How to estimate other images

[nomurakeiya]

Hello!
Thanks for good code!

I want to know How to estimate other images.

The part to train the model and verify it is done↓
{"mae": 8.784251246271253, "mse": 14.054188578956241, "ssim": 0.9816682679202778}

As a next step I would like to estimate the density map from the images not included in the Shanghai dataset.

[tempdata73]

Hey there!

By estimate do you mean that you want to put the model in production? If so, then you're gonna have to create your own function for loading and preprocessing the images. You might want to take a look into the create_dataloader function in the utils/augmentation.py file; in line 73 I define the preprocessing pipeline that images must get through in order to be a valid input for the model. Everything else is dedicated to loading and processing the targets which you won't need if the model will be put in production. Finally, you'll just need to load the model:

model = CSRNet(training=False)
model.load_state_dict(torch.load(PATH))
model.eval()

If you mean that you want to retrain the model using other datasets then in utils/dataset.py you'll have to define a subclass of Base and define the load_groundtruth and generate_density_map methods. However I'm not 100% sure if that will work as that part of the code is ugly. Probably this weekend I'll try and fix it so that it can work for other datasets.

[nomurakeiya]

Thank you so much for your polite answer!!

By estimate do you mean that you want to put the model in production?

That's right. I want to estimate the density map from the image I took and measure the number of people!!

file; in line 73 I define the preprocessing pipeline that images must get through in order to be a valid input for the model. Everything else is dedicated to loading and processing the targets which you won't need if the model will be put in production. Finally, you'll just need to load the model:

I'm sorry, I couldn't understand...
Do you mean to add these three lines of code somewhere?
I want to know how to create my own function for loading and preprocessing the images!!

If you mean that you want to retrain the model using other datasets then in utils/dataset.py you'll have to define a subclass of Base and define the load_groundtruth and generate_density_map methods. However I'm not 100% sure if that will work as that part of the code is ugly. Probably this weekend I'll try and fix it so that it can work for other datasets.

I would like to do this as well if possible.
Thanks for rewriting the code.

And I have a question,
Is data (.mat, .h5) required in addition to the image when estimating the density map with the trained model?
If possible, I would like to enter only the image to estimate the density map.

Thank you so much for your answer!!

[tempdata73]

Do you mean to add these three lines of code somewhere?

Yes, you'll have to create your own file where the images are loaded and processed using the preprocessing transformer defined in line 73. Then the model will have to be loaded and the images fed into it. I haven't used PyTorch in a while so I don't remember if you'll need a DataLoader or something of the sort in order to pass the images into the model. Since I created this repo for specifically training the model, I didn't create a "test with your own images" file, but it would be a nice add-on for the next update.

I would like to do this as well if possible.

I will try and update the code this weekend so that it accepts the other datasets mentioned in the README, however I can't promise anything.

Is data (.mat, .h5) required in addition to the image when estimating the density map with the trained model?

No, the .h5 files are the targets used for training the model; as I recall that's how the Shanghai dataset came so I had to transform them into images and postprocess them. You'll just need your own images.

[nomurakeiya]

I didn't create a "test with your own images" file, but it would be a nice add-on for the next update.

I will try and update the code this weekend so that it accepts the other datasets mentioned in the README, however I can't promise anything.

I understand everything.
Your technical skills are amazing, so I'm looking forward to the next update!!

You'll just need your own images.

I see, thank you!