Handle None gradients in nn.utils.clip_grad_norm

[monajalal]

I get this error:

python train.py --batch-size 20 --rnn_type GRU --cuda --gpu 1 --lr 0.0001 --mdl RNN --clip_norm 1 --opt Adam
/scratch/sjn-p2/anaconda/anaconda2/lib/python2.7/site-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.
  from ._conv import register_converters as _register_converters
Using TensorFlow backend.
There are 2 CUDA devices
Setting torch GPU to 1
Using device:1 
Stored Environment:['term_len', 'word_index', 'glove', 'max_len', 'train', 'dev', 'test', 'index_word']
Loaded environment
Creating Model...
Setting Pretrained Embeddings
Initialized GRU model
Starting training
Namespace(aggregation='mean', attention_width=5, batch_size=20, clip_norm=1, cuda=True, dataset='Restaurants', dev=1, dropout_prob=0.5, embedding_size=300, epochs=50, eval=1, gpu=1, hidden_layer_size=300, l2_reg=0.0, learn_rate=0.0001, log=1, maxlen=0, mode='term', model_type='RNN', opt='Adam', pretrained=1, rnn_direction='uni', rnn_layers=1, rnn_size=300, rnn_type='GRU', seed=1111, term_model='mean', toy=False, trainable=1)
/scratch2/debate_tweets/sentiment/pytorch_sentiment_rnn/models/rnn.py:51: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument.
  decoded = self.softmax(decoded)
Traceback (most recent call last):
  File "train.py", line 343, in <module>
    exp.train()
  File "train.py", line 326, in train
    loss = self.train_batch(i)
  File "train.py", line 303, in train_batch
    coeff = clip_gradient(self.mdl, self.args.clip_norm)
  File "train.py", line 35, in clip_gradient
    modulenorm = p.grad.data.norm()
AttributeError: 'NoneType' object has no attribute 'data'
[jalal@goku pytorch_sentiment_rnn]$ 

for train.py file in https://github.com/vanzytay/pytorch_sentiment_rnn I have followed all the steps in the readme up to here. What do you think should be fixed?

When submitting a bug report, please include the following information (where relevant):

• OS: CentOS Linux release 7.4.1708 (Core)
• PyTorch version: 0.3.1.post2
• How you installed PyTorch (conda, pip, source): conda install -c pytorch pytorch
• Python version: Python 2.7.14 |Anaconda custom (64-bit)| (default, Dec 7 2017, 17:05:42)
• CUDA/cuDNN version: CUDA Version 8.0.61
• GPU models and configuration: GP102 [GeForce GTX 1080 Ti], driver=nvidia latency=0
• GCC version (if compiling from source): [GCC 7.2.0] on linux2

[monajalal]

vanzytay/pytorch_sentiment_rnn#6

[zou3519]

What this says is p.grad is None. It's possible that p (whatever it is) wasn't used in the gradient computation, or there was no backwards pass applied.

[monajalal]

Well, I understand that however, this seems like a problem with PyTorch as the people who have used the repo following the commands provided didn't have this error. Possibly a recent problem with PyTorch actually.

[zou3519]

The last commit on the page is from Jan 24, 2017. Pytorch definitely has changed a lot since then. If you have specific questions about how to use pytorch, please ask on our forums: https://discuss.pytorch.org/

[soumith]

closed via @zou3519 's comment.

[apaszke]

I think the error is still legitimate. We should handle None .grad attributes correctly in clip_grad_norm (by assuming the norm is 0). Right now we fail like in the posted stack trace.

[zou3519]

@apaszke I think we do handle None .grad attributes correctly in clip_grad_norm: https://github.com/pytorch/pytorch/blob/master/torch/nn/utils/clip_grad.py#L18.

The traceback @monajalal posted implies that the code uses its own clip_gradient method:

Traceback (most recent call last):
  File "train.py", line 343, in <module>
    exp.train()
  File "train.py", line 326, in train
    loss = self.train_batch(i)
  File "train.py", line 303, in train_batch
    coeff = clip_gradient(self.mdl, self.args.clip_norm)
  File "train.py", line 35, in clip_gradient
    modulenorm = p.grad.data.norm()

@monajalal If you replace clip_gradient with some usage of torch.nn.utils.clip_grad_norm this particular error should go away.