Epoch 1: improved checkpoint (val_loss=0.1428)

README.md

@@ -0,0 +1,22 @@
+---
+library_name: pytorch
+tags:
+- image-inpainting
+- masked-image-modeling
+- bidirectional-transformer
+- custom-model
+---
+
+# Better Inpainting Transformer
+
+This repository contains the improved inpainting model.
+
+## Files
+- `pytorch_model.bin`: inference weights
+- `config.json`: config
+- `training_metadata.json`: epoch and val loss for best uploaded checkpoint
+- `model_last.pt`: resumable training checkpoint
+- `model_best.pt`: best local checkpoint
+
+## Notes
+Recreate the `BetterInpaintTransformer` class and load the state dict manually.

config.json

@@ -0,0 +1,41 @@
+{
+  "backbone": "custom",
+  "hf_repo": "ImageInpainters/liBERTy3.0",
+  "download_model": true,
+  "upload_model": true,
+  "upload_epochs": 1,
+  "data_root": "/kaggle/input/datasets/awsaf49/coco-2017-dataset/coco2017",
+  "train_dir": "/kaggle/input/datasets/awsaf49/coco-2017-dataset/coco2017/train2017",
+  "val_dir": "/kaggle/input/datasets/awsaf49/coco-2017-dataset/coco2017/val2017",
+  "train_ann_file": "/kaggle/input/datasets/awsaf49/coco-2017-dataset/coco2017/annotations/instances_train2017.json",
+  "val_ann_file": "/kaggle/input/datasets/awsaf49/coco-2017-dataset/coco2017/annotations/instances_val2017.json",
+  "img_size": 128,
+  "patch_size": 8,
+  "batch_size": 16,
+  "val_batch_size": 8,
+  "num_workers": 2,
+  "epochs": 20,
+  "dim": 192,
+  "depth": 6,
+  "heads": 6,
+  "mlp_ratio": 4.0,
+  "dropout": 0.1,
+  "decoder_dim": 192,
+  "lr": 0.0002,
+  "encoder_lr_scale": 0.1,
+  "weight_decay": 0.0001,
+  "valid_weight": 0.0,
+  "boundary_weight": 0.05,
+  "grad_weight": 0.05,
+  "boundary_kernel_size": 7,
+  "object_sample_prob": 0.6,
+  "dilate_object_mask_px": 4,
+  "min_object_area_ratio": 0.01,
+  "max_object_area_ratio": 0.4,
+  "min_paste_scale": 0.7,
+  "max_paste_scale": 1.3,
+  "pretrained_model_name": "google/vit-base-patch16-224-in21k",
+  "freeze_pretrained_embeddings": true,
+  "freeze_pretrained_blocks": 8,
+  "output_dir": "/kaggle/working/outputs_inpaint_bert"
+}

model_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4ee4696263dc999a74de6aab4a192d281696e7fdcfbfdc0a6092df623b3c643
+size 136758554

model_last.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c4cde11407cd8767c88e10fc470e00604a0eaee813a12fb80e837aefc7480df
+size 136758554

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81c99b6b60bb99b1cde7b026fd3c392df72fb54d299f08966cc6132e5d2ac495
+size 45582127

training_metadata.json

@@ -0,0 +1,5 @@
+{
+  "epoch": 1,
+  "val_loss": 0.14278978335857392,
+  "backbone": "custom"
+}