When using the pre-trained MT-DNN ckpt, the training loss does not converge

[heshenghuan]

The init_checkpoint is trained by using scripts/run_mt_dnn.sh, and only the random seed changed, the train loss:

05/12/2021 10:37:28 Total number of params: 110085136
05/12/2021 10:37:28 At epoch 0
05/12/2021 10:37:29 Task [ 1] updates[     1] train loss[0.11215] remaining[6:25:06]
05/12/2021 10:39:55 Task [ 1] updates[   500] train loss[0.07299] remaining[2:22:14]
05/12/2021 10:42:20 Task [ 4] updates[  1000] train loss[0.07097] remaining[2:19:29]
05/12/2021 10:44:45 Task [ 1] updates[  1500] train loss[0.07218] remaining[2:16:38]
05/12/2021 10:47:10 Task [ 4] updates[  2000] train loss[0.07355] remaining[2:14:01]
05/12/2021 10:49:37 Task [ 3] updates[  2500] train loss[0.07262] remaining[2:11:59]
05/12/2021 10:52:06 Task [ 1] updates[  3000] train loss[0.07277] remaining[2:10:00]
05/12/2021 10:54:33 Task [ 3] updates[  3500] train loss[0.07201] remaining[2:07:40]
05/12/2021 10:57:00 Task [ 1] updates[  4000] train loss[0.07130] remaining[2:05:15]
05/12/2021 10:59:25 Task [ 4] updates[  4500] train loss[0.07161] remaining[2:02:43]
05/12/2021 11:01:50 Task [ 4] updates[  5000] train loss[0.07407] remaining[2:00:12]
05/12/2021 11:04:19 Task [ 4] updates[  5500] train loss[0.09848] remaining[1:57:56]
05/12/2021 11:06:44 Task [ 6] updates[  6000] train loss[0.23665] remaining[1:55:25]
05/12/2021 11:09:10 Task [ 1] updates[  6500] train loss[0.51240] remaining[1:52:58]
05/12/2021 11:11:37 Task [ 4] updates[  7000] train loss[0.82062] remaining[1:50:33]
05/12/2021 11:14:05 Task [ 1] updates[  7500] train loss[1.12256] remaining[1:48:10]
05/12/2021 11:16:29 Task [ 4] updates[  8000] train loss[1.40271] remaining[1:45:38]
05/12/2021 11:18:53 Task [ 1] updates[  8500] train loss[1.67136] remaining[1:43:05]
05/12/2021 11:21:19 Task [ 4] updates[  9000] train loss[1.93972] remaining[1:40:39]
05/12/2021 11:23:46 Task [ 1] updates[  9500] train loss[2.20053] remaining[1:38:15]
05/12/2021 11:26:12 Task [ 4] updates[ 10000] train loss[2.43337] remaining[1:35:49]
05/12/2021 11:28:38 Task [ 1] updates[ 10500] train loss[2.67121] remaining[1:33:21]
05/12/2021 11:31:01 Task [ 4] updates[ 11000] train loss[2.90422] remaining[1:30:51]
05/12/2021 11:33:27 Task [ 4] updates[ 11500] train loss[3.13861] remaining[1:28:24]
05/12/2021 11:35:52 Task [ 1] updates[ 12000] train loss[3.35614] remaining[1:25:57]
05/12/2021 11:38:16 Task [ 4] updates[ 12500] train loss[3.56738] remaining[1:23:28]
05/12/2021 11:40:42 Task [ 4] updates[ 13000] train loss[3.78074] remaining[1:21:03]
05/12/2021 11:43:07 Task [ 1] updates[ 13500] train loss[3.98890] remaining[1:18:35]
05/12/2021 11:45:33 Task [ 1] updates[ 14000] train loss[4.19893] remaining[1:16:09]
05/12/2021 11:47:57 Task [ 6] updates[ 14500] train loss[4.38943] remaining[1:13:42]
05/12/2021 11:50:25 Task [ 1] updates[ 15000] train loss[4.59797] remaining[1:11:18]
05/12/2021 11:52:51 Task [ 4] updates[ 15500] train loss[4.79801] remaining[1:08:52]
05/12/2021 11:55:15 Task [ 1] updates[ 16000] train loss[4.99510] remaining[1:06:25]
05/12/2021 11:57:41 Task [ 1] updates[ 16500] train loss[5.18970] remaining[1:03:59]
05/12/2021 12:00:07 Task [ 3] updates[ 17000] train loss[5.38753] remaining[1:01:33]

[namisan]

Learning rate is too big?

[heshenghuan]

Learning rate is too big?

It seems like a random seed problem. If default random seed was chosen, the training loss increasing problem appeared(even with a smaller learning rate). Choose another random seed is useful for this problem.

[heshenghuan]

Another issuse:

mt-dnn/train.py

Line 398 in 471f717

opt.update(config)

This line seems will overwrite the training params when using a pre-trained mt-dnn model.

[namisan]

@heshenghuan , I reran the script with different random seeds and didnot hit the bug as your mentioned.
I'm wondering which pretrained model is used in your experiments.

Yes, config in mt-dnn should be the same as the pretrained config. If I remember correctly, I removed other unrelated args.