absalama opened a new issue #11393: Validation Accuracy is higher than training 
accuracy. 
URL: https://github.com/apache/incubator-mxnet/issues/11393
 
 
   I am training Imagenet (1k) on alexnet. I used the im2rec tool to split 5% 
of the training data to be used by the validation phase. The results are two 
set of records files (I use chunks) one set for training and on set for 
validation. 
   
   The log shows the following: 
   
   ```
   top_k_accuracy_5=0.159258    cross-entropy=5.465672
   INFO:root:Epoch[0] Batch [9400]      Speed: 395.94 samples/sec       
accuracy=0.052461       top_k_accuracy_5=0.160195       cross-entropy=5.454822
   **INFO:root:Epoch[0] Train-accuracy=0.056324
   INFO:root:Epoch[0] Train-top_k_accuracy_5=0.165848
   INFO:root:Epoch[0] Train-cross-entropy=5.416635
   INFO:root:Epoch[0] Time cost=3079.019
   INFO:root:Saved checkpoint to 
"mxnet_alexnet_single_gpu_all_data_set_256-0001.params"
   INFO:root:Epoch[0] `Validation-accuracy=0.078869
   INFO:root:Epoch[0] Validation-top_k_accuracy_5=0.216859
   INFO:root:Epoch[0] Validation-cross-entropy=5.142231**
   ```
   
   The validation here is higher than training accuracy and this increases with 
further epochs (Until the time I write this issue the epoch 10 , the the 
validation is higher than the training accuracy with around 7%). 
   
   **The commands used for data preprocessing:**
   `python3 im2rec.py --list --recursive  --chunks 1024 --train-ratio 0.95  
${IMAGENET_ROOT}/record_io_all_raw_data/metadata-train256/imagenet1k 
${IMAGENET_EXTRACTED}/train  
   `
   `python3 im2rec.py  --resize 256 --quality 95 --num-thread 16  
${IMAGENET_ROOT}/record_io_all_raw_data/metadata-train256/imagenet1k 
${IMAGENET_EXTRACTED}/train`
   
   python3 im2rec.py --resize 256 --quality 95 --num-thread 16  
`${IMAGENET_ROOT}/record_io_all_raw_data/metadata-val256/imagenet1k 
${IMAGENET_EXTRACTED}/train`
   
   **The arguments used for training:** 
   
   ```
    Namespace(batch_size=128, benchmark=0, data_nthreads=4, 
data_train='/work/projects/Project00755/datasets/imagenet/record_io_all_raw_data/train256/',
 data_train_
   idx='', 
data_val='/work/projects/Project00755/datasets/imagenet/record_io_all_raw_data/val256/',
 data_val_idx='', disp_batches=200, dtype='float32', gc_threshold=0.5, 
gc_type='none', gpus='0'
   , image_shape='3,227,227', initializer='default', kv_store='device', 
load_epoch=None, loss='ce', lr=0.01, lr_factor=0.1, lr_step_epochs='30,60', 
macrobatch_size=0, max_random_aspect_ratio=0.2
   5, max_random_h=36, max_random_l=50, max_random_rotate_angle=10, 
max_random_s=50, max_random_scale=1, max_random_shear_ratio=0.1, 
min_random_scale=1, model_prefix='mxnet_alexnet_single_gpu_al
   l_data_set_256', mom=0.9, monitor=0, network='alexnet', num_classes=1000, 
num_epochs=80, num_examples=1216718, num_layers=8, optimizer='sgd', pad_size=0, 
random_crop=1, random_mirror=1, rgb_m
   ean='123.68,116.779,103.939', save_period=1, test_io=0, top_k=5, 
warmup_epochs=5, warmup_strategy='linear', wd=0.0005)
   ```
   
   Any help will be appreciated. 
   Thanks 
   
   
   

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
[email protected]


With regards,
Apache Git Services

Reply via email to