type": "table", "image_path": "79c2a82a8142ebff0f5c8f234ade0e15387f8b272764352eb9d76ada50cb3a0e.jpg" } ] } ], "index": 14.5, "virtual_lines": [ { "bbox": [ 184, 384, 426, 400.0 ], "spans": [], "index": 13 }, { "bbox": [ 184, 400.0, 426, 416.0 ], "spans": [], "index": 14 }, { "bbox": [ 184, 416.0, 426, 432.0 ], "spans": [], "index": 15 }, { "bbox": [ 184, 432.0, 426, 448.0 ], "spans": [], "index": 16 } ] } ], "index": 13.25 }, { "type": "title", "bbox": [ 107, 475, 212, 489 ], "lines": [ { "bbox": [ 105, 474, 214, 491 ], "spans": [ { "bbox": [ 105, 474, 214, 491 ], "score": 1.0, "content": "C EXAMPLE CODE", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "title", "bbox": [ 107, 500, 154, 512 ], "lines": [ { "bbox": [ 105, 499, 155, 514 ], "spans": [ { "bbox": [ 105, 499, 155, 514 ], "score": 1.0, "content": "C.1 PEG", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 107, 521, 505, 566 ], "lines": [ { "bbox": [ 105, 520, 505, 534 ], "spans": [ { "bbox": [ 105, 520, 505, 534 ], "score": 1.0, "content": "In the simplest form, we use a single depth-wise convolution and show its usage in Transformer by", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 533, 505, 545 ], "spans": [ { "bbox": [ 105, 533, 505, 545 ], "score": 1.0, "content": "the following PyTorch snippet. Through experiments, we find that such a simple design (i.e., depth-", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 543, 505, 555 ], "spans": [ { "bbox": [ 105, 544, 127, 555 ], "score": 1.0, "content": "wise", "type": "text" }, { "bbox": [ 127, 543, 147, 554 ], "score": 0.87, "content": "3 \\times 3", "type": "inline_equation" }, { "bbox": [ 147, 544, 505, 555 ], "score": 1.0, "content": ") readily achieves on par or even better performance than the recent SOTAs. We give the", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 555, 271, 567 ], "spans": [ { "bbox": [ 106, 555, 271, 567 ], "score": 1.0, "content": "torch implementation example in Alg. 1.", "type": "text" } ], "index": 22 } ], "index": 20.5 }, { "type": "title", "bbox": [ 108, 581, 232, 594 ], "lines": [ { "bbox": [ 105, 581, 234, 596 ], "spans": [ { "bbox": [ 105, 581, 234, 596 ], "score": 1.0, "content": "D MORE DISCUSSIONS", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "title", "bbox": [ 107, 606, 357, 617 ], "lines": [ { "bbox": [ 106, 606, 358, 619 ], "spans": [ { "bbox": [ 106, 606, 358, 619 ], "score": 1.0, "content": "D.1 WHY RPE WORKS LESS WELL THAN ABSOLUTE PE?", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 626, 505, 682 ], "lines": [ { "bbox": [ 106, 627, 505, 639 ], "spans": [ { "bbox": [ 106, 627, 505, 639 ], "score": 1.0, "content": "As mentioned in Section 5.3 (main text), RPE is inferior to the absolute positional encoding. It", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 637, 505, 651 ], "spans": [ { "bbox": [ 105, 637, 505, 651 ], "score": 1.0, "content": "is because RPE does not encode any absolute position information. Also discussed in Section B.3", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 649, 506, 662 ], "spans": [ { "bbox": [ 106, 649, 506, 662 ], "score": 1.0, "content": "(main text), absolute position information is also important even for ImageNet classification as it is", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 659, 505, 673 ], "spans": [ { "bbox": [ 105, 659, 505, 673 ], "score": 1.0, "content": "needed to determine which object is at the center of the image. Note that there might be multiple", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 671, 457, 684 ], "spans": [ { "bbox": [ 105, 671, 457, 684 ], "score": 1.0, "content": "objects in an image, and the label of an image is the category of the object at the center.", "type": "text" } ], "index": 29 } ], "index": 27 }, { "type": "text", "bbox": [ 108, 687, 504, 732 ], "lines": [ { "bbox": [ 106, 687, 505, 700 ], "spans": [ { "bbox": [ 106, 687, 505, 700 ], "score": 1.0, "content": "Additionally, although RPE becomes popular recently, it is often jointly used with absolute posi-", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 699, 505, 711 ], "spans": [ { "bbox": [ 106, 699, 505, 711 ], "score": 1.0, "content": "tional encodings (e.g., in ConViT (d’Ascoli et al., 2021)), or the absolute position information is", "type": "text" } ], "index": 31 }, { "bbox": [ 104, 708, 506, 724 ], "spans": [ { "bbox": [ 104, 708, 506, 724 ], "score": 1.0, "content": "leaked in other ways (e.g., convolution paddings in CoAtNet (Dai et al., 2021)). This further sug-", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 721, 290, 732 ], "spans": [ { "bbox": [ 106, 721, 290, 732 ], "score": 1.0, "content": "gests absolute position information is crucial.", "type": "text" } ], "index": 33 } ], "index": 31.5 } ], "page_idx": 15, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 301, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 14 } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 110, 117, 502, 241 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 80, 502, 102 ], "group_id": 0, "lines": [ { "bbox": [ 106, 77, 504, 95 ], "spans": [ { "bbox": [ 106, 77, 504, 95 ], "score": 1.0, "content": "Table 13. Our method boosts the performance of PVT on ImageNet classification, ADE20K seg-", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 91, 236, 104 ], "spans": [ { "bbox": [ 106, 91, 236, 104 ], "score": 1.0, "content": "mentation and COCO detection", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table_body", "bbox": [ 110, 117, 502, 241 ], "group_id": 0, "lines": [ { "bbox": [ 110, 117, 502, 241 ], "spans": [ { "bbox": [ 110, 117, 502, 241 ], "score": 0.983, "html": "
BackboneImageNetSemantic FPN on ADE20KRetinaNet on COCO
Params (M)Top-1 (%)Params (M)mIoU (%）Params (M)mAP (%,1x)mAP (%,3×,+MS)
ResNet-18 (He et al.,2016)1269.81632.92131.835.4
PVT-tiny (Wang et al., 2021)1375.01735.72336.739.4
PVT-tiny+PEG1377.31738.02338.041.8
PVT-tiny+GAP1375.91736.02336.939.7
PVT-tiny+PEG+GAP1378.11738.82338.741.8
PVT-small (Wang et al., 2021)2579.82839.83440.442.2
PVT-small+PEG+GAP2581.22844.33443.045.2
PVT-Medium (Wang et al.,2021)4481.24841.65441.943.2
PVT-Medium+PEG+GAP4482.74844.95444.346.4
", "type": "table", "image_path": "6a5c348be4de1f0ebc180f4d83f11cd2431597c4136565afdef8f00598fb7f96.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 110, 117, 502, 158.33333333333334 ], "spans": [], "index": 2 }, { "bbox": [ 110, 158.33333333333334, 502, 199.66666666666669 ], "spans": [], "index": 3 }, { "bbox": [ 110, 199.66666666666669, 502, 241.00000000000003 ], "spans": [], "index": 4 } ] } ], "index": 1.75 }, { "type": "title", "bbox": [ 108, 264, 296, 276 ], "lines": [ { "bbox": [ 106, 264, 296, 277 ], "spans": [ { "bbox": [ 106, 264, 296, 277 ], "score": 1.0, "content": "B.7 ABLATION ON OTHER FORMS OF PEG", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 107, 285, 505, 351 ], "lines": [ { "bbox": [ 106, 284, 505, 298 ], "spans": [ { "bbox": [ 106, 284, 505, 298 ], "score": 1.0, "content": "We explore several forms of PEG based on the tiny model, which change the type of convolution,", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 294, 505, 310 ], "spans": [ { "bbox": [ 105, 294, 505, 310 ], "score": 1.0, "content": "kernel size and layers. The inserted position is 0. The result is shown in Table 14. When we use large", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 307, 505, 320 ], "spans": [ { "bbox": [ 106, 307, 144, 320 ], "score": 1.0, "content": "kernel of", "type": "text" }, { "bbox": [ 145, 307, 164, 318 ], "score": 0.89, "content": "7 \\times 7", "type": "inline_equation" }, { "bbox": [ 164, 307, 505, 320 ], "score": 1.0, "content": "or dense convolution, the performance improvement is limited. Stacking more layers", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 318, 505, 331 ], "spans": [ { "bbox": [ 106, 318, 505, 331 ], "score": 1.0, "content": "of depth-wise convolution doesn’t bring significant improvement. Therefore, we use the simplest", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 328, 505, 342 ], "spans": [ { "bbox": [ 105, 328, 505, 342 ], "score": 1.0, "content": "form as our default implementation. It indicates that this design is enough to provide good position", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 339, 159, 352 ], "spans": [ { "bbox": [ 106, 339, 159, 352 ], "score": 1.0, "content": "information.", "type": "text" } ], "index": 11 } ], "index": 8.5, "bbox_fs": [ 105, 284, 505, 352 ] }, { "type": "table", "bbox": [ 184, 384, 426, 448 ], "blocks": [ { "type": "table_caption", "bbox": [ 120, 359, 489, 371 ], "group_id": 1, "lines": [ { "bbox": [ 119, 357, 490, 373 ], "spans": [ { "bbox": [ 119, 357, 405, 373 ], "score": 1.0, "content": "Table 14. Other forms of PEG. The simple form of a single depth-wise", "type": "text" }, { "bbox": [ 405, 360, 424, 370 ], "score": 0.86, "content": "3 \\times 3", "type": "inline_equation" }, { "bbox": [ 425, 357, 490, 373 ], "score": 1.0, "content": "is good enough.", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "table_body", "bbox": [ 184, 384, 426, 448 ], "group_id": 1, "lines": [ { "bbox": [ 184, 384, 426, 448 ], "spans": [ { "bbox": [ 184, 384, 426, 448 ], "score": 0.98, "html": "
VariantsModelTop-1 Acc (%)
1 Depthwise Conv 3×3tiny72.4
1 Depthwise Conv 7×7tiny72.5
4 *(Depthwise Conv 3×3+BN+ReLU)tiny72.4
1 Dense Conv 3×3tiny72.3
4 * (Dense Conv 3×3+BN+ReLU)tiny72.5
", "type": "table", "image_path": "79c2a82a8142ebff0f5c8f234ade0e15387f8b272764352eb9d76ada50cb3a0e.jpg" } ] } ], "index": 14.5, "virtual_lines": [ { "bbox": [ 184, 384, 426, 400.0 ], "spans": [], "index": 13 }, { "bbox": [ 184, 400.0, 426, 416.0 ], "spans": [], "index": 14 }, { "bbox": [ 184, 416.0, 426, 432.0 ], "spans": [], "index": 15 }, { "bbox": [ 184, 432.0, 426, 448.0 ], "spans": [], "index": 16 } ] } ], "index": 13.25 }, { "type": "title", "bbox": [ 107, 475, 212, 489 ], "lines": [ { "bbox": [ 105, 474, 214, 491 ], "spans": [ { "bbox": [ 105, 474, 214, 491 ], "score": 1.0, "content": "C EXAMPLE CODE", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "title", "bbox": [ 107, 500, 154, 512 ], "lines": [ { "bbox": [ 105, 499, 155, 514 ], "spans": [ { "bbox": [ 105, 499, 155, 514 ], "score": 1.0, "content": "C.1 PEG", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 107, 521, 505, 566 ], "lines": [ { "bbox": [ 105, 520, 505, 534 ], "spans": [ { "bbox": [ 105, 520, 505, 534 ], "score": 1.0, "content": "In the simplest form, we use a single depth-wise convolution and show its usage in Transformer by", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 533, 505, 545 ], "spans": [ { "bbox": [ 105, 533, 505, 545 ], "score": 1.0, "content": "the following PyTorch snippet. Through experiments, we find that such a simple design (i.e., depth-", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 543, 505, 555 ], "spans": [ { "bbox": [ 105, 544, 127, 555 ], "score": 1.0, "content": "wise", "type": "text" }, { "bbox": [ 127, 543, 147, 554 ], "score": 0.87, "content": "3 \\times 3", "type": "inline_equation" }, { "bbox": [ 147, 544, 505, 555 ], "score": 1.0, "content": ") readily achieves on par or even better performance than the recent SOTAs. We give the", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 555, 271, 567 ], "spans": [ { "bbox": [ 106, 555, 271, 567 ], "score": 1.0, "content": "torch implementation example in Alg. 1.", "type": "text" } ], "index": 22 } ], "index": 20.5, "bbox_fs": [ 105, 520, 505, 567 ] }, { "type": "title", "bbox": [ 108, 581, 232, 594 ], "lines": [ { "bbox": [ 105, 581, 234, 596 ], "spans": [ { "bbox": [ 105, 581, 234, 596 ], "score": 1.0, "content": "D MORE DISCUSSIONS", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "title", "bbox": [ 107, 606, 357, 617 ], "lines": [ { "bbox": [ 106, 606, 358, 619 ], "spans": [ { "bbox": [ 106, 606, 358, 619 ], "score": 1.0, "content": "D.1 WHY RPE WORKS LESS WELL THAN ABSOLUTE PE?", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 626, 505, 682 ], "lines": [ { "bbox": [ 106, 627, 505, 639 ], "spans": [ { "bbox": [ 106, 627, 505, 639 ], "score": 1.0, "content": "As mentioned in Section 5.3 (main text), RPE is inferior to the absolute positional encoding. It", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 637, 505, 651 ], "spans": [ { "bbox": [ 105, 637, 505, 651 ], "score": 1.0, "content": "is because RPE does not encode any absolute position information. Also discussed in Section B.3", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 649, 506, 662 ], "spans": [ { "bbox": [ 106, 649, 506, 662 ], "score": 1.0, "content": "(main text), absolute position information is also important even for ImageNet classification as it is", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 659, 505, 673 ], "spans": [ { "bbox": [ 105, 659, 505, 673 ], "score": 1.0, "content": "needed to determine which object is at the center of the image. Note that there might be multiple", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 671, 457, 684 ], "spans": [ { "bbox": [ 105, 671, 457, 684 ], "score": 1.0, "content": "objects in an image, and the label of an image is the category of the object at the center.", "type": "text" } ], "index": 29 } ], "index": 27, "bbox_fs": [ 105, 627, 506, 684 ] }, { "type": "text", "bbox": [ 108, 687, 504, 732 ], "lines": [ { "bbox": [ 106, 687, 505, 700 ], "spans": [ { "bbox": [ 106, 687, 505, 700 ], "score": 1.0, "content": "Additionally, although RPE becomes popular recently, it is often jointly used with absolute posi-", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 699, 505, 711 ], "spans": [ { "bbox": [ 106, 699, 505, 711 ], "score": 1.0, "content": "tional encodings (e.g., in ConViT (d’Ascoli et al., 2021)), or the absolute position information is", "type": "text" } ], "index": 31 }, { "bbox": [ 104, 708, 506, 724 ], "spans": [ { "bbox": [ 104, 708, 506, 724 ], "score": 1.0, "content": "leaked in other ways (e.g., convolution paddings in CoAtNet (Dai et al., 2021)). This further sug-", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 721, 290, 732 ], "spans": [ { "bbox": [ 106, 721, 290, 732 ], "score": 1.0, "content": "gests absolute position information is crucial.", "type": "text" } ], "index": 33 } ], "index": 31.5, "bbox_fs": [ 104, 687, 506, 732 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 82, 261, 94 ], "lines": [ { "bbox": [ 106, 81, 263, 96 ], "spans": [ { "bbox": [ 106, 81, 263, 96 ], "score": 1.0, "content": "Algorithm 1 PyTorch snippet of PEG.", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 106, 104, 498, 338 ], "lines": [ { "bbox": [ 105, 101, 161, 111 ], "spans": [ { "bbox": [ 105, 101, 161, 111 ], "score": 1.0, "content": "import torch", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 109, 200, 119 ], "spans": [ { "bbox": [ 106, 109, 200, 119 ], "score": 1.0, "content": "import torch.nn as nn", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 116, 212, 126 ], "spans": [ { "bbox": [ 105, 116, 212, 126 ], "score": 1.0, "content": "class VisionTransformer:", "type": "text" } ], "index": 3 }, { "bbox": [ 112, 123, 421, 133 ], "spans": [ { "bbox": [ 112, 123, 191, 133 ], "score": 1.0, "content": "def __init__(layer", "type": "text" }, { "bbox": [ 192, 124, 208, 131 ], "score": 0.72, "content": "_ { \\mathrm { S } } = 1 2", "type": "inline_equation" }, { "bbox": [ 209, 123, 216, 133 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 216, 124, 247, 132 ], "score": 0.28, "content": "{ \\mathrm { d i m } } { = } 1 9 2", "type": "inline_equation" }, { "bbox": [ 248, 123, 277, 133 ], "score": 1.0, "content": ", nhead", "type": "text" }, { "bbox": [ 277, 124, 286, 131 ], "score": 0.57, "content": "^ { \\underline { { { \\textstyle \\dag } } } } = 3", "type": "inline_equation" }, { "bbox": [ 286, 123, 328, 133 ], "score": 1.0, "content": ", img_size", "type": "text" }, { "bbox": [ 328, 124, 345, 132 ], "score": 0.34, "content": "_ { : = 2 2 4 }", "type": "inline_equation" }, { "bbox": [ 346, 123, 421, 133 ], "score": 1.0, "content": ", patch_size=16):", "type": "text" } ], "index": 4 }, { "bbox": [ 119, 131, 228, 140 ], "spans": [ { "bbox": [ 119, 131, 182, 140 ], "score": 1.0, "content": "self.pos_block", "type": "text" }, { "bbox": [ 182, 132, 191, 138 ], "score": 0.79, "content": "=", "type": "inline_equation" }, { "bbox": [ 191, 131, 228, 140 ], "score": 1.0, "content": "PEG(dim)", "type": "text" } ], "index": 5 }, { "bbox": [ 118, 138, 497, 147 ], "spans": [ { "bbox": [ 118, 138, 169, 147 ], "score": 1.0, "content": "self.blocks", "type": "text" }, { "bbox": [ 170, 140, 178, 145 ], "score": 0.61, "content": "=", "type": "inline_equation" }, { "bbox": [ 178, 138, 412, 147 ], "score": 1.0, "content": "nn.ModuleList([TransformerEncoderLayer(dim, nhead, dim", "type": "text" }, { "bbox": [ 413, 139, 419, 145 ], "score": 0.39, "content": "\\mathbf { \\nabla } _ { \\cdot } \\star \\mathbf { \\nabla } _ { \\cdot }", "type": "inline_equation" }, { "bbox": [ 419, 138, 497, 147 ], "score": 1.0, "content": "4) for _ in range(", "type": "text" } ], "index": 6 }, { "bbox": [ 138, 145, 180, 155 ], "spans": [ { "bbox": [ 138, 145, 180, 155 ], "score": 1.0, "content": "layers)])", "type": "text" } ], "index": 7 }, { "bbox": [ 118, 152, 370, 162 ], "spans": [ { "bbox": [ 118, 152, 191, 162 ], "score": 1.0, "content": "self.patch_embed", "type": "text" }, { "bbox": [ 191, 153, 200, 159 ], "score": 0.75, "content": "=", "type": "inline_equation" }, { "bbox": [ 200, 152, 370, 162 ], "score": 1.0, "content": "PatchEmbed(img_size, patch_size, dim*4)", "type": "text" } ], "index": 8 }, { "bbox": [ 111, 158, 245, 169 ], "spans": [ { "bbox": [ 111, 158, 228, 169 ], "score": 1.0, "content": "def forward_features(self,", "type": "text" }, { "bbox": [ 229, 160, 236, 167 ], "score": 0.47, "content": "\\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 237, 158, 245, 169 ], "score": 1.0, "content": "):", "type": "text" } ], "index": 9 }, { "bbox": [ 117, 165, 209, 176 ], "spans": [ { "bbox": [ 117, 165, 157, 176 ], "score": 1.0, "content": "B, C, H,", "type": "text" }, { "bbox": [ 157, 167, 181, 174 ], "score": 0.6, "content": "W ~ = ~ \\mathrm { ~ x ~ }", "type": "inline_equation" }, { "bbox": [ 181, 165, 209, 176 ], "score": 1.0, "content": ".shape", "type": "text" } ], "index": 10 }, { "bbox": [ 117, 173, 272, 183 ], "spans": [ { "bbox": [ 117, 173, 178, 183 ], "score": 1.0, "content": "x, patch_size", "type": "text" }, { "bbox": [ 178, 175, 186, 181 ], "score": 0.58, "content": "=", "type": "inline_equation" }, { "bbox": [ 187, 173, 272, 183 ], "score": 1.0, "content": "self.patch_embed(x)", "type": "text" } ], "index": 11 }, { "bbox": [ 117, 180, 300, 191 ], "spans": [ { "bbox": [ 117, 180, 136, 191 ], "score": 1.0, "content": "_H,", "type": "text" }, { "bbox": [ 137, 182, 164, 189 ], "score": 0.42, "content": "\\_ \\mathrm { ~ \\tt ~ H ~ } = \\mathrm { ~ \\tt ~ H ~ }", "type": "inline_equation" }, { "bbox": [ 165, 180, 300, 191 ], "score": 1.0, "content": "// patch_size, W // patch_size", "type": "text" } ], "index": 12 }, { "bbox": [ 119, 189, 301, 197 ], "spans": [ { "bbox": [ 119, 189, 135, 195 ], "score": 0.58, "content": "\\qquad \\times \\quad =", "type": "inline_equation" }, { "bbox": [ 135, 189, 288, 197 ], "score": 1.0, "content": "torch.cat((self.cls_tokens, x), dim", "type": "text" }, { "bbox": [ 288, 189, 297, 196 ], "score": 0.33, "content": "^ { = 1 }", "type": "inline_equation" }, { "bbox": [ 297, 189, 301, 197 ], "score": 1.0, "content": ")", "type": "text" } ], "index": 13 }, { "bbox": [ 118, 194, 281, 205 ], "spans": [ { "bbox": [ 118, 194, 281, 205 ], "score": 1.0, "content": "for i, blk in enumerate(self.blocks):", "type": "text" } ], "index": 14 }, { "bbox": [ 125, 203, 168, 211 ], "spans": [ { "bbox": [ 125, 203, 158, 211 ], "score": 1.0, "content": "x = blk", "type": "text" }, { "bbox": [ 158, 203, 168, 210 ], "score": 0.27, "content": "( \\times )", "type": "inline_equation" } ], "index": 15 }, { "bbox": [ 126, 210, 171, 218 ], "spans": [ { "bbox": [ 126, 210, 143, 218 ], "score": 1.0, "content": "if i", "type": "text" }, { "bbox": [ 143, 211, 166, 217 ], "score": 0.47, "content": "\\quad . = = 0", "type": "inline_equation" }, { "bbox": [ 166, 210, 171, 218 ], "score": 1.0, "content": ":", "type": "text" } ], "index": 16 }, { "bbox": [ 132, 217, 258, 226 ], "spans": [ { "bbox": [ 132, 218, 148, 225 ], "score": 0.44, "content": "\\times \\quad =", "type": "inline_equation" }, { "bbox": [ 148, 217, 258, 226 ], "score": 1.0, "content": "self.pos_block(x, _H, _W)", "type": "text" } ], "index": 17 }, { "bbox": [ 118, 223, 183, 234 ], "spans": [ { "bbox": [ 118, 223, 183, 234 ], "score": 1.0, "content": "return x[:, 0]", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 245, 199, 254 ], "spans": [ { "bbox": [ 105, 245, 199, 254 ], "score": 1.0, "content": "class PEG(nn.Module):", "type": "text" } ], "index": 19 }, { "bbox": [ 112, 252, 297, 262 ], "spans": [ { "bbox": [ 112, 253, 130, 262 ], "score": 1.0, "content": "def", "type": "text" }, { "bbox": [ 135, 252, 208, 262 ], "score": 1.0, "content": "_init__(self, dim", "type": "text" }, { "bbox": [ 208, 253, 218, 260 ], "score": 0.41, "content": "^ { 1 = 2 }", "type": "inline_equation" }, { "bbox": [ 218, 252, 271, 262 ], "score": 1.0, "content": "\\textsc{56},", "type": "text" }, { "bbox": [ 272, 253, 287, 260 ], "score": 0.42, "content": "\\mathrm { k } = 3", "type": "inline_equation" }, { "bbox": [ 288, 252, 297, 262 ], "score": 1.0, "content": "):", "type": "text" } ], "index": 20 }, { "bbox": [ 117, 258, 466, 270 ], "spans": [ { "bbox": [ 117, 258, 156, 270 ], "score": 1.0, "content": "self.pos", "type": "text" }, { "bbox": [ 156, 261, 167, 267 ], "score": 0.6, "content": "=", "type": "inline_equation" }, { "bbox": [ 167, 258, 330, 270 ], "score": 1.0, "content": "nn.Conv2d(dim, dim, k, 1, k//2, groups", "type": "text" }, { "bbox": [ 331, 261, 336, 267 ], "score": 0.32, "content": "=", "type": "inline_equation" }, { "bbox": [ 336, 258, 466, 270 ], "score": 1.0, "content": "dim) # Only for demo use, more", "type": "text" } ], "index": 21 }, { "bbox": [ 139, 267, 313, 276 ], "spans": [ { "bbox": [ 139, 267, 313, 276 ], "score": 1.0, "content": "complicated functions are effective too.", "type": "text" } ], "index": 22 }, { "bbox": [ 111, 273, 231, 284 ], "spans": [ { "bbox": [ 111, 273, 231, 284 ], "score": 1.0, "content": "def forward(self, x, H, W):", "type": "text" } ], "index": 23 }, { "bbox": [ 118, 281, 196, 291 ], "spans": [ { "bbox": [ 118, 281, 145, 291 ], "score": 1.0, "content": "B, N,", "type": "text" }, { "bbox": [ 146, 282, 168, 289 ], "score": 0.36, "content": "{ \\mathrm { ~ \\small ~ \\mathscr ~ { ~ C ~ } ~ } } = { \\mathrm { ~ \\small ~ x ~ } }", "type": "inline_equation" }, { "bbox": [ 169, 281, 196, 291 ], "score": 1.0, "content": ".shape", "type": "text" } ], "index": 24 }, { "bbox": [ 118, 288, 302, 298 ], "spans": [ { "bbox": [ 118, 288, 217, 298 ], "score": 1.0, "content": "cls_token, feat_tokens", "type": "text" }, { "bbox": [ 218, 290, 225, 296 ], "score": 0.33, "content": "=", "type": "inline_equation" }, { "bbox": [ 226, 288, 302, 298 ], "score": 1.0, "content": "x[:, 0], x[:, 1:]", "type": "text" } ], "index": 25 }, { "bbox": [ 118, 295, 370, 305 ], "spans": [ { "bbox": [ 118, 295, 170, 305 ], "score": 1.0, "content": "feat_tokens", "type": "text" }, { "bbox": [ 170, 297, 178, 303 ], "score": 0.73, "content": "=", "type": "inline_equation" }, { "bbox": [ 178, 295, 370, 305 ], "score": 1.0, "content": "feat_tokens.transpose(1, 2).view(B, C, H, W)", "type": "text" } ], "index": 26 }, { "bbox": [ 119, 303, 290, 312 ], "spans": [ { "bbox": [ 119, 304, 135, 311 ], "score": 0.74, "content": "\\qquad \\times \\quad =", "type": "inline_equation" }, { "bbox": [ 136, 303, 231, 312 ], "score": 1.0, "content": "self.pos(feat_tokens)", "type": "text" }, { "bbox": [ 231, 304, 237, 311 ], "score": 0.41, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 238, 303, 290, 312 ], "score": 1.0, "content": "feat_tokens", "type": "text" } ], "index": 27 }, { "bbox": [ 119, 309, 259, 319 ], "spans": [ { "bbox": [ 119, 312, 135, 318 ], "score": 0.73, "content": "\\qquad \\times \\quad =", "type": "inline_equation" }, { "bbox": [ 135, 309, 259, 319 ], "score": 1.0, "content": "x.flatten(2).transpose(1, 2)", "type": "text" } ], "index": 28 }, { "bbox": [ 119, 317, 331, 326 ], "spans": [ { "bbox": [ 119, 318, 135, 325 ], "score": 0.73, "content": "\\qquad \\times \\quad =", "type": "inline_equation" }, { "bbox": [ 135, 317, 331, 326 ], "score": 1.0, "content": "torch.cat((cls_token.unsqueeze(1), x), dim=1)", "type": "text" } ], "index": 29 }, { "bbox": [ 119, 325, 157, 334 ], "spans": [ { "bbox": [ 119, 325, 157, 334 ], "score": 1.0, "content": "return x", "type": "text" } ], "index": 30 } ], "index": 15.5 }, { "type": "title", "bbox": [ 109, 366, 298, 377 ], "lines": [ { "bbox": [ 106, 365, 300, 379 ], "spans": [ { "bbox": [ 106, 365, 300, 379 ], "score": 1.0, "content": "D.2 COMPARISON TO LAMBDA NETWORKS", "type": "text" } ], "index": 31 } ], "index": 31 }, { "type": "text", "bbox": [ 107, 388, 505, 465 ], "lines": [ { "bbox": [ 106, 388, 505, 401 ], "spans": [ { "bbox": [ 106, 388, 505, 401 ], "score": 1.0, "content": "Our work is also related to Lambda Networks (Bello, 2021) which uses 2D relative positional encod-", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 398, 505, 413 ], "spans": [ { "bbox": [ 105, 398, 505, 413 ], "score": 1.0, "content": "ings. We evaluate its lambda module with an embedding size of 128, where we denote its encoding", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 409, 505, 423 ], "spans": [ { "bbox": [ 105, 409, 505, 423 ], "score": 1.0, "content": "scheme as RPE2D-d128. Noticeably, this configuration has about 5.9M parameters (comparable to", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 420, 505, 435 ], "spans": [ { "bbox": [ 105, 420, 221, 435 ], "score": 1.0, "content": "DeiT-tiny) but only obtains", "type": "text" }, { "bbox": [ 221, 421, 248, 432 ], "score": 0.86, "content": "6 8 . 7 \\%", "type": "inline_equation" }, { "bbox": [ 248, 420, 505, 435 ], "score": 1.0, "content": ". We attribute its failure to the limited ability in capturing the", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 432, 505, 444 ], "spans": [ { "bbox": [ 105, 432, 505, 444 ], "score": 1.0, "content": "correct positional information. After all, lambda layers are designed with the help of many CNN", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 443, 504, 455 ], "spans": [ { "bbox": [ 105, 443, 504, 455 ], "score": 1.0, "content": "backbones components such as down-sampling to form various stages, to replace ordinary convolu-", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 455, 402, 466 ], "spans": [ { "bbox": [ 106, 455, 402, 466 ], "score": 1.0, "content": "tions in ResNet (He et al., 2016). In contrast, CPVT is transformer-based.", "type": "text" } ], "index": 38 } ], "index": 35 }, { "type": "title", "bbox": [ 109, 483, 282, 494 ], "lines": [ { "bbox": [ 106, 482, 284, 496 ], "spans": [ { "bbox": [ 106, 482, 284, 496 ], "score": 1.0, "content": "D.3 QUALITATIVE ANALYSIS OF CPVT", "type": "text" } ], "index": 39 } ], "index": 39 }, { "type": "text", "bbox": [ 107, 505, 505, 571 ], "lines": [ { "bbox": [ 106, 505, 505, 518 ], "spans": [ { "bbox": [ 106, 505, 505, 518 ], "score": 1.0, "content": "Thus far, we have shown that PEG can have better performance than the original positional encod-", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 516, 505, 528 ], "spans": [ { "bbox": [ 105, 516, 505, 528 ], "score": 1.0, "content": "ings. However, because PEG provides the position in an implicit way, it is interesting to see if PEG", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 527, 505, 540 ], "spans": [ { "bbox": [ 105, 527, 505, 540 ], "score": 1.0, "content": "can indeed provide the position information as the original positional encodings. Here we inves-", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 538, 504, 551 ], "spans": [ { "bbox": [ 105, 538, 464, 551 ], "score": 1.0, "content": "tigate this by visualizing the attention weights of the transformers. Specifically, given a", "type": "text" }, { "bbox": [ 465, 538, 504, 549 ], "score": 0.88, "content": "2 2 4 \\times 2 2 4", "type": "inline_equation" } ], "index": 43 }, { "bbox": [ 105, 549, 505, 561 ], "spans": [ { "bbox": [ 105, 549, 155, 561 ], "score": 1.0, "content": "image (i.e.", "type": "text" }, { "bbox": [ 155, 549, 184, 560 ], "score": 0.89, "content": "1 4 \\times 1 4", "type": "inline_equation" }, { "bbox": [ 185, 549, 390, 561 ], "score": 1.0, "content": "patches), the score matrix within a single head is", "type": "text" }, { "bbox": [ 390, 549, 429, 560 ], "score": 0.89, "content": "1 9 6 \\times 1 9 6", "type": "inline_equation" }, { "bbox": [ 429, 549, 505, 561 ], "score": 1.0, "content": ". We visualize the", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 560, 377, 572 ], "spans": [ { "bbox": [ 106, 560, 377, 572 ], "score": 1.0, "content": "normalized self-attention score matrix of the second encoder block.", "type": "text" } ], "index": 45 } ], "index": 42.5 }, { "type": "text", "bbox": [ 107, 577, 505, 644 ], "lines": [ { "bbox": [ 105, 576, 505, 590 ], "spans": [ { "bbox": [ 105, 576, 505, 590 ], "score": 1.0, "content": "We first visualize the attention weights of DeiT with the original positional encodings. As shown in", "type": "text" } ], "index": 46 }, { "bbox": [ 106, 588, 505, 601 ], "spans": [ { "bbox": [ 106, 588, 505, 601 ], "score": 1.0, "content": "Figure 5 (middle), the diagonal element interacts strongly with its local neighbors but weakly with", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 599, 506, 612 ], "spans": [ { "bbox": [ 105, 599, 506, 612 ], "score": 1.0, "content": "those far-away elements, which suggests that DeiT with the original positional encodings learn to", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 610, 505, 623 ], "spans": [ { "bbox": [ 105, 610, 505, 623 ], "score": 1.0, "content": "attend the local neighbors of each patch. After the positional encodings are removed (denoted by", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 620, 506, 634 ], "spans": [ { "bbox": [ 105, 620, 506, 634 ], "score": 1.0, "content": "DeiT w/o PE), all the patches produce similar attention weights and fail to attend to the patches near", "type": "text" } ], "index": 50 }, { "bbox": [ 106, 632, 231, 645 ], "spans": [ { "bbox": [ 106, 632, 231, 645 ], "score": 1.0, "content": "themselves, see Figure 5 (left).", "type": "text" } ], "index": 51 } ], "index": 48.5 }, { "type": "text", "bbox": [ 107, 649, 504, 682 ], "lines": [ { "bbox": [ 105, 648, 505, 662 ], "spans": [ { "bbox": [ 105, 648, 505, 662 ], "score": 1.0, "content": "Finally, we show the attention weights of our CPVT model with PEG. As shown in Figure 5 (right),", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 659, 506, 673 ], "spans": [ { "bbox": [ 105, 659, 506, 673 ], "score": 1.0, "content": "like the original positional encodings, the model with PEG can also learn a similar attention pattern,", "type": "text" } ], "index": 53 }, { "bbox": [ 106, 671, 444, 683 ], "spans": [ { "bbox": [ 106, 671, 444, 683 ], "score": 1.0, "content": "which indicates that the proposed PEG can provide the position information as well.", "type": "text" } ], "index": 54 } ], "index": 53 }, { "type": "text", "bbox": [ 107, 687, 504, 732 ], "lines": [ { "bbox": [ 106, 687, 506, 699 ], "spans": [ { "bbox": [ 106, 687, 506, 699 ], "score": 1.0, "content": "We illustrate the attention scores in several encoder blocks of DeiT (Touvron et al., 2020) and CPVT", "type": "text" } ], "index": 55 }, { "bbox": [ 105, 699, 506, 711 ], "spans": [ { "bbox": [ 105, 699, 506, 711 ], "score": 1.0, "content": "in the Fig. 6. It shows both methods learn similar locality patterns. As attention scores are computed", "type": "text" } ], "index": 56 }, { "bbox": [ 105, 710, 506, 721 ], "spans": [ { "bbox": [ 105, 710, 506, 721 ], "score": 1.0, "content": "over the tokens projected in different subspaces (Q and K), they do not necessarily show a strict", "type": "text" } ], "index": 57 }, { "bbox": [ 105, 720, 498, 734 ], "spans": [ { "bbox": [ 105, 720, 498, 734 ], "score": 1.0, "content": "diagonal pattern, where some may have slight shift, see DeiT in Fig. 6c and CPVT of Fig. 5 right.", "type": "text" } ], "index": 58 } ], "index": 56.5 } ], "page_idx": 16, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "17", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 82, 261, 94 ], "lines": [ { "bbox": [ 106, 81, 263, 96 ], "spans": [ { "bbox": [ 106, 81, 263, 96 ], "score": 1.0, "content": "Algorithm 1 PyTorch snippet of PEG.", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "list", "bbox": [ 106, 104, 498, 338 ], "lines": [ { "bbox": [ 105, 101, 161, 111 ], "spans": [ { "bbox": [ 105, 101, 161, 111 ], "score": 1.0, "content": "import torch", "type": "text" } ], "index": 1, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 106, 109, 200, 119 ], "spans": [ { "bbox": [ 106, 109, 200, 119 ], "score": 1.0, "content": "import torch.nn as nn", "type": "text" } ], "index": 2, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 105, 116, 212, 126 ], "spans": [ { "bbox": [ 105, 116, 212, 126 ], "score": 1.0, "content": "class VisionTransformer:", "type": "text" } ], "index": 3, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 112, 123, 421, 133 ], "spans": [ { "bbox": [ 112, 123, 191, 133 ], "score": 1.0, "content": "def __init__(layer", "type": "text" }, { "bbox": [ 192, 124, 208, 131 ], "score": 0.72, "content": "_ { \\mathrm { S } } = 1 2", "type": "inline_equation" }, { "bbox": [ 209, 123, 216, 133 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 216, 124, 247, 132 ], "score": 0.28, "content": "{ \\mathrm { d i m } } { = } 1 9 2", "type": "inline_equation" }, { "bbox": [ 248, 123, 277, 133 ], "score": 1.0, "content": ", nhead", "type": "text" }, { "bbox": [ 277, 124, 286, 131 ], "score": 0.57, "content": "^ { \\underline { { { \\textstyle \\dag } } } } = 3", "type": "inline_equation" }, { "bbox": [ 286, 123, 328, 133 ], "score": 1.0, "content": ", img_size", "type": "text" }, { "bbox": [ 328, 124, 345, 132 ], "score": 0.34, "content": "_ { : = 2 2 4 }", "type": "inline_equation" }, { "bbox": [ 346, 123, 421, 133 ], "score": 1.0, "content": ", patch_size=16):", "type": "text" } ], "index": 4, "is_list_end_line": true }, { "bbox": [ 119, 131, 228, 140 ], "spans": [ { "bbox": [ 119, 131, 182, 140 ], "score": 1.0, "content": "self.pos_block", "type": "text" }, { "bbox": [ 182, 132, 191, 138 ], "score": 0.79, "content": "=", "type": "inline_equation" }, { "bbox": [ 191, 131, 228, 140 ], "score": 1.0, "content": "PEG(dim)", "type": "text" } ], "index": 5, "is_list_end_line": true }, { "bbox": [ 118, 138, 497, 147 ], "spans": [ { "bbox": [ 118, 138, 169, 147 ], "score": 1.0, "content": "self.blocks", "type": "text" }, { "bbox": [ 170, 140, 178, 145 ], "score": 0.61, "content": "=", "type": "inline_equation" }, { "bbox": [ 178, 138, 412, 147 ], "score": 1.0, "content": "nn.ModuleList([TransformerEncoderLayer(dim, nhead, dim", "type": "text" }, { "bbox": [ 413, 139, 419, 145 ], "score": 0.39, "content": "\\mathbf { \\nabla } _ { \\cdot } \\star \\mathbf { \\nabla } _ { \\cdot }", "type": "inline_equation" }, { "bbox": [ 419, 138, 497, 147 ], "score": 1.0, "content": "4) for _ in range(", "type": "text" } ], "index": 6 }, { "bbox": [ 138, 145, 180, 155 ], "spans": [ { "bbox": [ 138, 145, 180, 155 ], "score": 1.0, "content": "layers)])", "type": "text" } ], "index": 7, "is_list_end_line": true }, { "bbox": [ 118, 152, 370, 162 ], "spans": [ { "bbox": [ 118, 152, 191, 162 ], "score": 1.0, "content": "self.patch_embed", "type": "text" }, { "bbox": [ 191, 153, 200, 159 ], "score": 0.75, "content": "=", "type": "inline_equation" }, { "bbox": [ 200, 152, 370, 162 ], "score": 1.0, "content": "PatchEmbed(img_size, patch_size, dim*4)", "type": "text" } ], "index": 8, "is_list_end_line": true }, { "bbox": [ 111, 158, 245, 169 ], "spans": [ { "bbox": [ 111, 158, 228, 169 ], "score": 1.0, "content": "def forward_features(self,", "type": "text" }, { "bbox": [ 229, 160, 236, 167 ], "score": 0.47, "content": "\\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 237, 158, 245, 169 ], "score": 1.0, "content": "):", "type": "text" } ], "index": 9, "is_list_end_line": true }, { "bbox": [ 117, 165, 209, 176 ], "spans": [ { "bbox": [ 117, 165, 157, 176 ], "score": 1.0, "content": "B, C, H,", "type": "text" }, { "bbox": [ 157, 167, 181, 174 ], "score": 0.6, "content": "W ~ = ~ \\mathrm { ~ x ~ }", "type": "inline_equation" }, { "bbox": [ 181, 165, 209, 176 ], "score": 1.0, "content": ".shape", "type": "text" } ], "index": 10, "is_list_end_line": true }, { "bbox": [ 117, 173, 272, 183 ], "spans": [ { "bbox": [ 117, 173, 178, 183 ], "score": 1.0, "content": "x, patch_size", "type": "text" }, { "bbox": [ 178, 175, 186, 181 ], "score": 0.58, "content": "=", "type": "inline_equation" }, { "bbox": [ 187, 173, 272, 183 ], "score": 1.0, "content": "self.patch_embed(x)", "type": "text" } ], "index": 11, "is_list_end_line": true }, { "bbox": [ 117, 180, 300, 191 ], "spans": [ { "bbox": [ 117, 180, 136, 191 ], "score": 1.0, "content": "_H,", "type": "text" }, { "bbox": [ 137, 182, 164, 189 ], "score": 0.42, "content": "\\_ \\mathrm { ~ \\tt ~ H ~ } = \\mathrm { ~ \\tt ~ H ~ }", "type": "inline_equation" }, { "bbox": [ 165, 180, 300, 191 ], "score": 1.0, "content": "// patch_size, W // patch_size", "type": "text" } ], "index": 12, "is_list_end_line": true }, { "bbox": [ 119, 189, 301, 197 ], "spans": [ { "bbox": [ 119, 189, 135, 195 ], "score": 0.58, "content": "\\qquad \\times \\quad =", "type": "inline_equation" }, { "bbox": [ 135, 189, 288, 197 ], "score": 1.0, "content": "torch.cat((self.cls_tokens, x), dim", "type": "text" }, { "bbox": [ 288, 189, 297, 196 ], "score": 0.33, "content": "^ { = 1 }", "type": "inline_equation" }, { "bbox": [ 297, 189, 301, 197 ], "score": 1.0, "content": ")", "type": "text" } ], "index": 13, "is_list_end_line": true }, { "bbox": [ 118, 194, 281, 205 ], "spans": [ { "bbox": [ 118, 194, 281, 205 ], "score": 1.0, "content": "for i, blk in enumerate(self.blocks):", "type": "text" } ], "index": 14, "is_list_end_line": true }, { "bbox": [ 125, 203, 168, 211 ], "spans": [ { "bbox": [ 125, 203, 158, 211 ], "score": 1.0, "content": "x = blk", "type": "text" }, { "bbox": [ 158, 203, 168, 210 ], "score": 0.27, "content": "( \\times )", "type": "inline_equation" } ], "index": 15, "is_list_end_line": true }, { "bbox": [ 126, 210, 171, 218 ], "spans": [ { "bbox": [ 126, 210, 143, 218 ], "score": 1.0, "content": "if i", "type": "text" }, { "bbox": [ 143, 211, 166, 217 ], "score": 0.47, "content": "\\quad . = = 0", "type": "inline_equation" }, { "bbox": [ 166, 210, 171, 218 ], "score": 1.0, "content": ":", "type": "text" } ], "index": 16, "is_list_end_line": true }, { "bbox": [ 132, 217, 258, 226 ], "spans": [ { "bbox": [ 132, 218, 148, 225 ], "score": 0.44, "content": "\\times \\quad =", "type": "inline_equation" }, { "bbox": [ 148, 217, 258, 226 ], "score": 1.0, "content": "self.pos_block(x, _H, _W)", "type": "text" } ], "index": 17, "is_list_end_line": true }, { "bbox": [ 118, 223, 183, 234 ], "spans": [ { "bbox": [ 118, 223, 183, 234 ], "score": 1.0, "content": "return x[:, 0]", "type": "text" } ], "index": 18, "is_list_end_line": true }, { "bbox": [ 105, 245, 199, 254 ], "spans": [ { "bbox": [ 105, 245, 199, 254 ], "score": 1.0, "content": "class PEG(nn.Module):", "type": "text" } ], "index": 19, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 112, 252, 297, 262 ], "spans": [ { "bbox": [ 112, 253, 130, 262 ], "score": 1.0, "content": "def", "type": "text" }, { "bbox": [ 135, 252, 208, 262 ], "score": 1.0, "content": "_init__(self, dim", "type": "text" }, { "bbox": [ 208, 253, 218, 260 ], "score": 0.41, "content": "^ { 1 = 2 }", "type": "inline_equation" }, { "bbox": [ 218, 252, 271, 262 ], "score": 1.0, "content": "\\textsc{56},", "type": "text" }, { "bbox": [ 272, 253, 287, 260 ], "score": 0.42, "content": "\\mathrm { k } = 3", "type": "inline_equation" }, { "bbox": [ 288, 252, 297, 262 ], "score": 1.0, "content": "):", "type": "text" } ], "index": 20, "is_list_end_line": true }, { "bbox": [ 117, 258, 466, 270 ], "spans": [ { "bbox": [ 117, 258, 156, 270 ], "score": 1.0, "content": "self.pos", "type": "text" }, { "bbox": [ 156, 261, 167, 267 ], "score": 0.6, "content": "=", "type": "inline_equation" }, { "bbox": [ 167, 258, 330, 270 ], "score": 1.0, "content": "nn.Conv2d(dim, dim, k, 1, k//2, groups", "type": "text" }, { "bbox": [ 331, 261, 336, 267 ], "score": 0.32, "content": "=", "type": "inline_equation" }, { "bbox": [ 336, 258, 466, 270 ], "score": 1.0, "content": "dim) # Only for demo use, more", "type": "text" } ], "index": 21, "is_list_end_line": true }, { "bbox": [ 139, 267, 313, 276 ], "spans": [ { "bbox": [ 139, 267, 313, 276 ], "score": 1.0, "content": "complicated functions are effective too.", "type": "text" } ], "index": 22, "is_list_end_line": true }, { "bbox": [ 111, 273, 231, 284 ], "spans": [ { "bbox": [ 111, 273, 231, 284 ], "score": 1.0, "content": "def forward(self, x, H, W):", "type": "text" } ], "index": 23, "is_list_end_line": true }, { "bbox": [ 118, 281, 196, 291 ], "spans": [ { "bbox": [ 118, 281, 145, 291 ], "score": 1.0, "content": "B, N,", "type": "text" }, { "bbox": [ 146, 282, 168, 289 ], "score": 0.36, "content": "{ \\mathrm { ~ \\small ~ \\mathscr ~ { ~ C ~ } ~ } } = { \\mathrm { ~ \\small ~ x ~ } }", "type": "inline_equation" }, { "bbox": [ 169, 281, 196, 291 ], "score": 1.0, "content": ".shape", "type": "text" } ], "index": 24, "is_list_end_line": true }, { "bbox": [ 118, 288, 302, 298 ], "spans": [ { "bbox": [ 118, 288, 217, 298 ], "score": 1.0, "content": "cls_token, feat_tokens", "type": "text" }, { "bbox": [ 218, 290, 225, 296 ], "score": 0.33, "content": "=", "type": "inline_equation" }, { "bbox": [ 226, 288, 302, 298 ], "score": 1.0, "content": "x[:, 0], x[:, 1:]", "type": "text" } ], "index": 25, "is_list_end_line": true }, { "bbox": [ 118, 295, 370, 305 ], "spans": [ { "bbox": [ 118, 295, 170, 305 ], "score": 1.0, "content": "feat_tokens", "type": "text" }, { "bbox": [ 170, 297, 178, 303 ], "score": 0.73, "content": "=", "type": "inline_equation" }, { "bbox": [ 178, 295, 370, 305 ], "score": 1.0, "content": "feat_tokens.transpose(1, 2).view(B, C, H, W)", "type": "text" } ], "index": 26, "is_list_end_line": true }, { "bbox": [ 119, 303, 290, 312 ], "spans": [ { "bbox": [ 119, 304, 135, 311 ], "score": 0.74, "content": "\\qquad \\times \\quad =", "type": "inline_equation" }, { "bbox": [ 136, 303, 231, 312 ], "score": 1.0, "content": "self.pos(feat_tokens)", "type": "text" }, { "bbox": [ 231, 304, 237, 311 ], "score": 0.41, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 238, 303, 290, 312 ], "score": 1.0, "content": "feat_tokens", "type": "text" } ], "index": 27, "is_list_end_line": true }, { "bbox": [ 119, 309, 259, 319 ], "spans": [ { "bbox": [ 119, 312, 135, 318 ], "score": 0.73, "content": "\\qquad \\times \\quad =", "type": "inline_equation" }, { "bbox": [ 135, 309, 259, 319 ], "score": 1.0, "content": "x.flatten(2).transpose(1, 2)", "type": "text" } ], "index": 28, "is_list_end_line": true }, { "bbox": [ 119, 317, 331, 326 ], "spans": [ { "bbox": [ 119, 318, 135, 325 ], "score": 0.73, "content": "\\qquad \\times \\quad =", "type": "inline_equation" }, { "bbox": [ 135, 317, 331, 326 ], "score": 1.0, "content": "torch.cat((cls_token.unsqueeze(1), x), dim=1)", "type": "text" } ], "index": 29, "is_list_end_line": true }, { "bbox": [ 119, 325, 157, 334 ], "spans": [ { "bbox": [ 119, 325, 157, 334 ], "score": 1.0, "content": "return x", "type": "text" } ], "index": 30, "is_list_end_line": true } ], "index": 15.5, "bbox_fs": [ 105, 101, 497, 334 ] }, { "type": "title", "bbox": [ 109, 366, 298, 377 ], "lines": [ { "bbox": [ 106, 365, 300, 379 ], "spans": [ { "bbox": [ 106, 365, 300, 379 ], "score": 1.0, "content": "D.2 COMPARISON TO LAMBDA NETWORKS", "type": "text" } ], "index": 31 } ], "index": 31 }, { "type": "text", "bbox": [ 107, 388, 505, 465 ], "lines": [ { "bbox": [ 106, 388, 505, 401 ], "spans": [ { "bbox": [ 106, 388, 505, 401 ], "score": 1.0, "content": "Our work is also related to Lambda Networks (Bello, 2021) which uses 2D relative positional encod-", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 398, 505, 413 ], "spans": [ { "bbox": [ 105, 398, 505, 413 ], "score": 1.0, "content": "ings. We evaluate its lambda module with an embedding size of 128, where we denote its encoding", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 409, 505, 423 ], "spans": [ { "bbox": [ 105, 409, 505, 423 ], "score": 1.0, "content": "scheme as RPE2D-d128. Noticeably, this configuration has about 5.9M parameters (comparable to", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 420, 505, 435 ], "spans": [ { "bbox": [ 105, 420, 221, 435 ], "score": 1.0, "content": "DeiT-tiny) but only obtains", "type": "text" }, { "bbox": [ 221, 421, 248, 432 ], "score": 0.86, "content": "6 8 . 7 \\%", "type": "inline_equation" }, { "bbox": [ 248, 420, 505, 435 ], "score": 1.0, "content": ". We attribute its failure to the limited ability in capturing the", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 432, 505, 444 ], "spans": [ { "bbox": [ 105, 432, 505, 444 ], "score": 1.0, "content": "correct positional information. After all, lambda layers are designed with the help of many CNN", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 443, 504, 455 ], "spans": [ { "bbox": [ 105, 443, 504, 455 ], "score": 1.0, "content": "backbones components such as down-sampling to form various stages, to replace ordinary convolu-", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 455, 402, 466 ], "spans": [ { "bbox": [ 106, 455, 402, 466 ], "score": 1.0, "content": "tions in ResNet (He et al., 2016). In contrast, CPVT is transformer-based.", "type": "text" } ], "index": 38 } ], "index": 35, "bbox_fs": [ 105, 388, 505, 466 ] }, { "type": "title", "bbox": [ 109, 483, 282, 494 ], "lines": [ { "bbox": [ 106, 482, 284, 496 ], "spans": [ { "bbox": [ 106, 482, 284, 496 ], "score": 1.0, "content": "D.3 QUALITATIVE ANALYSIS OF CPVT", "type": "text" } ], "index": 39 } ], "index": 39 }, { "type": "text", "bbox": [ 107, 505, 505, 571 ], "lines": [ { "bbox": [ 106, 505, 505, 518 ], "spans": [ { "bbox": [ 106, 505, 505, 518 ], "score": 1.0, "content": "Thus far, we have shown that PEG can have better performance than the original positional encod-", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 516, 505, 528 ], "spans": [ { "bbox": [ 105, 516, 505, 528 ], "score": 1.0, "content": "ings. However, because PEG provides the position in an implicit way, it is interesting to see if PEG", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 527, 505, 540 ], "spans": [ { "bbox": [ 105, 527, 505, 540 ], "score": 1.0, "content": "can indeed provide the position information as the original positional encodings. Here we inves-", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 538, 504, 551 ], "spans": [ { "bbox": [ 105, 538, 464, 551 ], "score": 1.0, "content": "tigate this by visualizing the attention weights of the transformers. Specifically, given a", "type": "text" }, { "bbox": [ 465, 538, 504, 549 ], "score": 0.88, "content": "2 2 4 \\times 2 2 4", "type": "inline_equation" } ], "index": 43 }, { "bbox": [ 105, 549, 505, 561 ], "spans": [ { "bbox": [ 105, 549, 155, 561 ], "score": 1.0, "content": "image (i.e.", "type": "text" }, { "bbox": [ 155, 549, 184, 560 ], "score": 0.89, "content": "1 4 \\times 1 4", "type": "inline_equation" }, { "bbox": [ 185, 549, 390, 561 ], "score": 1.0, "content": "patches), the score matrix within a single head is", "type": "text" }, { "bbox": [ 390, 549, 429, 560 ], "score": 0.89, "content": "1 9 6 \\times 1 9 6", "type": "inline_equation" }, { "bbox": [ 429, 549, 505, 561 ], "score": 1.0, "content": ". We visualize the", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 560, 377, 572 ], "spans": [ { "bbox": [ 106, 560, 377, 572 ], "score": 1.0, "content": "normalized self-attention score matrix of the second encoder block.", "type": "text" } ], "index": 45 } ], "index": 42.5, "bbox_fs": [ 105, 505, 505, 572 ] }, { "type": "text", "bbox": [ 107, 577, 505, 644 ], "lines": [ { "bbox": [ 105, 576, 505, 590 ], "spans": [ { "bbox": [ 105, 576, 505, 590 ], "score": 1.0, "content": "We first visualize the attention weights of DeiT with the original positional encodings. As shown in", "type": "text" } ], "index": 46 }, { "bbox": [ 106, 588, 505, 601 ], "spans": [ { "bbox": [ 106, 588, 505, 601 ], "score": 1.0, "content": "Figure 5 (middle), the diagonal element interacts strongly with its local neighbors but weakly with", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 599, 506, 612 ], "spans": [ { "bbox": [ 105, 599, 506, 612 ], "score": 1.0, "content": "those far-away elements, which suggests that DeiT with the original positional encodings learn to", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 610, 505, 623 ], "spans": [ { "bbox": [ 105, 610, 505, 623 ], "score": 1.0, "content": "attend the local neighbors of each patch. After the positional encodings are removed (denoted by", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 620, 506, 634 ], "spans": [ { "bbox": [ 105, 620, 506, 634 ], "score": 1.0, "content": "DeiT w/o PE), all the patches produce similar attention weights and fail to attend to the patches near", "type": "text" } ], "index": 50 }, { "bbox": [ 106, 632, 231, 645 ], "spans": [ { "bbox": [ 106, 632, 231, 645 ], "score": 1.0, "content": "themselves, see Figure 5 (left).", "type": "text" } ], "index": 51 } ], "index": 48.5, "bbox_fs": [ 105, 576, 506, 645 ] }, { "type": "text", "bbox": [ 107, 649, 504, 682 ], "lines": [ { "bbox": [ 105, 648, 505, 662 ], "spans": [ { "bbox": [ 105, 648, 505, 662 ], "score": 1.0, "content": "Finally, we show the attention weights of our CPVT model with PEG. As shown in Figure 5 (right),", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 659, 506, 673 ], "spans": [ { "bbox": [ 105, 659, 506, 673 ], "score": 1.0, "content": "like the original positional encodings, the model with PEG can also learn a similar attention pattern,", "type": "text" } ], "index": 53 }, { "bbox": [ 106, 671, 444, 683 ], "spans": [ { "bbox": [ 106, 671, 444, 683 ], "score": 1.0, "content": "which indicates that the proposed PEG can provide the position information as well.", "type": "text" } ], "index": 54 } ], "index": 53, "bbox_fs": [ 105, 648, 506, 683 ] }, { "type": "text", "bbox": [ 107, 687, 504, 732 ], "lines": [ { "bbox": [ 106, 687, 506, 699 ], "spans": [ { "bbox": [ 106, 687, 506, 699 ], "score": 1.0, "content": "We illustrate the attention scores in several encoder blocks of DeiT (Touvron et al., 2020) and CPVT", "type": "text" } ], "index": 55 }, { "bbox": [ 105, 699, 506, 711 ], "spans": [ { "bbox": [ 105, 699, 506, 711 ], "score": 1.0, "content": "in the Fig. 6. It shows both methods learn similar locality patterns. As attention scores are computed", "type": "text" } ], "index": 56 }, { "bbox": [ 105, 710, 506, 721 ], "spans": [ { "bbox": [ 105, 710, 506, 721 ], "score": 1.0, "content": "over the tokens projected in different subspaces (Q and K), they do not necessarily show a strict", "type": "text" } ], "index": 57 }, { "bbox": [ 105, 720, 498, 734 ], "spans": [ { "bbox": [ 105, 720, 498, 734 ], "score": 1.0, "content": "diagonal pattern, where some may have slight shift, see DeiT in Fig. 6c and CPVT of Fig. 5 right.", "type": "text" } ], "index": 58 } ], "index": 56.5, "bbox_fs": [ 105, 687, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 155, 81, 455, 167 ], "blocks": [ { "type": "image_body", "bbox": [ 155, 81, 455, 167 ], "group_id": 0, "lines": [ { "bbox": [ 155, 81, 455, 167 ], "spans": [ { "bbox": [ 155, 81, 455, 167 ], "score": 0.96, "type": "image", "image_path": "19c1477edf28cf26db81ecd53fd8947ea0edbe4a9868e6368aab289ef778d7ce.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 155, 81, 455, 109.66666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 155, 109.66666666666667, 455, 138.33333333333334 ], "spans": [], "index": 1 }, { "bbox": [ 155, 138.33333333333334, 455, 167.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 178, 504, 223 ], "group_id": 0, "lines": [ { "bbox": [ 106, 177, 504, 191 ], "spans": [ { "bbox": [ 106, 177, 504, 191 ], "score": 1.0, "content": "Figure 5. Normalized attention scores (first head) of the second encoder block of DeiT without po-", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 189, 504, 201 ], "spans": [ { "bbox": [ 106, 189, 504, 201 ], "score": 1.0, "content": "sition encoding (DeiT w/o PE), DeiT (Touvron et al., 2020), and CPVT on the same input sequence.", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 200, 504, 213 ], "spans": [ { "bbox": [ 105, 200, 504, 213 ], "score": 1.0, "content": "Position encodings are key to developing a schema of locality in lower layers of DeiT. Meantime,", "type": "text" } ], "index": 5 }, { "bbox": [ 107, 212, 426, 223 ], "spans": [ { "bbox": [ 107, 212, 426, 223 ], "score": 1.0, "content": "CPVT profits from conditional encodings and follows a similar locality pattern.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 }, { "type": "image", "bbox": [ 115, 235, 482, 419 ], "blocks": [ { "type": "image_body", "bbox": [ 115, 235, 482, 419 ], "group_id": 1, "lines": [ { "bbox": [ 115, 235, 482, 419 ], "spans": [ { "bbox": [ 115, 235, 482, 419 ], "score": 0.974, "type": "image", "image_path": "0a11531d79480ac414e128993ffaedb8cbf0d456688b90d3ed26f02be66c7b97.jpg" } ] } ], "index": 8, "virtual_lines": [ { "bbox": [ 115, 235, 482, 296.3333333333333 ], "spans": [], "index": 7 }, { "bbox": [ 115, 296.3333333333333, 482, 357.66666666666663 ], "spans": [], "index": 8 }, { "bbox": [ 115, 357.66666666666663, 482, 418.99999999999994 ], "spans": [], "index": 9 } ] }, { "type": "image_caption", "bbox": [ 108, 428, 504, 463 ], "group_id": 1, "lines": [ { "bbox": [ 106, 428, 505, 441 ], "spans": [ { "bbox": [ 106, 428, 505, 441 ], "score": 1.0, "content": "Figure 6. Normalized attention scores (the second and third head) of the second and third encoder", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 439, 504, 452 ], "spans": [ { "bbox": [ 106, 439, 504, 452 ], "score": 1.0, "content": "block of DeiT (Touvron et al., 2020), and CPVT on the same input sequence. DeiT and CPVT share", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 450, 391, 464 ], "spans": [ { "bbox": [ 106, 450, 391, 464 ], "score": 1.0, "content": "similar locality patterns that are aligned diagonally (some might shift).", "type": "text" } ], "index": 12 } ], "index": 11 } ], "index": 9.5 }, { "type": "title", "bbox": [ 107, 484, 309, 495 ], "lines": [ { "bbox": [ 106, 483, 311, 496 ], "spans": [ { "bbox": [ 106, 483, 311, 496 ], "score": 1.0, "content": "D.4 COMPARISON WITH OTHER APPROACHES", "type": "text" } ], "index": 13 } ], "index": 13 }, { "type": "text", "bbox": [ 107, 504, 505, 581 ], "lines": [ { "bbox": [ 105, 504, 506, 517 ], "spans": [ { "bbox": [ 105, 504, 506, 517 ], "score": 1.0, "content": "We further compare our method with other approaches such as CvT (Wu et al., 2021), ConViT", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 515, 506, 528 ], "spans": [ { "bbox": [ 105, 515, 506, 528 ], "score": 1.0, "content": "(d’Ascoli et al., 2021) and CoAtNet (Dai et al., 2021) on ImageNet validation set in Table 15. To", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 526, 505, 539 ], "spans": [ { "bbox": [ 105, 526, 505, 539 ], "score": 1.0, "content": "make fair comparisons, we categorize these methods into two groups: plain and pyramid models.", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 538, 505, 550 ], "spans": [ { "bbox": [ 105, 538, 505, 550 ], "score": 1.0, "content": "Since our models are primarily for plain models, we adapt our methods on two popular pyramid", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 547, 505, 561 ], "spans": [ { "bbox": [ 105, 547, 443, 561 ], "score": 1.0, "content": "frameworks PVT and Swin. Our CPVT-S-GAP slightly outperforms ConViT-S by", "type": "text" }, { "bbox": [ 444, 549, 466, 559 ], "score": 0.85, "content": "0 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 466, 547, 505, 561 ], "score": 1.0, "content": "with 4M", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 558, 506, 572 ], "spans": [ { "bbox": [ 105, 558, 506, 572 ], "score": 1.0, "content": "fewer parameters and 0.8G fewer FLOPs. When equipped with pyramid designs, our methods are", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 571, 261, 582 ], "spans": [ { "bbox": [ 106, 571, 261, 582 ], "score": 1.0, "content": "still comparable to CvT and CoAtNet.", "type": "text" } ], "index": 20 } ], "index": 17 }, { "type": "text", "bbox": [ 106, 594, 505, 693 ], "lines": [ { "bbox": [ 106, 593, 505, 606 ], "spans": [ { "bbox": [ 106, 593, 505, 606 ], "score": 1.0, "content": "Comparison with DeiT w/ Convolutional Projection. Note CvT uses a depth-wise convolution", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 604, 505, 618 ], "spans": [ { "bbox": [ 106, 604, 117, 618 ], "score": 1.0, "content": "in", "type": "text" }, { "bbox": [ 117, 605, 141, 617 ], "score": 0.33, "content": "\\scriptstyle q - k - v", "type": "inline_equation" }, { "bbox": [ 141, 604, 505, 618 ], "score": 1.0, "content": "projection which they call it Convolutional Projection. Instead of using it in all layers, we", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 615, 506, 629 ], "spans": [ { "bbox": [ 105, 615, 506, 629 ], "score": 1.0, "content": "put only one of such design into DeiT-tiny and train such a model from scratch under strictly con-", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 627, 504, 639 ], "spans": [ { "bbox": [ 106, 627, 504, 639 ], "score": 1.0, "content": "trolled settings. We insert it in the position 0 as in our method. The result is shown in Table 16. This", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 637, 506, 650 ], "spans": [ { "bbox": [ 106, 637, 221, 650 ], "score": 1.0, "content": "CvT-flavored DeiT achieves", "type": "text" }, { "bbox": [ 221, 638, 248, 648 ], "score": 0.86, "content": "7 0 . 6 \\%", "type": "inline_equation" }, { "bbox": [ 249, 637, 506, 650 ], "score": 1.0, "content": "top-1 accuracy on ImageNet validation set, which is lower than", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 648, 505, 662 ], "spans": [ { "bbox": [ 105, 648, 127, 662 ], "score": 1.0, "content": "ours", "type": "text" }, { "bbox": [ 127, 649, 160, 660 ], "score": 0.88, "content": "( 7 2 . 4 \\% )", "type": "inline_equation" }, { "bbox": [ 160, 648, 207, 662 ], "score": 1.0, "content": ". Note that", "type": "text" }, { "bbox": [ 208, 650, 214, 660 ], "score": 0.59, "content": "q", "type": "inline_equation" }, { "bbox": [ 214, 648, 505, 662 ], "score": 1.0, "content": "-k-v projections in CvT utilize three depthwise convolutions, therefore,", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 657, 506, 675 ], "spans": [ { "bbox": [ 105, 657, 506, 675 ], "score": 1.0, "content": "this setting has more parameters than ours. This attests the difference of CvT and CPVT, verifying", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 671, 505, 683 ], "spans": [ { "bbox": [ 106, 671, 505, 683 ], "score": 1.0, "content": "our advantage by learning better position encodings other than inserting them in all layers to have", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 682, 397, 694 ], "spans": [ { "bbox": [ 106, 682, 397, 694 ], "score": 1.0, "content": "the ability to capture local context and to remove ambiguity in attention.", "type": "text" } ], "index": 29 } ], "index": 25 } ], "page_idx": 17, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 763 ], "spans": [ { "bbox": [ 299, 750, 313, 763 ], "score": 1.0, "content": "18", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 155, 81, 455, 167 ], "blocks": [ { "type": "image_body", "bbox": [ 155, 81, 455, 167 ], "group_id": 0, "lines": [ { "bbox": [ 155, 81, 455, 167 ], "spans": [ { "bbox": [ 155, 81, 455, 167 ], "score": 0.96, "type": "image", "image_path": "19c1477edf28cf26db81ecd53fd8947ea0edbe4a9868e6368aab289ef778d7ce.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 155, 81, 455, 109.66666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 155, 109.66666666666667, 455, 138.33333333333334 ], "spans": [], "index": 1 }, { "bbox": [ 155, 138.33333333333334, 455, 167.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 178, 504, 223 ], "group_id": 0, "lines": [ { "bbox": [ 106, 177, 504, 191 ], "spans": [ { "bbox": [ 106, 177, 504, 191 ], "score": 1.0, "content": "Figure 5. Normalized attention scores (first head) of the second encoder block of DeiT without po-", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 189, 504, 201 ], "spans": [ { "bbox": [ 106, 189, 504, 201 ], "score": 1.0, "content": "sition encoding (DeiT w/o PE), DeiT (Touvron et al., 2020), and CPVT on the same input sequence.", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 200, 504, 213 ], "spans": [ { "bbox": [ 105, 200, 504, 213 ], "score": 1.0, "content": "Position encodings are key to developing a schema of locality in lower layers of DeiT. Meantime,", "type": "text" } ], "index": 5 }, { "bbox": [ 107, 212, 426, 223 ], "spans": [ { "bbox": [ 107, 212, 426, 223 ], "score": 1.0, "content": "CPVT profits from conditional encodings and follows a similar locality pattern.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 }, { "type": "image", "bbox": [ 115, 235, 482, 419 ], "blocks": [ { "type": "image_body", "bbox": [ 115, 235, 482, 419 ], "group_id": 1, "lines": [ { "bbox": [ 115, 235, 482, 419 ], "spans": [ { "bbox": [ 115, 235, 482, 419 ], "score": 0.974, "type": "image", "image_path": "0a11531d79480ac414e128993ffaedb8cbf0d456688b90d3ed26f02be66c7b97.jpg" } ] } ], "index": 8, "virtual_lines": [ { "bbox": [ 115, 235, 482, 296.3333333333333 ], "spans": [], "index": 7 }, { "bbox": [ 115, 296.3333333333333, 482, 357.66666666666663 ], "spans": [], "index": 8 }, { "bbox": [ 115, 357.66666666666663, 482, 418.99999999999994 ], "spans": [], "index": 9 } ] }, { "type": "image_caption", "bbox": [ 108, 428, 504, 463 ], "group_id": 1, "lines": [ { "bbox": [ 106, 428, 505, 441 ], "spans": [ { "bbox": [ 106, 428, 505, 441 ], "score": 1.0, "content": "Figure 6. Normalized attention scores (the second and third head) of the second and third encoder", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 439, 504, 452 ], "spans": [ { "bbox": [ 106, 439, 504, 452 ], "score": 1.0, "content": "block of DeiT (Touvron et al., 2020), and CPVT on the same input sequence. DeiT and CPVT share", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 450, 391, 464 ], "spans": [ { "bbox": [ 106, 450, 391, 464 ], "score": 1.0, "content": "similar locality patterns that are aligned diagonally (some might shift).", "type": "text" } ], "index": 12 } ], "index": 11 } ], "index": 9.5 }, { "type": "title", "bbox": [ 107, 484, 309, 495 ], "lines": [ { "bbox": [ 106, 483, 311, 496 ], "spans": [ { "bbox": [ 106, 483, 311, 496 ], "score": 1.0, "content": "D.4 COMPARISON WITH OTHER APPROACHES", "type": "text" } ], "index": 13 } ], "index": 13 }, { "type": "text", "bbox": [ 107, 504, 505, 581 ], "lines": [ { "bbox": [ 105, 504, 506, 517 ], "spans": [ { "bbox": [ 105, 504, 506, 517 ], "score": 1.0, "content": "We further compare our method with other approaches such as CvT (Wu et al., 2021), ConViT", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 515, 506, 528 ], "spans": [ { "bbox": [ 105, 515, 506, 528 ], "score": 1.0, "content": "(d’Ascoli et al., 2021) and CoAtNet (Dai et al., 2021) on ImageNet validation set in Table 15. To", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 526, 505, 539 ], "spans": [ { "bbox": [ 105, 526, 505, 539 ], "score": 1.0, "content": "make fair comparisons, we categorize these methods into two groups: plain and pyramid models.", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 538, 505, 550 ], "spans": [ { "bbox": [ 105, 538, 505, 550 ], "score": 1.0, "content": "Since our models are primarily for plain models, we adapt our methods on two popular pyramid", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 547, 505, 561 ], "spans": [ { "bbox": [ 105, 547, 443, 561 ], "score": 1.0, "content": "frameworks PVT and Swin. Our CPVT-S-GAP slightly outperforms ConViT-S by", "type": "text" }, { "bbox": [ 444, 549, 466, 559 ], "score": 0.85, "content": "0 . 2 \\%", "type": "inline_equation" }, { "bbox": [ 466, 547, 505, 561 ], "score": 1.0, "content": "with 4M", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 558, 506, 572 ], "spans": [ { "bbox": [ 105, 558, 506, 572 ], "score": 1.0, "content": "fewer parameters and 0.8G fewer FLOPs. When equipped with pyramid designs, our methods are", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 571, 261, 582 ], "spans": [ { "bbox": [ 106, 571, 261, 582 ], "score": 1.0, "content": "still comparable to CvT and CoAtNet.", "type": "text" } ], "index": 20 } ], "index": 17, "bbox_fs": [ 105, 504, 506, 582 ] }, { "type": "text", "bbox": [ 106, 594, 505, 693 ], "lines": [ { "bbox": [ 106, 593, 505, 606 ], "spans": [ { "bbox": [ 106, 593, 505, 606 ], "score": 1.0, "content": "Comparison with DeiT w/ Convolutional Projection. Note CvT uses a depth-wise convolution", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 604, 505, 618 ], "spans": [ { "bbox": [ 106, 604, 117, 618 ], "score": 1.0, "content": "in", "type": "text" }, { "bbox": [ 117, 605, 141, 617 ], "score": 0.33, "content": "\\scriptstyle q - k - v", "type": "inline_equation" }, { "bbox": [ 141, 604, 505, 618 ], "score": 1.0, "content": "projection which they call it Convolutional Projection. Instead of using it in all layers, we", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 615, 506, 629 ], "spans": [ { "bbox": [ 105, 615, 506, 629 ], "score": 1.0, "content": "put only one of such design into DeiT-tiny and train such a model from scratch under strictly con-", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 627, 504, 639 ], "spans": [ { "bbox": [ 106, 627, 504, 639 ], "score": 1.0, "content": "trolled settings. We insert it in the position 0 as in our method. The result is shown in Table 16. This", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 637, 506, 650 ], "spans": [ { "bbox": [ 106, 637, 221, 650 ], "score": 1.0, "content": "CvT-flavored DeiT achieves", "type": "text" }, { "bbox": [ 221, 638, 248, 648 ], "score": 0.86, "content": "7 0 . 6 \\%", "type": "inline_equation" }, { "bbox": [ 249, 637, 506, 650 ], "score": 1.0, "content": "top-1 accuracy on ImageNet validation set, which is lower than", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 648, 505, 662 ], "spans": [ { "bbox": [ 105, 648, 127, 662 ], "score": 1.0, "content": "ours", "type": "text" }, { "bbox": [ 127, 649, 160, 660 ], "score": 0.88, "content": "( 7 2 . 4 \\% )", "type": "inline_equation" }, { "bbox": [ 160, 648, 207, 662 ], "score": 1.0, "content": ". Note that", "type": "text" }, { "bbox": [ 208, 650, 214, 660 ], "score": 0.59, "content": "q", "type": "inline_equation" }, { "bbox": [ 214, 648, 505, 662 ], "score": 1.0, "content": "-k-v projections in CvT utilize three depthwise convolutions, therefore,", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 657, 506, 675 ], "spans": [ { "bbox": [ 105, 657, 506, 675 ], "score": 1.0, "content": "this setting has more parameters than ours. This attests the difference of CvT and CPVT, verifying", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 671, 505, 683 ], "spans": [ { "bbox": [ 106, 671, 505, 683 ], "score": 1.0, "content": "our advantage by learning better position encodings other than inserting them in all layers to have", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 682, 397, 694 ], "spans": [ { "bbox": [ 106, 682, 397, 694 ], "score": 1.0, "content": "the ability to capture local context and to remove ambiguity in attention.", "type": "text" } ], "index": 29 } ], "index": 25, "bbox_fs": [ 105, 593, 506, 694 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 169, 223, 442, 337 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 175, 505, 209 ], "group_id": 0, "lines": [ { "bbox": [ 106, 175, 505, 189 ], "spans": [ { "bbox": [ 106, 175, 505, 189 ], "score": 1.0, "content": "Table 15. Performance comparison with other approaches such as CvT (Wu et al., 2021), ConViT", "type": "text" } ], "index": 0 }, { "bbox": [ 107, 187, 504, 199 ], "spans": [ { "bbox": [ 107, 187, 504, 199 ], "score": 1.0, "content": "(d’Ascoli et al., 2021) and CoAtNet (Dai et al., 2021) on ImageNet validation set. All the models", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 198, 484, 211 ], "spans": [ { "bbox": [ 106, 198, 399, 211 ], "score": 1.0, "content": "are trained on ImageNet-1k dataset and tested on the validation set using", "type": "text" }, { "bbox": [ 399, 198, 439, 208 ], "score": 0.89, "content": "2 2 4 \\times 2 2 4", "type": "inline_equation" }, { "bbox": [ 439, 198, 484, 211 ], "score": 1.0, "content": "resolution.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "table_body", "bbox": [ 169, 223, 442, 337 ], "group_id": 0, "lines": [ { "bbox": [ 169, 223, 442, 337 ], "spans": [ { "bbox": [ 169, 223, 442, 337 ], "score": 0.981, "html": "
ModelTypeParamsFLOPsTop-1 Acc(%)
DeiT-small (Touvron et al., 2020)ConViT-S (d'Ascoli et al.,2021)CPVT-S-GAP (ours)PlainPlainPlain22M27M23M4.6G5.4G4.6G79.981.381.5
CoAtNet-0 (Dai et al., 2021)CvT-13 (Wu et al., 2021)PVT-small (Wang et al., 2021)PVT-small+PEG+GAPSwin-tiny (Liu et al.,2021)Swin-tiny+PEG+GAPPyramidPyramidPyramidPyramidPyramidPyramid25M20M25M25M29M4.2G4.5G3.8G3.8G4.5G81.681.679.881.281.382.3
29M4.5G
", "type": "table", "image_path": "92a0883388c61f8609d4cd2d321d19b1837a9453c336019225ca6e91ff9d7de4.jpg" } ] } ], "index": 4, "virtual_lines": [ { "bbox": [ 169, 223, 442, 261.0 ], "spans": [], "index": 3 }, { "bbox": [ 169, 261.0, 442, 299.0 ], "spans": [], "index": 4 }, { "bbox": [ 169, 299.0, 442, 337.0 ], "spans": [], "index": 5 } ] } ], "index": 2.5 }, { "type": "table", "bbox": [ 174, 587, 437, 630 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 538, 504, 572 ], "group_id": 1, "lines": [ { "bbox": [ 107, 539, 503, 551 ], "spans": [ { "bbox": [ 107, 539, 503, 551 ], "score": 1.0, "content": "Table 16. Comparison with positional encoding in CvT (Wu et al., 2021) on ImageNet validation set.", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 549, 503, 562 ], "spans": [ { "bbox": [ 106, 549, 464, 562 ], "score": 1.0, "content": "All the models are trained on ImageNet-1k dataset and tested on the validation set using", "type": "text" }, { "bbox": [ 464, 550, 503, 561 ], "score": 0.87, "content": "2 2 4 \\times 2 2 4", "type": "inline_equation" } ], "index": 7 }, { "bbox": [ 105, 560, 153, 573 ], "spans": [ { "bbox": [ 105, 560, 153, 573 ], "score": 1.0, "content": "resolution.", "type": "text" } ], "index": 8 } ], "index": 7 }, { "type": "table_body", "bbox": [ 174, 587, 437, 630 ], "group_id": 1, "lines": [ { "bbox": [ 174, 587, 437, 630 ], "spans": [ { "bbox": [ 174, 587, 437, 630 ], "score": 0.972, "html": "
ModelParamsInsert PositionTop-1 Acc (%)
CPVT-Ti DeiT+ Convolutional Projection5681320 56853520 072.4 70.6
", "type": "table", "image_path": "5fc96495858a148a2706ce4d6ef25f7ce93fdb93652fb7c2ecf894f53719bf2d.jpg" } ] } ], "index": 10, "virtual_lines": [ { "bbox": [ 174, 587, 437, 601.3333333333334 ], "spans": [], "index": 9 }, { "bbox": [ 174, 601.3333333333334, 437, 615.6666666666667 ], "spans": [], "index": 10 }, { "bbox": [ 174, 615.6666666666667, 437, 630.0000000000001 ], "spans": [], "index": 11 } ] } ], "index": 8.5 } ], "page_idx": 18, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2023", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "19", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 169, 223, 442, 337 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 175, 505, 209 ], "group_id": 0, "lines": [ { "bbox": [ 106, 175, 505, 189 ], "spans": [ { "bbox": [ 106, 175, 505, 189 ], "score": 1.0, "content": "Table 15. Performance comparison with other approaches such as CvT (Wu et al., 2021), ConViT", "type": "text" } ], "index": 0 }, { "bbox": [ 107, 187, 504, 199 ], "spans": [ { "bbox": [ 107, 187, 504, 199 ], "score": 1.0, "content": "(d’Ascoli et al., 2021) and CoAtNet (Dai et al., 2021) on ImageNet validation set. All the models", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 198, 484, 211 ], "spans": [ { "bbox": [ 106, 198, 399, 211 ], "score": 1.0, "content": "are trained on ImageNet-1k dataset and tested on the validation set using", "type": "text" }, { "bbox": [ 399, 198, 439, 208 ], "score": 0.89, "content": "2 2 4 \\times 2 2 4", "type": "inline_equation" }, { "bbox": [ 439, 198, 484, 211 ], "score": 1.0, "content": "resolution.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "table_body", "bbox": [ 169, 223, 442, 337 ], "group_id": 0, "lines": [ { "bbox": [ 169, 223, 442, 337 ], "spans": [ { "bbox": [ 169, 223, 442, 337 ], "score": 0.981, "html": "
ModelTypeParamsFLOPsTop-1 Acc(%)
DeiT-small (Touvron et al., 2020)ConViT-S (d'Ascoli et al.,2021)CPVT-S-GAP (ours)PlainPlainPlain22M27M23M4.6G5.4G4.6G79.981.381.5
CoAtNet-0 (Dai et al., 2021)CvT-13 (Wu et al., 2021)PVT-small (Wang et al., 2021)PVT-small+PEG+GAPSwin-tiny (Liu et al.,2021)Swin-tiny+PEG+GAPPyramidPyramidPyramidPyramidPyramidPyramid25M20M25M25M29M4.2G4.5G3.8G3.8G4.5G81.681.679.881.281.382.3
29M4.5G
", "type": "table", "image_path": "92a0883388c61f8609d4cd2d321d19b1837a9453c336019225ca6e91ff9d7de4.jpg" } ] } ], "index": 4, "virtual_lines": [ { "bbox": [ 169, 223, 442, 261.0 ], "spans": [], "index": 3 }, { "bbox": [ 169, 261.0, 442, 299.0 ], "spans": [], "index": 4 }, { "bbox": [ 169, 299.0, 442, 337.0 ], "spans": [], "index": 5 } ] } ], "index": 2.5 }, { "type": "table", "bbox": [ 174, 587, 437, 630 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 538, 504, 572 ], "group_id": 1, "lines": [ { "bbox": [ 107, 539, 503, 551 ], "spans": [ { "bbox": [ 107, 539, 503, 551 ], "score": 1.0, "content": "Table 16. Comparison with positional encoding in CvT (Wu et al., 2021) on ImageNet validation set.", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 549, 503, 562 ], "spans": [ { "bbox": [ 106, 549, 464, 562 ], "score": 1.0, "content": "All the models are trained on ImageNet-1k dataset and tested on the validation set using", "type": "text" }, { "bbox": [ 464, 550, 503, 561 ], "score": 0.87, "content": "2 2 4 \\times 2 2 4", "type": "inline_equation" } ], "index": 7 }, { "bbox": [ 105, 560, 153, 573 ], "spans": [ { "bbox": [ 105, 560, 153, 573 ], "score": 1.0, "content": "resolution.", "type": "text" } ], "index": 8 } ], "index": 7 }, { "type": "table_body", "bbox": [ 174, 587, 437, 630 ], "group_id": 1, "lines": [ { "bbox": [ 174, 587, 437, 630 ], "spans": [ { "bbox": [ 174, 587, 437, 630 ], "score": 0.972, "html": "
ModelParamsInsert PositionTop-1 Acc (%)
CPVT-Ti DeiT+ Convolutional Projection5681320 56853520 072.4 70.6
", "type": "table", "image_path": "5fc96495858a148a2706ce4d6ef25f7ce93fdb93652fb7c2ecf894f53719bf2d.jpg" } ] } ], "index": 10, "virtual_lines": [ { "bbox": [ 174, 587, 437, 601.3333333333334 ], "spans": [], "index": 9 }, { "bbox": [ 174, 601.3333333333334, 437, 615.6666666666667 ], "spans": [], "index": 10 }, { "bbox": [ 174, 615.6666666666667, 437, 630.0000000000001 ], "spans": [], "index": 11 } ] } ], "index": 8.5 } ] } ], "_backend": "pipeline", "_version_name": "2.2.2" }