问下运行情况正常吗

[pageedward]

使用这份代码在im2latex 100k 上面训练，得到如下的指标：和论文上的Em76相差有些大，anyone meet the same situation？ Eval: BLEU-4 88.89 - EM 37.86 - Edit 91.88 - perplexity -1.14

[Root970103]

你这个训练感觉已经不错了 BLEU值都比较高了 我的EM 到后来就变0了（哭 顺便问下，你使用的就是默认参数咩

[ziyaxuanyi]

这个模型训练出来，有多大啊

[chenyangMl]

Posting a train-log for who wants to know performance of this model .
train-set 7.5w val-set 8.3k, using the default training-config but batch_size = 6.

• Training: - 3669s - loss: 1.6814 - perplexity: 8.9318 - lr: 0.0001
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 69.10 - EM 9.93 - Edit 73.04 - perplexity -1.60
• New best score (-1.60)!
• Saved model in results/demo0/model.weights/
• Elapsed time: 4461.13, lr: 0.00010 Epoch 2/15 12543/12543 [==============================] - 2364s - loss: 0.3243 - perplexity: 1.4122 - lr: 0.0001
• Training: - 2364s - loss: 0.3243 - perplexity: 1.4122 - lr: 0.0001
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 61.36 - EM 7.14 - Edit 63.19 - perplexity -1.70
• Elapsed time: 3286.70, lr: 0.00010 Epoch 3/15 12543/12543 [==============================] - 2343s - loss: 0.5157 - perplexity: 1.8175 - lr: 0.0010
• Training: - 2343s - loss: 0.5157 - perplexity: 1.8175 - lr: 0.0010
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 63.57 - EM 7.57 - Edit 66.20 - perplexity -1.68
• Elapsed time: 3176.76, lr: 0.00100 Epoch 4/15 12543/12543 [==============================] - 2341s - loss: 0.5249 - perplexity: 1.9761 - lr: 0.0010
• Training: - 2341s - loss: 0.5249 - perplexity: 1.9761 - lr: 0.0010
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 58.75 - EM 5.89 - Edit 62.85 - perplexity -1.79
• Elapsed time: 3176.96, lr: 0.00100 Epoch 5/15 12543/12543 [==============================] - 2341s - loss: 1.0391 - perplexity: 4.5142 - lr: 0.0010
• Training: - 2341s - loss: 1.0391 - perplexity: 4.5142 - lr: 0.0010
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 16.77 - EM 0.00 - Edit 26.80 - perplexity -6.52
• Elapsed time: 3330.05, lr: 0.00100 Epoch 6/15 12543/12543 [==============================] - 2339s - loss: 1.2535 - perplexity: 4.3205 - lr: 0.0010
• Training: - 2339s - loss: 1.2535 - perplexity: 4.3205 - lr: 0.0010
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 40.13 - EM 0.39 - Edit 42.14 - perplexity -2.78
• Elapsed time: 3329.48, lr: 0.00100 Epoch 7/15 12543/12543 [==============================] - 2342s - loss: 1.2953 - perplexity: 4.6243 - lr: 0.0009
• Training: - 2342s - loss: 1.2953 - perplexity: 4.6243 - lr: 0.0009
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 41.91 - EM 0.69 - Edit 44.38 - perplexity -2.69
• Elapsed time: 3285.01, lr: 0.00072 Epoch 8/15 12543/12543 [==============================] - 2344s - loss: 1.2576 - perplexity: 4.1392 - lr: 0.0006
• Training: - 2344s - loss: 1.2576 - perplexity: 4.1392 - lr: 0.0006
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 29.63 - EM 0.01 - Edit 36.10 - perplexity -4.38
• Elapsed time: 3322.23, lr: 0.00052 Epoch 9/15 12543/12543 [==============================] - 2344s - loss: 1.4169 - perplexity: 4.7844 - lr: 0.0004
• Training: - 2344s - loss: 1.4169 - perplexity: 4.7844 - lr: 0.0004
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 21.44 - EM 0.04 - Edit 30.39 - perplexity -5.16
• Elapsed time: 3297.64, lr: 0.00037 Epoch 10/15 12543/12543 [==============================] - 2343s - loss: 1.1873 - perplexity: 3.6239 - lr: 0.0003
• Training: - 2343s - loss: 1.1873 - perplexity: 3.6239 - lr: 0.0003
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 42.70 - EM 0.51 - Edit 45.73 - perplexity -2.63
• Elapsed time: 3272.69, lr: 0.00027 Epoch 11/15 12543/12543 [==============================] - 2345s - loss: 0.8784 - perplexity: 2.4891 - lr: 0.0002
• Training: - 2345s - loss: 0.8784 - perplexity: 2.4891 - lr: 0.0002
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 46.11 - EM 1.51 - Edit 49.44 - perplexity -2.32
• Elapsed time: 3244.99, lr: 0.00019 Epoch 12/15 12543/12543 [==============================] - 2344s - loss: 0.6840 - perplexity: 2.0360 - lr: 0.0002
• Training: - 2344s - loss: 0.6840 - perplexity: 2.0360 - lr: 0.0002
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 52.00 - EM 2.51 - Edit 55.21 - perplexity -2.07
• Elapsed time: 3229.79, lr: 0.00014 Epoch 13/15 12543/12543 [==============================] - 2345s - loss: 1.0375 - perplexity: 2.9436 - lr: 0.0001
• Training: - 2345s - loss: 1.0375 - perplexity: 2.9436 - lr: 0.0001
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 38.67 - EM 0.47 - Edit 44.49 - perplexity -2.77
• Elapsed time: 3268.07, lr: 0.00010 Epoch 14/15 12543/12543 [==============================] - 2344s - loss: 0.9637 - perplexity: 2.6928 - lr: 0.0001
• Training: - 2344s - loss: 0.9637 - perplexity: 2.6928 - lr: 0.0001
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 41.50 - EM 0.39 - Edit 44.36 - perplexity -2.67
• Elapsed time: 3291.70, lr: 0.00010 Epoch 15/15 12543/12543 [==============================] - 2344s - loss: 0.7661 - perplexity: 2.2090 - lr: 0.0001
• Training: - 2344s - loss: 0.7661 - perplexity: 2.2090 - lr: 0.0001
• Evaluating...Loaded 8370 formulas from results/demo0/formulas_val/ref.txt Loaded 8370 formulas from results/demo0/formulas_val/hyp_0.txt
• Eval: BLEU-4 53.78 - EM 2.71 - Edit 56.13 - perplexity -2.02
• Elapsed time: 3229.17, lr: 0.00010

[chenyangMl]

这个模型训练出来，有多大啊

103M

[ziyaxuanyi]

这个模型训练出来，有多大啊

103M

请问您也在做数学公式识别吗？有没有什么好的方案或者开源工程推荐呢？