Closed MinionAttack closed 2 years ago
@MinionAttack Hi, have you figured out the issue?
I just trained a biaffine-sdp model using albert-base-v2
, and the results appeared to be normal.
2022-02-23 14:00:28 INFO BiaffineSemanticDependencyModel(
(encoder): TransformerEmbedding(albert-base-v2, n_layers=4, n_out=768, stride=256, pooling=mean, pad_index=0, requires_grad=True)
(encoder_dropout): Dropout(p=0.1, inplace=False)
(edge_mlp_d): MLP(n_in=768, n_out=600, dropout=0.25)
(edge_mlp_h): MLP(n_in=768, n_out=600, dropout=0.25)
(label_mlp_d): MLP(n_in=768, n_out=600, dropout=0.33)
(label_mlp_h): MLP(n_in=768, n_out=600, dropout=0.33)
(edge_attn): Biaffine(n_in=600, n_out=2, bias_x=True, bias_y=True)
(label_attn): Biaffine(n_in=600, n_out=60, bias_x=True, bias_y=True)
(criterion): CrossEntropyLoss()
)
2022-02-23 14:00:28 INFO Loading the data
2022-02-23 14:01:32 INFO
train: Dataset(n_sentences=33916, n_batches=322, n_buckets=32)
dev: Dataset(n_sentences=1692, n_batches=33, n_buckets=32)
test: Dataset(n_sentences=1692, n_batches=32, n_buckets=32)
2022-02-23 14:01:32 INFO Epoch 1 / 10:
2022-02-23 14:03:10 INFO lr: 5.0000e-05 - loss: 0.6832 - UP: 51.01% UR: 35.75% UF: 42.04% P: 47.96% R: 33.61% F: 39.52%
2022-02-23 14:03:11 INFO dev: loss: 0.2957 - UP: 21.90% UR: 87.71% UF: 35.05% P: 20.66% R: 82.74% F: 33.06%
2022-02-23 14:03:13 INFO test: loss: 0.2911 - UP: 21.90% UR: 87.71% UF: 35.05% P: 20.66% R: 82.74% F: 33.06%
2022-02-23 14:03:16 INFO 0:01:40.838080s elapsed (saved)
2022-02-23 14:03:16 INFO Epoch 2 / 10:
2022-02-23 14:04:53 INFO lr: 4.4444e-05 - loss: 0.0169 - UP: 75.56% UR: 67.14% UF: 71.10% P: 73.67% R: 65.46% F: 69.32%
2022-02-23 14:04:55 INFO dev: loss: 0.0547 - UP: 87.28% UR: 81.35% UF: 84.21% P: 85.79% R: 79.96% F: 82.77%
2022-02-23 14:04:57 INFO test: loss: 0.0519 - UP: 87.28% UR: 81.35% UF: 84.21% P: 85.79% R: 79.96% F: 82.77%
2022-02-23 14:04:59 INFO 0:01:41.048991s elapsed (saved)
2022-02-23 14:04:59 INFO Epoch 3 / 10:
2022-02-23 14:06:36 INFO lr: 3.8889e-05 - loss: 0.0143 - UP: 84.24% UR: 76.87% UF: 80.39% P: 82.88% R: 75.63% F: 79.09%
2022-02-23 14:06:38 INFO dev: loss: 0.0463 - UP: 87.70% UR: 83.84% UF: 85.73% P: 86.49% R: 82.68% F: 84.54%
2022-02-23 14:06:40 INFO test: loss: 0.0431 - UP: 87.70% UR: 83.84% UF: 85.73% P: 86.49% R: 82.68% F: 84.54%
2022-02-23 14:06:42 INFO 0:01:40.983313s elapsed (saved)
2022-02-23 14:06:42 INFO Epoch 4 / 10:
2022-02-23 14:08:19 INFO lr: 3.3333e-05 - loss: 0.2192 - UP: 88.32% UR: 83.40% UF: 85.79% P: 87.28% R: 82.41% F: 84.78%
2022-02-23 14:08:21 INFO dev: loss: 0.0485 - UP: 83.48% UR: 86.43% UF: 84.93% P: 82.34% R: 85.26% F: 83.77%
2022-02-23 14:08:23 INFO test: loss: 0.0459 - UP: 83.48% UR: 86.43% UF: 84.93% P: 82.34% R: 85.26% F: 83.77%
2022-02-23 14:08:23 INFO 0:01:40.920955s elapsed
2022-02-23 14:08:23 INFO Epoch 5 / 10:
2022-02-23 14:10:00 INFO lr: 2.7778e-05 - loss: 0.0097 - UP: 87.49% UR: 83.32% UF: 85.35% P: 86.56% R: 82.44% F: 84.45%
2022-02-23 14:10:02 INFO dev: loss: 0.0443 - UP: 89.97% UR: 81.80% UF: 85.69% P: 88.89% R: 80.82% F: 84.66%
2022-02-23 14:10:04 INFO test: loss: 0.0402 - UP: 89.97% UR: 81.80% UF: 85.69% P: 88.89% R: 80.82% F: 84.66%
2022-02-23 14:10:06 INFO 0:01:41.001333s elapsed (saved)
Nope, I haven't given it much importance, it happens randomly and maybe if you let it do more iterations it would appear. I want to think that they are rounding or decimal errors because it doesn't make sense to me that some epochs are right and in others it doesn't match.
This issue is stale because it has been open for 30 days with no activity.
This issue was closed because it has been inactive for 7 days since being marked as stale.
Hi, I am training some models and have noticed some strange behaviour. If I specify the DEV file for the
validation
andtest
phase, I generally get the same score when training a model, but sometimes (randomly?) the score doesn't match.Below I show an example of a training with albert-base-v2 the differences are minimal but in some (rare) cases with other models I have seen them a bit bigger. For this part of the traning log file, the affected epochs are: 4, 5, 7, 9, 12, 15, 17, 18, 20, 22, 23, 24, 25, 26, 27, 28, 30, 33, 35, 39, 40, 43, 45, 46, 49, 50 ,51, 54, 56, 57, 58 and 59.
Is it a rounding error or a bug?
Regards.