| Name | rtx-deep JSON |
| Version |
1.4.6
JSON |
| download |
| home_page | |
| Summary | Deep AI modules developed by MOGO RTX team |
| upload_time | 2023-08-31 10:45:53 |
| maintainer | |
| docs_url | None |
| author | Andy |
| requires_python | >=3.10, <3.11 |
| license | GPLv3 |
| keywords |
|
| VCS |
|
| bugtrack_url |
|
| requirements |
No requirements were recorded.
|
| Travis-CI |
No Travis.
|
| coveralls test coverage |
No coveralls.
|
## `rtx_deep`: Deep AI modules developed by MOGO RTX team, aims to accelerate the distributed training, int8-aware distributed training, distributed evaluation and inference, model tracing and optimization, and TensorRT deployment.
#### 1 Dependency
```bash
torch>=1.10.0
tensorrt>=7.0
graphviz
```
#### 2 Installation
```bash
pip3 install graphviz
apt-get install graphviz
python3 setup.py install
```
#### 3 Examples
##### 3.1 Graph Tracing and Model Optimization
```bash
import torch
import torch.nn as nn
import torch.nn.functional as F
import rtx_deep
class conv3x3_bn_relu(nn.Module):
def __init__(self, in_planes, out_planes, stride=1, dilation=1, groups=1):
super(conv3x3_bn_relu, self).__init__()
self.net = nn.Sequential(
nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=dilation, dilation=dilation, groups=groups, bias=False),
nn.BatchNorm2d(out_planes),
nn.ReLU(inplace=True)
)
def forward(self, x):
x1 = self.net(x)
return x1
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
self.net = nn.Sequential(
conv3x3_bn_relu(64, 64),
conv3x3_bn_relu(64, 64)
)
def forward(self, x):
x1 = self.net(x)
return x1
model = Model()
model.eval()
model.cuda()
input_data = torch.randn(1, 64, 1024, 1024).cuda()
# graph tracing
model_fx = rtx_deep.graph_tracer.ad_trace.graph_trace(model, function_name=None)
# Model Optimization
# conduct graph tracing in graph_optim_from_module automatically
model_fx_optim = rtx_deep.graph_tracer.graph_utils.graph_optim_from_module(model, function_name=None, sample_inputs=(input_data,))
```
##### 3.2 Quantization-Aware Training
```bash
import torch
import torch.nn as nn
import torch.nn.functional as F
import rtx_deep
class conv3x3_bn_relu(nn.Module):
def __init__(self, in_planes, out_planes, stride=1, dilation=1, groups=1):
super(conv3x3_bn_relu, self).__init__()
self.net = nn.Sequential(
nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=dilation, dilation=dilation, groups=groups, bias=False),
nn.BatchNorm2d(out_planes),
nn.ReLU(inplace=True)
)
def forward(self, x):
x1 = self.net(x)
return x1
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
self.net = nn.Sequential(
conv3x3_bn_relu(64, 64),
conv3x3_bn_relu(64, 64)
)
def forward(self, x):
x1 = self.net(x)
return x1
model = Model()
model.eval()
model.cuda()
input_data = torch.randn(1, 64, 1024, 1024).cuda()
# Model Optimization
# conduct graph tracing in graph_optim_from_module automatically
model_fx_optim = rtx_deep.graph_tracer.graph_utils.graph_optim_from_module(model, function_name=None)
# qat
model_qat = rtx_deep.quant_lib.quant_utils.prepare_qat(model_fx_optim,
sample_inputs=[input_data],
observe_config_dic=dict(averaging_constant=0.05),
quant_config_dic=dict(quant_min=-127, quant_max=127, is_symmetric=True, is_quant=True),
disable_prefix=[])
# vis model network
rtx_deep.graph_tracer.vis_model.vis(model_fx_optim, './model_fx_optim.png')
rtx_deep.graph_tracer.vis_model.vis(model_qat, './model_qat.png')
# qat training
...
```
##### 3.3 TensorRT Deployment
```bash
import torch
import torch.nn as nn
import torch.nn.functional as F
import rtx_deep
import rtx_deep_plugin
from rtx_deep.deploy_lib.convert_trt import InputTensor, torch2trt
class conv3x3_bn_relu(nn.Module):
def __init__(self, in_planes, out_planes, stride=1, dilation=1, groups=1):
super(conv3x3_bn_relu, self).__init__()
self.net = nn.Sequential(
nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=dilation, dilation=dilation, groups=groups, bias=False),
nn.BatchNorm2d(out_planes),
nn.ReLU(inplace=True)
)
def forward(self, x):
x1 = self.net(x)
return x1
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
self.net = nn.Sequential(
conv3x3_bn_relu(64, 64),
conv3x3_bn_relu(64, 64)
)
def forward(self, x):
x1 = self.net(x)
x2 = rtx_deep_plugin.max_op(x1, dim=1)
return x2
model = Model()
model.eval()
model.cuda()
input_data = torch.randn(1, 64, 1024, 1024).cuda()
# Model Optimization
# conduct graph tracing in graph_optim_from_module automatically
model_fx_optim = rtx_deep.graph_tracer.graph_utils.graph_optim_from_module(model, function_name=None, sample_inputs=(input_data,))
# TensorRT Deployment
model_trt = torch2trt(
model=model_fx_optim,
input_specs=[InputTensor(input_data, 'input_data')],
output_names=['max_value', 'max_index'],
fp16_mode=True,
#dla_core=0,
strict_type_constraints=True,
explicit_precision=True
)
# vis tensorrt network
rtx_deep.deploy_lib.tools.vis_trt.vis(model_trt.network, 'test.png')
error = model(input_data)[0] - model_trt(input_data)[0]
print(error.abs().max())
```
Raw data
{
"_id": null,
"home_page": "",
"name": "rtx-deep",
"maintainer": "",
"docs_url": null,
"requires_python": ">=3.10, <3.11",
"maintainer_email": "",
"keywords": "",
"author": "Andy",
"author_email": "",
"download_url": "",
"platform": null,
"description": "## `rtx_deep`: Deep AI modules developed by MOGO RTX team, aims to accelerate the distributed training, int8-aware distributed training, distributed evaluation and inference, model tracing and optimization, and TensorRT deployment.\n\n#### 1 Dependency\n```bash\ntorch>=1.10.0\ntensorrt>=7.0\ngraphviz\n```\n\n#### 2 Installation\n```bash\npip3 install graphviz\napt-get install graphviz\npython3 setup.py install\n```\n\n#### 3 Examples\n##### 3.1 Graph Tracing and Model Optimization\n```bash\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nimport rtx_deep\n\n\nclass conv3x3_bn_relu(nn.Module):\n def __init__(self, in_planes, out_planes, stride=1, dilation=1, groups=1):\n super(conv3x3_bn_relu, self).__init__()\n self.net = nn.Sequential(\n nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=dilation, dilation=dilation, groups=groups, bias=False),\n nn.BatchNorm2d(out_planes),\n nn.ReLU(inplace=True)\n )\n \n def forward(self, x):\n x1 = self.net(x)\n return x1\n\n\nclass Model(nn.Module):\n def __init__(self):\n super(Model, self).__init__()\n self.net = nn.Sequential(\n conv3x3_bn_relu(64, 64),\n conv3x3_bn_relu(64, 64)\n )\n \n def forward(self, x):\n x1 = self.net(x)\n return x1\n\nmodel = Model()\nmodel.eval()\nmodel.cuda()\n\ninput_data = torch.randn(1, 64, 1024, 1024).cuda()\n\n# graph tracing\nmodel_fx = rtx_deep.graph_tracer.ad_trace.graph_trace(model, function_name=None)\n\n# Model Optimization\n# conduct graph tracing in graph_optim_from_module automatically\nmodel_fx_optim = rtx_deep.graph_tracer.graph_utils.graph_optim_from_module(model, function_name=None, sample_inputs=(input_data,))\n```\n\n##### 3.2 Quantization-Aware Training\n```bash\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nimport rtx_deep\n\n\nclass conv3x3_bn_relu(nn.Module):\n def __init__(self, in_planes, out_planes, stride=1, dilation=1, groups=1):\n super(conv3x3_bn_relu, self).__init__()\n self.net = nn.Sequential(\n nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=dilation, dilation=dilation, groups=groups, bias=False),\n nn.BatchNorm2d(out_planes),\n nn.ReLU(inplace=True)\n )\n \n def forward(self, x):\n x1 = self.net(x)\n return x1\n\n\nclass Model(nn.Module):\n def __init__(self):\n super(Model, self).__init__()\n self.net = nn.Sequential(\n conv3x3_bn_relu(64, 64),\n conv3x3_bn_relu(64, 64)\n )\n \n def forward(self, x):\n x1 = self.net(x)\n return x1\n\nmodel = Model()\nmodel.eval()\nmodel.cuda()\n\ninput_data = torch.randn(1, 64, 1024, 1024).cuda()\n\n# Model Optimization\n# conduct graph tracing in graph_optim_from_module automatically\nmodel_fx_optim = rtx_deep.graph_tracer.graph_utils.graph_optim_from_module(model, function_name=None)\n\n# qat\nmodel_qat = rtx_deep.quant_lib.quant_utils.prepare_qat(model_fx_optim,\n sample_inputs=[input_data],\n observe_config_dic=dict(averaging_constant=0.05),\n quant_config_dic=dict(quant_min=-127, quant_max=127, is_symmetric=True, is_quant=True),\n disable_prefix=[])\n\n\n# vis model network\nrtx_deep.graph_tracer.vis_model.vis(model_fx_optim, './model_fx_optim.png')\nrtx_deep.graph_tracer.vis_model.vis(model_qat, './model_qat.png')\n\n# qat training\n...\n```\n\n##### 3.3 TensorRT Deployment\n```bash\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nimport rtx_deep\nimport rtx_deep_plugin\nfrom rtx_deep.deploy_lib.convert_trt import InputTensor, torch2trt\n\nclass conv3x3_bn_relu(nn.Module):\n def __init__(self, in_planes, out_planes, stride=1, dilation=1, groups=1):\n super(conv3x3_bn_relu, self).__init__()\n self.net = nn.Sequential(\n nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=dilation, dilation=dilation, groups=groups, bias=False),\n nn.BatchNorm2d(out_planes),\n nn.ReLU(inplace=True)\n )\n \n def forward(self, x):\n x1 = self.net(x)\n return x1\n\n\nclass Model(nn.Module):\n def __init__(self):\n super(Model, self).__init__()\n self.net = nn.Sequential(\n conv3x3_bn_relu(64, 64),\n conv3x3_bn_relu(64, 64)\n )\n \n def forward(self, x):\n x1 = self.net(x)\n x2 = rtx_deep_plugin.max_op(x1, dim=1)\n return x2\n\nmodel = Model()\nmodel.eval()\nmodel.cuda()\n\ninput_data = torch.randn(1, 64, 1024, 1024).cuda()\n\n# Model Optimization\n# conduct graph tracing in graph_optim_from_module automatically\nmodel_fx_optim = rtx_deep.graph_tracer.graph_utils.graph_optim_from_module(model, function_name=None, sample_inputs=(input_data,))\n\n# TensorRT Deployment\nmodel_trt = torch2trt(\n model=model_fx_optim,\n input_specs=[InputTensor(input_data, 'input_data')],\n output_names=['max_value', 'max_index'],\n fp16_mode=True,\n #dla_core=0,\n strict_type_constraints=True,\n explicit_precision=True\n)\n\n# vis tensorrt network\nrtx_deep.deploy_lib.tools.vis_trt.vis(model_trt.network, 'test.png')\n\nerror = model(input_data)[0] - model_trt(input_data)[0]\nprint(error.abs().max())\n```\n\n",
"bugtrack_url": null,
"license": "GPLv3",
"summary": "Deep AI modules developed by MOGO RTX team",
"version": "1.4.6",
"project_urls": null,
"split_keywords": [],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "7e536cce5e9d6911f1a070bd816cd92ae15cd148f30dc85e2a7d1267c3193687",
"md5": "c9d8ee1886d2cda3507c91f686f81b60",
"sha256": "38d1eb2f3a5ac30580d4b5fa01633b8140693dd71a10301254b9e794e7cf6483"
},
"downloads": -1,
"filename": "rtx_deep-1.4.6-py310-none-any.whl",
"has_sig": false,
"md5_digest": "c9d8ee1886d2cda3507c91f686f81b60",
"packagetype": "bdist_wheel",
"python_version": "py310",
"requires_python": ">=3.10, <3.11",
"size": 81773,
"upload_time": "2023-08-31T10:45:53",
"upload_time_iso_8601": "2023-08-31T10:45:53.055253Z",
"url": "https://files.pythonhosted.org/packages/7e/53/6cce5e9d6911f1a070bd816cd92ae15cd148f30dc85e2a7d1267c3193687/rtx_deep-1.4.6-py310-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "644b70d339e9ae1af7fafadba3fa4aaa55269229fde3731edefc546e68423a16",
"md5": "4c6c100ba11276bd03fd1c83a7ada0b4",
"sha256": "7fe0f511679aebf8377694d0ca4898a05a508e632b5a2723b47c7a4439642485"
},
"downloads": -1,
"filename": "rtx_deep-1.4.6-py311-none-any.whl",
"has_sig": false,
"md5_digest": "4c6c100ba11276bd03fd1c83a7ada0b4",
"packagetype": "bdist_wheel",
"python_version": "py311",
"requires_python": ">=3.10, <3.11",
"size": 126270,
"upload_time": "2023-08-31T10:45:55",
"upload_time_iso_8601": "2023-08-31T10:45:55.486051Z",
"url": "https://files.pythonhosted.org/packages/64/4b/70d339e9ae1af7fafadba3fa4aaa55269229fde3731edefc546e68423a16/rtx_deep-1.4.6-py311-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "30df2896f82d5bcfc3707ba46eccace37ad26774946c24aeb6ff873d9d0fbe23",
"md5": "e2e112816f82c78d05ef540e20c3f1e4",
"sha256": "2173a5ab5d38c6f2fe9fbff51dc73f63769d47a4ec0967bcbd2acfd1c32eccea"
},
"downloads": -1,
"filename": "rtx_deep-1.4.6-py36-none-any.whl",
"has_sig": false,
"md5_digest": "e2e112816f82c78d05ef540e20c3f1e4",
"packagetype": "bdist_wheel",
"python_version": "py36",
"requires_python": ">=3.10, <3.11",
"size": 81054,
"upload_time": "2023-08-31T10:45:57",
"upload_time_iso_8601": "2023-08-31T10:45:57.537058Z",
"url": "https://files.pythonhosted.org/packages/30/df/2896f82d5bcfc3707ba46eccace37ad26774946c24aeb6ff873d9d0fbe23/rtx_deep-1.4.6-py36-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "ee4eb72993d646249864d002801e7639a78c4b118991b91ecef061a6c72acfba",
"md5": "9433cb892e98c415e08dc860eb0505ef",
"sha256": "0262ebe97c03010d62b0c84dced187324f029f203917124b17d9b4e271c137ef"
},
"downloads": -1,
"filename": "rtx_deep-1.4.6-py37-none-any.whl",
"has_sig": false,
"md5_digest": "9433cb892e98c415e08dc860eb0505ef",
"packagetype": "bdist_wheel",
"python_version": "py37",
"requires_python": ">=3.10, <3.11",
"size": 81474,
"upload_time": "2023-08-31T10:45:59",
"upload_time_iso_8601": "2023-08-31T10:45:59.278012Z",
"url": "https://files.pythonhosted.org/packages/ee/4e/b72993d646249864d002801e7639a78c4b118991b91ecef061a6c72acfba/rtx_deep-1.4.6-py37-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "bfacdaf978843b3e42c1678ec726dc284685bdf15ab042b76e93f12b447549f0",
"md5": "b9d17ddefa3813d742576be20526380b",
"sha256": "eca4919f63083b74d34fd05196d1df9d9bf737c9e24911a030e976e06f029fa8"
},
"downloads": -1,
"filename": "rtx_deep-1.4.6-py38-none-any.whl",
"has_sig": false,
"md5_digest": "b9d17ddefa3813d742576be20526380b",
"packagetype": "bdist_wheel",
"python_version": "py38",
"requires_python": ">=3.10, <3.11",
"size": 81422,
"upload_time": "2023-08-31T10:46:01",
"upload_time_iso_8601": "2023-08-31T10:46:01.763543Z",
"url": "https://files.pythonhosted.org/packages/bf/ac/daf978843b3e42c1678ec726dc284685bdf15ab042b76e93f12b447549f0/rtx_deep-1.4.6-py38-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "10a911d8e7531263a84739a39361eed75adfc1e7b3e3d5d10018ec20420e11e0",
"md5": "eb83bc07d305029949d9b273b7db75e3",
"sha256": "84eb6006195f124b658b8a7c45c3d9ffac7c40616d65202518fe1a9f56268856"
},
"downloads": -1,
"filename": "rtx_deep-1.4.6-py39-none-any.whl",
"has_sig": false,
"md5_digest": "eb83bc07d305029949d9b273b7db75e3",
"packagetype": "bdist_wheel",
"python_version": "py39",
"requires_python": ">=3.10, <3.11",
"size": 81175,
"upload_time": "2023-08-31T10:46:03",
"upload_time_iso_8601": "2023-08-31T10:46:03.757742Z",
"url": "https://files.pythonhosted.org/packages/10/a9/11d8e7531263a84739a39361eed75adfc1e7b3e3d5d10018ec20420e11e0/rtx_deep-1.4.6-py39-none-any.whl",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2023-08-31 10:45:53",
"github": false,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"lcname": "rtx-deep"
}
