Pretraining Sen12ms

[taeil]

The first thing I think we should try is using 1x1 convolutions before the input (1x1 conv, batch norm, relu), which will map the different number of bands to a common number of input features. Then our final model will have different input layers for the different types of input images. The Data loader would then have to return a dictionary instead of a concatenated tensor and mayo the given image into the associated input type.

Colorado's change with custom data loader:
Berkeley-Data/OpenSelfSup#3

[taeil]

Changed to Tensor and tried pipeline.. now error saying it must be 3 channels :(

  File "/home/ubuntu/anaconda3/envs/hpt/lib/python3.7/site-packages/torchvision/transforms/functional_tensor.py", line 215, in rgb_to_grayscale
    raise TypeError("Input image tensor should 3 channels, but found {}".format(c))
TypeError: Input image tensor should 3 channels, but found 2

[taeil]

some of pytorch transformation that such as color, grayscale requires 3 channel tensor input.

[taeil]

result looks strange...
https://wandb.ai/cal-capstone/hpt/runs/2p9a1ijm?workspace=user-taeil

[taeil]

35K samples does not convert very well.

[taeil]

running with 35K samples again since they don't actually looks different based on the amount of steps.

[taeil]

ran on full dataset with 20 epochs pretty good.

[taeil]

With 200 epochs on full dataset: