TF + PyTorch installed (same env): RuntimeError: Given input size: (128x1x16). Calculated output size: (128x0x8). Output size is too small"

[sohamyedgaonkar]

Bug description

Whenever I am trying to run the basic flow the readme page for my own image .
This is what error is hitting up again and again
even when i am trying on local computer and Google collab

RuntimeError Traceback (most recent call last)
in <cell line: 0>()
7 doc = DocumentFile.from_images(r"/content/data-set.png")
8 # Analyze
----> 9 result = model(doc)
10 # Display the result (requires matplotlib & mplcursors to be installed)
11 result.show()

19 frames
/usr/local/lib/python3.11/dist-packages/torch/nn/functional.py in _max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode, return_indices)
828 if stride is None:
829 stride = torch.jit.annotate(List[int], [])
--> 830 return torch.max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode)
831
832

RuntimeError: Given input size: (128x1x16). Calculated output size: (128x0x8). Output size is too small

Code snippet to reproduce the bug

from doctr.io import DocumentFile
from doctr.models import ocr_predictor

model = ocr_predictor(pretrained=True)

Image

doc = DocumentFile.from_images(r"c:\Users\DELL\Desktop\data-set.png")

Analyze

result = model(doc)

Display the result (requires matplotlib & mplcursors to be installed)

result.show()

Error traceback

in <cell line: 0>()
7 doc = DocumentFile.from_images(r"/content/data-set.png")
8 # Analyze
----> 9 result = model(doc)
10 # Display the result (requires matplotlib & mplcursors to be installed)
11 result.show()

19 frames
/usr/local/lib/python3.11/dist-packages/torch/nn/functional.py in _max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode, return_indices)
828 if stride is None:
829 stride = torch.jit.annotate(List[int], [])
--> 830 return torch.max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode)
831
832

RuntimeError: Given input size: (128x1x16). Calculated output size: (128x0x8). Output size is too small

Environment

Google collab after running !pip install "python-doctr[tf]"

Deep Learning backend

is_tf_available: True
is_torch_available: True

[felixdittrich92]

Hi @sohamyedgaonkar 👋🏼,

Thanks for reporting would it be possible to share an example that we can reproduce the issue ?

[sohamyedgaonkar]

yeah sure

this was the image i was working on

[felixdittrich92]

Mh I can't reproduce it 🤔

from doctr.io import DocumentFile
from doctr.models import ocr_predictor

doc = DocumentFile.from_images("/home/felix/Desktop/testi.png")
model = ocr_predictor(pretrained=True)
res = model(doc)
res.show()
json_res = res.export()

Tested with TF and PyTorch backend

[felixdittrich92]

Which pytorch version do you use ?

[sohamyedgaonkar]

Name: torch
Version: 2.2.1

Can you please check on google colab
i was facing the same issue on it

[felixdittrich92]

Name: torch Version: 2.2.1

Can you please check on google colab i was facing the same issue on it

On it and yes here I can reproduce the issue ...strange

[felixdittrich92]

For the moment you can use onnxtr (same as doctr but prod optimized and doesn't requires pytorch or tensorflow)

https://colab.research.google.com/drive/1Q4dNqywh-q-GILKiIznD6P9gUa4df50r?usp=sharing

https://github.com/felixdittrich92/OnnxTR

[felixdittrich92]

Ok got it

It's an dependency issue if both pytorch and tensorflow are installed ..protobuff forces anything strange

Uninstalling tensorflow before solves the issue

!pip3 uninstall -y tensorflow
!pip3 install python-doctr[torch,viz]

[sohamyedgaonkar]

great ! thanks for your timely response .
it Will really help me
thanks for both onnxtr and issue

[felixdittrich92]

Will keep the issue open if someone else has the same issue ... unfortunately I can't do anything from my end because it conflicts from TF dependencies

[XanderBrooks]

Hey, I believe I found a work around without uninstalling tensorflow.

I was also getting the "RuntimeError: Given input size: (128x1x16). Calculated output size: (128x0x8). Output size is too small" error for some images that are a regular size (750-1600px width/height)

Sometimes I was able to run an image, but the blocks repeatedly came out empty -- I tried 8+ images and pdfs.

I have both Pytorch and Tensorflow installed

PyTorch version: 2.6.0.dev20241126+cu118
CUDA available: True
TensorFlow version: 2.18.0

Solution:

import os
os.environ["USE_TF"] = "0"

Run this BEFORE initializing the model. This simply disables tensorflow temporarily. Give it a shot, I hope it helps!