I'm seeing this:

TypeError: Don't know how to create Python object for N5torch8autograd8VariableE

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<stdin>", line 15, in add_nodes
  File "<stdin>", line 15, in add_nodes
  File "<stdin>", line 15, in add_nodes
  File "<stdin>", line 15, in add_nodes
  File "<stdin>", line 11, in add_nodes

For reference with line numbers, I have the function as:

>>> def add_nodes(var):
...         print(var)
...         print("didthatone")
...         if var not in seen:
...             if isinstance(var, Variable):
...                 value = '('+(', ').join(['%d'% v for v in var.size()])+')'
...                 dot.node(str(id(var)), str(value), fillcolor='lightblue')
...             else:
...                 dot.node(str(id(var)), str(type(var).__name__))
...             seen.add(var)
...             if hasattr(var, 'previous_functions'):
...                 for u in var.previous_functions:
...                     print(u)
...                     dot.edge(str(id(u[0])), str(id(var)))
...                     add_nodes(u[0])

Hey, so I am facing the same issue as issue #14 . Although what you said is correct about the model.state_dict(), the error about grad_fn arises from var, which is the output variable of the model and not the params dictionary. Any ideas on how to fix this?

As for the model.state_dict() part, I think it can be converted to the required (name, Variable) format using;

params = model.state_dict()
for k in sorted(params.keys()):
	v = params[k]
	params[k] = autograd.Variable(v, requires_grad=True)

I printed this out, and it seemed fine.

Need help to remove the var.grad_fn attribute error though. Thanks.

The readme explains that all the models listed so far have bn params integrated into the conv weights, for performance, thus making these models not useful for fine-tuning. Fine-tuning is, perhaps, the most common thing that people use pre-trained models for, so it would be great it you could do one of the following:

• Explain how to separate out the bn params in the existing models
• Show how to re-create the models so that others can do so without the bn param integration
• Make every model available in both integrated and standard format

Many thanks for this great project!

Hi,

thanks a lot for making that function available! However, when I try calling make_dot(loss) I get the error below.

Any idea how that could be fixed?

Thanks again!

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-10-b7f0e150648d> in <module>()
      1 from visualize import make_dot
----> 2 make_dot(loss)

/home/carnd/basic_pytorch/visualize.py in make_dot(var, params)
     55                     dot.edge(str(id(t)), str(id(var)))
     56                     add_nodes(t)
---> 57     add_nodes(var.grad_fn)
     58     return dot

/home/carnd/basic_pytorch/visualize.py in add_nodes(var)
     50                     if u[0] is not None:
     51                         dot.edge(str(id(u[0])), str(id(var)))
---> 52                         add_nodes(u[0])
     53             if hasattr(var, 'saved_tensors'):
     54                 for t in var.saved_tensors:

/home/carnd/basic_pytorch/visualize.py in add_nodes(var)
     50                     if u[0] is not None:
     51                         dot.edge(str(id(u[0])), str(id(var)))
---> 52                         add_nodes(u[0])
     53             if hasattr(var, 'saved_tensors'):
     54                 for t in var.saved_tensors:

/home/carnd/basic_pytorch/visualize.py in add_nodes(var)
     50                     if u[0] is not None:
     51                         dot.edge(str(id(u[0])), str(id(var)))
---> 52                         add_nodes(u[0])
     53             if hasattr(var, 'saved_tensors'):
     54                 for t in var.saved_tensors:

/home/carnd/basic_pytorch/visualize.py in add_nodes(var)
     50                     if u[0] is not None:
     51                         dot.edge(str(id(u[0])), str(id(var)))
---> 52                         add_nodes(u[0])
     53             if hasattr(var, 'saved_tensors'):
     54                 for t in var.saved_tensors:

/home/carnd/basic_pytorch/visualize.py in add_nodes(var)
     50                     if u[0] is not None:
     51                         dot.edge(str(id(u[0])), str(id(var)))
---> 52                         add_nodes(u[0])
     53             if hasattr(var, 'saved_tensors'):
     54                 for t in var.saved_tensors:

/home/carnd/basic_pytorch/visualize.py in add_nodes(var)
     51                         dot.edge(str(id(u[0])), str(id(var)))
     52                         add_nodes(u[0])
---> 53             if hasattr(var, 'saved_tensors'):
     54                 for t in var.saved_tensors:
     55                     dot.edge(str(id(t)), str(id(var)))

RuntimeError: Trying to backward through the graph a second time, but the buffers have already been freed. Specify retain_graph=True when calling backward the first time.

Hi @szagoruyko,

First of all thanks for converting the wide resnet model to tensorflow :)

I have some question about the details in WRN50 model:

1. The Input image, is it RGB, BGR, or others? I saw the inputs.permute(0,2,3,1) and its kinda unintuitive.
2. What about image mean?

As pytorch autograd has been refactored, old visualizer doesn't work, so this fixes and adds new nodes:

• blue nodes are the ones that need grad
• orange nodes hold tensors in save_for_backward, useful for tracking memory usage For now I don't add None nodes for nodes that don't need grad, which would be nice to fix, but requires changes in autograd engine.

fixes https://github.com/szagoruyko/functional-zoo/issues/7 will merge as soon as https://github.com/pytorch/pytorch/pull/1506 is in

Hello, thanks for sharing your code!

I'd like to know what preprocessing you've used to produce the ImageNet weights. Can you confirm that you used meanstd normalization?

All the best, Carl

I try to export the wide-resnet-50-2 in tensorflow by using your code in https://github.com/szagoruyko/functional-zoo/blob/master/wide-resnet-50-2-export.ipynb But it seems like I cannot find out where to download the paramters, can you tell me where the .pth is ?

when I use following code to visualize the torch pretrained resnet18 model, I get AttributeError: grad_fn, at line: add_nodes(var.grad_fn)

input=torch.randn(1,3,224,224)  
model=models.resnet18(pretrained=True)
for param in model.parameters():
     param.requires_grad = True
params = model.state_dict()
y=model(Variable(input))
print y
dot=make_dot(y, params)

HDF5-DIAG: Error detected in HDF5 (1.8.16) thread 139655733122880:
  #000: ../../../src/H5G.c line 467 in H5Gopen2(): unable to open group
    major: Symbol table
    minor: Can't open object
  #001: ../../../src/H5Gint.c line 320 in H5G__open_name(): group not found
    major: Symbol table
    minor: Object not found
  #002: ../../../src/H5Gloc.c line 430 in H5G_loc_find(): can't find object
    major: Symbol table
    minor: Object not found
  #003: ../../../src/H5Gtraverse.c line 861 in H5G_traverse(): internal path traversal failed
    major: Symbol table
    minor: Object not found
  #004: ../../../src/H5Gtraverse.c line 641 in H5G_traverse_real(): traversal operator failed
    major: Symbol table
    minor: Callback failed
  #005: ../../../src/H5Gloc.c line 385 in H5G_loc_find_cb(): object 'data_0' doesn't exist
    major: Symbol table
    minor: Object not found
terminate called after throwing an instance of 'H5::FileIException'
Aborted (core dumped)

I am trying to read the weights of yolo network ./hdf5_app /home/myst/yolo/yolo.h5

In the latest version of PyTorch, the 'previous_functions' attribute has been renamed to 'next_functions' and the 'creator' attribute has been renamed to 'grad_fn'. This PR accounts for this while not breaking existing functionality.

File "/usr/local/lib/python3.6/dist-packages/hickle/hickle.py", line 526, in load raise RuntimeError("Cannot open file. This file was likely" RuntimeError: Cannot open file. This file was likely created with Python 2 and an old hickle version.

How can we debug in this situation above? It is problem of using python3 but unsupported from hickle.

hickle.py

 elif VER_MAJOR == 2:
                    if six.PY2:
                        warnings.warn("Hickle file appears to be old version (v2), attempting legacy loading...")
                        import hickle_legacy2
                        return hickle_legacy2.load(fileobj, safe=safe)
                    else:
                        raise RuntimeError("Cannot open file. This file was likely"
                                           " created with Python 2 and an old hickle version.")
                # There is an unfortunate period of time where hickle 2.1.0 claims VERSION = int(3)
                # For backward compatibility we really need to catch this.
                # Actual hickle v3 files are versioned as A.B.C (e.g. 3.1.0)

When trying to load the weights of the pretrained wide resnet for ImageNet, using Tensorflow, I get the error message: "RuntimeError: Cannot open file. This file was likely created with Python 2 and an old hickle version."

I haven't found any solution to this. Can you help me find weights that can be used with python 3.6? Thanks, Jenny

Hi,

In the "Tensorflow model definition" section, the parameters are taken from torch: (params = {k: v.numpy() for k,v in torch.load('wide-resnet-50-2-export.pth').items()})

How can I run this without using torch? Thanks!

To extract features from an arbitrary layer with say, VGG19, given an input image "elephant.jpg", the code in keras+TF is simply:

from keras.applications.vgg19 import VGG19 from keras.preprocessing import image from keras.applications.vgg19 import preprocess_input from keras.models import Model import numpy as np

base_model = VGG19(weights='imagenet') model = Model(input=base_model.input, output=base_model.get_layer('block4_pool').output)

img_path = 'elephant.jpg'

img = image.load_img(img_path, target_size=(224, 224)) x = image.img_to_array(img) x = np.expand_dims(x, axis=0) x = preprocess_input(x)

block4_pool_features = model.predict(x)

This can be turned into a function to support multiple calls. How is this achieved in PyTorch?