Printing pretrained torch model options

[engahmed1190]

I am trying to print the torch model options of training, I have used the provided model model.t7

I am using this code to print the option

require 'torch'
require 'nn'

require 'InstanceNormalization'

--[[
Prints the options that were used to train a a feedforward model.
--]]

local cmd = torch.CmdLine()
cmd:option('-model', 'models/instance_norm/candy.t7')
local opt = cmd:parse(arg)

print('Loading model from ' .. opt.model)
local checkpoint = torch.load(opt.model)

for k, v in pairs(checkpoint.opt) do
  if type(v) == 'table' then
    v = table.concat(v, ',')
  end
  print(string.format('%s: %s', k, v))
end

but I am getting this error:

/src/torch/install/bin/luajit: print_options.lua:22: bad argument #1 to 'pairs' (table expected, got nil)
stack traceback:
    [C]: in function 'pairs'
    print_options.lua:22: in main chunk
    [C]: in function 'dofile'
    .../src/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk
    [C]: at 0x00406670

[engahmed1190]

this is the model that you provided for reference , it loaded perfectly find but the main issue with the evaluate method .

nn.Sequential {
  [input -> (0) -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> (9) -> (10) -> (11) -> (12) -> (13) -> (14) -> (15) -> (16) -> (17) -> (18) -> (19) -> (20) -> (21) -> (22) -> (23) -> output]
  (0): TorchObject(nn.TVLoss, {'_type': 'torch.FloatTensor', 'strength': 0, 'x_diff': 
  ( 0 ,.,.) = 
    0.0000  0.0000  0.0000  ...  -0.0039  0.0078  0.0000
    0.0000  0.0000  0.0000  ...   0.0039  0.0000  0.0000
    0.0000  0.0000  0.0000  ...   0.0078 -0.0157  0.0000
             ...             ⋱             ...          
   -0.0196  0.0157 -0.0196  ...   0.0078  0.0588  0.0392
   -0.0196  0.0196 -0.0235  ...   0.0235  0.0745  0.1451
   -0.0078  0.0000  0.0118  ...   0.0431  0.1882  0.1569

  ( 1 ,.,.) = 
    0.0000  0.0000  0.0000  ...  -0.0039  0.0078  0.0000
    0.0000  0.0000  0.0000  ...   0.0039  0.0000  0.0000
    0.0000  0.0000  0.0000  ...   0.0078 -0.0078  0.0000
             ...             ⋱             ...          
   -0.0196  0.0157 -0.0196  ...   0.0078  0.0431  0.0235
   -0.0196  0.0196 -0.0235  ...   0.0196  0.0588  0.1373
   -0.0078  0.0000  0.0118  ...   0.0353  0.1765  0.1451

  ( 2 ,.,.) = 
    0.0000  0.0000  0.0000  ...  -0.0039  0.0078  0.0000
    0.0000  0.0000  0.0000  ...   0.0039  0.0000  0.0000
    0.0000  0.0000  0.0000  ...   0.0078 -0.0078  0.0000
             ...             ⋱             ...          
   -0.0196  0.0157 -0.0196  ...   0.0078  0.0471  0.0235
   -0.0196  0.0196 -0.0235  ...   0.0118  0.0510  0.1216
   -0.0078  0.0039  0.0118  ...   0.0157  0.1647  0.1176
  [torch.FloatTensor of size 3x511x511]
  , 'gradInput': [torch.FloatTensor with no dimension]
  , 'y_diff': 
  ( 0 ,.,.) = 
    0.0000  0.0000  0.0000  ...   0.0000  0.0078  0.0000
    0.0000  0.0000  0.0000  ...   0.0078  0.0118 -0.0039
    0.0000  0.0000  0.0000  ...  -0.0118 -0.0078  0.0000
             ...             ⋱             ...          
    0.0039  0.0039  0.0078  ...   0.0157  0.0314  0.0471
   -0.0157 -0.0039 -0.0235  ...   0.0863  0.1059  0.2196
    0.0039  0.0039  0.0039  ...   0.1020  0.1961  0.1451

  ( 1 ,.,.) = 
    0.0000  0.0000  0.0000  ...   0.0000  0.0078  0.0000
    0.0000  0.0000  0.0000  ...   0.0000  0.0039 -0.0039
    0.0000  0.0000  0.0000  ...  -0.0118 -0.0078  0.0000
             ...             ⋱             ...          
    0.0039  0.0039  0.0078  ...   0.0196  0.0314  0.0471
   -0.0157 -0.0039 -0.0235  ...   0.0706  0.0863  0.2039
    0.0039  0.0039  0.0039  ...   0.0941  0.1922  0.1451

  ( 2 ,.,.) = 
    0.0000  0.0000  0.0000  ...   0.0000  0.0078  0.0000
    0.0000  0.0000  0.0000  ...   0.0000  0.0039 -0.0039
    0.0000  0.0000  0.0000  ...  -0.0196 -0.0078  0.0000
             ...             ⋱             ...          
    0.0039  0.0039  0.0078  ...   0.0039  0.0078  0.0118
   -0.0196 -0.0078 -0.0235  ...   0.0275  0.0314  0.1451
    0.0039  0.0039  0.0039  ...   0.0745  0.1725  0.0980
  [torch.FloatTensor of size 3x511x511]
  , 'train': True, 'output': [torch.FloatTensor with no dimension]
  })
  (1): nn.SpatialReplicationPadding(4, 4, 4, 4)
  (2): nn.SpatialConvolution(3 -> 32, 9x9)
  (3): nn.InstanceNormalization
  (4): nn.ReLU
  (5): nn.SpatialConvolution(32 -> 64, 3x3, 2, 2, 1, 1)
  (6): nn.InstanceNormalization
  (7): nn.ReLU
  (8): nn.SpatialConvolution(64 -> 128, 3x3, 2, 2, 1, 1)
  (9): nn.InstanceNormalization
  (10): nn.ReLU
  (11): nn.Sequential {
    [input -> (0) -> (1) -> output]
    (0): torch.legacy.nn.ConcatTable.ConcatTable {
      input
        |`-> (0): nn.Identity
        |`-> (1): nn.Sequential {
               [input -> (0) -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> output]
               (0): nn.SpatialReplicationPadding(1, 1, 1, 1)
               (1): nn.SpatialConvolution(128 -> 128, 3x3)
               (2): nn.InstanceNormalization
               (3): nn.ReLU
               (4): nn.SpatialReplicationPadding(1, 1, 1, 1)
               (5): nn.SpatialConvolution(128 -> 128, 3x3)
               (6): nn.InstanceNormalization
             }
         +. -> output
    }
    (1): nn.CAddTable
  }
  (12): nn.Sequential {
    [input -> (0) -> (1) -> output]
    (0): torch.legacy.nn.ConcatTable.ConcatTable {
      input
        |`-> (0): nn.Identity
        |`-> (1): nn.Sequential {
               [input -> (0) -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> output]
               (0): nn.SpatialReplicationPadding(1, 1, 1, 1)
               (1): nn.SpatialConvolution(128 -> 128, 3x3)
               (2): nn.InstanceNormalization
               (3): nn.ReLU
               (4): nn.SpatialReplicationPadding(1, 1, 1, 1)
               (5): nn.SpatialConvolution(128 -> 128, 3x3)
               (6): nn.InstanceNormalization
             }
         +. -> output
    }
    (1): nn.CAddTable
  }
  (13): nn.Sequential {
    [input -> (0) -> (1) -> output]
    (0): torch.legacy.nn.ConcatTable.ConcatTable {
      input
        |`-> (0): nn.Identity
        |`-> (1): nn.Sequential {
               [input -> (0) -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> output]
               (0): nn.SpatialReplicationPadding(1, 1, 1, 1)
               (1): nn.SpatialConvolution(128 -> 128, 3x3)
               (2): nn.InstanceNormalization
               (3): nn.ReLU
               (4): nn.SpatialReplicationPadding(1, 1, 1, 1)
               (5): nn.SpatialConvolution(128 -> 128, 3x3)
               (6): nn.InstanceNormalization
             }
         +. -> output
    }
    (1): nn.CAddTable
  }
  (14): nn.Sequential {
    [input -> (0) -> (1) -> output]
    (0): torch.legacy.nn.ConcatTable.ConcatTable {
      input
        |`-> (0): nn.Identity
        |`-> (1): nn.Sequential {
               [input -> (0) -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> output]
               (0): nn.SpatialReplicationPadding(1, 1, 1, 1)
               (1): nn.SpatialConvolution(128 -> 128, 3x3)
               (2): nn.InstanceNormalization
               (3): nn.ReLU
               (4): nn.SpatialReplicationPadding(1, 1, 1, 1)
               (5): nn.SpatialConvolution(128 -> 128, 3x3)
               (6): nn.InstanceNormalization
             }
         +. -> output
    }
    (1): nn.CAddTable
  }
  (15): nn.Sequential {
    [input -> (0) -> (1) -> output]
    (0): torch.legacy.nn.ConcatTable.ConcatTable {
      input
        |`-> (0): nn.Identity
        |`-> (1): nn.Sequential {
               [input -> (0) -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> output]
               (0): nn.SpatialReplicationPadding(1, 1, 1, 1)
               (1): nn.SpatialConvolution(128 -> 128, 3x3)
               (2): nn.InstanceNormalization
               (3): nn.ReLU
               (4): nn.SpatialReplicationPadding(1, 1, 1, 1)
               (5): nn.SpatialConvolution(128 -> 128, 3x3)
               (6): nn.InstanceNormalization
             }
         +. -> output
    }
    (1): nn.CAddTable
  }
  (16): nn.SpatialFullConvolution(128 -> 64, 3x3, 2, 2, 1, 1, 1, 1)
  (17): nn.InstanceNormalization
  (18): nn.ReLU
  (19): nn.SpatialFullConvolution(64 -> 32, 3x3, 2, 2, 1, 1, 1, 1)
  (20): nn.InstanceNormalization
  (21): nn.ReLU
  (22): nn.SpatialReplicationPadding(1, 1, 1, 1)
  (23): nn.SpatialConvolution(32 -> 3, 3x3)
}

the model is loaded perfectly fine but when i am trying to evaluate model , i get this error

 Traceback (most recent call last):
  File "convert-fast-neural-style.py", line 176, in <module>
    main()
  File "convert-fast-neural-style.py", line 162, in main
    unknown_layer_converter_fn=convert_instance_norm
  File "/usr/local/lib/python2.7/dist-packages/torch2coreml/_torch_converter.py", line 194, in convert
    print (model.evaluate())
  File "/usr/local/lib/python2.7/dist-packages/torch/legacy/nn/Container.py", line 39, in evaluate
    self.applyToModules(lambda m: m.evaluate())
  File "/usr/local/lib/python2.7/dist-packages/torch/legacy/nn/Container.py", line 26, in applyToModules
    func(module)
  File "/usr/local/lib/python2.7/dist-packages/torch/legacy/nn/Container.py", line 39, in <lambda>
    self.applyToModules(lambda m: m.evaluate())
TypeError: 'NoneType' object is not callable

what is the main reason behind the evaluation to bring this NoneType error. is their any other may that is equivalent to m.evaluate()

for reference also this is the Model i am trying to evaluate

[DmitryUlyanov]

The file convert-fast-neural-style.py does not belong to this repo, so I cannot help you...

[engahmed1190]

i can provided for you but the part i am having an error with is is this this is the load function

def load_torch_model(path):
    model = load_lua(path, unknown_classes=True)
    replace_module(
        model,
        lambda m: isinstance(m, TorchObject) and
        m.torch_typename() == 'nn.InstanceNormalization',
        create_instance_norm
    )
    replace_module(
        model,
        lambda m: isinstance(m, SpatialFullConvolution),
        fix_full_conv
    )
    return model

this is the create

def convert_instance_norm(builder, name, layer, input_names, output_names):
    if not isinstance(layer, InstanceNormalization):
        raise TypeError('Unsupported type {}'.format(layer,))

    epsilon = layer.eps
    weight = layer.weight.numpy()
    bias = None
    if layer.bias is not None:
        bias = layer.bias.numpy()

    builder.add_batchnorm(
        name=name,
        channels=weight.shape[0],
        gamma=weight,
        beta=bias,
        compute_mean_var=True,
        instance_normalization=True,
        input_name=input_names[0],
        output_name=output_names[0],
        epsilon=epsilon
    )

    return output_names

this is the full fix conv

def fix_full_conv(m):
    m.finput = None
    m.fgradInput = None
    return m

when i use function evaluate i got the previous error , my final aim is to transfer this model to coreml can you help ?

[DmitryUlyanov]

It seems you need to have instance norm defined in nn/Legacy in pytorch. I haven't ever tried to load torch models to pytorch, so I don't know how to do it.