Pretrained Resnet 34 "KeyError: '0.1.weight'"

[I-C-Karakozis]

Describe the bug I changed my ResNet code to have more flexibility in what I am trying to do and the converter no longer works (see below for more information and relevant past Issue posted). Any ideas for why this is the case?

To Reproduce

import math
import torch
import torchvision
import torch.nn as nn
import torch.nn.functional as F
import torch.backends.cudnn as cudnn

def initLinear(linear, val=None):
    if val is None:
        fan = linear.in_features + linear.out_features
        spread = math.sqrt(2.0) * math.sqrt(2.0/fan)
    else:
        spread = val
    linear.weight.data.uniform_(-spread, spread)
    linear.bias.data.uniform_(-spread, spread)
    return

class ResNet34_Pretrained(nn.Module):
    def base_size(self): return 512
    def rep_size(self): return 1024

    def __init__(self, num_classes, pretrained=True):
        super(ResNet34_Pretrained, self).__init__()
        self.resnet = torchvision.models.resnet34(pretrained=pretrained)

        # define layers
        self.num_classes = num_classes
        self.linear1 = nn.Linear(7*7*self.base_size(), self.rep_size())
        self.linear2 = nn.Linear(self.rep_size(), self.num_classes)
        self.dropout2d = nn.Dropout2d(.5)
        self.dropout = nn.Dropout(.5)
        self.relu = nn.LeakyReLU()
        initLinear(self.linear1)
        initLinear(self.linear2)

    def forward(self, out0):
        x = self.resnet.conv1(out0)
        x = self.resnet.bn1(x)
        x = self.resnet.relu(x)
        out1 = self.resnet.maxpool(x)

        out2 = self.resnet.layer1(out1)
        out3 = self.resnet.layer2(out2)
        out4 = self.resnet.layer3(out3)
        out5 = self.resnet.layer4(out4)

        x = self.dropout2d(out5)

        features = self.dropout(
                    self.relu(self.linear1(x.view(-1, 7*7*self.base_size())))
        )
        out = self.linear2(features)
        return out

def load_resnet34(n_classes):
    # load network architecture
    net = ResNet34(n_classes)

    # determine processing device
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    net = net.to(device)
    if device == 'cuda':
        # net = torch.nn.DataParallel(net)
        cudnn.benchmark = True

    return net, device

Logs The error message:

graph node: ResNet34_Pretrained/Conv2d node id: 233 type: onnx::Conv inputs: ['0', '1'] outputs: ['ResNet34_Pretrained/Conv2d'] name in state_dict: 0.1 attrs: {'dilations': [1, 1], 'group': 1, 'kernel_shape': [7, 7], 'pads': [3, 3, 3, 3], 'strides': [2, 2]} is_terminal: False Converting convolution ... Traceback (most recent call last): File "CEM.py", line 151, in main(args) File "CEM.py", line 84, in main model = pytorch_to_keras(model, orig_img, [batch_shape[1:]], names='short', verbose=True) File "/mnt/c/Users/Yannis/Desktop/Fairness_Blackout/venv/lib/python3.6/site-packages/pytorch2keras/converter.py", line 315, in pytorch_to_keras names File "/mnt/c/Users/Yannis/Desktop/Fairness_Blackout/venv/lib/python3.6/site-packages/pytorch2keras/convolution_layers.py", line 35, in convert_conv if len(weights[weights_name].numpy().shape) == 5: # 3D conv KeyError: '0.1.weight'

Environment (please complete the following information):

• Ubuntu for Windows
• Python 3.6.7
• Pytorch 1.0.1.post2
• pytorch2keras 0.1.14

Additional context Last bug report by me: https://github.com/nerox8664/pytorch2keras/issues/50 I was using a similar model.

[I-C-Karakozis]

Fixed it. Here is the solution for anyone running into the same problem (it was very similar to my previous error):

def initLinear(linear, val=None):
    if val is None:
        fan = linear.in_features + linear.out_features
        spread = math.sqrt(2.0) * math.sqrt(2.0/fan)
    else:
        spread = val
    linear.weight.data.uniform_(-spread, spread)
    linear.bias.data.uniform_(-spread, spread)
    return

class ResNet34_Pretrained(nn.Module):
    def base_size(self): return 512
    def rep_size(self): return 1024

    def __init__(self, n_classes):
        super(ResNet34_Pretrained, self).__init__()
        self.resnet = torchvision.models.resnet34(pretrained=True)
        self.conv1 = self.resnet.conv1
        self.bn1 = self.resnet.bn1
        self.layer1 = self.resnet.layer1
        self.layer2 = self.resnet.layer2
        self.layer3 = self.resnet.layer3
        self.layer4 = self.resnet.layer4

        # define layers
        self.n_classes = n_classes
        self.linear1 = nn.Linear(7 * 7 * self.base_size(), self.rep_size())
        self.linear2 = nn.Linear(self.rep_size(), self.n_classes)
        self.dropout2d = nn.Dropout2d(.5)
        self.dropout = nn.Dropout(.5)
        self.relu = nn.LeakyReLU()

        # initialize linear layers
        initLinear(self.linear1)
        initLinear(self.linear2)

    def forward(self, out0):
        x = self.conv1(out0)
        x = self.bn1(x)
        x = self.resnet.relu(x)
        x = self.resnet.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = self.dropout2d(x)

        x = self.relu(self.linear1(x.view(-1, 7*7*self.base_size())))
        x = self.dropout(x.clone())
        cls_scores = self.linear2(x.clone())

        return cls_scores

def load_resnet34(n_classes, pretrained=False):
    # load network architecture
    if pretrained: net = ResNet34_Pretrained(n_classes)
    else: net = ResNet34(n_classes)

    # determine processing device
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    net = net.to(device)
    if device == 'cuda':
        # net = torch.nn.DataParallel(net)
        cudnn.benchmark = True

    return net, device

I also had to revert to pytorch2keras 0.1.14 and torch 0.4.1.