Closed
Milestone
DAN-P4: Improve data loading and preprocessing
Unstarted Issues (open and unassigned)
0
Ongoing Issues (open and assigned)
0
Completed Issues (closed)
26
- Fix call to seed_worker
- Fix valid batch size to 1
- Set default value to None for --image_max_width during prediction
- Add a prediction test for process_image function
- Debug/remove PiecewiseAffine augmentation transform
- Load image using torch + use training pre-processing function during prediction
- Merge DatasetManager / GenericDataset / OCRDatasetManager / OCRDataset classes
- Pre-process the images immediately after loading them.
- Use default mean and std values?
- Compute mean and std only if training from scratch
- Indicate python version compatibility
- Remove the remove_linebreaks parameter from training configuration
- Remove DPIAdjusting transform
- Use a single padding method
- Directly read images using torch
- Utils pairwise function can be replaced by itertools pariwise
- Simplify mean and std computation
- Remove randint, rand, rand_uniform and round_floats from utils.py
- Use torchvision functions / transforms for data augmentation
- Remove normalize parameter from training configuration
- Remove padding value and padding token parameters from training configuration
- Remove add_eot and add_sot parameters from training configuration
- Remove width_divisor and height_divisor parameters from training configuration
- Extract data in sub-resolution
- Move OCR utils functions to utils.py
- Remove deepcopy from OCR code
Loading
Loading
Loading