So i input a picture in script tflite_inference.py
Then i go to subdirectory "result", i see a file like my picture but with the text highlighted in red.
No problem with what i see.
My problem is with the file "res_*.txt" that is with the picture: the file is empty. How to get the text ?
Is it a bug ?
What variable stores this data in the script?

Thank you very much for your answer

RuntimeError                              Traceback (most recent call last)
[<ipython-input-43-65b5f67ee40d>](https://localhost:8080/#) in <module>()
     13 
     14 x = x.detach().numpy()
---> 15 y, feature = run_tflite_model(x)
     16 
     17 y = torch.from_numpy(y)

1 frames
[/usr/local/lib/python3.7/dist-packages/tensorflow/lite/python/interpreter.py](https://localhost:8080/#) in allocate_tensors(self)
    512   def allocate_tensors(self):
    513     self._ensure_safe()
--> 514     return self._interpreter.AllocateTensors()
    515 
    516   def _safe_to_run(self):

RuntimeError: tensorflow/lite/kernels/space_to_batch_nd.cc:83 final_dim_size % block_shape[dim] != 0 (4 != 0)Node number 66 (SPACE_TO_BATCH_ND) failed to prepare.Failed to apply the default TensorFlow Lite delegate indexed at 0.

How can i fix it?

Hi @tulasiram58827

Thank you for your great work. When I run your colab notebook CRAFT_TFLITE.ipynb. I get the following error -

ConverterError: /usr/local/lib/python3.6/dist-packages/tensorflow/python/saved_model/load.py:890:0: error: 'tf.MaxPool' op is neither a custom op nor a flex op

Do you know how to get around it ? Thanks

Hi @tulasiram58827, I was wondering if it is possible to change the input format of the CRAFT, from [b c h w] to [b h w c], before converting it to tflite? By doing so, the input image need no reformatting (x.permute) when running the inference.

Any help would be very appreciated. Thanks!

Hi,
thanks for this amazing work,
i'm running craft in 111 ms on windows, but i want to get maximum speed up , so i want to use onnx and i'm new to that,
but i'm confused that which scripts should i use to test?
( i want to use craft.onnx model so i dont need to produce it )

in below code line it is .pth version of craft.
parser.add_argument('--trained_model', default='../models/craft_mlt_25k.pth', type=str, help='pretrained model')

I don't see the other conversion recipes you took in the conversion notebook. Could you include them?

You can also consider clearing the unnecessary outputs from the notebooks to keep them clean.

I have run savedmodel_to_tflite.ipynb and got this error:
RuntimeError: Input to reshape is a tensor with 3840000 values, but the requested shape has 401408
(while executing 'Reshape' via Eager)Node number 450 (TfLiteFlexDelegate) failed to invoke.
Thank you so much

Although I converted craft model to tflite successfully and results seems pretty similar with the original model there are latency problems with converted tflite model.

Sample image used for testing.
Before Conversion:

1. Model Size -- 80MB
2. Time to Run -- 0.27s

After Converting to TFLITE

1. Model Size -- 20MB
2. Time to Run -- 25.1s

Hardware Used:

model name : Intel(R) Core(TM) i5-9300H CPU @ 2.40GHz

All the models used available in the repo.

@sayakpaul

Hi @tulasiram58827 thank you for the awesome work on converting craft to tflite, I am trying to predict images from craft on android devices. I am having trouble converting input images and apply preprocessing before running inference on the model.

Bro can you please guide me on how shall i proceed with converting the following code :
`
image = loadImage(image_path)
image = cv2.resize(image, dsize=(800, 1280), interpolation=cv2.INTER_LINEAR)
img_resized, target_ratio, size_heatmap = resize_aspect_ratio(image, canvas_size, interpolation=cv2.INTER_LINEAR, mag_ratio=mag_ratio)
ratio_h = ratio_w = 1 / target_ratio

preprocessing

x = normalizeMeanVariance(img_resized)
x = torch.from_numpy(x).permute(2, 0, 1) # [h, w, c] to [c, h, w]
x = Variable(x.unsqueeze(0)) # [c, h, w] to [b, c, h, w]

forward pass

x = x.cpu().detach().numpy()
`