Number of parameters?

[underfitting]

Has anyone compared number of parameters?
I tried some good accuracy models, but they are very slow.

In terms of both accuracy and speed, which model is the best one to choose?

[psavine42]

Made the following function awhile ago, and the resulting dictionary sitting around. Has number of parameters, and last layer names which I was using to pull apart the torchvision models to get feature outputs, as well as the input size to last layer (most is not verified), but the number of parameters should be straightforward.

def get_n_params(model):
    pp=0
    for p in list(model.parameters()):
        nn = 1
        for s in list(p.size()):
            nn = nn*s
        pp += nn
    return pp

def get_stats():
    res = collections.defaultdict(dict)
    for x in pretrainedmodels.model_names:
        try:
            m = pm.__dict__[x](num_classes=1000, pretrained='imagenet')
            num_params = get_n_params(m)
            name_last, layer = list(m._modules.items())[-1]
            res[x]['n_params'] = num_params
            res[x]['feature_l_name'] = name_last
            res[x]['model_out_size'] = layer.in_features
        except:
            print('err', x)
    return res

_models = {
    'squeezenet1_0': { 'n_params': 1248424, 'feature_l_name': 'last_conv'},
    'squeezenet1_1': {'n_params': 1235496, 'feature_l_name': 'last_conv'},

    'vgg11_bn': { 'n_params': 123642856, 'feature_l_name': 'last_linear', 'model_out_size': 4096},
    'vgg13_bn': { 'n_params': 123642856, 'feature_l_name': 'last_linear', 'model_out_size': 4096},
    'vgg16_bn': { 'n_params': 123642856, 'feature_l_name': 'last_linear', 'model_out_size': 4096},
    'vgg19_bn': { 'n_params': 123642856, 'feature_l_name': 'last_linear', 'model_out_size': 4096},

    'vgg11': { 'n_params': 132863336, 'feature_l_name': 'last_linear', 'model_out_size': 4096},
    'vgg13': {'n_params': 123642856, 'feature_l_name': 'last_linear', 'model_out_size': 4096},
    'vgg16': { 'n_params': 123642856, 'feature_l_name': 'last_linear', 'model_out_size': 4096},
    'vgg19': { 'n_params': 123642856, 'feature_l_name': 'last_linear', 'model_out_size': 4096},

    'densenet121': { 'n_params': 7978856, 'feature_l_name': 'last_linear', 'model_out_size': 1024},
    'densenet161': {'n_params': 28681000, 'feature_l_name': 'last_linear', 'model_out_size': 2208},
    'densenet169':{ 'n_params': 14149480, 'feature_l_name': 'last_linear', 'model_out_size': 1664},
    'densenet201': { 'n_params': 20013928, 'feature_l_name': 'last_linear', 'model_out_size': 1920},


    'inceptionresnetv2': { 'n_params': 55843464, 'feature_l_name': 'last_linear', 'model_out_size': 1536},
    'inceptionv3': { 'n_params': 27161264, 'feature_l_name': 'last_linear', 'model_out_size': 2048},
    'inceptionv4': { 'n_params': 42679816, 'feature_l_name': 'last_linear', 'model_out_size': 1536},
    'bninception': {'n_params': 11295240, 'feature_l_name': 'last_linear', 'model_out_size': 1024},
    'xception': { 'n_params': 22855952, 'feature_l_name': 'last_linear', 'model_out_size': 2048},

    'dpn68': { 'n_params': 12611602, 'feature_l_name': 'classifier'},
    'dpn98': { 'n_params': 61570728, 'feature_l_name': 'classifier'},
    'dpn131': {'n_params': 79254504, 'feature_l_name': 'classifier'},

    'alexnet': { 'n_params': 61100840, 'feature_l_name': 'last_linear', 'model_out_size': 4096},

    'resnet18': { 'n_params': 11689512, 'feature_l_name': 'last_linear', 'model_out_size': 512},
    'resnet34': {'n_params': 21797672, 'feature_l_name': 'last_linear', 'model_out_size': 512},
    'resnet50': {'n_params': 25557032, 'feature_l_name': 'last_linear', 'model_out_size': 2048},
    'resnet101': { 'n_params': 44549160, 'feature_l_name': 'last_linear', 'model_out_size': 2048},
    'resnet152': { 'n_params': 60192808, 'feature_l_name': 'last_linear', 'model_out_size': 2048},

    'resnext101_32x4d': { 'n_params': 44177704, 'feature_l_name': 'last_linear', 'model_out_size': 2048},
    'resnext101_64x4d': { 'n_params': 83455272, 'feature_l_name': 'last_linear', 'model_out_size': 2048},
    'senet154': { 'n_params': 115088984, 'feature_l_name': 'last_linear', 'model_out_size': 2048},
    'fbresnet152': {'n_params': 60268520, 'feature_l_name': 'last_linear', 'model_out_size': 2048},

    'se_resnet50': { 'n_params': 28088024, 'feature_l_name': 'last_linear', 'model_out_size': 2048},
    'se_resnet101': { 'n_params': 49326872, 'feature_l_name': 'last_linear', 'model_out_size': 2048},

    'se_resnet152': { 'n_params': 66821848, 'feature_l_name': 'last_linear', 'model_out_size': 2048},
    'se_resnext50_32x4d': { 'n_params': 27559896, 'feature_l_name': 'last_linear',  'model_out_size': 2048},
    'se_resnext101_32x4d': {'n_params': 48955416, 'feature_l_name': 'last_linear', 'model_out_size': 2048},

    'nasnetalarge': {'n_params': 88753150, 'feature_l_name': 'last_linear', 'model_out_size': 4032},
    'nasnetamobile': {'n_params': 5289978, 'feature_l_name': 'last_linear', 'model_out_size': 1056},
}

[underfitting]

Thank you @psavine42 ! It's so useful.
I simply calculated score = (acc@1)/(log10(n_params)) and top5 score models are below:

model name | n_params | acc@1
-------------------------------
nasnetamobile | 5289978 | 74.1
densenet121 | 7978856 | 74.6
xception | 22855952 | 78.9
dpn68 | 12611602 | 75.9
se_resnext50_32x4d | 27559896 | 79.1