How to use CLIP-ReID as feature extractor? about clip-reid

[awarebayes]

def forward_override(self, x: torch.Tensor, cv_emb = None, old_forward = None):
    _, image_features, image_features_proj = old_forward(x, cv_emb)
    return torch.cat([image_features[:,0], image_features_proj[:,0]], dim=1)
    
def  main():

    model_path = cfg.TEST.WEIGHT
    batch_size = cfg.TEST.IMS_PER_BATCH
    output_dir = cfg.OUTPUT_DIR
    output_name = cfg.MODEL.OUTPUT_NAME

    train_loader, train_loader_normal, val_loader, num_query, num_classes, camera_num, view_num = make_dataloader(cfg)
    model = make_model(cfg, num_class=num_classes, camera_num=camera_num, view_num = view_num)

    logger.info(f"Loading model for eval from {model_path}. Batch size = {batch_size}")

    model.load_param(model_path)
    model = model.image_encoder

    old_forward = model.forward
    model.forward = lambda *args, **kwargs: forward_override(model, old_forward=old_forward, *args, **kwargs)

    device = torch.device('cuda:0')
    model = model.eval().to(device)

    image_dir = Path(args.test_dir)
    dataset = ReidDataset(image_dir, cfg)
    loader = DataLoader(dataset, batch_size=batch_size, shuffle=False, num_workers=8, pin_memory=True)
    for i, (images, person_ids) in loader:
            images = images.to(device)
            vectors = model(images)

[awarebayes]

We are working on some fixes on the repo.
If you are using cosine distance for your descriptors, just change euclidean_dist to cosine_dist, used for ranking.
https://github.com/Syliz517/CLIP-ReID/blob/master/loss/triplet_loss.py#L123C26-L123C26

We have found it makes metrics worse, but separation with cosine distances is way better. Can show you the graphs of average distances between same vs others

[mikel-brostrom]

Thanks for the example code @awarebayes! Really appreciate it

[mikel-brostrom]

If you are using cosine distance for your descriptors, just change euclidean_dist to cosine_dist, used for ranking

Have you visualized the feature learning process? Would make sense if the features tend to cluster like beams shooting out of origo. Would be enlightening if you could share the graphs ✨

[mikel-brostrom]

Let me know when this is in usable state 😄 . I have been trying to get the MSMT17_clipreid_12x12sie_ViT model working with no luck.

Traceback (most recent call last):
  File "/home/mikel.brostrom/CLIP-ReID/feature_extraction_test.py", line 52, in <module>
    main()
  File "/home/mikel.brostrom/CLIP-ReID/feature_extraction_test.py", line 40, in main
    model.load_param(model_path)
  File "/home/mikel.brostrom/CLIP-ReID/model/make_model.py", line 121, in load_param
    self.state_dict()[i.replace('module.', '')].copy_(param_dict[i])
RuntimeError: The size of tensor a (2) must match the size of tensor b (15) at non-singleton dimension 0

You can reproduce this with:

import os
import torch
import yaml
import argparse
from config import cfg_base as cfg
from model.make_model import make_model

def forward_override(self, x: torch.Tensor, cv_emb = None, old_forward = None):
    _, image_features, image_features_proj = old_forward(x, cv_emb)
    return torch.cat([image_features[:,0], image_features_proj[:,0]], dim=1)
    
def  main():
    parser = argparse.ArgumentParser(description="ReID Baseline Training")
    parser.add_argument(
        "--config_file",
        default="/home/mikel.brostrom/CLIP-ReID/MSMT17_clipreid_12x12sie_ViT-B-16_60_test_log.yml",
        help="path to config file",
        type=str
    )
    parser.add_argument(
        "opts",
        help="Modify config options using the command-line",
        default=None,
        nargs=argparse.REMAINDER
    )

    args = parser.parse_args()

    if args.config_file != "":
        cfg.merge_from_file(args.config_file)
    cfg.merge_from_list(args.opts)
    cfg.freeze()

    model = make_model(cfg, num_class=1501, camera_num=2, view_num = 1)

    model_path = 'MSMT17_clipreid_12x12sie_ViT-B-16_60.pth'
    batch_size= 1
    print(f"Loading model for eval from {model_path}. Batch size = {batch_size}")

    model.load_param(model_path)
    model = model.image_encoder

    old_forward = model.forward
    model.forward = lambda *args, **kwargs: forward_override(model, old_forward=old_forward, *args, **kwargs)

    device = torch.device('cuda:0')
    model = model.eval().to(device)

    
main()

Had to delete some stuff from the configs as well to get to the weight loading part...

[awarebayes]

You can override config parameters as:

--config_file
src/configs/person/vit_clipreid.yml
--test_dir
"./dataset/reid/market1501/bounding_box_test"
MODEL.DEVICE_ID
"'0'"
DATASETS.NAMES
"market1501"
DATASETS.ROOT_DIR
"./dataset/reid"
TEST.WEIGHT
output/clipreid.pt
SOLVER.SEED
42

[awarebayes]

As for the graphs, I am doing one vs rest cosine distance distribution.

When ranking with cosine:

When ranking with euclidean:

Do you have any examples on how to 'visualize the feature learning process'?

[mikel-brostrom]

Do you have any examples on how to 'visualize the feature learning process'?

https://github.com/KaiyangZhou/pytorch-center-loss

[awarebayes]

They have a bottleneck of size 2, which they visualize. I dont think we can have that with person reid

self.fc1 = nn.Linear(128*3*3, 2)
self.prelu_fc1 = nn.PReLU()
self.fc2 = nn.Linear(2, num_classes)

[mikel-brostrom]

You can override config parameters as:

Thx, but it seems that the loaded ViT-B-16 is not matching the one in MSMT17_clipreid_12x12sie_ViT-B-16_60.pth

[awarebayes]

import os
import torch
import yaml
import argparse

from PIL import Image

from config import cfg
import torchvision.transforms as T

from model.make_model import make_model


def forward_override(self, x: torch.Tensor, cv_emb=None, old_forward=None):
    _, image_features, image_features_proj = old_forward(x, cv_emb)
    return torch.cat([image_features[:, 0], image_features_proj[:, 0]], dim=1)


def main():
    parser = argparse.ArgumentParser(description="ReID Baseline Training")
    parser.add_argument(
        "--config_file",
        default="",
        help="path to config file",
        type=str
    )
    parser.add_argument(
        "opts",
        help="Modify config options using the command-line",
        default=None,
        nargs=argparse.REMAINDER
    )

    args = parser.parse_args()

    if args.config_file != "":
        cfg.merge_from_file(args.config_file)
    cfg.merge_from_list(args.opts)
    cfg.freeze()

    model = make_model(cfg, num_class=751, camera_num=2, view_num=1)

    model_path = '/home/mscherbina/Downloads/Market1501_clipreid_ViT-B-16_60.pth'
    batch_size = 1
    print(f"Loading model for eval from {model_path}. Batch size = {batch_size}")

    model.load_param(model_path)
    model = model.image_encoder

    old_forward = model.forward
    model.forward = lambda *args, **kwargs: forward_override(model, old_forward=old_forward, *args, **kwargs)

    device = torch.device('cuda:0')
    model = model.eval().to(device)

    transforms = T.Compose([
        T.Resize([256, 128]),
        T.ToTensor(),
        T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
    ])

    image_dir = '....'
    for i in os.listdir(image_dir):
        img = Image.open(f"{image_dir}/{i}").convert("RGB")
        img = transforms(img).to(device)
        img = img.unsqueeze(0)
        with torch.no_grad():
            feats = model(img)
            print("i", i, feats.size())




main()

Tested with ViT-CLIP-ReID Market

With that you can keep config the default one

from config import cfg

[mikel-brostrom]

I changed the following content in:

config/defaults.py

-_C.MODEL.NAME = 'resnet50'
+_C.MODEL.NAME = 'ViT-B-16'

and adapted model_path in the code above. I am on master with no modifications. Run the above script as:

python feature_extraction_test.py

but get the following error

Loading model for eval from /home/mikel.brostrom/CLIP-ReID/MSMT17_clipreid_12x12sie_ViT-B-16_60.pth. Batch size = 1
Traceback (most recent call last):
  File "/home/mikel.brostrom/CLIP-ReID/feature_extraction_test.py", line 73, in <module>
    main()
  File "/home/mikel.brostrom/CLIP-ReID/feature_extraction_test.py", line 48, in main
    model.load_param(model_path)
  File "/home/mikel.brostrom/CLIP-ReID/model/make_model.py", line 121, in load_param
    self.state_dict()[i.replace('module.', '')].copy_(param_dict[i])
KeyError: 'cv_embed'

[awarebayes]

I changed load_param to be

    def load_param(self, trained_path):
        param_dict = torch.load(trained_path)
        for i in self.state_dict():
            self.state_dict()[i.replace('module.', '')].copy_(param_dict[i])
        print('Loading pretrained model from {}'.format(trained_path))

[mikel-brostrom]

Ok, think I am loading the SIE-OLP model. Let me try the one you linked

[mikel-brostrom]

Loading Market1501_clipreid_ViT-B-16_60.pth now and getting

Loading model for eval from /home/mikel.brostrom/CLIP-ReID/Market1501_clipreid_ViT-B-16_60.pth. Batch size = 1
Traceback (most recent call last):
  File "/home/mikel.brostrom/CLIP-ReID/feature_extraction_test.py", line 73, in <module>
    main()
  File "/home/mikel.brostrom/CLIP-ReID/feature_extraction_test.py", line 48, in main
    model.load_param(model_path)
  File "/home/mikel.brostrom/CLIP-ReID/model/make_model.py", line 121, in load_param
    self.state_dict()[i.replace('module.', '')].copy_(param_dict[i])
RuntimeError: The size of tensor a (193) must match the size of tensor b (129) at non-singleton dimension 0

[mikel-brostrom]

Still same issue after updating load_param

[awarebayes]

Hmm, which key is it?
Also do you have telegram, we could go there

[mikel-brostrom]

Sorry, no telegram, only whatsapp 😞

[mikel-brostrom]

Added a print in load_param:

def load_param(self, trained_path):
        param_dict = torch.load(trained_path)
        for i in self.state_dict():
            print(i)
            self.state_dict()[i.replace('module.', '')].copy_(param_dict[i])
        print('Loading pretrained model from {}'.format(trained_path))

It stops at positional_embedding:

bottleneck.weight
bottleneck.bias
bottleneck.running_mean
bottleneck.running_var
bottleneck.num_batches_tracked
bottleneck_proj.weight
bottleneck_proj.bias
bottleneck_proj.running_mean
bottleneck_proj.running_var
bottleneck_proj.num_batches_tracked
image_encoder.class_embedding
image_encoder.positional_embedding
Traceback (most recent call last):
  File "/home/mikel.brostrom/CLIP-ReID/feature_extraction_test.py", line 73, in <module>
    main()
  File "/home/mikel.brostrom/CLIP-ReID/feature_extraction_test.py", line 48, in main
    model.load_param(model_path)
  File "/home/mikel.brostrom/CLIP-ReID/model/make_model.py", line 122, in load_param
    self.state_dict()[i.replace('module.', '')].copy_(param_dict[i])
RuntimeError: The size of tensor a (193) must match the size of tensor b (129) at non-singleton dimension 0

[awarebayes]

I believe config is to blame:

        self.h_resolution = int((cfg.INPUT.SIZE_TRAIN[0]-16)//cfg.MODEL.STRIDE_SIZE[0] + 1)
        self.w_resolution = int((cfg.INPUT.SIZE_TRAIN[1]-16)//cfg.MODEL.STRIDE_SIZE[1] + 1)
        self.vision_stride_size = cfg.MODEL.STRIDE_SIZE[0]
        clip_model = load_clip_to_cpu(self.model_name, self.h_resolution, self.w_resolution, self.vision_stride_size)
        clip_model.to("cuda")

Config I load with is https://gist.github.com/awarebayes/271beb52cabc9cf0bc77f592764e1b62, maybe change it.

[mikel-brostrom]

Stride is 16 in both confs...

[awarebayes]

At the end of the day I have these values here:
Ignore 457, its #persons in my dataset

[awarebayes]

By the way I noticed ViT with tensorrt + fp16 does not work any slower than resnet50

[mikel-brostrom]

ok. Fixed.

# Size of the image during training
_C.INPUT.SIZE_TRAIN = [256, 128]
# Size of the image during test
_C.INPUT.SIZE_TEST = [256, 128]

instead of

# Size of the image during training
_C.INPUT.SIZE_TRAIN = [384, 128]
# Size of the image during test
_C.INPUT.SIZE_TEST = [384, 128]

[mikel-brostrom]

Thx for you patience 😄.

[mikel-brostrom]

I am kind of surprised by the inference time. For a torch.Size([1, 3, 256, 128]) input, I got 0.058 seconds averaged over 1000 images on an old NVIDIA p2000. Didn't though it would be that low, even taking into consideration the small input size.

[mikel-brostrom]

0.37 seconds for an torch.Size([10, 3, 256, 128]) input. In case anybody is interested 😄.

[mikel-brostrom]

Are you responsible for this repo now @awarebayes? If this is the case would you be okay with me implementing these models here: https://github.com/mikel-brostrom/yolo_tracking. @Syliz517 ? I intend to use them for associating detections across frames both by visual means only, as well as motion plus appearance information.

[mikel-brostrom]

By the way I noticed ViT with tensorrt + fp16 does not work any slower than resnet50

Any slower than the regular pt resnet weights? Do you have any plans of adding a TensorRT converter to this repo?

[mikel-brostrom]

As for the graphs, I am doing one vs rest cosine distance distribution.

Clearly cosine distance separates the embeddings better. I am however not entirely sure why the large tails on the distributions.

[awarebayes]

I can help you with implementing clip reid into yolo tracking.
I just got tasked to fix this repo at work.
TensorRT converter is unnessesary since it has ONNX, and ONNX converts to tensorrt natively using trt-exec

[mikel-brostrom]

I have it working in real-time now

[mikel-brostrom]

git clone https://github.com/mikel-brostrom/yolo_tracking
git checkout clip-reid
python examples/track.py --reid-model clip_market1501.pt
# or
python examples/track.py --reid-model clip_duke.pt

haven't evaluated on any MOT dataset yet

[awarebayes]

I dont know if keeping this guys repo issues section as a communication medium is a good idea, but whatever.
Have you tried using video reid nets?

https://paperswithcode.com/task/video-based-person-re-identification

They basically work the same way, but during learning, they take an average of the tracklet's features instead of individual image's.

[mikel-brostrom]

I dont know if keeping this guys repo issues section as a communication medium is a good idea

I think I have telegram, just uninstalled it because didn't use it. Let me check

Have you tried using video reid nets?

Interesting, never heard of this! Could be quite heavy though for real-time as crop-outs are stacked. This is however handled to some extent in modern multi-object trackers as individual feature-maps are averaged over several frames. Like here:
https://github.com/NirAharon/BoT-SORT/blob/251985436d6712aaf682aaaf5f71edb4987224bd/tracker/bot_sort.py#L41

[Syliz517]

Hi! Would first of all like to know whether your are okay with me implementing these models here: https://github.com/mikel-brostrom/yolo_tracking. Then I would also like to know if there is any easy way of extracting features with these models. Keep up the great work!

Of course, I notice you had some problems before, have you resolved them now?

[mikel-brostrom]

Thanks for your response @Syliz517! @awarebayes helped me out to get it working 😄