"load weight error" and "Not compiled with GPU support"

[FJR-Nancy]

While torch.load('mq-glip-l') , there is a RuntimeError: "PytorchStreamReader failed reading zip archive: not a ZIP archive". mq-glip-t is the same as well. mq-glip-l and mq-glip-t are downloaded from https://drive.google.com/file/d/1O_eb1LrlNqpEsoxD23PAIxW8WB6sGoBO/view and https://drive.google.com/file/d/1n0_D-tisqN5v-IESUEIGzMuO-9wolXiu/view

[YifanXu74]

Hi,
I didn't encounter any problem when following these steps:

1. Download mq-glip-t from google drive.
2. Load the checkpoint: ckpt = torch.load("mq-glip-t", map_location="cpu").

The problem might be due to incomplete file downloads. Please verify the file sizes and redownload the checkpoint files. The expected file sizes are as follows:
mq-glip-l: 1.8G
mq-glip-t: 1.1G
Let me know if you need any further assistance.

[FJR-Nancy]

The download is complete and the downloaded file sizes are as follows:
mq-glip-l: 2.0G
mq-glip-t: 1.2G
By the way, my python version is 3.8, because 3.9 is incompatible with my environment. Maybe this is the reason why the weight could not be loaded.

[YifanXu74]

Hi, I didn't encounter any problem when following these steps:

1. Download mq-glip-t from google drive.
2. Load the checkpoint: ckpt = torch.load("mq-glip-t", map_location="cpu").

The problem might be due to incomplete file downloads. Please verify the file sizes and redownload the checkpoint files. The expected file sizes are as follows: mq-glip-l: 1.8G mq-glip-t: 1.1G Let me know if you need any further assistance.

I tested the mentioned process on three different environments: python3.8&torch1.5.0, python3.8&torch2.0.1, and python3.8&torch2.1.0. In all cases, the process executed successfully.
Could you provide the details of your implementation environment, including the versions of python and torch, and a detailed procedure to reproduce the error?

[FJR-Nancy]

I also tried python 3.9 in another environment but it's all the same. I also downloaded the weights several times with different device, but nothing changed. My environment is Python 3.8/3.9 with torch 2.0.1+cu117. Just tried "torch.load('mq-glip-l')" and failed.

[YifanXu74]

This is very strange, and at the moment, I'm unsure how to resolve it. I tested it with different devices and there were no such issues. Additionally, I have successfully utilized the checkpoints in the past to reproduce the model, so I am certain that the released checkpoints are valid.

[FJR-Nancy]

Sorry, I found the problem is that some data is lost while transferring the weight to cluster. So the checkpoints are valid.

However, there is another problem in finetuning-free evaluation for custom dataset. "RuntimeError: Not compiled with GPU support" is reported in _C.modulated_deform_conv_forward() of deform_conv.py. Is only CPU supported by now?

[YifanXu74]

The code is GPU-compatible. This error might be due to your environment, such as an incorrect CUDA installation or using a CPU version of torch.

[FJR-Nancy]

The GPU problem is solved following microsoft/GLIP#41.

But while finetuning, an error "too many values to unpack" happens in "for iteration, (images, targets, idxs, positive_map, positive_map_eval, greenlight_map) in enumerate(data_loader, start_iter):" (line 92 of maskrcnn_benchmark/engine/trainer.py). What is this problem?

[YifanXu74]

Hi, this issue seems unrelated to the previous topic. In order to better track and address it, would you mind opening a new issue to discuss this problem? It would be helpful if you could provide detailed instructions or commands to reproduce the error you encountered. And I will handle this issue tomorrow.

[FJR-Nancy]

ok