Skip to content

Commit

Permalink
更新
Browse files Browse the repository at this point in the history
  • Loading branch information
liuxinwei committed Apr 23, 2024
1 parent 9c0124b commit a5029c0
Show file tree
Hide file tree
Showing 24 changed files with 1,038 additions and 12 deletions.
1 change: 1 addition & 0 deletions doc/tutorial/index.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,7 @@ TorchScript/index
quant/index
fx/index
intermediate/index
onnxscript/index
onnx-op/index
compile/index
ultralytics/index
Expand Down
32 changes: 20 additions & 12 deletions doc/tutorial/onnx-op/export.ipynb
View file
Original file line number Diff line number Diff line change
Expand Up @@ -139,22 +139,30 @@
"cell_type": "code",
"execution_count": 4,
"metadata": {},
"outputs": [],
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"/media/pc/data/tmp/cache/conda/envs/xin/lib/python3.12/site-packages/torch/onnx/utils.py:1548: OnnxExporterWarning: Exporting to ONNX opset version 18 is not supported. by 'torch.onnx.export()'. The highest opset version supported is 17. To use a newer opset version, consider 'torch.onnx.dynamo_export()'. Note that dynamo_export() is in preview. Please report errors with dynamo_export() as Github issues to https://github.com/pytorch/pytorch/issues.\n",
" warnings.warn(\n"
]
}
],
"source": [
"# Input to the model\n",
"x = torch.randn(batch_size, 1, 224, 224, requires_grad=True)\n",
"torch_out = torch_model(x)\n",
"\n",
"# Export the model\n",
"torch.onnx.export(torch_model, # model being run\n",
" x, # model input (or a tuple for multiple inputs)\n",
" \"super_resolution.onnx\", # where to save the model (can be a file or file-like object)\n",
" export_params=True, # store the trained parameter weights inside the model file\n",
" opset_version=10, # the ONNX version to export the model to\n",
" do_constant_folding=True, # whether to execute constant folding for optimization\n",
" input_names = ['input'], # the model's input names\n",
" output_names = ['output'], # the model's output names\n",
" dynamic_axes={'input' : {0 : 'batch_size'}, # variable length axes\n",
"# 导出模型\n",
"torch.onnx.export(torch_model, # torch 模型\n",
" x, # 模型输入或者对于多个输入，使用元组\n",
" \"super_resolution.onnx\", # 模型保存的位置（可以是文件或类似文件的对象）\n",
" export_params=True, # 将训练后的参数权重存储在模型文件内\n",
" opset_version=18, # 导出模型的 ONNX 版本\n",
" do_constant_folding=True, # 是否执行常量折叠以进行优化\n",
" input_names = ['data'], # 模型的输入名称\n",
" output_names = ['output'], # 模型的输出名称\n",
" dynamic_axes={'data' : {0 : 'batch_size'}, # 可变长度的轴\n",
" 'output' : {0 : 'batch_size'}})"
]
},
Expand Down
20 changes: 20 additions & 0 deletions doc/tutorial/onnx-op/faqs/GatherND.ipynb
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,20 @@
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# ONNX GatherND\n",
"\n",
"参考：[GatherND](https://onnx.ai/onnx/operators/onnx__GatherND.html#gathernd)"
]
}
],
"metadata": {
"language_info": {
"name": "python"
}
},
"nbformat": 4,
"nbformat_minor": 2
}
76 changes: 76 additions & 0 deletions doc/tutorial/onnx-op/faqs/L2.ipynb
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,76 @@
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# onnx L2 范数"
]
},
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [],
"source": [
"import torch\n",
"from torch import nn\n",
"\n",
"class Model(nn.Module):\n",
" def forward(self, x):\n",
" x = torch.norm(x, p=2)\n",
" return x\n",
"\n",
"x = torch.rand(1, 3, 32, 32)\n",
"\n",
"torch_model = Model()\n",
"# 导出模型\n",
"torch.onnx.export(\n",
" torch_model, # torch 模型\n",
" x, # 模型输入或者对于多个输入，使用元组\n",
" \"L2.onnx\", # 模型保存的位置（可以是文件或类似文件的对象）\n",
" export_params=True, # 将训练后的参数权重存储在模型文件内\n",
" opset_version=17, # 导出模型的 ONNX 版本\n",
" do_constant_folding=True, # 是否执行常量折叠以进行优化\n",
" input_names = ['data'], # 模型的输入名称\n",
" output_names = ['output'], # 模型的输出名称\n",
" dynamic_axes={'data' : {0 : 'batch_size'}, # 可变长度的轴\n",
" 'output' : {0 : 'batch_size'}}\n",
")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"![](images/L2.jpg)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "xin",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.12.2"
}
},
"nbformat": 4,
"nbformat_minor": 2
}
149 changes: 149 additions & 0 deletions doc/tutorial/onnx-op/faqs/ScatterND.ipynb
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,149 @@
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# onnx ScatterND\n",
"\n",
"参考：[ScatterND](https://onnx.ai/onnx/operators/onnx__ScatterND.html)\n",
"\n",
"ScatterND 算子涉及三个输入：秩为 `r`（`r >= 1`）的 `data` 张量，秩为 `q`（`q >= 1`）的 `indices` 张量，以及秩为 `q + r - indices.shape[-1] - 1` 的 `updates` 张量。该算子的输出是通过创建 `data` 输入的副本，然后根据 ``indices`` 指定的特定位置，将其值更新为 `updates` 张量中指定的值来产生的。它的输出形状与 `data` 的形状相同。\n",
"\n",
"`indices` 是整数张量。设 `k` 为索引形状的最后一维，即 `indices.shape[-1]`。`indices` 被视为由 `k` 元组组成的 `(q-1)` 维张量，其中每个 `k` 元组都是对 `data` 的偏索引（partial-index）。因此， `k` 的值最多可以等于数据的秩。当 `k` 等于 `data` 的秩时，每个更新项指定了对张量单个元素的更新。当 `k` 小于 `data` 的秩时，每个更新项指定了对张量切片的更新。索引值可以是负数，按照从末尾开始倒数的通常惯例，但需要在有效范围内。\n",
"\n",
"`updates` 被视为替换切片值的(q-1)维张量。因此，`updates` 形状的前(q-1)个维度必须与索引形状的前(q-1)个维度匹配。`updates` 的其余维度对应于替换切片值的维度。每个替换切片值是 (r-k) 维张量，对应于 `data` 的尾部 (r-k) 个维度。因此，`updates` 的形状必须等于 `indices.shape[0:q-1] ++ data.shape[k:r-1]`，其中 `++` 表示形状的连接。\n",
"\n",
"输出通过以下方程计算：\n",
"\n",
"```\n",
"output = np.copy(data)\n",
"update_indices = indices.shape[:-1]\n",
"for idx in np.ndindex(update_indices):\n",
" output[indices[idx]] = updates[idx]\n",
"```"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"上述循环中的迭代顺序未指定。特别是，索引中不应有重复项：也就是说，如果 `idx1 != idx2`，那么 `indices[idx1] != indices[idx2]`。这确保了输出值不依赖于迭代顺序。\n",
"\n",
"缩减（`reduction`）允许指定可选的缩减操作，该操作将所有 `updates` 张量中的值应用于指定 `indices` 的 `output`。在将 `reduction` 设置为“none”的情况下，索引中不应有重复项：也就是说，如果 `idx1 != idx2`，那么 `indices[idx1] != indices[idx2]`。这确保了输出值不依赖于迭代顺序。\n",
"\n",
"当 `reduction` 设置为某个缩减函数 `f` 时，`output` 按以下方式计算：\n",
"\n",
"```\n",
"output = np.copy(data)\n",
"update_indices = indices.shape[:-1]\n",
"for idx in np.ndindex(update_indices):\n",
" output[indices[idx]] = f(output[indices[idx]], updates[idx])\n",
"```\n",
"\n",
"其中 `f` 是指定的加法（`+`），乘法（`*`），最大值（`max`）或最小值（`min`）。\n",
"\n",
"这个算子是 `GatherND` 的逆运算。"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"示例1：\n",
"```\n",
"data = [1, 2, 3, 4, 5, 6, 7, 8]\n",
"indices = [[4], [3], [1], [7]]\n",
"updates = [9, 10, 11, 12]\n",
"output = [1, 11, 3, 10, 9, 6, 7, 12]\n",
"```"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"示例2：\n",
"\n",
"```\n",
"data = [[[1, 2, 3, 4], [5, 6, 7, 8], [8, 7, 6, 5], [4, 3, 2, 1]],\n",
" [[1, 2, 3, 4], [5, 6, 7, 8], [8, 7, 6, 5], [4, 3, 2, 1]],\n",
" [[8, 7, 6, 5], [4, 3, 2, 1], [1, 2, 3, 4], [5, 6, 7, 8]],\n",
" [[8, 7, 6, 5], [4, 3, 2, 1], [1, 2, 3, 4], [5, 6, 7, 8]]]\n",
"indices = [[0], [2]]\n",
"updates = [[[5, 5, 5, 5], [6, 6, 6, 6], [7, 7, 7, 7], [8, 8, 8, 8]],\n",
" [[1, 1, 1, 1], [2, 2, 2, 2], [3, 3, 3, 3], [4, 4, 4, 4]]]\n",
"output = [[[5, 5, 5, 5], [6, 6, 6, 6], [7, 7, 7, 7], [8, 8, 8, 8]],\n",
" [[1, 2, 3, 4], [5, 6, 7, 8], [8, 7, 6, 5], [4, 3, 2, 1]],\n",
" [[1, 1, 1, 1], [2, 2, 2, 2], [3, 3, 3, 3], [4, 4, 4, 4]],\n",
" [[8, 7, 6, 5], [4, 3, 2, 1], [1, 2, 3, 4], [5, 6, 7, 8]]]\n",
"```"
]
},
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [],
"source": [
"import torch\n",
"from torch import nn\n",
"\n",
"class Model(nn.Module):\n",
" def forward(self, x):\n",
" x[0] = x[0] + 1\n",
" return x\n",
"\n",
"x = torch.rand(192, 36)\n",
"\n",
"torch_model = Model()\n",
"# 导出模型\n",
"torch.onnx.export(\n",
" torch_model, # torch 模型\n",
" x, # 模型输入或者对于多个输入，使用元组\n",
" \"ScatterND.onnx\", # 模型保存的位置（可以是文件或类似文件的对象）\n",
" export_params=True, # 将训练后的参数权重存储在模型文件内\n",
" opset_version=17, # 导出模型的 ONNX 版本\n",
" do_constant_folding=True, # 是否执行常量折叠以进行优化\n",
" input_names = ['data'], # 模型的输入名称\n",
" output_names = ['output'], # 模型的输出名称\n",
" # dynamic_axes={'data' : {0 : 'batch_size'}, # 可变长度的轴\n",
" # 'output' : {0 : 'batch_size'}}\n",
")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"![模型结构](images/ScatterND.jpg)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "xin",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.12.2"
}
},
"nbformat": 4,
"nbformat_minor": 2
}
77 changes: 77 additions & 0 deletions doc/tutorial/onnx-op/faqs/bn.ipynb
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,77 @@
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# BN"
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"outputs": [],
"source": [
"import torch\n",
"from torch import nn\n",
"\n",
"class Model(nn.Module):\n",
" def __init__(self):\n",
" super().__init__()\n",
" self.conv1 = nn.Conv2d(3, 16, 3, 1, 1, bias=True, groups=1)\n",
" self.bn = nn.BatchNorm2d(16)\n",
"\n",
" def forward(self, x):\n",
" x = self.conv1(x)\n",
" x = self.bn(x)\n",
" return x\n",
"\n",
"x = torch.rand(1, 3, 32, 32)\n",
"\n",
"torch_model = Model()\n",
"# 导出模型\n",
"torch.onnx.export(\n",
" torch_model, # torch 模型\n",
" x, # 模型输入或者对于多个输入，使用元组\n",
" \"bn.onnx\", # 模型保存的位置（可以是文件或类似文件的对象）\n",
" export_params=True, # 将训练后的参数权重存储在模型文件内\n",
" opset_version=17, # 导出模型的 ONNX 版本\n",
" do_constant_folding=False, # 是否执行常量折叠以进行优化\n",
" input_names = ['data'], # 模型的输入名称\n",
" output_names = ['output'], # 模型的输出名称\n",
" dynamic_axes={'data' : {0 : 'batch_size'}, # 可变长度的轴\n",
" 'output' : {0 : 'batch_size'}}\n",
")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "xin",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.12.2"
}
},
"nbformat": 4,
"nbformat_minor": 2
}
Loading

0 comments on commit a5029c0

Please sign in to comment.