Train-accuracy =1.000000

[Agoniii]

Environment info

Operating System: centos 7.1
MXNet version: v0.9
Python versio: v2.7

Error Message:

When I train with the googlenet, the Train_accuracy =1.000000 all the time. anyone can help?

INFO:root:Epoch[4] Batch [20] Speed: 436.48 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [40] Speed: 418.70 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [60] Speed: 419.75 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [80] Speed: 418.40 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [100] Speed: 421.43 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [120] Speed: 419.66 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [140] Speed: 415.42 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [160] Speed: 415.30 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [180] Speed: 416.12 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [200] Speed: 420.10 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [220] Speed: 416.36 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [240] Speed: 416.22 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [260] Speed: 418.99 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [280] Speed: 415.44 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [300] Speed: 414.69 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [320] Speed: 415.61 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [340] Speed: 417.83 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [360] Speed: 412.90 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [380] Speed: 417.75 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [400] Speed: 416.55 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [420] Speed: 416.80 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [440] Speed: 415.29 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [460] Speed: 415.40 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [480] Speed: 415.84 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [500] Speed: 417.56 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [520] Speed: 424.66 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [540] Speed: 416.23 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [560] Speed: 414.90 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [580] Speed: 423.35 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [600] Speed: 408.58 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [620] Speed: 418.28 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [640] Speed: 424.21 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [660] Speed: 418.61 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [680] Speed: 412.77 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [700] Speed: 415.54 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [720] Speed: 416.24 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Batch [740] Speed: 415.85 samples/sec Train-accuracy=1.000000
INFO:root:Epoch[4] Resetting Data Iterator
INFO:root:Epoch[4] Time cost=113.786

Steps to reproduce

1. The data is ILSVRC-2012 50000 .jpg images.
2. prepare the .lst files

python tools/im2rec.py --list True --recursive True --train-ratio 0.8 example/image-classification/data/image-50000 ../imagenet_images/JPEG50000

3. generate the rec file

python tools/im2rec.py --resize 227 --quality 95 --num-thread 16  example/image-classification/data/image-50000 ../imagenet_images/JPEG50000

4. train_image5w.py is modified based on the train_imagenet.py

import os
import argparse
import logging
logging.basicConfig(level=logging.DEBUG)
from common import find_mxnet, data, fit
import mxnet as mx

if __name__ == '__main__':
    # parse args
    parser = argparse.ArgumentParser(description="train cifar10",
                                     formatter_class=argparse.ArgumentDefaultsHelpFormatter)
    fit.add_fit_args(parser)
    data.add_data_args(parser)
    data.add_data_aug_args(parser)
    # use a large aug level
    data.set_data_aug_level(parser, 3)
    parser.set_defaults(
        # network
        network          = 'resnet',
        num_layers       = 50,
        # data
        num_classes      = 1000,
        num_examples     = 50000,
        #num_examples     = 1281167,
        image_shape      = '3,224,224',
        min_random_scale = 1.143, # if input image has min size k, suggest to use
                              # 256.0/x, e.g. 0.533 for 480
        # train
        num_epochs       = 80,
        lr_step_epochs   = '30,60',
    )
    args = parser.parse_args()
    # load network
    from importlib import import_module
    net = import_module('symbol.'+args.network)
    sym = net.get_symbol(**vars(args))
    # train
    fit.fit(args, sym, data.get_rec_iter)

5. train

python train_image5w.py \
 --batch-size 64 --lr 0.05 --lr-factor .94 --gpus 0,1,2,3 --num-epoch 60 --network googlenet \
 --data-train data/image-50000_train.rec --data-val  data/image-50000_val.rec

[Godricly]

I guess your input image intensity is now in range [0, 255]. Try rescaling the input image to [-1.0, 1.0].

[Agoniii]

@Godricly Thanks for your reply. Could you tell me how to the input image to [-1.0, 1.0] ?

[Agoniii]

And my .lst file is as follows, the second argument is 0.000000. Does it affect the result?

15695 0.000000 ILSVRC2012_val_00019427.JPEG
35606 0.000000 ILSVRC2012_val_00038662.JPEG
28510 0.000000 ILSVRC2012_val_00022692.JPEG
33382 0.000000 ILSVRC2012_val_00002164.JPEG
19970 0.000000 ILSVRC2012_val_00006132.JPEG
10581 0.000000 ILSVRC2012_val_00004564.JPEG
7239 0.000000 ILSVRC2012_val_00033902.JPEG
4040 0.000000 ILSVRC2012_val_00021363.JPEG
25354 0.000000 ILSVRC2012_val_00022234.JPEG
5724 0.000000 ILSVRC2012_val_00001899.JPEG
43001 0.000000 ILSVRC2012_val_00027081.JPEG
4472 0.000000 ILSVRC2012_val_00042036.JPEG
22487 0.000000 ILSVRC2012_val_00044006.JPEG
43588 0.000000 ILSVRC2012_val_00023252.JPEG
14546 0.000000 ILSVRC2012_val_00024613.JPEG
35123 0.000000 ILSVRC2012_val_00030380.JPEG
14105 0.000000 ILSVRC2012_val_00028619.JPEG
6570 0.000000 ILSVRC2012_val_00031830.JPEG
4058 0.000000 ILSVRC2012_val_00028800.JPEG
24310 0.000000 ILSVRC2012_val_00023230.JPEG
8333 0.000000 ILSVRC2012_val_00042119.JPEG
15673 0.000000 ILSVRC2012_val_00037226.JPEG
31441 0.000000 ILSVRC2012_val_00027397.JPEG
13310 0.000000 ILSVRC2012_val_00009035.JPEG
13417 0.000000 ILSVRC2012_val_00039050.JPEG
34483 0.000000 ILSVRC2012_val_00006357.JPEG
27787 0.000000 ILSVRC2012_val_00033303.JPEG
24195 0.000000 ILSVRC2012_val_00027245.JPEG
48802 0.000000 ILSVRC2012_val_00042124.JPEG
16379 0.000000 ILSVRC2012_val_00019032.JPEG
42437 0.000000 ILSVRC2012_val_00035997.JPEG
......

[Godricly]

There is a scale parameter in mxnet.io.ImageRecordIter which u can rescale the intensity. you need to add this parameter to this file if you are running the example. Also remember to subtract the mean before doing so.
I think the label should be ok.

[phunterlau]

This issue is closed due to lack of activity in the last 90 days. Feel free to reopen if this is still an active issue. Thanks!