Neural network graphs and training metrics for PyTorch, Tensorflow, and Keras.

Last update: Dec 31, 2022

Overview

HiddenLayer

A lightweight library for neural network graphs and training metrics for PyTorch, Tensorflow, and Keras.

HiddenLayer is simple, easy to extend, and works great with Jupyter Notebook. It's not intended to replace advanced tools, such as TensorBoard, but rather for cases where advanced tools are too big for the task. HiddenLayer was written by Waleed Abdulla and Phil Ferriere, and is licensed under the MIT License.

1. Readable Graphs

Use HiddenLayer to render a graph of your neural network in Jupyter Notebook, or to a pdf or png file. See Jupyter notebook examples for TensorFlow, PyTorch, and Keras.

The graphs are designed to communicate the high-level architecture. Therefore, low-level details are hidden by default (e.g. weight initialization ops, gradients, internal ops of common layer types, ...etc.). HiddenLayer also folds commonly used sequences of layers together. For example, the Convolution -> RELU -> MaxPool sequence is very common, so they get merged into one box for simplicity.

Customizing Graphs

The rules for hiding and folding nodes are fully customizable. You can use graph expressions and transforms to add your own rules. For example, this rule folds all the nodes of a bottleneck block of a ResNet101 into one node.

    # Fold bottleneck blocks
    ht.Fold("((ConvBnRelu > ConvBnRelu > ConvBn) | ConvBn) > Add > Relu", 
            "BottleneckBlock", "Bottleneck Block"),

2. Training Metrics in Jupyter Notebook

If you run training experiments in Jupyter Notebook then you might find this useful. You can use it to plot loss and accuracy, histograms of weights, or visualize activations of a few layers.

Outside Jupyter Notebook:

You can use HiddenLayer outside Jupyter Notebook as well. In a Python script running from command line, it'll open a separate window for the metrics. And if you're on a server without a GUI, you can save snapshots of the graphs to png files for later inspection. See history_canvas.py for an example of this use case.

3. Hackable

HiddenLayer is a small library. It covers the basics, but you'll likely need to extend it for your own use case. For example, say you want to represent the model accuracy as a pie chart rather than a plot. This can be done by extending the Canvas class and adding a new method as such:

class MyCanvas(hl.Canvas):
    """Extending Canvas to add a pie chart method."""
    def draw_pie(self, metric):
        # set square aspect ratio
        self.ax.axis('equal')
        # Get latest value of the metric
        value = np.clip(metric.data[-1], 0, 1)
        # Draw pie chart
        self.ax.pie([value, 1-value], labels=["Accuracy", ""])

See the pytorch_train.ipynb or tf_train.ipynb for an example.

The keras_train.ipynb notebook contains an actual training example that illustrates how to create a custom Canvas to plot a confusion matrix alongside validation metrics:

Demos

PyTorch:

pytorch_graph.ipynb: This notebook shows how to generate graphs for a few popular Pytorch models.
pytorch_train.ipynb: Explains tracking and displaying training metrics.
history_canvas.py: An example of using HiddenLayer without a GUI.

TensorFlow:

tf_graph.ipynb: This notebook illustrates how to generate graphs for various TF SLIM models.
tf_train.ipynb: Demonstrates tracking and visualizing training metrics with TensorFlow.
history_canvas.py: An example of using HiddenLayer without a GUI.

Keras:

keras_graph.ipynb: This notebook illustrates how to generate graphs for various Keras models.
keras_train.ipynb: Demonstrates model graphing, visualization of training metrics, and how to create a custom Keras callback that uses a subclassed Canvas in order to plot a confusion matrix at the end of each training epoch.

Contributing

HiddenLayer is released under the MIT license. Feel free to extend it or customize it for your needs. If you discover bugs, which is likely since this is an early release, please do report them or submit a pull request.

If you like to contribute new features, here are a few things we wanted to add but never got around to it:

Support for older versions of Python. Currently, it's only tested on Python 3.6.
Optimization to support logging big experiments.

Installation

1. Prerequisites

a. Python3, Numpy, Matplotlib, and Jupyter Notebook.
b. Either TensorFlow or PyTorch
c. GraphViz and its Python wrapper to generate network graphs. The easiest way to install it is

If you use Conda:
```
conda install graphviz python-graphviz
```
Otherwise:
- Install GraphViz
- Then install the Python wrapper for GraphViz using pip:
```
pip3 install graphviz
```

2. Install HiddenLayer

a. Clone From GitHub (Developer Mode)

Use this if you want to edit or customize the library locally.

# Clone the repository
git clone [email protected]:waleedka/hiddenlayer.git
cd hiddenlayer

# Install in dev mode
pip install -e .

b. Using PIP ("stable" release)

pip install hiddenlayer

c. Install to your `site-packages` directly from GitHub

Use the following if you just want to install the latest version of the library:

pip install git+https://github.com/waleedka/hiddenlayer.git

Comments

get_trace_graph private in Pytorch 1.4

system: Python 3.7.5 Pytorch 1.4 torchvision 0.5

code and error: import hiddenlayer as hl

model= vgg16(pretrained=True).features hl.build_graph(model, torch.zeros([1,3,224,224]))

" 69 # Run the Pytorch graph to get a trace and generate a graph from it ---> 70 trace, out = torch.jit.get_trace_graph(model, args) 71 torch.onnx._optimize_trace(trace, torch.onnx.OperatorExportTypes.ONNX) 72 torch_graph = trace.graph()"

AttributeError: module 'torch.jit' has no attribute 'get_trace_graph'

Context: Since Pytorch version 1.4 (this pull request) the function jit.get_trace_graph is private (jit._get_trace_graph).

opened by nanohanno 13
'torch._C.Value' object has no attribute 'uniqueName'
import torch import torchvision.models import hiddenlayer as hl # VGG16 with BatchNorm model = torchvision.models.vgg16() # Build HiddenLayer graph # Jupyter Notebook renders it automatically hl.build_graph(model, torch.zeros([1, 3, 224, 224]))

python 3.6 torch 1.3.1 hl 0.2

AttributeError Traceback (most recent call last) in 7 # Build HiddenLayer graph 8 # Jupyter Notebook renders it automatically ----> 9 hl.build_graph(model, torch.zeros([1, 3, 224, 224]))

~/anaconda3/envs/py36/lib/python3.6/site-packages/hiddenlayer-0.2-py3.6.egg/hiddenlayer/graph.py in build_graph(model, args, input_names, transforms, framework_transforms) 141 from .pytorch_builder import import_graph, FRAMEWORK_TRANSFORMS 142 assert args is not None, "Argument args must be provided for Pytorch models." --> 143 import_graph(g, model, args) 144 elif framework == "tensorflow": 145 from .tf_builder import import_graph, FRAMEWORK_TRANSFORMS

~/anaconda3/envs/py36/lib/python3.6/site-packages/hiddenlayer-0.2-py3.6.egg/hiddenlayer/pytorch_builder.py in import_graph(hl_graph, model, args, input_names, verbose) 88 shape = get_shape(torch_node) 89 # Add HL node ---> 90 hl_node = Node(uid=pytorch_id(torch_node), name=None, op=op, 91 output_shape=shape, params=params) 92 hl_graph.add_node(hl_node)

~/anaconda3/envs/py36/lib/python3.6/site-packages/hiddenlayer-0.2-py3.6.egg/hiddenlayer/pytorch_builder.py in pytorch_id(node) 43 # After ONNX simplification, the scopeName is not unique anymore 44 # so append node outputs to guarantee uniqueness ---> 45 return node.scopeName() + "/outputs/" + "/".join([o.uniqueName() for o in node.outputs()]) 46 47

~/anaconda3/envs/py36/lib/python3.6/site-packages/hiddenlayer-0.2-py3.6.egg/hiddenlayer/pytorch_builder.py in (.0) 43 # After ONNX simplification, the scopeName is not unique anymore 44 # so append node outputs to guarantee uniqueness ---> 45 return node.scopeName() + "/outputs/" + "/".join([o.uniqueName() for o in node.outputs()]) 46 47

AttributeError: 'torch._C.Value' object has no attribute 'uniqueName'
opened by woodg07 5
hiddenlayer cannot identify that my module is indeed a torch.nn.Module
I have a personalised model class called let's say MyNet(Net), and it inherits from Net(nn.Module).

When I call hl.build_graph(model, ...), hiddenlayer then raises the exception:

ValueError: model input param must be a PyTorch, TensorFlow, or Keras-with-TensorFlow-backend model.

When I put everything inside only one class it works...
opened by gmunizc 5
Bugfix/get trace graph

Builds on the earlier fix by jccurtis. This pull request fixes issue #66 . It makes hiddenlayer work with Pytorch 1.4, where get_trace_graph became private and ONNX formatting appears to have changed.

opened by nanohanno 4
module 'torch.onnx' has no attribute 'OperatorExportTypes'

I run this code in Jupyter Notebook，but one error occurs: ` import torch

import torchvision.models

import hiddenlayer as hl

model = torchvision.models.vgg16()

hl.build_graph(model, torch.zeros([1, 3, 224, 224]))

`

AttributeError: module 'torch.onnx' has no attribute 'OperatorExportTypes'

And I run the code under Ubuntu16.04, pytorch 0.4.0

opened by geekac 4
module 'torch.jit' has no attribute '_get_trace_graph'

Hi，thank you for your amazing job I try to use your work to visualize my own model When I run : #my model import hiddenlayer as hl hl.build_graph(model, torch.zeros([1, 3, 224, 224])) but I got: AttributeError: module 'torch.jit' has no attribute '_get_trace_graph' Could you tell me how can I slove it? Thank you

opened by zhongqiu1245 3
"torch._C.Value has no attribute 'uniqueName'" Error running with PyTorch 1.2

PyTorch Version: '1.2.0a' Python: 3.6.8 Exception has occurred: AttributeError 'torch._C.Value' object has no attribute 'uniqueName' File "hiddenlayer/hiddenlayer/pytorch_builder.py", line 45, in <listcomp> return node.scopeName() + "/outputs/" + "/".join([o.uniqueName() for o in node.outputs()]) File "hiddenlayer/hiddenlayer/pytorch_builder.py", line 45, in pytorch_id return node.scopeName() + "/outputs/" + "/".join([o.uniqueName() for o in node.outputs()]) File "hiddenlayer/hiddenlayer/pytorch_builder.py", line 90, in import_graph hl_node = Node(uid=pytorch_id(torch_node), name=None, op=op, File "hiddenlayer/hiddenlayer/graph.py", line 143, in build_graph import_graph(g, model, args) File "visualizer.py", line 20, in <module> graph = hl.build_graph(model, input)

Works well with older version of PyTorch (0.4.1).

opened by cted18 3
adaptive_avg_pool2d does not exist

Hi, I met this problem when I want to visualize densenet, error shows below: c:\python35\lib\site-packages\torch\onnx\utils.py:446: UserWarning: ONNX export failed on ATen operator adaptive_avg_pool2d because torch.onnx.symbolic.adaptive_avg_pool2d does not exist .format(op_name, op_name)) Actually, there is no adaptive_avg_pool2d in my densenet, only nn.AvgPool2d() exists.

opened by zhouyuangan 3

support yolov5

hello, I want to draw the graph of yolov5 by folowing code:

import torch

model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True)
# print("==>> model: ", model)

import hiddenlayer as hl
graph = hl.build_graph(model, torch.zeros([1, 3, 512, 512]))

but it reports error:

c:\ProgramData\Anaconda3\lib\site-packages\hiddenlayer\pytorch_builder.py in import_graph(hl_graph, model, args, input_names, verbose)
     69     # Run the Pytorch graph to get a trace and generate a graph from it
     70     trace, out = torch.jit._get_trace_graph(model, args)
---> 71     torch_graph = torch.onnx._optimize_trace(trace, torch.onnx.OperatorExportTypes.ONNX)
...
RuntimeError: Unsupported: ONNX export of index_put in opset 9. Please try opset version 11.

some issues (https://github.com/onnx/onnx/issues/3057, https://github.com/pytorch/pytorch/issues/46237) say place opset_version=11 in torch.onnx.export(), but here i can not find torch.onnx.export(), so i don't know how to fix this error.

opened by wwdok 2

How to display branch nodes / parallel blocks

I am confused about how to transform the following structure which has branches coming out from a node. This is what I tried hl.transforms.Fold("""Conv > LeakyRelu > Conv > Concat > LeakyRelu """, "Fire","FireBlock")

opened by talhaanwarch 2
TypeError: zeros(): argument 'out' (position 2) must be Tensor, not list
I want to visualizer a model with 3-input and have problem with feed 3-input. This model have three input, this filgure:

I conferenced this ideal of Waleed (https://github.com/waleedka)and I visualier 3 patch with 3 input [ 32,6, 25, 25] [ 32,6, 51, 51],[32, 6, 75, 75] with code line:

hl.build_graph(model, torch.zeros([32,6, 25, 25], [32,6, 51, 51],[32, 6, 75, 75]))

But my code had the error: TypeError: zeros(): argument 'out' (position 2) must be Tensor, not list. How do I fix this problem? (I also tried many ways, example: #hl.build_graph(net, torch.zeros( 32,6, 25, 25), torch.zeros(32, 6, 51, 51), torch.zeros(32, 6, 75, 75)) #hl.build_graph(net, torch.zeros(( 32,6, 25, 25), ( 32,6, 51, 51), ( 32,6, 75, 75))) #hl.build_graph(net, torch.zeros([( 32,6, 25, 25)],[( 32,6, 51, 51)],[( 32,6, 75, 75)]) )

In addition, I successed with the first path with code line: hl.build_graph(model, torch.zeros([32,6, 25, 25]). And had figure: ..................... 2) And question more about window size of model visualize. Can we show full model visualize one window or save it ? as we see these below picture, I must to stride many time in oder to take a my model.

Thank you.
opened by tphankr 2

How to plot bert model? (Transfomer models)

I try to plot bert model using this package. But I unable to do it.

Code:

from transformers import AutoModel, AutoTokenizer
model = AutoModel.from_pretrained("bert-base-uncased")
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
inputs = tokenizer("Hello world!", return_tensors="pt")

After that how to plot it?

import hiddenlayer as hl
hl.build_graph(model, inputs[0])

opened by indramal 0

Error while trying to run the example

TypeError Traceback (most recent call last) in 4 # Build HiddenLayer graph 5 # Jupyter Notebook renders it automatically ----> 6 hl.build_graph(model, torch.zeros([1, 3, 224, 224]))

3 frames /usr/local/lib/python3.7/dist-packages/torch/onnx/utils.py in _optimize_graph(graph, operator_export_type, _disable_torch_constant_prop, fixed_batch_size, params_dict, dynamic_axes, input_names, module) 276 # Unpack quantized weights for conv and linear ops and insert into graph. 277 _C._jit_pass_onnx_unpack_quantized_weights( --> 278 graph, params_dict, symbolic_helper.is_caffe2_aten_fallback() 279 ) 280 if symbolic_helper.is_caffe2_aten_fallback():

TypeError: _jit_pass_onnx_unpack_quantized_weights(): incompatible function arguments. The following argument types are supported: 1. (arg0: torch::jit::Graph, arg1: Dict[str, IValue], arg2: bool) -> Dict[str, IValue]

Invoked with: graph(%input.1 : Float(1, 3, 224, 224, strides=[150528, 50176, 224, 1], requires_grad=0, device=cpu), %1 : Float(64, 3, 3, 3, strides=[27, 9, 3, 1], requires_grad=1, device=cpu), %2 : Float(64, strides=[1], requires_grad=1, device=cpu), %3 : Float(64, 64, 3, 3, strides=[576, 9, 3, 1], requires_grad=1, device=cpu), %4 : Float(64, strides=[1], requires_grad=1, device=cpu), %5 : Float(128, 64, 3, 3, strides=[576, 9, 3, 1], requires_grad=1, device=cpu), %6 : Float(128, strides=[1], requires_grad=1, device=cpu), %7 : Float(128, 128, 3, 3, strides=[1152, 9, 3, 1], requires_grad=1, device=cpu), %8 : Float(128, strides=[1], requires_grad=1, device=cpu), %9 : Float(256, 128, 3, 3, strides=[1152, 9, 3, 1], requires_grad=1, device=cpu), %10 : Float(256, strides=[1], requires_grad=1, device=cpu), %11 : Float(256, 256, 3, 3, strides=[2304, 9, 3, 1], requires_grad=1, device=cpu), %12 : Float(256, strides=[1], requires_grad=1, device=cpu), %13 : Float(256, 256, 3, 3, strides=[2304, 9, 3, 1], requires_grad=1, device=cpu), %14 : Float(256, strides=[1], requires_grad=1, device=cpu), %15 : Float(512, 256, 3, 3, strides=[2304, 9, 3, 1], requires_grad=1, device=cpu), %16 : Float(512, strides=[1], requires_grad=1, device=cpu), %17 : Float(512, 512, 3, 3, strides=[4608, 9, 3, 1], requires_grad=1, device=cpu), %18 : Float(512, strides=[1], requires_grad=1, device=cpu), %19 : Float(512, 512, 3, 3, strides=[4608, 9, 3, 1], requires_grad=1, device=cpu), %20 : Float(512, strides=[1], requires_grad=1, device=cpu), %21 : Float(512, 512, 3, 3, strides=[4608, 9, 3, 1], requires_grad=1, device=cpu), %22 : Float(512, strides=[1], requires_grad=1, device=cpu), %23 : Float(512, 512, 3, 3, strides=[4608, 9, 3, 1], requires_grad=1, device=cpu), %24 : Float(512, strides=[1], requires_grad=1, device=cpu), %25 : Float(512, 512, 3, 3, strides=[4608, 9, 3, 1], requires_grad=1, device=cpu), %26 : Float(512, strides=[1], requires_grad=1, device=cpu), %27 : Float(4096, 25088, strides=[25088, 1], requires_grad=1, device=cpu), %28 : Float(4096, strides=[1], requires_grad=1, device=cpu), %29 : Float(4096, 4096, strides=[4096, 1], requires_grad=1, device=cpu), %30 : Float(4096, strides=[1], requires_grad=1, device=cpu), %31 : Float(1000, 4096, strides=[4096, 1], requires_grad=1, device=cpu), %32 : Float(1000, strides=[1], requires_grad=1, device=cpu)): %459 : int[] = prim::Constantvalue=[1, 1] %460 : int[] = prim::Constantvalue=[1, 1] %461 : int[] = prim::Constantvalue=[1, 1] %108 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %462 : int[] = prim::Constantvalue=[0, 0] %112 : int = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %113 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %114 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %115 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %116 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %input.3 : Float(1, 64, 224, 224, strides=[3211264, 50176, 224, 1], requires_grad=0, device=cpu) = aten::_convolution(%input.1, %1, %2, %459, %460, %461, %108, %462, %112, %113, %114, %115, %116) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %532 : Float(1, 64, 224, 224, strides=[3211264, 50176, 224, 1], requires_grad=1, device=cpu) = aten::relu(%input.3) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %463 : int[] = prim::Constantvalue=[1, 1] %464 : int[] = prim::Constantvalue=[1, 1] %465 : int[] = prim::Constantvalue=[1, 1] %128 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %466 : int[] = prim::Constantvalue=[0, 0] %132 : int = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %133 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %134 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %135 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %136 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %input.7 : Float(1, 64, 224, 224, strides=[3211264, 50176, 224, 1], requires_grad=0, device=cpu) = aten::_convolution(%532, %3, %4, %463, %464, %465, %128, %466, %132, %133, %134, %135, %136) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %533 : Float(1, 64, 224, 224, strides=[3211264, 50176, 224, 1], requires_grad=1, device=cpu) = aten::relu(%input.7) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %467 : int[] = prim::Constantvalue=[2, 2] %468 : int[] = prim::Constantvalue=[2, 2] %469 : int[] = prim::Constantvalue=[0, 0] %470 : int[] = prim::Constantvalue=[1, 1] %151 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:782:0 %input.9 : Float(1, 64, 112, 112, strides=[802816, 12544, 112, 1], requires_grad=1, device=cpu) = aten::max_pool2d(%533, %467, %468, %469, %470, %151) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:782:0 %471 : int[] = prim::Constantvalue=[1, 1] %472 : int[] = prim::Constantvalue=[1, 1] %473 : int[] = prim::Constantvalue=[1, 1] %162 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %474 : int[] = prim::Constantvalue=[0, 0] %166 : int = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %167 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %168 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %169 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %170 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %input.11 : Float(1, 128, 112, 112, strides=[1605632, 12544, 112, 1], requires_grad=0, device=cpu) = aten::_convolution(%input.9, %5, %6, %471, %472, %473, %162, %474, %166, %167, %168, %169, %170) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %534 : Float(1, 128, 112, 112, strides=[1605632, 12544, 112, 1], requires_grad=1, device=cpu) = aten::relu(%input.11) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %475 : int[] = prim::Constantvalue=[1, 1] %476 : int[] = prim::Constantvalue=[1, 1] %477 : int[] = prim::Constantvalue=[1, 1] %182 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %478 : int[] = prim::Constantvalue=[0, 0] %186 : int = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %187 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %188 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %189 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %190 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %input.15 : Float(1, 128, 112, 112, strides=[1605632, 12544, 112, 1], requires_grad=0, device=cpu) = aten::_convolution(%534, %7, %8, %475, %476, %477, %182, %478, %186, %187, %188, %189, %190) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %535 : Float(1, 128, 112, 112, strides=[1605632, 12544, 112, 1], requires_grad=1, device=cpu) = aten::relu(%input.15) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %479 : int[] = prim::Constantvalue=[2, 2] %480 : int[] = prim::Constantvalue=[2, 2] %481 : int[] = prim::Constantvalue=[0, 0] %482 : int[] = prim::Constantvalue=[1, 1] %205 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:782:0 %input.17 : Float(1, 128, 56, 56, strides=[401408, 3136, 56, 1], requires_grad=1, device=cpu) = aten::max_pool2d(%535, %479, %480, %481, %482, %205) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:782:0 %483 : int[] = prim::Constantvalue=[1, 1] %484 : int[] = prim::Constantvalue=[1, 1] %485 : int[] = prim::Constantvalue=[1, 1] %216 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %486 : int[] = prim::Constantvalue=[0, 0] %220 : int = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %221 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %222 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %223 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %224 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %input.19 : Float(1, 256, 56, 56, strides=[802816, 3136, 56, 1], requires_grad=0, device=cpu) = aten::_convolution(%input.17, %9, %10, %483, %484, %485, %216, %486, %220, %221, %222, %223, %224) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %536 : Float(1, 256, 56, 56, strides=[802816, 3136, 56, 1], requires_grad=1, device=cpu) = aten::relu(%input.19) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %487 : int[] = prim::Constantvalue=[1, 1] %488 : int[] = prim::Constantvalue=[1, 1] %489 : int[] = prim::Constantvalue=[1, 1] %236 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %490 : int[] = prim::Constantvalue=[0, 0] %240 : int = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %241 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %242 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %243 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %244 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %input.23 : Float(1, 256, 56, 56, strides=[802816, 3136, 56, 1], requires_grad=0, device=cpu) = aten::_convolution(%536, %11, %12, %487, %488, %489, %236, %490, %240, %241, %242, %243, %244) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %537 : Float(1, 256, 56, 56, strides=[802816, 3136, 56, 1], requires_grad=1, device=cpu) = aten::relu(%input.23) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %491 : int[] = prim::Constantvalue=[1, 1] %492 : int[] = prim::Constantvalue=[1, 1] %493 : int[] = prim::Constantvalue=[1, 1] %256 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %494 : int[] = prim::Constantvalue=[0, 0] %260 : int = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %261 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %262 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %263 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %264 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %input.27 : Float(1, 256, 56, 56, strides=[802816, 3136, 56, 1], requires_grad=0, device=cpu) = aten::_convolution(%537, %13, %14, %491, %492, %493, %256, %494, %260, %261, %262, %263, %264) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %538 : Float(1, 256, 56, 56, strides=[802816, 3136, 56, 1], requires_grad=1, device=cpu) = aten::relu(%input.27) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %495 : int[] = prim::Constantvalue=[2, 2] %496 : int[] = prim::Constantvalue=[2, 2] %497 : int[] = prim::Constantvalue=[0, 0] %498 : int[] = prim::Constantvalue=[1, 1] %279 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:782:0 %input.29 : Float(1, 256, 28, 28, strides=[200704, 784, 28, 1], requires_grad=1, device=cpu) = aten::max_pool2d(%538, %495, %496, %497, %498, %279) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:782:0 %499 : int[] = prim::Constantvalue=[1, 1] %500 : int[] = prim::Constantvalue=[1, 1] %501 : int[] = prim::Constantvalue=[1, 1] %290 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %502 : int[] = prim::Constantvalue=[0, 0] %294 : int = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %295 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %296 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %297 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %298 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %input.31 : Float(1, 512, 28, 28, strides=[401408, 784, 28, 1], requires_grad=0, device=cpu) = aten::_convolution(%input.29, %15, %16, %499, %500, %501, %290, %502, %294, %295, %296, %297, %298) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %539 : Float(1, 512, 28, 28, strides=[401408, 784, 28, 1], requires_grad=1, device=cpu) = aten::relu(%input.31) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %503 : int[] = prim::Constantvalue=[1, 1] %504 : int[] = prim::Constantvalue=[1, 1] %505 : int[] = prim::Constantvalue=[1, 1] %310 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %506 : int[] = prim::Constantvalue=[0, 0] %314 : int = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %315 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %316 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %317 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %318 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %input.35 : Float(1, 512, 28, 28, strides=[401408, 784, 28, 1], requires_grad=0, device=cpu) = aten::_convolution(%539, %17, %18, %503, %504, %505, %310, %506, %314, %315, %316, %317, %318) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %540 : Float(1, 512, 28, 28, strides=[401408, 784, 28, 1], requires_grad=1, device=cpu) = aten::relu(%input.35) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %507 : int[] = prim::Constantvalue=[1, 1] %508 : int[] = prim::Constantvalue=[1, 1] %509 : int[] = prim::Constantvalue=[1, 1] %330 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %510 : int[] = prim::Constantvalue=[0, 0] %334 : int = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %335 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %336 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %337 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %338 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %input.39 : Float(1, 512, 28, 28, strides=[401408, 784, 28, 1], requires_grad=0, device=cpu) = aten::_convolution(%540, %19, %20, %507, %508, %509, %330, %510, %334, %335, %336, %337, %338) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %541 : Float(1, 512, 28, 28, strides=[401408, 784, 28, 1], requires_grad=1, device=cpu) = aten::relu(%input.39) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %511 : int[] = prim::Constantvalue=[2, 2] %512 : int[] = prim::Constantvalue=[2, 2] %513 : int[] = prim::Constantvalue=[0, 0] %514 : int[] = prim::Constantvalue=[1, 1] %353 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:782:0 %input.41 : Float(1, 512, 14, 14, strides=[100352, 196, 14, 1], requires_grad=1, device=cpu) = aten::max_pool2d(%541, %511, %512, %513, %514, %353) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:782:0 %515 : int[] = prim::Constantvalue=[1, 1] %516 : int[] = prim::Constantvalue=[1, 1] %517 : int[] = prim::Constantvalue=[1, 1] %364 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %518 : int[] = prim::Constantvalue=[0, 0] %368 : int = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %369 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %370 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %371 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %372 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %input.43 : Float(1, 512, 14, 14, strides=[100352, 196, 14, 1], requires_grad=0, device=cpu) = aten::_convolution(%input.41, %21, %22, %515, %516, %517, %364, %518, %368, %369, %370, %371, %372) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %542 : Float(1, 512, 14, 14, strides=[100352, 196, 14, 1], requires_grad=1, device=cpu) = aten::relu(%input.43) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %519 : int[] = prim::Constantvalue=[1, 1] %520 : int[] = prim::Constantvalue=[1, 1] %521 : int[] = prim::Constantvalue=[1, 1] %384 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %522 : int[] = prim::Constantvalue=[0, 0] %388 : int = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %389 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %390 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %391 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %392 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %input.47 : Float(1, 512, 14, 14, strides=[100352, 196, 14, 1], requires_grad=0, device=cpu) = aten::_convolution(%542, %23, %24, %519, %520, %521, %384, %522, %388, %389, %390, %391, %392) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %543 : Float(1, 512, 14, 14, strides=[100352, 196, 14, 1], requires_grad=1, device=cpu) = aten::relu(%input.47) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %523 : int[] = prim::Constantvalue=[1, 1] %524 : int[] = prim::Constantvalue=[1, 1] %525 : int[] = prim::Constantvalue=[1, 1] %404 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %526 : int[] = prim::Constantvalue=[0, 0] %408 : int = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %409 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %410 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %411 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %412 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %input.51 : Float(1, 512, 14, 14, strides=[100352, 196, 14, 1], requires_grad=0, device=cpu) = aten::_convolution(%543, %25, %26, %523, %524, %525, %404, %526, %408, %409, %410, %411, %412) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py:454:0 %544 : Float(1, 512, 14, 14, strides=[100352, 196, 14, 1], requires_grad=1, device=cpu) = aten::relu(%input.51) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %527 : int[] = prim::Constantvalue=[2, 2] %528 : int[] = prim::Constantvalue=[2, 2] %529 : int[] = prim::Constantvalue=[0, 0] %530 : int[] = prim::Constantvalue=[1, 1] %427 : bool = prim::Constantvalue=0 # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:782:0 %input.53 : Float(1, 512, 7, 7, strides=[25088, 49, 7, 1], requires_grad=1, device=cpu) = aten::max_pool2d(%544, %527, %528, %529, %530, %427) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:782:0 %531 : int[] = prim::Constantvalue=[7, 7] %444 : Float(1, 512, 7, 7, strides=[25088, 49, 7, 1], requires_grad=1, device=cpu) = aten::adaptive_avg_pool2d(%input.53, %531) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1214:0 %445 : int = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torchvision/models/vgg.py:68:0 %446 : int = prim::Constantvalue=-1 # /usr/local/lib/python3.7/dist-packages/torchvision/models/vgg.py:68:0 %447 : Float(1, 25088, strides=[25088, 1], requires_grad=1, device=cpu) = aten::flatten(%444, %445, %446) # /usr/local/lib/python3.7/dist-packages/torchvision/models/vgg.py:68:0 %input.55 : Float(1, 4096, strides=[4096, 1], requires_grad=1, device=cpu) = aten::linear(%447, %27, %28) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/linear.py:114:0 %545 : Float(1, 4096, strides=[4096, 1], requires_grad=1, device=cpu) = aten::relu(%input.55) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %450 : float = prim::Constantvalue=0.5 # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1252:0 %451 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1252:0 %452 : Float(1, 4096, strides=[4096, 1], requires_grad=1, device=cpu) = aten::dropout(%545, %450, %451) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1252:0 %input.59 : Float(1, 4096, strides=[4096, 1], requires_grad=1, device=cpu) = aten::linear(%452, %29, %30) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/linear.py:114:0 %546 : Float(1, 4096, strides=[4096, 1], requires_grad=1, device=cpu) = aten::relu(%input.59) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1455:0 %455 : float = prim::Constantvalue=0.5 # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1252:0 %456 : bool = prim::Constantvalue=1 # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1252:0 %457 : Float(1, 4096, strides=[4096, 1], requires_grad=1, device=cpu) = aten::dropout(%546, %455, %456) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1252:0 %458 : Float(1, 1000, strides=[1000, 1], requires_grad=1, device=cpu) = aten::linear(%457, %31, %32) # /usr/local/lib/python3.7/dist-packages/torch/nn/modules/linear.py:114:0 return (%458) , None, False

opened by Junaid199f 0

Error when trying to print the Model

Hello thank you for sharing your lib.

I m running into an error: RuntimeError: Input type (torch.FloatTensor) and weight type (torch.cuda.FloatTensor) should be the same or input should be a MKLDNN tensor and weight is a dense tensor`

When I shift it to the CPU: hl.build_graph(self.online_net.to('cpu'), torch.zeros([1, 4, 32, 32]))

I m getting RuntimeError: Expected node type 'onnx::Constant' for argument 'rounding_mode' of node '_div_rounding_mode', got 'prim::Param'.

I will also print my code. Maybe somebody can figure out why its not working.

# -*- coding: utf-8 -*-
from __future__ import division
import math
import torch
from torch import nn
from torch.nn import functional as F
import functools
import operator
from torch.nn.utils import spectral_norm


# Factorised NoisyLinear layer with bias
class NoisyLinear(nn.Module):
  def __init__(self, in_features, out_features, std_init=0.3):
    super(NoisyLinear, self).__init__()
    self.in_features = in_features
    self.out_features = out_features
    self.std_init = std_init
    self.weight_mu = nn.Parameter(torch.empty(out_features, in_features))
    self.weight_sigma = nn.Parameter(torch.empty(out_features, in_features))
    self.register_buffer('weight_epsilon', torch.empty(out_features, in_features))
    self.bias_mu = nn.Parameter(torch.empty(out_features))
    self.bias_sigma = nn.Parameter(torch.empty(out_features))
    self.register_buffer('bias_epsilon', torch.empty(out_features))
    self.reset_parameters()
    self.reset_noise()

  def reset_parameters(self):
    mu_range = 1 / math.sqrt(self.in_features)
    self.weight_mu.data.uniform_(-mu_range, mu_range)
    self.weight_sigma.data.fill_(self.std_init / math.sqrt(self.in_features))
    self.bias_mu.data.uniform_(-mu_range, mu_range)
    self.bias_sigma.data.fill_(self.std_init / math.sqrt(self.out_features))

  def _scale_noise(self, size):
    x = torch.randn(size, device=self.weight_mu.device)
    return x.sign().mul_(x.abs().sqrt_())

  def reset_noise(self):
    epsilon_in = self._scale_noise(self.in_features)
    epsilon_out = self._scale_noise(self.out_features)
    self.weight_epsilon.copy_(epsilon_out.ger(epsilon_in))
    self.bias_epsilon.copy_(epsilon_out)

  def forward(self, input):
    if self.training:
      return F.linear(input, self.weight_mu + self.weight_sigma * self.weight_epsilon, self.bias_mu + self.bias_sigma * self.bias_epsilon)
    else:
      return F.linear(input, self.weight_mu, self.bias_mu)



class ResBlock(nn.Module):
    def __init__(self, in_channels, out_channels, downsample):
        super().__init__()
        if downsample:
            self.conv1 = nn.Conv2d(
                in_channels, out_channels, kernel_size=3, stride=2, padding=1)
            self.shortcut = nn.Sequential(

                nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=2)#,
                #nn.BatchNorm2d(out_channels)

            )
        else:
            self.conv1 = nn.Conv2d(
                in_channels, out_channels, kernel_size=3, stride=1, padding=1)
            self.shortcut = nn.Sequential()

        self.conv2 = nn.Conv2d(out_channels, out_channels,
                               kernel_size=3, stride=1, padding=1)

        #self.bn1 = nn.BatchNorm2d(out_channels)
        #self.bn2 = nn.BatchNorm2d(out_channels)

    def forward(self, input):
        shortcut = self.shortcut(input)
        input = nn.ReLU()(self.conv1(input))
        input = nn.ReLU()(self.conv2(input))
        input = input + shortcut
        return nn.ReLU()(input)


class Rainbow_ResNet(nn.Module):
  def __init__(self, args, action_space, resblock, repeat):
    super(Rainbow_ResNet, self).__init__()
    self.atoms = args.atoms
    self.action_space = action_space

    filters = [128, 128, 256, 512, 1024]
    self.layer0 = nn.Sequential(
      nn.Conv2d(4, 128, kernel_size=5, stride=1, padding=1),
      #nn.MaxPool2d(kernel_size=3, stride=2, padding=1),
      #nn.BatchNorm2d(64),

      nn.ReLU())

    self.layer1 = nn.Sequential()
    self.layer1.add_module('conv2_1', ResBlock(filters[0], filters[1], downsample=True))
    for i in range(1, repeat[0]):
            self.layer1.add_module('conv2_%d'%(i+1,), ResBlock(filters[1], filters[1], downsample=False))

    self.layer2 = nn.Sequential()

    self.layer2.add_module('conv3_1', ResBlock(filters[1], filters[2], downsample=True))

    for i in range(1, repeat[1]):
            self.layer2.add_module('conv3_%d' % (
                i+1,), ResBlock(filters[2], filters[2], downsample=False))


    #self.layer3 = nn.Sequential()
    #self.layer3.add_module('conv4_1', ResBlock(filters[2], filters[3], downsample=True))
    #for i in range(1, repeat[2]):
    #    self.layer3.add_module('conv4_%d' % (
    #        i+1,), ResBlock(filters[3], filters[3], downsample=False))

    #self.layer4 = nn.Sequential()
    #self.layer4.add_module('conv5_1', ResBlock(filters[3], filters[4], downsample=True))
    #for i in range(1, repeat[3]):
    #    self.layer4.add_module('conv5_%d'%(i+1,),ResBlock(filters[4], filters[4], downsample=False))

    #self.dense = nn.Sequential(spectral_norm(nn.Linear(12544, 1024)), nn.ReLU())
    self.fc_h_v = spectral_norm(nn.Linear(16384, 512))
    self.fc_h_a = spectral_norm(nn.Linear(16384, 512))

    self.fc_z_v = NoisyLinear(512, self.atoms, std_init=args.noisy_std)
    self.fc_z_a = NoisyLinear(512, action_space * self.atoms, std_init=args.noisy_std) 

  def forward(self, x, log=False):
    input = self.layer0(x)
    input = self.layer1(input)
    input = self.layer2(input)

    #input = self.layer3(input)
    #input = self.layer4(input)
    input = torch.flatten(input, start_dim=1)
    #input = self.dense(input)

    v_uuv = self.fc_z_v(F.relu(self.fc_h_v(input)))  # Value stream
    a_uuv = self.fc_z_a(F.relu(self.fc_h_a(input)))  # Advantage stream

    #v_uav, a_uav = v_uav.view(-1, 1, self.atoms), a_uav.view(-1, self.action_space, self.atoms)    
    v_uuv, a_uuv = v_uuv.view(-1, 1, self.atoms), a_uuv.view(-1, self.action_space, self.atoms)
    
    #q_uav = v_uav + a_uav - a_uav.mean(1, keepdim=True)  # Combine streams
    q_uuv = v_uuv + a_uuv - a_uuv.mean(1, keepdim=True)  # Combine streams

    if log:  # Use log softmax for numerical stability
      #q_uav = F.log_softmax(q_uav, dim=2)  # Log probabilities with action over second dimension
      q_uuv = F.log_softmax(q_uuv, dim=2)  # Log probabilities with action over second dimension
    else:
      #q_uav = F.softmax(q_uav, dim=2)  # Probabilities with action over second dimension
      q_uuv = F.softmax(q_uuv, dim=2)  # Probabilities with action over second dimension
    return  q_uuv #q_uav,
  def reset_noise(self):
    for name, module in self.named_children():
      if 'fc_z' in name:
        module.reset_noise()



  def reset_noise(self):
    for name, module in self.named_children():
      if 'fc_z' in name:
        module.reset_noise()

opened by Mateus224 0

TypeError for Pytorch Model

Hello, I am trying to use hiddenlayer to draw a pytorch model, I got some error coming out of onnx

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
/home/ubuntu/mstar/scripts/rlfh/visualization.ipynb Cell 4' in <cell line: 13>()
      [6](vscode-notebook-cell://ssh-remote%2Bmistd/home/ubuntu/mstar/scripts/rlfh/visualization.ipynb#ch0000003vscode-remote?line=5) #model = torchvision.models.vgg16()
      [8](vscode-notebook-cell://ssh-remote%2Bmistd/home/ubuntu/mstar/scripts/rlfh/visualization.ipynb#ch0000003vscode-remote?line=7) model = torch.nn.Sequential(
      [9](vscode-notebook-cell://ssh-remote%2Bmistd/home/ubuntu/mstar/scripts/rlfh/visualization.ipynb#ch0000003vscode-remote?line=8)     nn.Linear(10, 10),
     [10](vscode-notebook-cell://ssh-remote%2Bmistd/home/ubuntu/mstar/scripts/rlfh/visualization.ipynb#ch0000003vscode-remote?line=9)     nn.Linear(10, 2)
     [11](vscode-notebook-cell://ssh-remote%2Bmistd/home/ubuntu/mstar/scripts/rlfh/visualization.ipynb#ch0000003vscode-remote?line=10) )
---> [13](vscode-notebook-cell://ssh-remote%2Bmistd/home/ubuntu/mstar/scripts/rlfh/visualization.ipynb#ch0000003vscode-remote?line=12) hl.build_graph(model, torch.zeros([1, 10]))

File ~/py38/lib/python3.8/site-packages/hiddenlayer/graph.py:143, in build_graph(model, args, input_names, transforms, framework_transforms)
    [141](file:///home/ubuntu/py38/lib/python3.8/site-packages/hiddenlayer/graph.py?line=140)     from .pytorch_builder import import_graph, FRAMEWORK_TRANSFORMS
    [142](file:///home/ubuntu/py38/lib/python3.8/site-packages/hiddenlayer/graph.py?line=141)     assert args is not None, "Argument args must be provided for Pytorch models."
--> [143](file:///home/ubuntu/py38/lib/python3.8/site-packages/hiddenlayer/graph.py?line=142)     import_graph(g, model, args)
    [144](file:///home/ubuntu/py38/lib/python3.8/site-packages/hiddenlayer/graph.py?line=143) elif framework == "tensorflow":
    [145](file:///home/ubuntu/py38/lib/python3.8/site-packages/hiddenlayer/graph.py?line=144)     from .tf_builder import import_graph, FRAMEWORK_TRANSFORMS

File ~/py38/lib/python3.8/site-packages/hiddenlayer/pytorch_builder.py:71, in import_graph(hl_graph, model, args, input_names, verbose)
     [66](file:///home/ubuntu/py38/lib/python3.8/site-packages/hiddenlayer/pytorch_builder.py?line=65) def import_graph(hl_graph, model, args, input_names=None, verbose=False):
     [67](file:///home/ubuntu/py38/lib/python3.8/site-packages/hiddenlayer/pytorch_builder.py?line=66)     # TODO: add input names to graph
     [68](file:///home/ubuntu/py38/lib/python3.8/site-packages/hiddenlayer/pytorch_builder.py?line=67) 
     [69](file:///home/ubuntu/py38/lib/python3.8/site-packages/hiddenlayer/pytorch_builder.py?line=68)     # Run the Pytorch graph to get a trace and generate a graph from it
     [70](file:///home/ubuntu/py38/lib/python3.8/site-packages/hiddenlayer/pytorch_builder.py?line=69)     trace, out = torch.jit._get_trace_graph(model, args)
---> [71](file:///home/ubuntu/py38/lib/python3.8/site-packages/hiddenlayer/pytorch_builder.py?line=70)     torch_graph = torch.onnx._optimize_trace(trace, torch.onnx.OperatorExportTypes.ONNX)
     [73](file:///home/ubuntu/py38/lib/python3.8/site-packages/hiddenlayer/pytorch_builder.py?line=72)     # Dump list of nodes (DEBUG only)
     [74](file:///home/ubuntu/py38/lib/python3.8/site-packages/hiddenlayer/pytorch_builder.py?line=73)     if verbose:

File ~/py38/lib/python3.8/site-packages/torch/onnx/__init__.py:394, in _optimize_trace(graph, operator_export_type)
    [391](file:///home/ubuntu/py38/lib/python3.8/site-packages/torch/onnx/__init__.py?line=390) def _optimize_trace(graph, operator_export_type):
    [392](file:///home/ubuntu/py38/lib/python3.8/site-packages/torch/onnx/__init__.py?line=391)     from torch.onnx import utils
--> [394](file:///home/ubuntu/py38/lib/python3.8/site-packages/torch/onnx/__init__.py?line=393)     return utils._optimize_graph(graph, operator_export_type)

File ~/py38/lib/python3.8/site-packages/torch/onnx/utils.py:276, in _optimize_graph(graph, operator_export_type, _disable_torch_constant_prop, fixed_batch_size, params_dict, dynamic_axes, input_names, module)
    [274](file:///home/ubuntu/py38/lib/python3.8/site-packages/torch/onnx/utils.py?line=273) symbolic_helper._quantized_ops.clear()
    [275](file:///home/ubuntu/py38/lib/python3.8/site-packages/torch/onnx/utils.py?line=274) # Unpack quantized weights for conv and linear ops and insert into graph.
--> [276](file:///home/ubuntu/py38/lib/python3.8/site-packages/torch/onnx/utils.py?line=275) _C._jit_pass_onnx_unpack_quantized_weights(
    [277](file:///home/ubuntu/py38/lib/python3.8/site-packages/torch/onnx/utils.py?line=276)     graph, params_dict, symbolic_helper.is_caffe2_aten_fallback()
    [278](file:///home/ubuntu/py38/lib/python3.8/site-packages/torch/onnx/utils.py?line=277) )
    [279](file:///home/ubuntu/py38/lib/python3.8/site-packages/torch/onnx/utils.py?line=278) if symbolic_helper.is_caffe2_aten_fallback():
    [280](file:///home/ubuntu/py38/lib/python3.8/site-packages/torch/onnx/utils.py?line=279)     # Insert permutes before and after each conv op to ensure correct order.
    [281](file:///home/ubuntu/py38/lib/python3.8/site-packages/torch/onnx/utils.py?line=280)     _C._jit_pass_onnx_quantization_insert_permutes(graph, params_dict)

TypeError: _jit_pass_onnx_unpack_quantized_weights(): incompatible function arguments. The following argument types are supported:
    1. (arg0: torch::jit::Graph, arg1: Dict[str, IValue], arg2: bool) -> Dict[str, IValue]

Invoked with: graph(%0 : Float(1, 10, strides=[10, 1], requires_grad=0, device=cpu),
      %1 : Float(10, 10, strides=[10, 1], requires_grad=1, device=cpu),
      %2 : Float(10, strides=[1], requires_grad=1, device=cpu),
      %3 : Float(2, 10, strides=[10, 1], requires_grad=1, device=cpu),
      %4 : Float(2, strides=[1], requires_grad=1, device=cpu)):
  %15 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::linear(%0, %1, %2) # /home/ubuntu/py38/lib/python3.8/site-packages/torch/nn/modules/linear.py:114:0
  %16 : Float(1, 2, strides=[2, 1], requires_grad=1, device=cpu) = aten::linear(%15, %3, %4) # /home/ubuntu/py38/lib/python3.8/site-packages/torch/nn/modules/linear.py:114:0
  return (%16)
, None, False

runtime:

ubuntu 20.04, python 3.8, torch 1.13.0 (experimental), hiddenlayer 0.3

script to reproduce the error:

import hiddenlayer as hl
import torch
import torch.nn as nn
import torchvision.models

#model = torchvision.models.vgg16()

model = torch.nn.Sequential(
    nn.Linear(10, 10),
    nn.Linear(10, 2)
)

hl.build_graph(model, torch.zeros([1, 10]))

opened by hsl89 4

How to display dimensions and change font to Times

Good day! I just discovered this amazing package as I was trying to visualize the model that I made. May I know how can I display the dimensions (e.g., 100x3, 1x200x3, etc.) and also how can I make the font to be Times New Roman since I'll be transferring it within an academic paper. Thank you! Much power!

opened by egmaminta 1

Releases(v0.2)

v0.2(Dec 3, 2018)

This release adds Keras support and fixes several bugs.
Source code(tar.gz)
Source code(zip)

Neural network graphs and training metrics for PyTorch, Tensorflow, and Keras.

Related tags

Overview

HiddenLayer

1. Readable Graphs

2. Training Metrics in Jupyter Notebook

3. Hackable

Demos

Contributing

Installation

1. Prerequisites

2. Install HiddenLayer

a. Clone From GitHub (Developer Mode)

b. Using PIP ("stable" release)

c. Install to your site-packages directly from GitHub

Comments

runtime:

script to reproduce the error:

Releases(v0.2)

v0.2(Dec 3, 2018)

Owner

Waleed

Code and data for the paper "Hearing What You Cannot See"

Python Blood Vessel Topology Analysis

Open CV - Convert a picture to look like a cartoon sketch in python

Facebook AI Research Sequence-to-Sequence Toolkit written in Python.

A tiny, friendly, strong baseline code for Person-reID (based on pytorch).

A Physics-based Noise Formation Model for Extreme Low-light Raw Denoising (CVPR 2020 Oral & TPAMI 2021)

Code for KDD'20 "An Efficient Neighborhood-based Interaction Model for Recommendation on Heterogeneous Graph"

Pytorch reimplementation of the Mixer (MLP-Mixer: An all-MLP Architecture for Vision)

Self-supervised Augmentation Consistency for Adapting Semantic Segmentation (CVPR 2021)

torchbearer: A model fitting library for PyTorch

v objective diffusion inference code for JAX.

GeneralOCR is open source Optical Character Recognition based on PyTorch.

Clean and readable code for Decision Transformer: Reinforcement Learning via Sequence Modeling

VisualGPT: Data-efficient Adaptation of Pretrained Language Models for Image Captioning

VLG-Net: Video-Language Graph Matching Networks for Video Grounding

My published benchmark for a Kaggle Simulations Competition

Rotary Transformer

GNNAdvisor: An Efficient Runtime System for GNN Acceleration on GPUs

Deep Learning tutorials in jupyter notebooks.

Database Reasoning Over Text project for ACL paper

c. Install to your `site-packages` directly from GitHub