Hello, I've been working with Solaris lib and I've followed the Workshop Tutorial on Github for training a new pre-trained model, using my own dataset (RIO).
Unfortunately, my image size is 438 x 408, and it doesn't fit to the input size of my xdxd model. Then, following instructions from here, I add some zeros in my image using the trainning augmentation (PadIfNeed) to 480x480. All right, my model has trained for 3 epochs and everything looked fine...
Finally, I did a script to run inference on the new model, but I don't know why, this function causes an error due to my input shape and I don't know what to do, it's the same config file... In the image below you can check the error on my terminal
And I will paste my config file (which has been successfully trained). Anybody can help me with this error?
##########################################################################################
model_name: xdxd_spacenet4
Hello, I've been working with Solaris lib and I've followed the Workshop Tutorial on Github for training a new pre-trained model, using my own dataset (RIO).
Unfortunately, my image size is 438 x 408, and it doesn't fit to the input size of my xdxd model. Then, following instructions from here, I add some zeros in my image using the trainning augmentation (PadIfNeed) to 480x480. All right, my model has trained for 3 epochs and everything looked fine...
Finally, I did a script to run inference on the new model, but I don't know why, this function causes an error due to my input shape and I don't know what to do, it's the same config file... In the image below you can check the error on my terminal
And I will paste my config file (which has been successfully trained). Anybody can help me with this error? ########################################################################################## model_name: xdxd_spacenet4
model_path: train: true infer: true
pretrained: true nn_framework: torch batch_size: 2
data_specs: width: 480 height: 480 dtype: image_type: zscore rescale: false rescale_minima: auto rescale_maxima: auto channels: 3 label_type: mask is_categorical: false mask_channels: 1 val_holdout_frac: 0.025 data_workers:
training_data_csv: 'config/rio_train.csv' validation_data_csv: inference_data_csv: 'config/rio_test.csv'
training_augmentation: augmentations: PadIfNeeded: min_height: 480 # or whatever's the next biggest divisible-by-32 number min_width: 480 HorizontalFlip: p: 0.5 RandomRotate90: p: 0.5 RandomCrop: height: 480 width: 480 p: 1.0 Normalize: mean:
inference: window_step_size_x: window_step_size_y: output_dir: 'inference_out/'