train with imagenet or not

[alicera]

how to train the network with pretrain imagenet

how to train the network with any init weight

using --pretrained to archive the target that pretrain from imagenet ?

it will train with initial weights without --pretrained?

# Model parameters
group = parser.add_argument_group('Model parameters')
group.add_argument('--model', default='resnet50', type=str, metavar='MODEL',
                   help='Name of model to train (default: "resnet50")')
group.add_argument('--pretrained', action='store_true', default=False,
                   help='Start with pretrained version of specified network (if avail)')
group.add_argument('--initial-checkpoint', default='', type=str, metavar='PATH',
                   help='Initialize model from this checkpoint (default: none)')

[rwightman]

@alicera it will not use pretrained weights if you do not use the --pretrained flag for the train script (or pass pretrained=True to the create_model funtion) ... if you want any weights, you can use --initial-checkpoint, BUT, the weights have to match the target model (classifier size) as it loads after creation...

if they don't match you need disable strict checkpoint loading, and remove the head from the state_dict in the checkpoint, but that requies a bit of code hacking...

Another hack if you want to to behave like a pretrained weight, you can locally modify the pretrained cfg for the model and add a file= field which will take priority over the url and hf_hub_id source specifications.

ie for example in efficientnet.py if you had weights for b5, change

    'efficientnet_b5.in12k_ft_in1k': _cfg(
        hf_hub_id='timm/',
        input_size=(3, 448, 448), pool_size=(14, 14), crop_pct=1.0, crop_mode='squash'),

to

    'efficientnet_b5.in12k_ft_in1k': _cfg(
        file='path/to/checkpoint.pth',
        num_classes=num_classes_in_my_checkpoint,
        hf_hub_id='timm/',
        input_size=(3, 448, 448), pool_size=(14, 14), crop_pct=1.0, crop_mode='squash'),

[alicera]

OK, THANKS

--pretrained will use the imageNet pretrain weight to training