errors for charptb_discreteflow_af-scf

[yhgon]

Thanks for your sharing the awesome work.

I'm trying to reproduce your result on PTB dataset and baseline and charptb_discreteflow_af-af works well as below log but I got error for charptb_discreteflow_af-scf. could you check it?

NaN for loss and NameError: name 'cur_impatience' is not defined
thanks

• log1 charptb_baseline

python main.py --dataset ptb --run_name charptb_baseline_lstm --model_type baseline --dropout_p 0.1 --optim sgd --lr 20 --B_train 64 --B_val 64

Epoch 97 | train loss: 0.885, val loss: 0.976, val NLL (0): -0.000 | train kl: 0.000, val kl: 0.000 | kl_weight: 1.000, time: 113.07s/4.26s
* Learning rate dropping by a factor of 4
Epoch 98 | train loss: 0.885, val loss: 0.976, val NLL (0): -0.000 | train kl: 0.000, val kl: 0.000 | kl_weight: 1.000, time: 111.98s/4.26s
* Learning rate dropping by a factor of 4
Epoch 99 | train loss: 0.885, val loss: 0.976, val NLL (0): -0.000 | train kl: 0.000, val kl: 0.000 | kl_weight: 1.000, time: 113.57s/3.15s
* Learning rate dropping by a factor of 4

• *log2 charptb_discreteflow_af-af *

python main.py --dataset ptb --run_name charptb_discreteflow_af-af  --B_train 64 --B_val 64
Epoch 94 | train loss: 0.921, val loss: 1.021, val NLL (30): 0.983 | train kl: 0.875, val kl: 0.975 | kl_weight: 1.000, time: 789.37s/62.59s
* Learning rate dropping by a factor of 4
Epoch 95 | train loss: 0.921, val loss: 1.021, val NLL (0): -0.000 | train kl: 0.875, val kl: 0.974 | kl_weight: 1.000, time: 794.56s/19.22s
* Learning rate dropping by a factor of 4
Epoch 96 | train loss: 0.921, val loss: 1.021, val NLL (0): -0.000 | train kl: 0.875, val kl: 0.974 | kl_weight: 1.000, time: 788.00s/18.88s
* Learning rate dropping by a factor of 4
Epoch 97 | train loss: 0.921, val loss: 1.021, val NLL (0): -0.000 | train kl: 0.875, val kl: 0.974 | kl_weight: 1.000, time: 791.33s/19.06s
* Learning rate dropping by a factor of 4

• *log2 charptb_discreteflow_af-scf *

python main.py --dataset ptb --run_name charptb_discreteflow_af-scf --hiddenflow_scf_layers

Epoch 0 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.000, time: 762.07s/19.64s
Epoch 1 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.000, time: 758.30s/19.49s
Epoch 2 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.000, time: 760.94s/19.00s
Epoch 3 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.000, time: 760.13s/18.88s
Epoch 4 | train loss: nan, val loss: nan, val NLL (30): nan | train kl: nan, val kl: nan | kl_weight: 0.000, time: 759.14s/59.24s
Epoch 5 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.100, time: 756.40s/18.98s
Epoch 6 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.200, time: 746.24s/18.95s
Epoch 7 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.300, time: 759.76s/18.85s
Epoch 8 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.400, time: 757.32s/18.70s
Epoch 9 | train loss: nan, val loss: nan, val NLL (30): nan | train kl: nan, val kl: nan | kl_weight: 0.500, time: 751.98s/59.42s
Epoch 10 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.600, time: 747.67s/18.53s
Epoch 11 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.700, time: 750.80s/18.60s
Epoch 12 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.800, time: 750.70s/18.71s
Epoch 13 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.900, time: 750.79s/18.66s
Epoch 14 | train loss: nan, val loss: nan, val NLL (30): nan | train kl: nan, val kl: nan | kl_weight: 1.000, time: 750.22s/59.38s
Epoch 15 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 1.000, time: 746.29s/17.96s
Traceback (most recent call last):
  File "main.py", line 217, in <module>
    cur_impatience += 1
NameError: name 'cur_impatience' is not defined

[yhgon]

configure charptb_discreteflow_iaf-scf is also fail to learn with NaN

python main.py --dataset ptb --run_name charptb_discreteflow_iaf-scf --dropout_p 0 --hiddenflow_flow_layers 3 --hiddenflow_scf_layers --prior_type IAF
The `device` argument should be set by using `torch.device` or passing a string as an argument. This behavior will be deprecated soon and currently defaults to cpu.
The `device` argument should be set by using `torch.device` or passing a string as an argument. This behavior will be deprecated soon and currently defaults to cpu.
The `device` argument should be set by using `torch.device` or passing a string as an argument. This behavior will be deprecated soon and currently defaults to cpu.
/mnt/git/TextFlow/lstm_flow.py:157: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).
  y = torch.tensor(x) # [T, B, inp_dim]
/mnt/git/TextFlow/flows.py:188: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).
  y = torch.tensor(x)
Epoch 0 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.000, time: 4479.11s/133.32s
Epoch 1 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.000, time: 4474.15s/128.58s
Epoch 2 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.000, time: 4920.69s/145.28s
Epoch 3 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.000, time: 4716.16s/135.83s
Epoch 4 | train loss: nan, val loss: nan, val NLL (30): nan | train kl: nan, val kl: nan | kl_weight: 0.000, time: 4872.09s/296.83s
Epoch 5 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.100, time: 4514.57s/141.28s
Epoch 6 | train loss: nan, val loss: nan, val NLL (0): -0.000 | train kl: nan, val kl: nan | kl_weight: 0.200, time: 4950.72s/129.10s

[hyunjik11]

When running the code on pytorch 0.4.1 (as listed in the Dependencies section of README), the above NaN behaviour did not happen and learning seems stable (although I haven't checked whether the results in the paper are replicated yet)

However when running the code for pytorch version >= 1.0, I can confirm that the above NaN issue emerges. However there is a fix. I've looked into what was causing the NaN, and observed that the log_p_z in kl_loss was growing to be large negative values, eventually going to NaN.

Looking at the release notes for pytorch 1.0 (that shows what has changed from v0.4.1), it turns out that the change made to torch.tensor was causing the issue.

So for version 0.4.1, y = torch.tensor(x)was not returning a detached tensor for tensor inputs x (i.e. y.requires_grad = True if x.requires_grad = True) whereas it is detaching for version >= 1.0 (i.e. y.requires_grad = False). So this leads to incorrect gradients when calling obj_per_accum.backward() due to the calls to torch.tensor in SCFLayer.forward, AFLayer.generate and LSTM_AFLayer.generate. Replacing y=torch.tensor(x) with y=x.clone() seems to resolve the issue and lead to stable training.

[hyunjik11]

A side note for compatibility of code with pytorch 1.0 and 1.1: it seems like using in-place operations for tensors that lead to the training loss obj_per_accum can give rise to error messages such as RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation in pytorch v1.0 or v1.1, that doesn't have the in-place correctness checks implemented for version >= 1.2. This can be prevented by in-place operations with equivalent operations that create new tensors, in particular the lines that have in-place operations on loss in run_epoch of main.py. Namely use:

        indices = torch.arange(batch_data.shape[0]).view(-1, 1).to(device)  # [T, 1]
        loss_mask = indices >= lengths.view(1, -1)  # [T, B]
        loss_mask = loss_mask[:, :, None].repeat(1, 1, loss_.shape[-1])  # [T, B, s]
        
        loss = loss * (1 - loss_mask.float())
        kl_loss = kl_loss * (1 - loss_mask.float())

instead of the original code:

TextFlow/main.py

Lines 53 to 58 in 0eb8552

	indices = torch.arange(batch_data.shape[0]).view(-1, 1).to(device)
	loss_mask = indices >= lengths.view(1, -1)
	loss_mask = loss_mask[:, :, None].repeat(1, 1, loss.shape[-1])
	loss[loss_mask] = 0
	kl_loss[loss_mask] = 0

In general, I think it's recommended not to use in-place operations for autograd although for pytorch >= 1.2 the in-place correctness checks do confirm that you are getting correct gradients when there is no error message

[yhgon]

thanks for your confirmation and guide. I'll try it.