georgeretsi/pytorch-phocnet

PHOCNet implementation in Pytorch based on Sudholt's implementation

pytorch-phocnet's Issues

Where can we find 'almazan_cv_indices.npy'?

implementation result wrong

i tried to run your program in google colab withouth changing the parameter. and changing some of the lines but mostly regarding display like parser info, etc.
But the result when using gw datasets far different than the result in journal "PHOCNet: A Deep Convolutional Neural Network for Word Spotting in Handwritten Documents". so, where i do wrong?

the result from journal

the result from me running your program
default='60000:1e-4,100000:1e-5'
learning_rate_step = learning_rate_step_parser(default)
weight_decay = 0.00005
iter_size = 10
augmentation = True
batch_size = 1
phoc_unigram_levels=[1,2,4,8]
gpp_type='gpp',
pooling_levels=([1], [5])

[2018-11-14 10:06:15, INFO, PHOCNet-Experiment::train] --- Running PHOCNet Training ---
[2018-11-14 10:06:15, INFO, PHOCNet-Experiment::train] Loading dataset gw...
100%|¦¦¦¦¦¦¦¦¦¦| 4860/4860 [00:00<00:00, 7170.59it/s]
[2018-11-14 10:06:17, INFO, PHOCNet-Experiment::train] Preparing PHOCNet...
/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:87: UserWarning: nn.init.constant is now deprecated in favor of nn.init.constant_.
[2018-11-14 10:06:18, INFO, PHOCNet-Experiment::train] Training:
/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:146: UserWarning: invalid index of a 0-dim tensor. This will be an error in PyTorch 0.5. Use tensor.item() to convert a 0-dim tensor to a Python number
[2018-11-14 10:11:48, INFO, PHOCNet-Experiment::train] Iteration 500: 0.413078
[2018-11-14 10:17:15, INFO, PHOCNet-Experiment::train] Iteration 1000: 0.239131
[2018-11-14 10:22:40, INFO, PHOCNet-Experiment::train] Iteration 1500: 0.226796
[2018-11-14 10:28:08, INFO, PHOCNet-Experiment::train] Iteration 2000: 0.182755
[2018-11-14 10:28:08, INFO, PHOCNet-Experiment::train] Evaluating net after 2000 iterations
[2018-11-14 10:28:08, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.02it/s]
[2018-11-14 10:28:34, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 10:28:34, INFO, PHOCNet-Experiment::test] mAP: 10.78
[2018-11-14 10:34:01, INFO, PHOCNet-Experiment::train] Iteration 2500: 0.198502
[2018-11-14 10:39:30, INFO, PHOCNet-Experiment::train] Iteration 3000: 0.180421
[2018-11-14 10:44:53, INFO, PHOCNet-Experiment::train] Iteration 3500: 0.102496
[2018-11-14 10:50:19, INFO, PHOCNet-Experiment::train] Iteration 4000: 0.160037
[2018-11-14 10:50:19, INFO, PHOCNet-Experiment::train] Evaluating net after 4000 iterations
[2018-11-14 10:50:19, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 53.83it/s]
[2018-11-14 10:50:45, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 10:50:45, INFO, PHOCNet-Experiment::test] mAP: 8.96
[2018-11-14 10:56:11, INFO, PHOCNet-Experiment::train] Iteration 4500: 0.199652
[2018-11-14 11:01:36, INFO, PHOCNet-Experiment::train] Iteration 5000: 0.186553
[2018-11-14 11:07:04, INFO, PHOCNet-Experiment::train] Iteration 5500: 0.180660
[2018-11-14 11:12:28, INFO, PHOCNet-Experiment::train] Iteration 6000: 0.204796
[2018-11-14 11:12:28, INFO, PHOCNet-Experiment::train] Evaluating net after 6000 iterations
[2018-11-14 11:12:28, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:26<00:00, 46.68it/s]
[2018-11-14 11:12:55, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 11:12:55, INFO, PHOCNet-Experiment::test] mAP: 9.19
[2018-11-14 11:18:22, INFO, PHOCNet-Experiment::train] Iteration 6500: 0.204873
[2018-11-14 11:23:45, INFO, PHOCNet-Experiment::train] Iteration 7000: 0.113907
[2018-11-14 11:29:10, INFO, PHOCNet-Experiment::train] Iteration 7500: 0.105077
[2018-11-14 11:34:34, INFO, PHOCNet-Experiment::train] Iteration 8000: 0.178453
[2018-11-14 11:34:34, INFO, PHOCNet-Experiment::train] Evaluating net after 8000 iterations
[2018-11-14 11:34:34, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.21it/s]
[2018-11-14 11:34:59, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 11:35:00, INFO, PHOCNet-Experiment::test] mAP: 8.77
[2018-11-14 11:40:25, INFO, PHOCNet-Experiment::train] Iteration 8500: 0.166540
[2018-11-14 11:45:47, INFO, PHOCNet-Experiment::train] Iteration 9000: 0.149730
[2018-11-14 11:51:17, INFO, PHOCNet-Experiment::train] Iteration 9500: 0.176998
[2018-11-14 11:56:42, INFO, PHOCNet-Experiment::train] Iteration 10000: 0.158446
[2018-11-14 11:56:42, INFO, PHOCNet-Experiment::train] Evaluating net after 10000 iterations
[2018-11-14 11:56:42, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.36it/s]
[2018-11-14 11:57:07, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 11:57:08, INFO, PHOCNet-Experiment::test] mAP: 9.01
[2018-11-14 12:02:37, INFO, PHOCNet-Experiment::train] Iteration 10500: 0.090935
[2018-11-14 12:08:04, INFO, PHOCNet-Experiment::train] Iteration 11000: 0.170351
[2018-11-14 12:13:31, INFO, PHOCNet-Experiment::train] Iteration 11500: 0.181705
[2018-11-14 12:18:56, INFO, PHOCNet-Experiment::train] Iteration 12000: 0.103799
[2018-11-14 12:18:56, INFO, PHOCNet-Experiment::train] Evaluating net after 12000 iterations
[2018-11-14 12:18:56, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:26<00:00, 46.65it/s]
[2018-11-14 12:19:22, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 12:19:23, INFO, PHOCNet-Experiment::test] mAP: 9.06
[2018-11-14 12:24:47, INFO, PHOCNet-Experiment::train] Iteration 12500: 0.105887
[2018-11-14 12:30:10, INFO, PHOCNet-Experiment::train] Iteration 13000: 0.170190
[2018-11-14 12:35:37, INFO, PHOCNet-Experiment::train] Iteration 13500: 0.171008
[2018-11-14 12:41:01, INFO, PHOCNet-Experiment::train] Iteration 14000: 0.169650
[2018-11-14 12:41:01, INFO, PHOCNet-Experiment::train] Evaluating net after 14000 iterations
[2018-11-14 12:41:01, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.56it/s]
[2018-11-14 12:41:27, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 12:41:27, INFO, PHOCNet-Experiment::test] mAP: 9.23
[2018-11-14 12:46:52, INFO, PHOCNet-Experiment::train] Iteration 14500: 0.164980
[2018-11-14 12:52:13, INFO, PHOCNet-Experiment::train] Iteration 15000: 0.171670
[2018-11-14 12:57:41, INFO, PHOCNet-Experiment::train] Iteration 15500: 0.203490
[2018-11-14 13:03:04, INFO, PHOCNet-Experiment::train] Iteration 16000: 0.158424
[2018-11-14 13:03:04, INFO, PHOCNet-Experiment::train] Evaluating net after 16000 iterations
[2018-11-14 13:03:04, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.69it/s]
[2018-11-14 13:03:30, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 13:03:31, INFO, PHOCNet-Experiment::test] mAP: 9.46
[2018-11-14 13:08:55, INFO, PHOCNet-Experiment::train] Iteration 16500: 0.171116
[2018-11-14 13:14:18, INFO, PHOCNet-Experiment::train] Iteration 17000: 0.201872
[2018-11-14 13:19:44, INFO, PHOCNet-Experiment::train] Iteration 17500: 0.178268
[2018-11-14 13:25:15, INFO, PHOCNet-Experiment::train] Iteration 18000: 0.156980
[2018-11-14 13:25:15, INFO, PHOCNet-Experiment::train] Evaluating net after 18000 iterations
[2018-11-14 13:25:15, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.60it/s]
[2018-11-14 13:25:41, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 13:25:41, INFO, PHOCNet-Experiment::test] mAP: 9.87
[2018-11-14 13:31:09, INFO, PHOCNet-Experiment::train] Iteration 18500: 0.141147
[2018-11-14 13:36:36, INFO, PHOCNet-Experiment::train] Iteration 19000: 0.159306
[2018-11-14 13:42:02, INFO, PHOCNet-Experiment::train] Iteration 19500: 0.158914
[2018-11-14 13:47:31, INFO, PHOCNet-Experiment::train] Iteration 20000: 0.127444
[2018-11-14 13:47:31, INFO, PHOCNet-Experiment::train] Evaluating net after 20000 iterations
[2018-11-14 13:47:31, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.02it/s]
[2018-11-14 13:47:57, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 13:47:58, INFO, PHOCNet-Experiment::test] mAP: 10.03
[2018-11-14 13:53:23, INFO, PHOCNet-Experiment::train] Iteration 20500: 0.180828
[2018-11-14 13:58:49, INFO, PHOCNet-Experiment::train] Iteration 21000: 0.086803
[2018-11-14 14:04:13, INFO, PHOCNet-Experiment::train] Iteration 21500: 0.147529
[2018-11-14 14:09:33, INFO, PHOCNet-Experiment::train] Iteration 22000: 0.177185
[2018-11-14 14:09:33, INFO, PHOCNet-Experiment::train] Evaluating net after 22000 iterations
[2018-11-14 14:09:33, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.12it/s]
[2018-11-14 14:09:59, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 14:10:00, INFO, PHOCNet-Experiment::test] mAP: 10.26
[2018-11-14 14:15:28, INFO, PHOCNet-Experiment::train] Iteration 22500: 0.139572
[2018-11-14 14:20:54, INFO, PHOCNet-Experiment::train] Iteration 23000: 0.139268
[2018-11-14 14:26:21, INFO, PHOCNet-Experiment::train] Iteration 23500: 0.157128
[2018-11-14 14:31:46, INFO, PHOCNet-Experiment::train] Iteration 24000: 0.173044
[2018-11-14 14:31:46, INFO, PHOCNet-Experiment::train] Evaluating net after 24000 iterations
[2018-11-14 14:31:46, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.48it/s]
[2018-11-14 14:32:11, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 14:32:12, INFO, PHOCNet-Experiment::test] mAP: 10.63
[2018-11-14 14:37:44, INFO, PHOCNet-Experiment::train] Iteration 24500: 0.234703
[2018-11-14 14:43:08, INFO, PHOCNet-Experiment::train] Iteration 25000: 0.178952
[2018-11-14 14:48:35, INFO, PHOCNet-Experiment::train] Iteration 25500: 0.075044
[2018-11-14 14:54:00, INFO, PHOCNet-Experiment::train] Iteration 26000: 0.161672
[2018-11-14 14:54:00, INFO, PHOCNet-Experiment::train] Evaluating net after 26000 iterations
[2018-11-14 14:54:00, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 46.82it/s]
[2018-11-14 14:54:26, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 14:54:27, INFO, PHOCNet-Experiment::test] mAP: 10.56
[2018-11-14 14:59:58, INFO, PHOCNet-Experiment::train] Iteration 26500: 0.140049
[2018-11-14 15:05:23, INFO, PHOCNet-Experiment::train] Iteration 27000: 0.249265
[2018-11-14 15:10:50, INFO, PHOCNet-Experiment::train] Iteration 27500: 0.215417
[2018-11-14 15:16:16, INFO, PHOCNet-Experiment::train] Iteration 28000: 0.169030
[2018-11-14 15:16:16, INFO, PHOCNet-Experiment::train] Evaluating net after 28000 iterations
[2018-11-14 15:16:16, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 46.92it/s]
[2018-11-14 15:16:42, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 15:16:43, INFO, PHOCNet-Experiment::test] mAP: 10.82
[2018-11-14 15:22:11, INFO, PHOCNet-Experiment::train] Iteration 28500: 0.075680
[2018-11-14 15:27:34, INFO, PHOCNet-Experiment::train] Iteration 29000: 0.168502
[2018-11-14 15:33:04, INFO, PHOCNet-Experiment::train] Iteration 29500: 0.074602
[2018-11-14 15:38:31, INFO, PHOCNet-Experiment::train] Iteration 30000: 0.072901
[2018-11-14 15:38:31, INFO, PHOCNet-Experiment::train] Evaluating net after 30000 iterations
[2018-11-14 15:38:31, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.01it/s]
[2018-11-14 15:38:57, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 15:38:58, INFO, PHOCNet-Experiment::test] mAP: 11.03
[2018-11-14 15:44:24, INFO, PHOCNet-Experiment::train] Iteration 30500: 0.176869
[2018-11-14 15:49:47, INFO, PHOCNet-Experiment::train] Iteration 31000: 0.153867
[2018-11-14 15:55:13, INFO, PHOCNet-Experiment::train] Iteration 31500: 0.158284
[2018-11-14 16:00:38, INFO, PHOCNet-Experiment::train] Iteration 32000: 0.145576
[2018-11-14 16:00:38, INFO, PHOCNet-Experiment::train] Evaluating net after 32000 iterations
[2018-11-14 16:00:38, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.06it/s]
[2018-11-14 16:01:04, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 16:01:05, INFO, PHOCNet-Experiment::test] mAP: 11.61
[2018-11-14 16:06:31, INFO, PHOCNet-Experiment::train] Iteration 32500: 0.080178
[2018-11-14 16:12:00, INFO, PHOCNet-Experiment::train] Iteration 33000: 0.163559
[2018-11-14 16:17:27, INFO, PHOCNet-Experiment::train] Iteration 33500: 0.267429
[2018-11-14 16:22:54, INFO, PHOCNet-Experiment::train] Iteration 34000: 0.144429
[2018-11-14 16:22:54, INFO, PHOCNet-Experiment::train] Evaluating net after 34000 iterations
[2018-11-14 16:22:54, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.18it/s]
[2018-11-14 16:23:20, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 16:23:20, INFO, PHOCNet-Experiment::test] mAP: 11.57
[2018-11-14 16:28:46, INFO, PHOCNet-Experiment::train] Iteration 34500: 0.069810
[2018-11-14 16:34:13, INFO, PHOCNet-Experiment::train] Iteration 35000: 0.166294
[2018-11-14 16:39:39, INFO, PHOCNet-Experiment::train] Iteration 35500: 0.151384
[2018-11-14 16:45:04, INFO, PHOCNet-Experiment::train] Iteration 36000: 0.212405
[2018-11-14 16:45:04, INFO, PHOCNet-Experiment::train] Evaluating net after 36000 iterations
[2018-11-14 16:45:04, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:26<00:00, 46.73it/s]
[2018-11-14 16:45:30, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 16:45:31, INFO, PHOCNet-Experiment::test] mAP: 11.60
[2018-11-14 16:50:56, INFO, PHOCNet-Experiment::train] Iteration 36500: 0.170672
[2018-11-14 16:56:21, INFO, PHOCNet-Experiment::train] Iteration 37000: 0.067887
[2018-11-14 17:01:48, INFO, PHOCNet-Experiment::train] Iteration 37500: 0.159329
[2018-11-14 17:07:13, INFO, PHOCNet-Experiment::train] Iteration 38000: 0.162292
[2018-11-14 17:07:13, INFO, PHOCNet-Experiment::train] Evaluating net after 38000 iterations
[2018-11-14 17:07:13, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.29it/s]
[2018-11-14 17:07:39, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 17:07:39, INFO, PHOCNet-Experiment::test] mAP: 12.00
[2018-11-14 17:13:05, INFO, PHOCNet-Experiment::train] Iteration 38500: 0.159215
[2018-11-14 17:18:29, INFO, PHOCNet-Experiment::train] Iteration 39000: 0.066484
[2018-11-14 17:24:00, INFO, PHOCNet-Experiment::train] Iteration 39500: 0.156148
[2018-11-14 17:29:28, INFO, PHOCNet-Experiment::train] Iteration 40000: 0.135394
[2018-11-14 17:29:28, INFO, PHOCNet-Experiment::train] Evaluating net after 40000 iterations
[2018-11-14 17:29:28, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 46.75it/s]
[2018-11-14 17:29:54, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 17:29:55, INFO, PHOCNet-Experiment::test] mAP: 12.20
[2018-11-14 17:35:19, INFO, PHOCNet-Experiment::train] Iteration 40500: 0.149086
[2018-11-14 17:40:46, INFO, PHOCNet-Experiment::train] Iteration 41000: 0.151285
[2018-11-14 17:46:09, INFO, PHOCNet-Experiment::train] Iteration 41500: 0.208820
[2018-11-14 17:51:34, INFO, PHOCNet-Experiment::train] Iteration 42000: 0.198136
[2018-11-14 17:51:34, INFO, PHOCNet-Experiment::train] Evaluating net after 42000 iterations
[2018-11-14 17:51:34, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 46.76it/s]
[2018-11-14 17:52:00, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 17:52:01, INFO, PHOCNet-Experiment::test] mAP: 12.45
[2018-11-14 17:57:26, INFO, PHOCNet-Experiment::train] Iteration 42500: 0.242783
[2018-11-14 18:02:51, INFO, PHOCNet-Experiment::train] Iteration 43000: 0.179466
[2018-11-14 18:08:14, INFO, PHOCNet-Experiment::train] Iteration 43500: 0.129665
[2018-11-14 18:13:39, INFO, PHOCNet-Experiment::train] Iteration 44000: 0.065423
[2018-11-14 18:13:39, INFO, PHOCNet-Experiment::train] Evaluating net after 44000 iterations
[2018-11-14 18:13:39, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 46.73it/s]
[2018-11-14 18:14:05, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 18:14:06, INFO, PHOCNet-Experiment::test] mAP: 12.30
[2018-11-14 18:19:27, INFO, PHOCNet-Experiment::train] Iteration 44500: 0.122817
[2018-11-14 18:24:51, INFO, PHOCNet-Experiment::train] Iteration 45000: 0.188579
[2018-11-14 18:30:21, INFO, PHOCNet-Experiment::train] Iteration 45500: 0.157796
[2018-11-14 18:35:43, INFO, PHOCNet-Experiment::train] Iteration 46000: 0.153816
[2018-11-14 18:35:43, INFO, PHOCNet-Experiment::train] Evaluating net after 46000 iterations
[2018-11-14 18:35:43, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.09it/s]
[2018-11-14 18:36:09, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 18:36:09, INFO, PHOCNet-Experiment::test] mAP: 12.94
[2018-11-14 18:41:36, INFO, PHOCNet-Experiment::train] Iteration 46500: 0.140441
[2018-11-14 18:47:01, INFO, PHOCNet-Experiment::train] Iteration 47000: 0.155333
[2018-11-14 18:52:28, INFO, PHOCNet-Experiment::train] Iteration 47500: 0.169566
[2018-11-14 18:57:50, INFO, PHOCNet-Experiment::train] Iteration 48000: 0.154580
[2018-11-14 18:57:50, INFO, PHOCNet-Experiment::train] Evaluating net after 48000 iterations
[2018-11-14 18:57:50, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:26<00:00, 46.45it/s]
[2018-11-14 18:58:17, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 18:58:17, INFO, PHOCNet-Experiment::test] mAP: 13.03
[2018-11-14 19:03:47, INFO, PHOCNet-Experiment::train] Iteration 48500: 0.177419
[2018-11-14 19:09:14, INFO, PHOCNet-Experiment::train] Iteration 49000: 0.170523
[2018-11-14 19:14:38, INFO, PHOCNet-Experiment::train] Iteration 49500: 0.162858
[2018-11-14 19:20:06, INFO, PHOCNet-Experiment::train] Iteration 50000: 0.165651
[2018-11-14 19:20:06, INFO, PHOCNet-Experiment::train] Evaluating net after 50000 iterations
[2018-11-14 19:20:06, INFO, PHOCNet-Experiment::test] Computing net output:
100%|¦¦¦¦¦¦¦¦¦¦| 1215/1215 [00:25<00:00, 47.01it/s]
[2018-11-14 19:20:32, INFO, PHOCNet-Experiment::test] Computing mAPs...
[2018-11-14 19:20:33, INFO, PHOCNet-Experiment::test] mAP: 13.28
[2018-11-14 19:20:33, INFO, PHOCNet-Experiment::train] Resetting data loader

Problem with GW dataset - folder "pages" does not exist

I'm running into an issue trying to reproduce training with the GW dataset:

Line 65 of experiments/cnn_ws_experiments/datasets/gw_alt.py is attempting to parse a folder named "pages" within the GW dataset. However this folder does not exist in the 1.0 version of the GW dataset downloaded from https://fki.tic.heia-fr.ch/databases/washington-database

It seems the dataset available from this link only has folders for word and line images: data/word_images_normalized and data/line_images_normalized

Which version of the GW dataset was used for this project, and is it still available anywhere?

no license, how about MIT

This project has no license, makes it hard for me to use in my organization. How about MIT license? I have submitted a PR at #3

GW Dataset

Hi,
where did you download the GW dataset? Did you use images from the complete document (--> 20 pages) or did you used line images (as provided here http://www.fki.inf.unibe.ch/databases/iam-historical-document-database/washington-database).
I can't find a dataset online with an 'gtp' annotation file.
Thank you for you help.

Isn't the training loss, binary cross entropy loss?

Since
loss = CosineLoss(size_average=False, use_sigmoid=True)
loss_val = loss(output, embedding)*args.batch_size
loss_val.backward()

Does this mean training is based on cosine loss?
As I understand from the paper, training loss is based on crossentropy.