I’ve trained a RaResNet50 from here: robust principles. I used this command: make experiments/RaResNet50/.done_train and then I would want to check the robustness by evaluating on CIFAR-10.

So I did this:

import foolbox as fb
import torch
import torch.nn as nn
import numpy as np
import os
import hydra

from robustarch.models.model import NormalizedConfigurableModel

MODEL_PATH = '/home/name/robust-principles/trained_models/RaResNet50_fat_train_eps5/phase3/model_best.pt'

def run_attack():

    mean = [0.485, 0.456, 0.406]
    std = [0.229, 0.224, 0.225]
    kwargs = {
        'stage_widths': [288, 576, 1120, 2160],
        'kernel': 3,
        'strides': [2, 2, 2, 2],
        'dilation': 1,
        'norm_layer': [
            hydra.utils.get_class('torch.nn.Identity'),
            hydra.utils.get_class('torch.nn.BatchNorm2d'),
            hydra.utils.get_class('torch.nn.BatchNorm2d')
        ],
        'activation_layer': [
            hydra.utils.get_class('torch.nn.SiLU'),
            hydra.utils.get_class('torch.nn.SiLU'),
            hydra.utils.get_class('torch.nn.SiLU')
        ],
        'group_widths': [36, 72, 140, 270],
        'bottleneck_multipliers': [0.25, 0.25, 0.25, 0.25],
        'downsample_norm': hydra.utils.get_class('torch.nn.BatchNorm2d'),
        'depths': [5, 8, 13, 1],
        'dense_ratio': None,
        'stem_type': hydra.utils.get_class('robustarch.models.model.Stem'),
        'stem_width': 96,
        'stem_kernel': 7,
        'stem_downsample_factor': 2,
        'stem_patch_size': None,
        'block_constructor': hydra.utils.get_class('robustarch.models.model.BottleneckBlock'),
        'ConvBlock': hydra.utils.get_class('robustarch.models.model.Conv2dNormActivation'),
        'se_ratio': 0.25,
        'se_activation': hydra.utils.get_class('torch.nn.ReLU'),
        'weight_init_type': 'resnet',
        'num_classes': 1000
    }
    # I got these configurations from printing them in the NormalizedConfigurableModel class when I started my training first

    model = NormalizedConfigurableModel(
        mean=mean,
        std=std,
        **kwargs
    )

    checkpoint = torch.load(MODEL_PATH)

    model.load_state_dict(checkpoint)
...

But I got this error:

File "/home/name/anaconda3/envs/ra-principles/lib/python3.9/site-packages/torch/nn/modules/module.py", line 2189, in load_state_dict
    raise RuntimeError('Error(s) in loading state_dict for {}:nt{}'.format(
RuntimeError: Error(s) in loading state_dict for NormalizedConfigurableModel:
        Missing key(s) in state_dict: "stem.stem.stem.0.weight", "stem.stem.stem.1.weight", "stem.stem.stem.1.bias",.....

it appears in this line of my code

model.load_state_dict(checkpoint)

The error message is a very big list of missing keys. What did I do wrong?