From 50f01980b61a8d9f8c228b7e9e11eb41337e615e Mon Sep 17 00:00:00 2001 From: FavorMylikes Date: Mon, 4 Dec 2023 20:14:50 +0800 Subject: [PATCH 1/2] use aria2c download model --- README.md | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index a7c2c9b..9b53d37 100644 --- a/README.md +++ b/README.md @@ -42,10 +42,15 @@ pip install -e CLIP/ ``` Download checkpoint under `zero123` through one of the following sources: - ``` https://huggingface.co/cvlab/zero123-weights/tree/main -wget https://cv.cs.columbia.edu/zero123/assets/$iteration.ckpt # iteration = [105000, 165000, 230000, 300000] +``` + +Use aria2c`` download models from huggingface + +``` +apt install aria2 +aria2c -x 10 --disable-ipv6 --input-file models.txt --dir zero123/ --continue ``` Note that we have released 4 model weights: 105000.ckpt, 165000.ckpt, 230000.ckpt, 300000.ckpt. By default, we use 105000.ckpt which is the checkpoint after finetuning 105000 iterations on objaverse. Naturally, checkpoints trained longer tend to overfit to training data and suffer in zero-shot generalization, though we didn't empirically verify this. 300000.ckpt is trained for around 6000 A100 hours. From 314aecee904c59e52e42cc6c3b4bb37bd216b87b Mon Sep 17 00:00:00 2001 From: FavorMylikes Date: Mon, 4 Dec 2023 20:16:33 +0800 Subject: [PATCH 2/2] use aria2c download models --- models.txt | 4 ++++ 1 file changed, 4 insertions(+) create mode 100644 models.txt diff --git a/models.txt b/models.txt new file mode 100644 index 0000000..c788615 --- /dev/null +++ b/models.txt @@ -0,0 +1,4 @@ +https://huggingface.co/cvlab/zero123-weights/resolve/main/105000.ckpt?download=true + out=105000.ckpt +https://huggingface.co/cvlab/zero123-weights/resolve/main/165000.ckpt?download=true + out=165000.ckpt \ No newline at end of file