465 questions
0
votes
1
answer
138
views
Using DataLoader for efficient model prediction
I'm trying to understand the role/utility of batch_size in torch beyond model training. I already have a trained model, where the batch_size was optimized as a hyperparameter. I want to use the model ...
- 190
0
votes
0
answers
69
views
Batching temporal graphs with Pytorch geometric data loader
I'm conducting research with temporal graph data using Pytorch-geometric.
I'm facing some issues of memory usage when making PyG data in dense format (with to_dense_batch() and to_dense_adj()).
I have ...
- 11
0
votes
0
answers
32
views
Significant overhead when calling DataLoader for a dataset within FastAPI endpoint using multiple processing
I am calling a machine learning model for a dataset that I have loaded using torch DataLoader:
class FilesDataset():
def __init__(self, path):
file_paths = glob.glob(os.path.join(path, "*....
- 367
0
votes
0
answers
32
views
PyTorch DataLoader gradually slowing down as training progresses
I noticed my dataset iteration gradually slows down as training progresses. I'm using an A100 Google Colab instance. I removed the model and all the training stuff to try to debug the dataset. With ...
- 501
1
vote
0
answers
42
views
Image Tensors Return As Zero When num_workers > 0
I am facing an issue with multiprocessing. I am trying to load my .pt data as dataloaders. Everything works fine when I set the num_workers = 0. But when I set it to a value greater than 0, the tensor ...
- 11
1
vote
0
answers
71
views
Error When Using Batch Size Greater Than 1 in PyTorch
I'm building a neural network to predict how an image will be partitioned during compression using VVC (Versatile Video Coding). The model takes a single Y-frame from a YUV420 image as input and uses ...
- 11
1
vote
1
answer
356
views
RuntimeError: Given groups=1, weight of size [64, 3, 3, 7, 7], expected input[1, 8, 3, 112, 112] to have 3 channels, but got 8 channels instead
import os
import shutil
import random
import torch
import torchvision.transforms as transforms
import cv2
import numpy as np
from torch.utils.data import Dataset, DataLoader
import torch.nn as nn
...
0
votes
0
answers
66
views
Pytorch DataLoader loops are slower than expected
I created a training loop with pytorch's TensorDataset and DataLoader classes, but encounter an interesting behavior. The progress intermittently halts every 10-15 batches with seemingly no reason. I ...
- 126
1
vote
0
answers
52
views
How to investigate memory consumption of pytorch_geometric data
I am working on a framework that uses pytorch_geometric graph data stored in the usual way in data.x and data.edge_index Additionally, the data loading process appends multiple other keys to that data ...
0
votes
1
answer
40
views
How to apply min-max scaling on a IterableDataset?
I'm using an iterableDataset because I have massive amounts of data. And since IterableDataset does not store all data in memory, we cannot directly compute min/max on the entire dataset before ...
- 13
0
votes
0
answers
177
views
Training stuck with num_workers > 0, but CPU is used instead of GPU with num_workers=0
I'm facing an issue with num_workers while training my model using PyTorch.
If I set num_workers = 0, the training starts, but the model is utilizing the CPU instead of the GPU. Although CUDA is ...
0
votes
0
answers
27
views
How to see what file Dask is working with at any time for stateful dataloader
Problem:
I am training an LLM for which my dataloader makes use of Dask to read in data. During LLM training, sometimes something breaks and you need to start again from the last checkpoint. Ideally ...
- 864
2
votes
0
answers
332
views
Why does my PyTorch DataLoader only use one CPU core despite setting num_workers>1?
I am trying to fine-tune BERT for a multi-label classification task (Jigsaw toxic comments). I created a custom dataset and DataLoader as follows:
class CustomDataSet(Dataset):
def __init__(...
- 67
0
votes
0
answers
32
views
How to modify batch data to reflect changes in original data in dataloader/pytorch?
I recently created a dataset class and am having trouble modifying the data in the batch so that it is reflected in future batches and original data
I have the following dataset class
class ...
-3
votes
1
answer
66
views
in testing dataset using dataloader , should we set shuffle=true or it doesn't matter?
I have a custom dataset (images of pizza,sushi and steak).
I'm using torch DataLoader for it , now when writing the test dataloader custom should we set shuffle=true or it just doesn't matter??
I ...
0
votes
0
answers
92
views
NVIDIA Jetson Orin FastAI2 model optimization with TensorRT and Torch2TRT incorrect Batch size
I have a Jetson Orin with the latest version of Jetpack 6.0 with CUDA 12 running on Ubuntu 22.04.
I have installed PyTorch and it has CUDA support installed:
Python 3.10.12 (main, Sep 11 2024, 15:47:...
- 368
0
votes
1
answer
156
views
Speeding up Dataset.__getitems__
I have a model with a forward function that receives optional parameters, like this:
class MyModel(nn.Module):
...
def forward(self, interactions: torch.Tensor, user_features: Optional[torch....
- 812
1
vote
0
answers
1k
views
Error loading image [SSL] record layer failure (_ssl.c:2578)
I am trying to use dataloaders in my code. I am implementing my code in aws sagemaker but for some reason when I use more than 0 num_workers for my dataloaders I get the error loading image [SSL] ...
-2
votes
1
answer
155
views
error in PyTorch dataloader with num_workers>0 in VSC under WSL
I want to utilize my GPU by adjusting the workers number, but I have a problem with the number of workers > 0.
test_loader = DataLoader(test_dataset, batch_size=32, shuffle=False, num_workers=0) - ...
- 581
0
votes
0
answers
57
views
The DataLoader in PyTorch modifies my data
I am trying to train my deep network on the MNIST dataset. When I try to upload the dataset to the dataloader and get the batched data through the iterator, I get modified data that differs from the ...
- 1
2
votes
1
answer
209
views
Resume training in Pytorch using previous persistent_workers state
How to save and restore persistent_workers state of DataLoader in order to resume training from a saved checkpoint.
First, I followed the steps in this discussion, so that the results are reproducible ...
- 371
0
votes
0
answers
49
views
TimeSeriesDataSet/TemporalFusionTransformer - found class 'NoneType' error, I think it is in the 'target_scale' array
I have tried with and without a normalizer. I am doing this with intraday financial data on one financial product, so only one group for now. The target data was already "normalized", in ...
- 59
0
votes
1
answer
31
views
What parameter do I need to change for it to match requirements?
I am trying to train a model based on a modified MNIST dataset so it classifies random images with label 10. I am constantly getting a TypeError.
transform = transforms.Compose([
transforms....
- 3
1
vote
1
answer
111
views
How do I implement training a neural network when generating the data on spot?
I would like to construct a surrogate model of a physics simulation. Thus I am able to generate the data by myself. The data itself is very big, so it makes sense to generate a few data samples (e.g. ...
0
votes
1
answer
213
views
PyTorch model training with DataLoader is too slow
I'm training a very small NN using the HAM10000 dataset. For loading the data I'm using the DataLoader that ships with PyTorch:
class CocoDetectionWithFilenames(CocoDetection):
def __init__(self, ...
- 3,940