---hardware_support: - NVIDIAtasks: - Image Classificationtraining_framework: PaddlePaddle---# MambaVision模型权重## 模型描述MambaVision 是 NVIDIA 提出的分层视觉骨干网络，在统一架构中结合卷积、自注意力与 Mamba 状态空间混合器。浅层卷积负责提取局部视觉特征，深层 Mamba/Attention Block 建模长程依赖，可用于 ImageNet 图像分类以及检测、分割等下游任务的视觉骨干。本模型空间提供全部 11 个官方分类规格由 PyTorch 转换得到的 PaddlePaddle 权重。转换只改变参数名称和张量布局，不重新训练或修改参数数值。- **论文**：[MambaVision: A Hybrid Mamba-Transformer Vision Backbone](https://arxiv.org/abs/2407.08083)- **原始代码**：[NVlabs/MambaVision](https://github.com/NVlabs/MambaVision)- **PaddleClas 代码**：配套的 MambaVision PaddleClas 实现- **PaddleClas 配置**：`ppcls/configs/ImageNet/MambaVision/`- **分类头**：全部权重均为 1,000 类；名称中的 `21K` 表示上游预训练来源，不表示当前分类头为 21,000 类。## 摘要| 项目 | 说明 | 结果 || --- | --- | --- || 权重转换 | 缺失键、冗余键、shape mismatch 均为 0 | 11/11 通过 || 随机输入前向 | PyTorch/Paddle logits 最大绝对差 `<1e-4` | 11/11 通过，最差 `1.520e-6` || 原生 PaddleClas 前向 | 纯 Paddle `parallel` scan 实现 logits 最大绝对差 `<1e-4` | 11/11 通过，最差 `1.639e-6` || ImageNet-1K 配对 | 同一 50,000 张图像的 Top-1 一致率 `>=99.8%` | 11/11 通过，均为 `100.0%` || 训练流程 | PaddleClas 训练入口可运行且 Loss 下降 | 11/11 完成 50 step |## 模型变体| 模型名称 | PaddleClas 入口 | 上游 Top-1 / Top-5 | 输入尺寸 | Stage 深度 | 参数量 || --- | --- | ---: | ---: | --- | ---: || MambaVision-T-1K | `MambaVision_T` | 82.3% / 96.2% | 224x224 | 1/3/8/4 | 31.8M || MambaVision-T2-1K | `MambaVision_T2` | 82.7% / 96.3% | 224x224 | 1/3/11/4 | 35.1M || MambaVision-S-1K | `MambaVision_S` | 83.3% / 96.5% | 224x224 | 3/3/7/5 | 50.1M || MambaVision-B-1K | `MambaVision_B` | 84.2% / 96.9% | 224x224 | 3/3/10/5 | 97.7M || MambaVision-B-21K | `MambaVision_B_21K` | 84.9% / 97.5% | 224x224 | 3/3/10/5 | 97.7M || MambaVision-L-1K | `MambaVision_L` | 85.0% / 97.1% | 224x224 | 3/3/10/5 | 227.9M || MambaVision-L-21K | `MambaVision_L_21K` | 86.1% / 97.9% | 224x224 | 3/3/10/5 | 227.9M || MambaVision-L2-1K | `MambaVision_L2` | 85.3% / 97.2% | 224x224 | 3/3/12/5 | 241.5M || MambaVision-L2-512-21K | `MambaVision_L2_512_21K` | 87.3% / 98.4% | 512x512 | 3/3/12/5 | 241.5M || MambaVision-L3-256-21K | `MambaVision_L3_256_21K` | 87.3% / 98.3% | 256x256 | 3/3/20/10 | 739.6M || MambaVision-L3-512-21K | `MambaVision_L3_512_21K` | 88.1% / 98.6% | 512x512 | 3/3/20/10 | 739.6M |以上“上游 Top-1/Top-5”为 NVIDIA 发布的 PyTorch 参考值。本文后续 50K 结果用于证明转换前后权重的评估一致性，不是使用 PaddlePaddle 从头训练的精度复现结果。## 预训练权重### 权重来源所有权重均从 NVIDIA 官方 MambaVision PyTorch 检查点转换而来。转换对卷积权重保持原布局，对 `Linear.weight` 执行 PyTorch `[out, in]` 到 Paddle `[in, out]` 转置，并严格检查全部参数键和 shape。| PaddleClas 入口 | 权重文件 | 文件大小 | SHA-256 前 12 位 || --- | --- | ---: | --- || `MambaVision_T` | `mambavision_tiny_1k.pdparams` | 122 MiB | `fcca1f7ea10b` || `MambaVision_T2` | `mambavision_tiny2_1k.pdparams` | 134 MiB | `d6c7281b3c40` || `MambaVision_S` | `mambavision_small_1k.pdparams` | 192 MiB | `14748c6c2328` || `MambaVision_B` | `mambavision_base_1k.pdparams` | 373 MiB | `1dd9377c2b35` || `MambaVision_B_21K` | `mambavision_base_21k.pdparams` | 373 MiB | `49dbf0e067ae` || `MambaVision_L` | `mambavision_large_1k.pdparams` | 870 MiB | `1448379ed967` || `MambaVision_L_21K` | `mambavision_large_21k.pdparams` | 870 MiB | `8e0f0f2b9f49` || `MambaVision_L2` | `mambavision_large2_1k.pdparams` | 922 MiB | `ca1a9c47882d` || `MambaVision_L2_512_21K` | `mambavision_L2_21k_240m_512.pdparams` | 922 MiB | `8c55870a42f2` || `MambaVision_L3_256_21K` | `mambavision_L3_21k_740m_256.pdparams` | 2.8 GiB | `6e8df2b7b781` || `MambaVision_L3_512_21K` | `mambavision_L3_21k_740m_512.pdparams` | 2.8 GiB | `3a5df64acaf4` |完整 SHA-256 位于 `SHA256SUMS`，结构化元数据位于 `manifest.json`。11 个权重合计 10,923,012,243 字节。## 使用方式### PaddleClas 中加载以下示例以 Tiny 为例，其他规格替换模型入口、输入尺寸和权重文件即可。当前 PaddleClas `release/2.6` 尚未内置 MambaVision，使用前需先安装配套的 MambaVision PaddleClas 实现。```pythonimport paddlefrom ppcls.arch.backbone.model_zoo.mambavision import MambaVision_Tmodel = MambaVision_T(pretrained=False, scan_impl="parallel")state_dict = paddle.load("mambavision_tiny_1k.pdparams")model.set_state_dict(state_dict)model.eval()x = paddle.randn([1, 3, 224, 224])with paddle.no_grad(): logits = model(x)print(logits.shape) # [1, 1000]```### 校验下载文件```bashsha256sum -c SHA256SUMSpython -m json.tool manifest.json >/dev/null```## 精度对齐### 1. 随机输入 Logits 对齐：PyTorch vs PaddlePaddlePyTorch 与 PaddlePaddle 加载同源权重，并使用相同的随机输入（seed=2026）。对齐阈值为最大绝对误差 `< 1e-4`。| 模型 | 输入尺寸 | 最大绝对误差 | 平均绝对误差 | 状态 || --- | ---: | ---: | ---: | --- || MambaVision-T-1K | 224x224 | 1.490e-6 | 2.590e-7 | PASS || MambaVision-T2-1K | 224x224 | 8.345e-7 | 1.389e-7 | PASS || MambaVision-S-1K | 224x224 | 6.855e-7 | 1.153e-7 | PASS || MambaVision-B-1K | 224x224 | 5.662e-7 | 1.295e-7 | PASS || MambaVision-B-21K | 224x224 | 1.132e-6 | 2.329e-7 | PASS || MambaVision-L-1K | 224x224 | 3.129e-7 | 7.971e-8 | PASS || MambaVision-L-21K | 224x224 | 7.413e-7 | 1.986e-7 | PASS || MambaVision-L2-1K | 224x224 | 3.278e-7 | 7.180e-8 | PASS || MambaVision-L2-512-21K | 512x512 | 1.520e-6 | 3.518e-7 | PASS || MambaVision-L3-256-21K | 256x256 | 9.537e-7 | 1.993e-7 | PASS || MambaVision-L3-512-21K | 512x512 | 8.233e-7 | 1.868e-7 | PASS |Tiny 还通过 128x128 和 256x320 两个额外尺寸的多尺度前向检查；全部随机输入用例中最差最大绝对误差为 `3.636e-6`。此外，PaddleClas 纯 Paddle `parallel` scan 实现已独立完成 11 个规格配对，最差最大绝对误差为 `1.639e-6`。### 2. ImageNet-1K 验证集PyTorch 与 PaddlePaddle 使用相同的 50,000 张 ImageNet-1K 验证图像、相同预处理 tensor 和同源权重。该检查验证权重转换和前向输出一致性，不代表使用 PaddlePaddle 从头训练复现论文精度。| 模型 | 图像数 | PyTorch Top-1 / Top-5 | Paddle Top-1 / Top-5 | Top-1 一致率 | 分类一致 | logits 最大误差 | logits 对齐 | 状态 || --- | ---: | ---: | ---: | ---: | --- | ---: | --- | --- || MambaVision-T-1K | 50,000 | 82.176% / 96.172% | 82.176% / 96.172% | 100.000% | PASS | 2.456e-05 | PASS | PASS || MambaVision-T2-1K | 50,000 | 82.636% / 96.272% | 82.636% / 96.272% | 100.000% | PASS | 2.402e-05 | PASS | PASS || MambaVision-S-1K | 50,000 | 83.232% / 96.502% | 83.232% / 96.502% | 100.000% | PASS | 1.287e-05 | PASS | PASS || MambaVision-B-1K | 50,000 | 84.204% / 96.848% | 84.204% / 96.848% | 100.000% | PASS | 3.159e-05 | PASS | PASS || MambaVision-B-21K | 50,000 | 84.876% / 97.478% | 84.876% / 97.478% | 100.000% | PASS | 2.003e-05 | PASS | PASS || MambaVision-L-1K | 50,000 | 84.954% / 97.078% | 84.954% / 97.078% | 100.000% | PASS | 3.934e-05 | PASS | PASS || MambaVision-L-21K | 50,000 | 86.140% / 97.968% | 86.140% / 97.968% | 100.000% | PASS | 4.673e-05 | PASS | PASS || MambaVision-L2-1K | 50,000 | 85.282% / 97.160% | 85.282% / 97.160% | 100.000% | PASS | 3.076e-05 | PASS | PASS || MambaVision-L2-512-21K | 50,000 | 87.114% / 98.256% | 87.114% / 98.256% | 100.000% | PASS | 8.440e-05 | PASS | PASS || MambaVision-L3-256-21K | 50,000 | 87.294% / 98.318% | 87.294% / 98.318% | 100.000% | PASS | 8.297e-05 | PASS | PASS || MambaVision-L3-512-21K | 50,000 | 87.822% / 98.452% | 87.822% / 98.452% | 100.000% | PASS | 7.239e-05 | PASS | PASS |当前完成 `11/11` 个规格的严格 50K 配对评估。11 个规格的逐图 Top-1 预测一致率均为 `100.0%`，且 logits 最大绝对差均低于 `1e-4`。### 3. 训练流程验证全部 11 个规格均通过 PaddleClas `tools/train.py` 对固定真实 ImageNet 批次训练 50 step。训练使用转换后的预训练权重、CrossEntropyLoss 和 AdamW（learning rate `1e-4`、weight decay `0.01`），并启用 `to_static`。以下逐步记录日志中的 Loss；该检查证明训练流程可执行且 Loss 下降，不等同于完整 ImageNet 训练或精度复现。#### MambaVision-T-1K- 输入尺寸：`224x224`- Batch Size：`32`| Step | Loss || ---: | ---: || 0 | 0.29934 || 1 | 0.21534 || 2 | 0.17814 || 3 | 0.13906 || 4 | 0.11292 || 5 | 0.09490 || 6 | 0.08194 || 7 | 0.07206 || 8 | 0.06423 || 9 | 0.06082 || 10 | 0.05564 || 11 | 0.05105 || 12 | 0.04716 || 13 | 0.04384 || 14 | 0.04094 || 15 | 0.03861 || 16 | 0.03635 || 17 | 0.03435 || 18 | 0.03259 || 19 | 0.03098 || 20 | 0.02954 || 21 | 0.02832 || 22 | 0.02712 || 23 | 0.02599 || 24 | 0.02496 || 25 | 0.02404 || 26 | 0.02315 || 27 | 0.02296 || 28 | 0.02217 || 29 | 0.02148 || 30 | 0.02079 || 31 | 0.02014 || 32 | 0.01955 || 33 | 0.01898 || 34 | 0.01845 || 35 | 0.01794 || 36 | 0.01746 || 37 | 0.01706 || 38 | 0.01663 || 39 | 0.01622 || 40 | 0.01585 || 41 | 0.01547 || 42 | 0.01512 || 43 | 0.01479 || 44 | 0.01446 || 45 | 0.01415 || 46 | 0.01385 || 47 | 0.01357 || 48 | 0.01330 || 49 | 0.01308 |#### MambaVision-T2-1K- 输入尺寸：`224x224`- Batch Size：`24`| Step | Loss || ---: | ---: || 0 | 0.38907 || 1 | 0.26223 || 2 | 0.19170 || 3 | 0.15396 || 4 | 0.12511 || 5 | 0.10554 || 6 | 0.09129 || 7 | 0.08038 || 8 | 0.07159 || 9 | 0.06452 || 10 | 0.05873 || 11 | 0.05389 || 12 | 0.04982 || 13 | 0.04630 || 14 | 0.04322 || 15 | 0.04056 || 16 | 0.03839 || 17 | 0.03627 || 18 | 0.03450 || 19 | 0.03278 || 20 | 0.03123 || 21 | 0.02981 || 22 | 0.02852 || 23 | 0.02734 || 24 | 0.02625 || 25 | 0.02524 || 26 | 0.02431 || 27 | 0.02345 || 28 | 0.02264 || 29 | 0.02190 || 30 | 0.02120 || 31 | 0.02054 || 32 | 0.01992 || 33 | 0.01933 || 34 | 0.01878 || 35 | 0.01827 || 36 | 0.01777 || 37 | 0.01731 || 38 | 0.01687 || 39 | 0.01645 || 40 | 0.01606 || 41 | 0.01569 || 42 | 0.01532 || 43 | 0.01498 || 44 | 0.01466 || 45 | 0.01435 || 46 | 0.01404 || 47 | 0.01375 || 48 | 0.01347 || 49 | 0.01320 |#### MambaVision-S-1K- 输入尺寸：`224x224`- Batch Size：`20`| Step | Loss || ---: | ---: || 0 | 0.48219 || 1 | 0.35910 || 2 | 0.26927 || 3 | 0.24154 || 4 | 0.19810 || 5 | 0.16766 || 6 | 0.14502 || 7 | 0.12751 || 8 | 0.11462 || 9 | 0.10341 || 10 | 0.09414 || 11 | 0.08644 || 12 | 0.07991 || 13 | 0.07429 || 14 | 0.06939 || 15 | 0.06509 || 16 | 0.06128 || 17 | 0.05789 || 18 | 0.05487 || 19 | 0.05214 || 20 | 0.04967 || 21 | 0.04745 || 22 | 0.04539 || 23 | 0.04351 || 24 | 0.04178 || 25 | 0.04018 || 26 | 0.03872 || 27 | 0.03734 || 28 | 0.03606 || 29 | 0.03486 || 30 | 0.03374 || 31 | 0.03269 || 32 | 0.03172 || 33 | 0.03079 || 34 | 0.02991 || 35 | 0.02909 || 36 | 0.02830 || 37 | 0.02764 || 38 | 0.02693 || 39 | 0.02626 || 40 | 0.02563 || 41 | 0.02502 || 42 | 0.02444 || 43 | 0.02389 || 44 | 0.02336 || 45 | 0.02285 || 46 | 0.02238 || 47 | 0.02191 || 48 | 0.02147 || 49 | 0.02104 |#### MambaVision-B-1K- 输入尺寸：`224x224`- Batch Size：`12`| Step | Loss || ---: | ---: || 0 | 0.75239 || 1 | 0.43447 || 2 | 0.30205 || 3 | 0.23011 || 4 | 0.18599 || 5 | 0.15542 || 6 | 0.13359 || 7 | 0.11701 || 8 | 0.10404 || 9 | 0.09370 || 10 | 0.08525 || 11 | 0.07816 || 12 | 0.07217 || 13 | 0.06703 || 14 | 0.06258 || 15 | 0.05868 || 16 | 0.05529 || 17 | 0.05222 || 18 | 0.04948 || 19 | 0.04700 || 20 | 0.04477 || 21 | 0.04274 || 22 | 0.04089 || 23 | 0.03918 || 24 | 0.03762 || 25 | 0.03618 || 26 | 0.03484 || 27 | 0.03360 || 28 | 0.03244 || 29 | 0.03136 || 30 | 0.03035 || 31 | 0.02941 || 32 | 0.02852 || 33 | 0.02768 || 34 | 0.02689 || 35 | 0.02614 || 36 | 0.02544 || 37 | 0.02477 || 38 | 0.02414 || 39 | 0.02354 || 40 | 0.02296 || 41 | 0.02242 || 42 | 0.02190 || 43 | 0.02140 || 44 | 0.02093 || 45 | 0.02048 || 46 | 0.02004 || 47 | 0.01963 || 48 | 0.01923 || 49 | 0.01885 |#### MambaVision-B-21K- 输入尺寸：`224x224`- Batch Size：`8`| Step | Loss || ---: | ---: || 0 | 0.73282 || 1 | 0.49616 || 2 | 0.34627 || 3 | 0.26447 || 4 | 0.21606 || 5 | 0.18132 || 6 | 0.15578 || 7 | 0.13673 || 8 | 0.12216 || 9 | 0.11021 || 10 | 0.10046 || 11 | 0.09214 || 12 | 0.08511 || 13 | 0.07905 || 14 | 0.07381 || 15 | 0.06924 || 16 | 0.06535 || 17 | 0.06174 || 18 | 0.05851 || 19 | 0.05561 || 20 | 0.05297 || 21 | 0.05058 || 22 | 0.04927 || 23 | 0.04724 || 24 | 0.04536 || 25 | 0.04363 || 26 | 0.04204 || 27 | 0.04055 || 28 | 0.03916 || 29 | 0.03787 || 30 | 0.03666 || 31 | 0.03552 || 32 | 0.03445 || 33 | 0.03345 || 34 | 0.03250 || 35 | 0.03160 || 36 | 0.03076 || 37 | 0.02995 || 38 | 0.02918 || 39 | 0.02846 || 40 | 0.02777 || 41 | 0.02712 || 42 | 0.02649 || 43 | 0.02589 || 44 | 0.02532 || 45 | 0.02483 || 46 | 0.02431 || 47 | 0.02380 || 48 | 0.02336 || 49 | 0.02289 |#### MambaVision-L-1K- 输入尺寸：`224x224`- Batch Size：`6`| Step | Loss || ---: | ---: || 0 | 1.69471 || 1 | 1.13642 || 2 | 0.85638 || 3 | 0.67056 || 4 | 0.54373 || 5 | 0.45630 || 6 | 0.39226 || 7 | 0.34385 || 8 | 0.30599 || 9 | 0.27560 || 10 | 0.25067 || 11 | 0.22983 || 12 | 0.21220 || 13 | 0.19707 || 14 | 0.18399 || 15 | 0.17253 || 16 | 0.16239 || 17 | 0.15364 || 18 | 0.14556 || 19 | 0.13830 || 20 | 0.13172 || 21 | 0.12592 || 22 | 0.12046 || 23 | 0.11544 || 24 | 0.11083 || 25 | 0.10657 || 26 | 0.10263 || 27 | 0.09897 || 28 | 0.09555 || 29 | 0.09237 || 30 | 0.08940 || 31 | 0.08660 || 32 | 0.08398 || 33 | 0.08151 || 34 | 0.07919 || 35 | 0.07699 || 36 | 0.07491 || 37 | 0.07294 || 38 | 0.07107 || 39 | 0.06930 || 40 | 0.06761 || 41 | 0.06600 || 42 | 0.06447 || 43 | 0.06300 || 44 | 0.06160 || 45 | 0.06027 || 46 | 0.05899 || 47 | 0.05776 || 48 | 0.05658 || 49 | 0.05545 |#### MambaVision-L-21K- 输入尺寸：`224x224`- Batch Size：`3`| Step | Loss || ---: | ---: || 0 | 1.22610 || 1 | 0.79174 || 2 | 0.55321 || 3 | 0.43295 || 4 | 0.34975 || 5 | 0.29215 || 6 | 0.25148 || 7 | 0.22036 || 8 | 0.19659 || 9 | 0.17730 || 10 | 0.16134 || 11 | 0.14800 || 12 | 0.13675 || 13 | 0.12703 || 14 | 0.11896 || 15 | 0.11157 || 16 | 0.10503 || 17 | 0.09921 || 18 | 0.09400 || 19 | 0.08931 || 20 | 0.08507 || 21 | 0.08123 || 22 | 0.07770 || 23 | 0.07447 || 24 | 0.07153 || 25 | 0.06879 || 26 | 0.06624 || 27 | 0.06389 || 28 | 0.06169 || 29 | 0.05963 || 30 | 0.05772 || 31 | 0.05592 || 32 | 0.05424 || 33 | 0.05265 || 34 | 0.05114 || 35 | 0.04973 || 36 | 0.04839 || 37 | 0.04711 || 38 | 0.04591 || 39 | 0.04476 || 40 | 0.04367 || 41 | 0.04264 || 42 | 0.04165 || 43 | 0.04071 || 44 | 0.03981 || 45 | 0.03894 || 46 | 0.03812 || 47 | 0.03733 || 48 | 0.03657 || 49 | 0.03584 |#### MambaVision-L2-1K- 输入尺寸：`224x224`- Batch Size：`5`| Step | Loss || ---: | ---: || 0 | 1.90336 || 1 | 1.30393 || 2 | 0.93078 || 3 | 0.71184 || 4 | 0.57377 || 5 | 0.47967 || 6 | 0.41176 || 7 | 0.36048 || 8 | 0.32075 || 9 | 0.28874 || 10 | 0.26258 || 11 | 0.24082 || 12 | 0.22234 || 13 | 0.20655 || 14 | 0.19279 || 15 | 0.18076 || 16 | 0.17019 || 17 | 0.16074 || 18 | 0.15230 || 19 | 0.14469 || 20 | 0.13781 || 21 | 0.13154 || 22 | 0.12583 || 23 | 0.12059 || 24 | 0.11577 || 25 | 0.11132 || 26 | 0.10720 || 27 | 0.10337 || 28 | 0.09981 || 29 | 0.09649 || 30 | 0.09338 || 31 | 0.09046 || 32 | 0.08772 || 33 | 0.08514 || 34 | 0.08271 || 35 | 0.08041 || 36 | 0.07824 || 37 | 0.07618 || 38 | 0.07423 || 39 | 0.07237 || 40 | 0.07061 || 41 | 0.06893 || 42 | 0.06733 || 43 | 0.06580 || 44 | 0.06434 || 45 | 0.06294 || 46 | 0.06160 || 47 | 0.06032 || 48 | 0.05909 || 49 | 0.05791 |#### MambaVision-L2-512-21K- 输入尺寸：`512x512`- Batch Size：`1`| Step | Loss || ---: | ---: || 0 | 6.95820 || 1 | 6.95413 || 2 | 6.95006 || 3 | 6.94599 || 4 | 6.94192 || 5 | 6.93785 || 6 | 6.93378 || 7 | 6.92971 || 8 | 6.92564 || 9 | 6.92157 || 10 | 6.91750 || 11 | 6.91342 || 12 | 6.90935 || 13 | 6.90528 || 14 | 6.90120 || 15 | 6.89713 || 16 | 6.89305 || 17 | 6.88897 || 18 | 6.88488 || 19 | 6.88080 || 20 | 6.87671 || 21 | 6.87262 || 22 | 6.86852 || 23 | 6.86442 || 24 | 6.86032 || 25 | 6.85621 || 26 | 6.85210 || 27 | 6.84798 || 28 | 6.84386 || 29 | 6.83973 || 30 | 6.83560 || 31 | 6.83146 || 32 | 6.82731 || 33 | 6.82316 || 34 | 6.81900 || 35 | 6.81483 || 36 | 6.81066 || 37 | 6.80647 || 38 | 6.80228 || 39 | 6.79808 || 40 | 6.79387 || 41 | 6.78965 || 42 | 6.78542 || 43 | 6.78118 || 44 | 6.77693 || 45 | 6.77267 || 46 | 6.76840 || 47 | 6.76412 || 48 | 6.75983 || 49 | 6.75552 |#### MambaVision-L3-256-21K- 输入尺寸：`256x256`- Batch Size：`1`| Step | Loss || ---: | ---: || 0 | 6.88040 || 1 | 6.87511 || 2 | 6.86983 || 3 | 6.86454 || 4 | 6.85925 || 5 | 6.85397 || 6 | 6.84868 || 7 | 6.84339 || 8 | 6.83809 || 9 | 6.83280 || 10 | 6.82750 || 11 | 6.82220 || 12 | 6.81690 || 13 | 6.81160 || 14 | 6.80629 || 15 | 6.80097 || 16 | 6.79566 || 17 | 6.79033 || 18 | 6.78501 || 19 | 6.77967 || 20 | 6.77434 || 21 | 6.76899 || 22 | 6.76364 || 23 | 6.75828 || 24 | 6.75292 || 25 | 6.74754 || 26 | 6.74216 || 27 | 6.73677 || 28 | 6.73137 || 29 | 6.72596 || 30 | 6.72054 || 31 | 6.71511 || 32 | 6.70967 || 33 | 6.70422 || 34 | 6.69876 || 35 | 6.69328 || 36 | 6.68779 || 37 | 6.68229 || 38 | 6.67678 || 39 | 6.67125 || 40 | 6.66570 || 41 | 6.66015 || 42 | 6.65457 || 43 | 6.64898 || 44 | 6.64338 || 45 | 6.63776 || 46 | 6.63212 || 47 | 6.62646 || 48 | 6.62079 || 49 | 6.61510 |#### MambaVision-L3-512-21K- 输入尺寸：`512x512`- Batch Size：`3`| Step | Loss || ---: | ---: || 0 | 1.56504 || 1 | 1.05278 || 2 | 0.74054 || 3 | 0.56457 || 4 | 0.46302 || 5 | 0.38692 || 6 | 0.33301 || 7 | 0.29163 || 8 | 0.25929 || 9 | 0.23339 || 10 | 0.21225 || 11 | 0.19458 || 12 | 0.17963 || 13 | 0.16680 || 14 | 0.15571 || 15 | 0.14598 || 16 | 0.13739 || 17 | 0.12976 || 18 | 0.12294 || 19 | 0.11679 || 20 | 0.11124 || 21 | 0.10619 || 22 | 0.10157 || 23 | 0.09734 || 24 | 0.09345 || 25 | 0.08986 || 26 | 0.08654 || 27 | 0.08345 || 28 | 0.08058 || 29 | 0.07789 || 30 | 0.07538 || 31 | 0.07302 || 32 | 0.07081 || 33 | 0.06873 || 34 | 0.06677 || 35 | 0.06491 || 36 | 0.06316 || 37 | 0.06150 || 38 | 0.05992 || 39 | 0.05842 || 40 | 0.05700 || 41 | 0.05564 || 42 | 0.05435 || 43 | 0.05312 || 44 | 0.05194 || 45 | 0.05081 || 46 | 0.04973 || 47 | 0.04869 || 48 | 0.04770 || 49 | 0.04674 |