paddleocr 2.9.1 rec_svtrv2_ch.yml 这个配置文件需要什么系统什么版本的paddle才能训练？需要多大显存？目前训练ResNet34模型没问题，一改这个配置文件就爆显存，或者自动退出训练 #23
Comments
|
If out of memory, you can lower first_bs: &bs 192 --> first_bs: &bs 96. |
|
Thank you for your reply. Is it possible to train with paddlepaddle 2.4.2, 8G card? |
|
You can try a smaller batchsize (e.g. first_bs: &bs 48) until it will run. |
|
I tried batch size ==1, still cannot train. |
|
When bs=192, it can run on the 24GB 3090 GPU. Have other settings been changed ? |
|
I did not change anything. 24GB for desktop ? |
|
Below setting works, thank you. Is below log normal? [2024/10/30 09:53:11] ppocr INFO: During the training process, after the 0th iteration, an evaluation is run every 2000 iterations |
|
The training speed is much slower than ResNet34? [2024/10/30 10:36:46] ppocr INFO: train dataloader has 1121727 iters |
|
yml 改了 cal_metric_during_train: True，可以显示了 acc: , norm_edit_dis: 0.864818 了 [2024/10/30 17:40:19] ppocr INFO: epoch: [1/2], global_step: 100, lr: 0.000000, acc: 0.000000, norm_edit_dis: 0.000001, CTCLoss: 65.103783, NRTRLoss: 7.208924, loss: 72.417091, avg_reader_cost: 0.20637 s, avg_batch_cost: 0.89114 s, avg_samples: 10.9, ips: 12.23155 samples/s, eta: 11 days, 13:38:42, max_mem_reserved: 4945 MB, max_mem_allocated: 4128 MB |
|
This requires manual configuration of the pre-trained model for SVTRv2. |
I train new model for Chinese, Korean, Japanese, English model, |
|
CTCLoss: 到了60左右，好像很难改善了。一般8G显卡要训练多久才能改善到10以内？ [2024/10/31 07:53:32] ppocr INFO: epoch: [1/2], global_step: 45800, lr: 0.000008, acc: 0.000000, norm_edit_dis: 0.000002, CTCLoss: 60.874081, NRTRLoss: 6.455424, loss: 67.166962, avg_reader_cost: 0.00234 s, avg_batch_cost: 0.39864 s, avg_samples: 5.94, ips: 14.90062 samples/s, eta: 9 days, 13:23:11, max_mem_reserved: 3465 MB, max_mem_allocated: 2774 MB |
|
For hardcase characters, do I need to add more images for them as ResNet34 after finished training with normal text or add them from the beginning? |
|
It is recommended to add them from the beginning. |
|
When I add 20% hardcases from the beginning, the acc. is very slow to improve. Another question is: under ResNet34, English letters in some images may be easily mis-recognized. Below is current progress. [2024/11/01 13:15:22] ppocr INFO: cur metric, acc: 0.0780839889889502, norm_edit_dis: 0.39439033382906086, fps: 43.12335596370007 |
|
That could be because it's the amount of data that's too small, causing the training not to converge. In general, training a model with a respectable accuracy requires 50w+ of data and training with 200 epoch using data augmentation. |
|
Ok. |
|
if 50w+ of data -> training with 200 epoch So, if 500W+data->20 epoch? |
|
Is it useful to increase accuracy if I rotate each character by a random angle on each image?
|
|
It is not clear if this is useful and ablation experiments need to be performed to validate it. |
|
How to perform ablation experiments ? if 50w+ of data -> training with 200 epoch So, if 500W+data->20 epoch? |
|
acc seems half of norm_edit_dis. Is this normal? [2024/11/03 06:16:08] ppocr INFO: best metric, acc: 0.4002624645652069, is_float16: False, norm_edit_dis: 0.7241033262538159, fps: 44.641736053984985, best_epoch: 1 |
|
There are thousands of fonts, is it necessary to include all fonts in model? |
|
Training speed seems similar to ResNet34. [2024/11/03 17:51:18] ppocr INFO: best metric, acc: 0.4960629888709778, is_float16: False, norm_edit_dis: 0.783771419196146, fps: 43.44646117132341, best_epoch: 1 |
|
When loading pretrained model, is it necessary to specify a learning rate? |
|
The computer automatically restarts during eval after training, what is the possible reason? any solution? |
paddleocr 2.9.1 rec_svtrv2_ch.yml 这个配置文件需要什么系统什么版本的paddle才能训练？需要多大显存？目前训练ResNet34模型没问题，一改这个配置文件就爆显存，或者自动退出训练
有没有办法将部分参数改成ResNet34，然后实现rec_svtrv2_ch架构？
The text was updated successfully, but these errors were encountered: