No works on GPU

[houpingx]

Try to run the code on GPU. But it seems still on CPU.

[wwyi1828]

me too

[uvipen]

I trained on GPU. Let check your gpu configuration. My code will automatically detect if gpu is available and if yes, the model will be trained on gpu

[TuringTrain]

Doesn't run on GPU for me as well, only CPU.
I tried both ways, calling .cuda() and .to('cuda') on the model and the tensors, no success.

The model and the tensors return is_cuda=True, still the GPU load is close to 0% and CPU load is 100%.

[kenny1109]

I found the reason why GPU load is so low that the dataloader processes data so slowly and GPU is always waiting.

[cindyhi]

I found the reason why GPU load is so low that the dataloader processes data so slowly and GPU is always waiting.

can you share your found ? I've been troubled by this problem too, the model and the tensors return is_cuda=True, still the GPU load is 0%.

[cindyhi]

I found the reason why GPU load is so low that the dataloader processes data so slowly and GPU is always waiting.

can you share your found ? I've been troubled by this problem too, the model and the tensors return is_cuda=True, still the GPU load is 0%.

i have set num_workers=8 and pin_memory=true in dataloader, but it doesn't work ,still the GPU load is 0%

[kenny1109]

you can try this: rewrite the __getitem__ function in dataset.py and just for getting items, not process data, and move the process operation to preprocess functions such as word segmentation, word encoding and so on.  I recommend that you can use the parameter 'collate_fn' in Dataloader and do some necessary operations like padding. My code put on below for reference. class MyData(data.Dataset): def __init__(self, root, state='train'): self.root = root with open(root, 'rb') as f: [self.train_x, self.train_y, self.val_x, self.val_y, self.test_x, self.test_y, self.word2index, self.label2index, self.n_words, self.n_labels, # self.train_sq_len, self.test_sq_len ] = pickle.load(f) if state is 'train': self.feature = self.train_x[:] self.label = self.train_y[:] elif state is 'test': self.feature = self.test_x self.label = self.test_y else: self.feature = self.val_x self.label = self.val_y def __getitem__(self, item): feature = self.feature[item] label = self.label[item] return feature, label def __len__(self): return len(self.feature) def collate_func(batch): feature = [] label = [] n_labels = 33 for i, j in batch: feature.append(i) label.append(j) feature, sentence_sq = DataProcess.pad_sentence(feature) feature, uttr_sq, sentence_sq = DataProcess.pad_utterance(feature, sentence_sq) label = DataProcess.change_label_shape(label, n_labels) return bt 在 2020年4月9日 +0800 AM9:25，cynthia <notifications@github.com>，写道：

…

> I found the reason why GPU load is so low that the dataloader processes data so slowly and GPU is always waiting. can you share your found ? I've been troubled by this problem too, the model and the tensors return is_cuda=True, still the GPU load is 0%. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

[cindyhi]

@kenny1109 thanks, it's very nice of you

[yongchaoyan]

@kenny1109
did you slove that problem,i met the same problem