kentaroy47 / efficientdet.pytorch

:neckbeard:Unofficial implementation of EfficientDet

Now training models..

backbone	resolution	VOCmAP	COCOmAP	Inference[ms]	model
EfficientnetB0(wo/BiFPN)	512	77.0	TBD
EfficientnetB0(w/BiFPN)	512	77.2	TBD
EfficientnetB2(wo/BiFPN)	768	TBD	TBD
EfficientnetB2(w/BiFPN)	768	TBD	TBD
EfficientnetB4(wo/BiFPN)	1024	TBD	TBD
EfficientnetB4(w/BiFPN)	1024	TBD	TBD
EfficientnetB6(wo/BiFPN)	1408	TBD	TBD
EfficientnetB6(w/BiFPN)	1408	TBD	TBD

efficientdet.pytorch

unofficial implementation of efficientdet

• supported efficientnet backbone (B0 for now)
• supported resnet backbone.
• supports BiFPN.
• supports VOC and COCO training.

requirements

pip install --upgrade efficientnet-pytorch

PASCAL-VOC training

• see train_efficientdet.ipynb
• see eval.ipynb for evaluation

COCO training

• see train_efficientdet.ipynb and configure MS-COCO dataset.

FP16 training

• see train_fp16_apex.ipynb

efficientdet.pytorch's Issues

Could you please provide the coco eval script?

Just as the title describe.

Working implementation for bifpn

You can find the working implementation for bifpn in the link mentioned below.
https://github.com/ViswanathaReddyGajjala/EfficientDet-Pytorch

why the BiFPN output channel is 256?

# using scaled BiFPN channels did not work in my implementation.
 # here, we use outc=256 for all settings.
outc = 256

In Google's paper the BiFPN channel is 64 in EfficientDet D0 and D2, it is 256 in your code, which confused me a lot. Please explain why you change it to 256. Thanks a lot.

ValueError: Cannot load file containing pickled data when allow_pickle=False

here is the error i got when execute python train.py --config configs/training/EcgResNet34.json

Trainer: Trainer1D <class 'trainers.trainers.Trainer1D'>
/usr/local/lib/python3.7/dist-packages/torch/utils/data/dataloader.py:481: UserWarning: This DataLoader will create 3 worker processes in total. Our suggested max number of worker in current system is 2, which is smaller than what this DataLoader is going to create. Please be aware that excessive worker creation might get DataLoader running slow or even freeze, lower the worker number to avoid potential slowness/freeze if necessary.
cpuset_checked))
Epoch - 1
Traceback (most recent call last):
File "train.py", line 20, in
trainer.loop()
File "/content/drive/MyDrive/researchHub/ecg-classification/trainers/base_trainer.py", line 141, in loop
self.train_epoch()
File "/content/drive/MyDrive/researchHub/ecg-classification/trainers/base_trainer.py", line 65, in train_epoch
for i, batch in enumerate(self.train_loader):
File "/usr/local/lib/python3.7/dist-packages/torch/utils/data/dataloader.py", line 521, in next
data = self._next_data()
File "/usr/local/lib/python3.7/dist-packages/torch/utils/data/dataloader.py", line 1203, in _next_data
return self._process_data(data)
File "/usr/local/lib/python3.7/dist-packages/torch/utils/data/dataloader.py", line 1229, in _process_data
data.reraise()
File "/usr/local/lib/python3.7/dist-packages/torch/_utils.py", line 434, in reraise
raise exception
ValueError: Caught ValueError in DataLoader worker process 0.
Original Traceback (most recent call last):
File "/usr/local/lib/python3.7/dist-packages/torch/utils/data/_utils/worker.py", line 287, in _worker_loop
data = fetcher.fetch(index)
File "/usr/local/lib/python3.7/dist-packages/torch/utils/data/_utils/fetch.py", line 49, in fetch
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/usr/local/lib/python3.7/dist-packages/torch/utils/data/_utils/fetch.py", line 49, in
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/content/drive/MyDrive/researchHub/ecg-classification/dataloaders/dataset1d.py", line 17, in getitem
img = np.load(self.data[index]["path"]).astype("float32")
File "/usr/local/lib/python3.7/dist-packages/numpy/lib/npyio.py", line 445, in load
raise ValueError("Cannot load file containing pickled data "
ValueError: Cannot load file containing pickled data when allow_pickle=False

Training_issue

Hello,
Could you please tell me which version of pytorch you used ? Or you can give the details of the environment. I am getting the tensor size issue while training with VOC dataset.

Best regards

I got 0.759map on voc with you code. can you provide your trained-model to me？

I got 0.759map on voc with you code. can you provide your trained-model to me？

ModuleNotFoundError: No module named 'efficientnet_pytorch'

File "train_efficientdet.py", line 111, in
from utils.efficientdet import EfficientDet
File "/EfficientDet/efficientdet.pytorch/utils/efficientdet.py", line 1, in
from efficientnet_pytorch import EfficientNet
ModuleNotFoundError: No module named 'efficientnet_pytorch'

Thanks!

the resolution for 79.0(map) with VOC2017 is 300 or 512?

the resolution for 79.0(map) with VOC2017 is 300 or 512?

Question training with own dataset and FPS

Hi Kentaroy,
Thank you for your great job.

• I would like train my own dataset. Do I change my dataset to COCO and there is different to modify the scale anchor?
• Do you have train with D1 and D2 and how is the result of FPS ans score?

Thank you again.

efficient net as feature extractor

Thanks for your torch implementation. I am interested in getting features of image same as we do in Resnet. Is it possible to do via your implementation? If yes, can you please let me know how to use that.
Like in resnet I can do like this way-

from model.res50 import ResNet
 self.encoder = ResNet()
self.fc = nn.Linear(hidden_dim, num_classes)

VOC mAP reasonable, when will D1 mAP release?

It seems this implementation is more reasonable then other implemetations. Have u trained D1 for now and how's that performance?