Epoch 1: improved checkpoint (val_loss=0.1466)

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 +1,31 @@
+{
+  "backbone": "hf_vit",
+  "hf_repo": "ImageInpainters/ViTInpainter",
+  "download_model": true,
+  "upload_model": true,
+  "upload_epochs": 1,
+  "img_size": 224,
+  "patch_size": 16,
+  "batch_size": 4,
+  "val_batch_size": 4,
+  "num_workers": 2,
+  "epochs": 20,
+  "dim": 192,
+  "depth": 6,
+  "heads": 6,
+  "mlp_ratio": 4.0,
+  "dropout": 0.1,
+  "decoder_dim": 256,
+  "lr": 0.0002,
+  "encoder_lr_scale": 0.1,
+  "weight_decay": 0.0001,
+  "valid_weight": 0.05,
+  "boundary_weight": 0.25,
+  "boundary_kernel_size": 7,
+  "pretrained_model_name": "/kaggle/input/datasets/rjawesome/pretrained-vit/vit_base_patch16_224",
+  "freeze_pretrained_embeddings": true,
+  "freeze_pretrained_blocks": 10,
+  "train_dir": "/kaggle/input/datasets/awsaf49/coco-2017-dataset/coco2017/train2017",
+  "val_dir": "/kaggle/input/datasets/awsaf49/coco-2017-dataset/coco2017/val2017",
+  "output_dir": "/kaggle/working/outputs_inpaint_bert"
+}

model_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a22d984aba8ac5619694f41f664820e3b72d46d7ef9830ae22a3b30bd410221b
+size 661197302

model_last.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0f489a35ad6ff4b5a0cda01cd053a22af15194b1f131030f4b1365ddc6bbf2c
+size 661197302

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7412834540fb8497bd8d5190906d2597abc1508108b877abbb7369c1012d53b9
+size 411436855

training_metadata.json

@@ -0,0 +1,5 @@
+{
+  "epoch": 1,
+  "val_loss": 0.1466214737147093,
+  "backbone": "hf_vit"
+}