HegdeSudarshan/BigEarthNetModels

Enable verbose error reporting in Gradio

Fix upsampling: use two-stage 2x2 pixel shuffle (30->60->120)

Fix u1 channels: 641024 for pixel_shuffle to output 64 channels

Fix channel mismatch: u2 layer should accept 16 channels from pixel_shuffle

Fix label loading: reverse mapping from label_indices.json

Fix ResNetClassifier: backbone is raw ResNet18, not Sequential wrapper

Fix ResNetClassifier to use 'backbone' matching saved weights

Add classifier.pth model file

Fix RFB_Modified: dense layers process concatenated [x, rfb_out] with 128+ input channels

Fix DenseLayer to use add_module for numbered .0 access matching state_dict

Final architecture fix based on state_dict analysis: DenseLayer .convs indexing, RFB_Branch correct kernels (3x3, 5x1, 7x1), RFB 64ch input, complete upsampling path, final RGB output

Fix remaining architecture issues: correct RFB_Branch kernels (1x5, 1x7), DenseLayer Sequential naming, RFB_Modified 160ch, rfb_fuse 64ch input

Fix model architecture to match trained weights - corrected RRDB, RFB_Branch, and RFB_Modified structures

Update README.md

Upload label_indices.json

Upload 5 files

initial commit