Cannot reproduce ResNet-18 listed in the table

[lychenyoko]

Hi,

I have tried to reproduce the result of ResNet-18 in your table which is:

Model	Version	Acc@1	Acc@5
ResNet18	Pytorch	70.142	89.274

However, I only get Acc@1 69.758 Acc@5 89.078

Here is my running results:

(base) [yl16@tigergpu pretrained-models.pytorch]$ python examples/imagenet_eval.py --data=/scratch/gpfs/yl16/ImageNet/Val_Img_Grouped/ -a resnet18 -b 200 -e
=> creating model 'resnet18'
=> using pre-trained parameters 'imagenet'
Images transformed from size 256 to [3, 224, 224]
Test: [0/250]   Time 26.461 (26.461)    Loss 0.6886 (0.6886)    Acc@1 79.000 (79.000)   Acc@5 96.000 (96.000)
Test: [10/250]  Time 5.830 (8.772)      Loss 0.9663 (0.7915)    Acc@1 73.000 (79.864)   Acc@5 92.500 (93.773)
Test: [20/250]  Time 2.226 (5.632)      Loss 0.5328 (0.9838)    Acc@1 85.500 (74.786)   Acc@5 94.500 (92.024)
Test: [30/250]  Time 12.578 (5.118)     Loss 0.9938 (0.9275)    Acc@1 72.500 (76.468)   Acc@5 90.000 (92.435)
Test: [40/250]  Time 2.317 (4.662)      Loss 1.0534 (0.8767)    Acc@1 77.000 (77.780)   Acc@5 91.000 (92.841)
Test: [50/250]  Time 11.165 (4.678)     Loss 0.9022 (0.9147)    Acc@1 69.500 (76.275)   Acc@5 95.000 (92.941)
Test: [60/250]  Time 2.194 (4.377)      Loss 1.1613 (0.9216)    Acc@1 59.500 (75.754)   Acc@5 97.000 (93.107)
Test: [70/250]  Time 11.250 (4.332)     Loss 1.2558 (0.9183)    Acc@1 59.500 (75.500)   Acc@5 92.500 (93.324)
Test: [80/250]  Time 2.183 (4.153)      Loss 0.3299 (0.9159)    Acc@1 93.000 (75.852)   Acc@5 98.000 (93.426)
Test: [90/250]  Time 9.864 (4.200)      Loss 0.6746 (0.9023)    Acc@1 85.000 (76.297)   Acc@5 94.000 (93.560)
Test: [100/250] Time 2.289 (4.058)      Loss 1.5319 (0.9136)    Acc@1 62.500 (76.178)   Acc@5 85.500 (93.396)
Test: [110/250] Time 8.061 (4.071)      Loss 1.4373 (0.9570)    Acc@1 62.000 (75.306)   Acc@5 88.000 (92.869)
Test: [120/250] Time 2.190 (3.967)      Loss 1.6624 (1.0095)    Acc@1 52.000 (74.335)   Acc@5 85.000 (92.165)
Test: [130/250] Time 6.493 (3.935)      Loss 1.3204 (1.0569)    Acc@1 70.500 (73.389)   Acc@5 89.000 (91.603)
Test: [140/250] Time 2.079 (3.855)      Loss 0.7964 (1.0799)    Acc@1 79.000 (72.940)   Acc@5 92.000 (91.259)
Test: [150/250] Time 4.494 (3.834)      Loss 1.9821 (1.1038)    Acc@1 55.500 (72.550)   Acc@5 78.500 (90.911)
Test: [160/250] Time 2.142 (3.772)      Loss 1.2508 (1.1303)    Acc@1 69.000 (72.000)   Acc@5 89.000 (90.537)
Test: [170/250] Time 4.506 (3.700)      Loss 1.5955 (1.1557)    Acc@1 61.500 (71.512)   Acc@5 85.500 (90.249)
Test: [180/250] Time 2.332 (3.669)      Loss 1.3661 (1.1676)    Acc@1 71.000 (71.298)   Acc@5 83.000 (90.122)
Test: [190/250] Time 5.764 (3.658)      Loss 1.6772 (1.1897)    Acc@1 65.500 (70.853)   Acc@5 84.000 (89.804)
Test: [200/250] Time 2.190 (3.668)      Loss 0.5099 (1.2041)    Acc@1 86.500 (70.669)   Acc@5 96.500 (89.597)
Test: [210/250] Time 2.213 (3.630)      Loss 1.7143 (1.2291)    Acc@1 61.000 (70.159)   Acc@5 80.000 (89.289)
Test: [220/250] Time 2.336 (3.598)      Loss 1.5360 (1.2413)    Acc@1 62.500 (69.876)   Acc@5 85.000 (89.113)
Test: [230/250] Time 7.709 (3.583)      Loss 1.6016 (1.2564)    Acc@1 63.000 (69.619)   Acc@5 89.000 (88.939)
Test: [240/250] Time 2.533 (3.575)      Loss 2.0250 (1.2488)    Acc@1 51.500 (69.739)   Acc@5 78.500 (89.025)
 * Acc@1 69.758 Acc@5 89.078

[Eric-qi]

I get the same results in ResNet18, and can not achieve the performance listed in table. Have you solved this problem?

[Fzz123]

您的邮件已收到!