I have tried to fine-tune the Resnet50_Final.pth on my own data with annotations, but the results look like the model was trained from the beginning without uploading pretrained weights. What could be a problem?
Current config below:
train:
training_dataset: ml/output/face_detection/random350_frames_faces/annotations_training/annotations.txt
network: resnet50 # Backbone network mobile0.25 or resnet50
num_workers: 4 # Number of workers used in dataloading
lr: 0.001 # initial learning rate
momentum: 0.9 # the gradient of the past steps to determine the direction to go
resume_net: ./weights/pretrained/Resnet50_Final.pth # path to pretrained weights
resume_epoch: 0 # for pretrained weights, epoch on which training was ended
weight_decay: 0.0005 # Weight decay for SGD
gamma: 0.1 # Gamma update for SGD
save_folder: ml/experiments/face_detection/weights/train_results
cfg_re50:
name: Resnet50
min_sizes: [[16, 32], [64, 128], [256, 512]]
steps: [8, 16, 32]
variance: [0.1, 0.2]
clip: false
loc_weight: 2.0
gpu_train: true
batch_size: 6
ngpu: 1
epoch: 100
decay1: 70
decay2: 90
image_size: 840
pretrain: true
return_layers:
layer2: 1
layer3: 2
layer4: 3
in_channel: 256
out_channel: 256
Hmm, actually, I didn't try it with resume_epoch, but I will.
Maybe someone has other thoughts?
I have tried to fine-tune the Resnet50_Final.pth on my own data with annotations, but the results look like the model was trained from the beginning without uploading pretrained weights. What could be a problem? Current config below:
Hmm, actually, I didn't try it with
resume_epoch
, but I will. Maybe someone has other thoughts?