clarkcga / multi-temporal-crop-classification-baseline Goto Github PK

I've had some headaches evaluating on another dataset since the implementation here in _generate_matrix assumes nodata is encoded as zero - however in my dataset 0 was a class and 255 was nodata.

Suggest either clarifying this expectation of the encoding, or allowing passing of a nodata value similar to here and here

Added this cell, might be useful to include:

# load previously trained model
checkpoint_path = "../output6/Unet_ep100/chkpt/Unet_final_state.pth"
checkpoint = torch.load(checkpoint_path, map_location=torch.device('cpu')) # or use GPU

# Remove 'module.' prefix if present (for nn.DataParallel compatibility)
new_state_dict = {k.replace('module.', ''): v for k, v in checkpoint.items()}
model.load_state_dict(new_state_dict)

From main readme:

However on running main.ipynb the metrics are None:

From the following config I assume there was a fair bit of experimentation performed to arrive at these parameters - are you able to shed light/into on the experiments run? I am seeking to compare Prithvi/Unet when both typical defaults are used, and when optimised.
Many thanks

use_skipAtt: false
train_dropout_rate: 0.15

optimizer: sam
LR: 0.011
LR_policy: PolynomialLR
criterion:
    name: TverskyFocalLoss
    weight:
    - 0.0182553
    - 0.03123664
    - 0.02590038
    - 0.03026126
    - 0.04142966
    - 0.04371284
    - 0.15352935
    - 0.07286951
    - 0.10277024
    - 0.10736637
    - 0.1447082
    - 0.17132445
    - 0.0566358
    ignore_index: 0
    gamma: 0.9

Hi
the default_config references a dataset:

train_dataset_name: chips_filtered_13_classes_complete

Can you confirm this is the exact dataset from HF, and not a modified version (implied by the reference to filtering)
Thanks

Issues I found.

1. The main issue if getting the data imported from HuggingFace or MLHub. I think ideally we should have clear instructions for all steps (including installing Large File Storage if needed). Ideally this should all be commands that can be typed in (no point and click). For me, it took a while to figure out how to install lfs, and then I still ran into below issue.

2. Config.yaml We should make it clear where to find the different options in this (i.e. in specific scripts). Sam also mentioned we might need to make updates for this.

3. Update README with new repo name. In the 'clone' command of Step 1.

4. Remove test code near end of main notebook. (Sam noted this).

I think these are the main issues. I also mentioned to Sam it might help to have examples for the commands in the README, but I'm not sure if that's typically done. (i.e. examples with specific file paths instead of <file_path>.

Setting gpu_devices=[0, 1, 2, 3] and calling compiled_model.fit I receive:

RuntimeError: module must have its parameters and buffers on device cuda:0 (device_ids[0]) but found one of them on device: cpu

Debugging

Image is not visible:

Running main.ipynb, on completing training and eval, metrics are None. Is this a code error or Jupyter? Note sure what I should be seeing here

A confusion matrix has been written but is wrong: