loss doesn't decrease and keep on 40 in my dataset

[guapizyq]

I train with my own dataset with the default script
my dataset has three types of objection with 40000 train sample and 6000 validation sample
the batch_size is 4, my loss decrease to 49 in the first epoch, but loss keep on 40 in the 13th epoch, the loss doesn't decrease anymore.
how should i change my learning rate and others parameter

[guapizyq]

it would be appreciated if you give some advise on this issue

[Adamdad]

Can you detect anything in the testset？if not，what is your learning rate,lr decay pacience?

[guapizyq]

I am sorry, it seems to my mistakes.
the first training is to get a stable loss
the second training without frozen layers is to get a lower loss?

[Adamdad]

correct

[guapizyq]

I got a lower loss than 40, but it still is 38
the result in my testdset looks ok, i do not konw how to decrease the loss

[zhangyufei1995]

@guapizyq @Adamdad I am very happy to discuss with you. What I want to ask is, 1. How much is the epoch setting of the red arrow here? 2, how much is the initial_epoch setting of the two black arrows? I look forward to your answer.

[zhangyufei1995]

And why is it divided into step-by-step training?the first training? the second training ?What is their role?

[Adamdad]

The first training part is for finetuning a model quickly. It freezes most layers, only to train on the last few layers. We can get an acceptable model for detection in a short period of time

The second training part is for getting a complete model. All the layers can be trained through this process.

Under most occasions, I only use the second part. Epoch under is not important here.

[gzz1529657064]

The first training part is for finetuning a model quickly. It freezes most layers, only to train on the last few layers. We can get an acceptable model for detection in a short period of time

The second training part is for getting a complete model. All the layers can be trained through this process.

Under most occasions, I only use the second part. Epoch under is not important here.

After model training, I have a model with size of 277M. It is bigger than YOLO-v3,Why?
Doesn't MobileNet reduce model parameters?
This is my training strategy in my dateset.

1. Unfreeze all of the layers
2. learning_rate = 0.001
3. load_pretrained=False
4. batch_size = 16