chuanqi1992 / gumbel-crf Goto Github PK

Yao Fu, Chuanqi Tan, Bin Bi, Mosha Chen, Yansong Feng, Alexander Rush. Latent Template Induction with Gumbel-CRFs. NeurIPS 2020. (pdf)

• Gumbel-FFBS (Algorithm 2 in the paper): src/modeling/structure/linear_crf.py function rsample.
• Core model (Figure 2 in the paper): src/modeling/latent_temp_crf_ar.py
• Training, validation, evaluation: src/controller.py
  • This file also serves as a general-purpose framework. New models can be added with minimum engineering. See src/modeling/rnnlm.py for an example
• Configuration: src/config.py

• src/modeling/rnnlm.py gives a minimum example of adding new models with the existing framework. The abstract functions that need overloading are defined in src/modeling/ftmodel.py

The following experiments are for reproducing Table 1 and Figure 3(B) in the paper.

nohup python main.py --model_name=latent_temp_crf_ar --dataset=e2e --task=density --model_version=1.0.3.1 --gpu_id=6 --latent_vocab_size=20 --z_beta=1e-3 --z_overlap_logits=False --use_copy=False --use_src_info=False --num_epoch=60 --validate_start_epoch=0 --num_sample_nll=100 --x_lambd_start_epoch=10 --x_lambd_anneal_epoch=2 --batch_size_train=100 --inspect_grad=False --inspect_model=True  > ../log/latent_temp_crf_ar.1.0.3.1  2>&1 & tail -f ../log/latent_temp_crf_ar.1.0.3.1

Parameters explained:

• gpu_id: change it to any index. I was using a 8-gpu so the range is 0-7
• latent_vocab_size: number of latent states. In Sam's and Lisa's paper they all use 50 but I find 20 suffice. Maybe bacause I have anonymized the entity.
• z_beta: beta parameter to controll entropy regularization. The convergence of the latent is quite sensitive to this parameter. If two low then constant posterior (always single state), if too high then uniform posterior. All the two are collapsed cases. I did not use any annealing since I always find posterior collapse can be addressed by a carefully tuned beta. See beta-VAE for more details.
• use_copy: if the decoder copy from the table. Should not be used for density estimation.
• use_src_info: if feed table embeddings to decoder. Should not be used for density estimation.
• num_epoch: convergence end at approximately 60-80. Did not use any learning rate warmup, annealing, or early stop (I tend to not use early stop)
• num_sample_nll: 100 is enough to get an low-variance estimate.
• x_lambd_start_epoch: which epoch to start word dropout. Dropout rate start from 1 and ends at 0, decreases linearly
• x_lambd_anneal_epoch: currently use 2 epochs for word dropout annealing.

Parameters not explained are not really used (so still need cleaning)

Training log explained:

• loss: total loss, however the decrease of this loss does NOT necessarily mean the model actually converges. The following values are all important for convergence monitoring
• p_log_prob: generative model probability. Should decrease
• p_log_prob_x: word part of generative model, easy to fit
• p_log_prob_z: latent part of the generative model. This one should also decrease to make the model converge. Note that the generative model can still converge without latent
• ent_z: entropy of the inference network. If too high (e.g., larger than 50) then the inference model is nearly uniform (i.e. does not converge). Also should not too low (e.g., smaller than 1.0), otherwise the inference model would converge to a constant state (i.e., collapse)
• z_sample_max: the maximum entry to relaxed z. should increase from around 0.5 to around 0.9. Then after we anneal tau it should be 1 (i.e., nearly hard sample). Note that tunning tau to 0.01 is enough for the relaxed sample to be near 1.0

Additionally, the controller also output the gradient of each part of the model (defined in torch_model_utils.py function print_grad(·)). This is the practice that I think everyone should do to make sure each component of the model would receive meaningful gradients because sometimes one may encounter gradient vanishing for certain layers of the model (e.g., the last layer of the inference network). Printing the gradient for all layers would help us to prevent that.

nohup python main.py --model_name=latent_temp_crf_ar --grad_estimator=score_func --dataset=e2e --task=density --model_version=2.0.0.1 --gpu_id=2 --latent_vocab_size=20 --z_beta=1.05 --z_gamma=0 --z_b0=0.1 --z_overlap_logits=False --use_copy=False --use_src_info=False --num_epoch=60 --validate_start_epoch=0 --batch_size_train=100 --num_sample_nll=100 --x_lambd_start_epoch=10 --x_lambd_anneal_epoch=2 > ../log/latent_temp_crf_ar.2.0.0.1  2>&1 & tail -f ../log/latent_temp_crf_ar.2.0.0.1

Parameters explained:

• z_b0: the constant baseline. Although there are many papers discussing what different baselines for variance reduction, this simplest baseline is ironically effective in the most cases. So a suggested practice is also use a constant baseline. It also serves as a scaling factor of the reward to make the gradient numerically centered at a desired scale (otherwise some baselines would reduce the scale of the reward close to 0)
• z_beta: the beta parameter, still important in this estimator.

nohup python main.py --model_name=latent_temp_crf_ar --dataset=e2e --task=density --model_version=1.5.0.0 --gpu_id=5 --latent_vocab_size=20 --z_beta=1e-3 --z_sample_method=pm --z_overlap_logits=False --use_copy=False --use_src_info=False --num_epoch=60 --validate_start_epoch=0 --num_sample_nll=100 --tau_anneal_epoch=60 --x_lambd_start_epoch=10 --x_lambd_anneal_epoch=2 --batch_size_train=100 --inspect_grad=False --inspect_model=True  > ../log/latent_temp_crf_ar.1.5.0.0  2>&1 & tail -f ../log/latent_temp_crf_ar.1.5.0.0

Note that the simple model formulation in the paper achieves SOTA performance on MSCOCO dataset (as a side product) and outperforms many previous methods with complicated techniques.

Yet we do not think unsupervised paraphrasing is easy. But we do think the inductive biased required by unsupervised paraphrasing and the evaluation should be seriously examined.

nohup python main.py --model_name=latent_temp_crf_ar --dataset=mscoco --task=generation --model_version=1.3.1.0 --gpu_id=0 --latent_vocab_size=50 --z_beta=1e-3 --z_overlap_logits=False --use_copy=True --use_src_info=True --num_epoch=40 --validate_start_epoch=0 --validation_criteria=b2 --num_sample_nll=100 --x_lambd_start_epoch=0 --x_lambd_anneal_epoch=10 --batch_size_train=100 --batch_size_eval=100 --inspect_grad=False --inspect_model=True --write_full_predictions=True > ../log/latent_temp_crf_ar.1.3.1.0  2>&1 & tail -f ../log/latent_temp_crf_ar.1.3.1.0

nohup python main.py --model_name=latent_temp_crf_ar --grad_estimator=score_func --dataset=mscoco --task=generation --model_version=2.5.0.0 --gpu_id=4 --latent_vocab_size=50 --z_beta=1.05 --z_gamma=0 --z_b0=0.1 --use_copy=True --use_src_info=True --num_epoch=40 --validate_start_epoch=0 --batch_size_train=100 --num_sample_nll=100 --x_lambd_start_epoch=10 --x_lambd_anneal_epoch=2 --validation_criteria=b4 --test_validate=true > ../log/latent_temp_crf_ar.2.5.0.0  2>&1 & tail -f ../log/latent_temp_crf_ar.2.5.0.0

For reproducing the results, use the following codes. For more practical formulation and modeling for this task, please check Lisa's posterior regularization paper. A good practice would be combining Gumbel-CRF with posterior regularization.

nohup python main.py --model_name=latent_temp_crf_ar --dataset=e2e --task=generation --model_version=1.2.0.1 --gpu_id=4 --latent_vocab_size=20 --z_beta=1e-3 --z_overlap_logits=False --use_copy=True --use_src_info=True --num_epoch=80 --validate_start_epoch=0 --validation_criteria=b2 --num_sample_nll=100 --x_lambd_start_epoch=0 --x_lambd_anneal_epoch=10 --batch_size_train=100 --inspect_grad=False --inspect_model=True --write_full_predictions=True --test_validate > ../log/latent_temp_crf_ar.1.2.0.1  2>&1 & tail -f ../log/latent_temp_crf_ar.1.2.0.1

nohup python main.py --model_name=latent_temp_crf_ar --grad_estimator=score_func --dataset=e2e --task=generation --model_version=2.2.0.1 --gpu_id=6 --latent_vocab_size=20 --z_beta=1.05 --z_gamma=0 --z_b0=0.1 --z_overlap_logits=False --use_copy=True --use_src_info=True --num_epoch=80 --validate_start_epoch=0 --validation_criteria=b4 --batch_size_train=100 --num_sample_nll=100 --x_lambd_start_epoch=0 --x_lambd_anneal_epoch=10 > ../log/latent_temp_crf_ar.2.2.0.1  2>&1 & tail -f ../log/latent_temp_crf_ar.2.2.0.1

Results are produced from template_manager.py with some other scripts and the current implementation is slow and a bit of hacky. I'm cleaning and speeding it up so currently it is not very well integrated (or not quite runnable). You should be able to observe the produced templates during training, also reproduce the results with a bit of engineering. But if you do encounter problems during coding, please contact me.

See DGM4NLP, Gradient Estimation section and Continuous Relaxtion section for more related works.