johannesu / nasnet-keras Goto Github PK

How can I use this implementation to train on my own data?

Hi, in paper (Figure4, ReductionCell) there are only 3 layer for Concate input , as you code (line 300 in nasnet.py) ,it should be
return Concatenate(.....)([add_2, add_3 add_4])
add_1 is not in Concate input!!!! please check it.

I tried to finetune NASNet like a Keras K.application model however, there is no weights kwarg. For example, for Xception it defaults to loading the imagenet pretrained weights.

I tried to use some of the functions provided in the package as follows,

import nasnet

origin = 'https://storage.googleapis.com/download.tensorflow.org/models/nasnet-a_large_04_10_2017.tar.gz'
fname='nasnet_large'
md5_hash='5286bdbb29bab27c4d3431c70f8becf9'
cache_dir = os.path.expanduser(os.path.join('~', '.keras', 'models'))

model = nasnet.NASNetA(include_top=False, input_shape=(img_height, img_width, 3))

nasnet.load_pretrained_weights(model, fname, origin, md5_hash,
                               skip_first_dense=True, cache_dir=cache_dir)

However, this throws the error,

ValueError: Layer weight shape (1, 1, 96, 8) not compatible with provided weight shape (1, 1, 96, 42)

What is going wrong here? Better still, it would be great to get a weights kwarg built in to make it fully follow the Keras API. Many thanks.

File "nasnet.py", line 229
dropout_rate=0.5) -> Model:
^
SyntaxError: invalid syntax
@johannesu

Hi, thank you for your sharing. And I am confused about penultimate_filters !!! Is it only used in Line 354(nasnet.py) ??