Do you have check points for MAWPS? about deductive-mwp

[allanj]

That would be five checkpoints as the experiments were conducted in five-fold. Do yo want that?

[ssh1419]

Yes, it would be great if I could use that.

[allanj]

Sorry, I don't think I keep that (only the log files are available) as limited space. But I can try to run the experiments again for you.

[ssh1419]

Oh, if that is possible, can you try it, please?

[allanj]

I'm currently on leave for 10 days. So probably update you later

[ssh1419]

Thank you!

I am working on MathQA as well now and this is what I got. Can you have a look at it? I cannot figure out what the problem is.

11/05/2022 13:37:06 - INFO - main - device = cuda:0
11/05/2022 13:37:06 - INFO - main - batch_size = 30
11/05/2022 13:37:06 - INFO - main - train_num = -1
11/05/2022 13:37:06 - INFO - main - dev_num = -1
11/05/2022 13:37:06 - INFO - main - test_num = -1
11/05/2022 13:37:06 - INFO - main - train_file = data/math23k/train23k_processed_nodup.json
11/05/2022 13:37:06 - INFO - main - dev_file = data/math23k/valid23k_processed_nodup.json
11/05/2022 13:37:06 - INFO - main - test_file = data/MathQA/mathqa_test_nodup_our_filtered.json
11/05/2022 13:37:06 - INFO - main - train_filtered_steps = None
11/05/2022 13:37:06 - INFO - main - test_filtered_steps = None
11/05/2022 13:37:06 - INFO - main - seed = 42
11/05/2022 13:37:06 - INFO - main - model_folder = mathqa_roberta-base_gru
11/05/2022 13:37:06 - INFO - main - bert_folder = none
11/05/2022 13:37:06 - INFO - main - bert_model_name = roberta-base
11/05/2022 13:37:06 - INFO - main - height = 10
11/05/2022 13:37:06 - INFO - main - train_max_height = 15
11/05/2022 13:37:06 - INFO - main - var_update_mode = gru
11/05/2022 13:37:06 - INFO - main - mode = test
11/05/2022 13:37:06 - INFO - main - learning_rate = 2e-05
11/05/2022 13:37:06 - INFO - main - max_grad_norm = 1.0
11/05/2022 13:37:06 - INFO - main - num_epochs = 1000
11/05/2022 13:37:06 - INFO - main - fp16 = 1
11/05/2022 13:37:06 - INFO - main - parallel = 0
11/05/2022 13:37:06 - INFO - main - cut_off = -100
11/05/2022 13:37:06 - INFO - main - print_error = 0
11/05/2022 13:37:06 - INFO - main - error_file = results/error.json
11/05/2022 13:37:06 - INFO - main - result_file = results/res.json
11/05/2022 13:37:07 - INFO - main - [Data Info] constant info: {'1': 0, 'PI': 1}
11/05/2022 13:37:07 - INFO - main - Testing the model now.
Some weights of UniversalModel_Roberta were not initialized from the model checkpoint at model_files/mathqa_roberta-base_gru and are newly initialized because the shapes did not match:

const_rep: found shape torch.Size([20, 768]) in the checkpoint and torch.Size([2, 768]) in the model instantiated

You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.

11/05/2022 13:37:14 - INFO - main - [Data Info] Reading test data

Tokenization: 0%| | 0/1605 [00:00<?, ?it/s]

Tokenization: 12%|█▏ | 196/1605 [00:00<00:00, 1957.17it/s][WARNING] find right var (100.0) invalid, returning FALSE
[WARNING] find right var (1.0) invalid, returning FALSE
[WARNING] find right var (2.0) invalid, returning FALSE
[WARNING] find right var (100.0) invalid, returning FALSE
[WARNING] find right var (10.0) invalid, returning FALSE
[WARNING] find right var (100.0) invalid, returning FALSE
[WARNING] find right var (1000.0) invalid, returning FALSE
[WARNING] find right var (3600.0) invalid, returning FALSE
[WARNING] find right var (100.0) invalid, returning FALSE
[WARNING] find right var (1.0) invalid, returning FALSE
[WARNING] find left_var (2.0) invalid, returning FALSE
[WARNING] find right var (2.0) invalid, returning FALSE
[WARNING] find right var (100.0) invalid, returning FALSE
[WARNING] find right var (4.0) invalid, returning FALSE
[WARNING] find right var (60.0) invalid, returning FALSE
[WARNING] find right var (1000.0) invalid, returning FALSE
[WARNING] find left_var (5.0) invalid, returning FALSE
[WARNING] find right var (2.0) invalid, returning FALSE
[WARNING] find right var (3.0) invalid, returning FALSE
[WARNING] find right var (0.5) invalid, returning FALSE
[WARNING] find right var (2.0) invalid, returning FALSE

...

Tokenization: 100%|██████████| 1605/1605 [00:00<00:00, 2300.72it/s]
11/05/2022 13:37:15 - INFO - src.data.universal_dataset - , total number instances: 468 (before filter: 1605), max num steps: 9
11/05/2022 13:37:15 - INFO - src.data.universal_dataset - filtered type counter: Counter({'cannot obtain the label sequence': 1131, 'larger than the max height 10': 6})
11/05/2022 13:37:15 - INFO - src.data.universal_dataset - number of instances removed: 1137
11/05/2022 13:37:15 - WARNING - src.data.universal_dataset - [WARNING] find duplication num: 2 (not removed)
11/05/2022 13:37:15 - INFO - src.data.universal_dataset - Counter({3: 139, 2: 93, 1: 73, 4: 72, 5: 54, 6: 20, 7: 10, 8: 5, 9: 2})
[WARNING] find right var (100.0) invalid, returning FALSE
[WARNING] find right var (3.6) invalid, returning FALSE
[WARNING] find right var (2.0) invalid, returning FALSE
[WARNING] find left_var (1.0) invalid, returning FALSE
[WARNING] find right var (2.0) invalid, returning FALSE
[WARNING] find right var (100.0) invalid, returning FALSE
[WARNING] find right var (1.0) invalid, returning FALSE
[WARNING] find right var (100.0) invalid, returning FALSE
...

--validation: 0%| | 0/16 [00:00<?, ?it/s]
--validation: 6%|▋ | 1/16 [00:01<00:23, 1.55s/it]
--validation: 19%|█▉ | 3/16 [00:01<00:06, 2.14it/s]
--validation: 25%|██▌ | 4/16 [00:01<00:04, 2.91it/s]
--validation: 38%|███▊ | 6/16 [00:02<00:02, 4.56it/s]
--validation: 50%|█████ | 8/16 [00:02<00:01, 6.12it/s]
--validation: 62%|██████▎ | 10/16 [00:02<00:00, 7.30it/s]
--validation: 75%|███████▌ | 12/16 [00:02<00:00, 7.60it/s]
--validation: 88%|████████▊ | 14/16 [00:02<00:00, 8.06it/s]
--validation: 100%|██████████| 16/16 [00:02<00:00, 9.29it/s]
--validation: 100%|██████████| 16/16 [00:02<00:00, 5.36it/s]
11/05/2022 13:37:18 - INFO - main - [Info] Equation accuracy: 23.49%, total: 468, corr: 377, adjusted_total: 1605
11/05/2022 13:37:18 - INFO - main - [Info] Value accuracy: 24.42%, total: 468, corr: 392, adjusted_total: 1605
11/05/2022 13:37:18 - INFO - main - [Info] step num: 3 Acc.:84.89 (118/139) val acc: 86.33 (120/139)
11/05/2022 13:37:18 - INFO - main - [Info] step num: 7 Acc.:70.00 (7/10) val acc: 70.00 (7/10)
11/05/2022 13:37:18 - INFO - main - [Info] step num: 4 Acc.:84.72 (61/72) val acc: 86.11 (62/72)
11/05/2022 13:37:18 - INFO - main - [Info] step num: 2 Acc.:82.80 (77/93) val acc: 86.02 (80/93)
11/05/2022 13:37:18 - INFO - main - [Info] step num: 1 Acc.:82.19 (60/73) val acc: 84.93 (62/73)
11/05/2022 13:37:18 - INFO - main - [Info] step num: 5 Acc.:62.96 (34/54) val acc: 70.37 (38/54)
11/05/2022 13:37:18 - INFO - main - [Info] step num: 6 Acc.:65.00 (13/20) val acc: 80.00 (16/20)
11/05/2022 13:37:18 - INFO - main - [Info] step num: 9 Acc.:100.00 (2/2) val acc: 100.00 (2/2)
11/05/2022 13:37:18 - INFO - main - [Info] step num: 8 Acc.:100.00 (5/5) val acc: 100.00 (5/5)

[allanj]

what's the error that you are having?

[ssh1419]

I tested the model with the checkpoints of MathQA and as you can see that 'tokenization' does not work well and accuracies are very low. Actually it worked completely well with Math23k(train/dev/test setting). Do you have any idea why it did not work with MathQA checkpoints?