Conversation
|
It looks like this PR includes changes unrelated to linear eval (perhaps from multi-gpu support?). Can you please fix that? |
I have put this pr before merging ICGAN, maybe that is the reason to have extra changes. I will update it today and put it up. |
| size (int): Image size. | ||
| """ | ||
| transform_list = [ | ||
| transforms.Resize(size=(size,size)), |
There was a problem hiding this comment.
For inference I believe SimCLR performs a random center crop before resizing. Can you please have a look at either their paper or the code to verify?
There was a problem hiding this comment.
In the code they didn't have any transformation rather than .ToTensor, but I thought this would fail on imagenet as we have various images sizes. However in the paper they mentioned "As preprocessing, all images were resized to 224 pixels along the shorter side using bicubic resampling, after which we took a 224 × 224 center crop." It is bit vague for me, should we first resize on smaller dimension and then centre crop? or should we first crop on smaller dimension and then resize?
| data_utils.CenterCropLongEdge(), | ||
| transforms.Resize((size, size)), | ||
| transforms.ToTensor(), | ||
| transforms.Normalize(self.config.norm_mean, self.config.norm_std), |
There was a problem hiding this comment.
Why did you remove this? Is it causing an error?
There was a problem hiding this comment.
I think this implementation should be available. Do you need additional capability beyond this? https://github.com/VectorInstitute/GenerativeSSL/blob/main/icgan/data_utils/utils.py#L29
| @@ -0,0 +1,51 @@ | |||
| #!/bin/bash | |||
|
|
|||
| #SBATCH --job-name=train_sunrgbd | |||
There was a problem hiding this comment.
I have added changes to run_simCLR.py in my PR that takes checkpoint_dir and creates a directory based on the time of running the training job. So let's revert these changes.
| from torch import nn | ||
| from torchvision import models | ||
|
|
||
| from ..exceptions.exceptions import InvalidBackboneError |
There was a problem hiding this comment.
Please use absolute paths.
| self.device_id = kwargs["device_id"] | ||
| self.writer = SummaryWriter() | ||
| # Create a directory to save the model checkpoints and logs | ||
| now = datetime.now() |
There was a problem hiding this comment.
I have added logic to run_simCLR.py that does something like this. So you have to either remove these parts, or remove my changes before merging.
| #SBATCH --nodes=1 | ||
| #SBATCH --gres=gpu:4 | ||
| #SBATCH --ntasks-per-node=4 | ||
| #SBATCH --ntasks-per-node=1 |
There was a problem hiding this comment.
Please keep in mind that this uses a single GPU per node.
| # “srun” executes the script <ntasks-per-node * nodes> times | ||
| srun python run_simCLR.py \ | ||
| # srun execute ntasks-per-node * nodes times | ||
| srun pythong run_simCLR.py \ |
There was a problem hiding this comment.
If the github repo from which we borrowed simclr implementation provides pretrained model checkpoints, can you please download one of them and test this script on it to see if the results match with what's reported by them?
afkanpour
left a comment
afkanpour
left a comment
There was a problem hiding this comment.
Do we still need this? If not, should we delete it?
Reviewable status: 0 of 19 files reviewed, 30 unresolved discussions (waiting on @sanaAyrml and @vahid0001)
2 participants
PR Type
[Feature ]
Short Description
I added evaluation module in this new PR which adds a classification head to pre-trained backbone. Then it freezes backbone and only train classification head and report accuracy over train and test data.
This change is