Generation of saliency maps

There is a problem in the code src/ptbench/scripts/generate_saliencymaps.py on line 256 and 258. To solve this problem, change "model" to "state_dict"

To reproduce the error:

ptbench train -vv densenet montgomery --batch-size=4 --epochs=1

ptbench generate-saliencymaps -vv densenet montgomery --accelerator="cuda" --weight=results/model_final_epoch.ckpt --output-folder=path/to/visualizations

assigned to @ogueler

Hello @cardosol

Can you tell me more about this error, or provide me the densenet model file with something like google drive? It is not reproducible for me, although there was a different issue at line 268 - 269 with the code for densenet models that I have now fixed. Are you per chance training the model without using pytorch lightning? The code here might not work with models trained on previous ptbench versions without the lightning module.

changed title from generation of generate saliency maps to Generation of saliency maps

changed the description

@ogueler: we trained and tested on the same branch. Let me try this later and report back.

@cardosol: I can't reproduce the problem as well. There is an issue with your checkout, IMO - the --accelerator option is gone on the newest branches (i.e. add-datamodule and add-datamodule-gradcam), and this makes me suspicious of your setup. This set of commands work for me in a freshly checked out source:

$ ptbench train -vv montgomery densenet --device="cuda" --batch-size=4 --cache-samples --epochs=1 --parallel=6
$ ptbench generate-saliencymaps -vv densenet montgomery --device="cuda" --weight=results/model_final_epoch.ckpt --output-folder=vis

What doesn't work for me, @ogueler, is a device swap - this means: if I train on cuda and generate saliency maps on the local cpu (i.e., suppress the --device option), this will generate an exception. Can you please fix the generate-saliencymaps script (and all scripts that require a --device for that matter), so that the work can be carried on correctly, independently of the chosen device? To do so, you may have to cast the model to the new device. You will find more information on this link from the pytorch documentation.

For info, the above test also does not work if I choose --device="mps" (running on macOS), and then no device (i.e. use "cpu").

mentioned in commit 3920e964

mentioned in commit 2a7150d6

mentioned in commit 05c24a29

closed with commit 2a7150d6

closed with merge request !12 (merged)