Transfer learning with Resnet50 fail with Exception about deep-learning-models HOT 5 CLOSED

I solved the problem after reading through source code for Resnet50.py. If anyone encounter the same problem, please add the input_shape parameter when you call Resnet50():

model = ResNet50(include_top=False, weights='imagenet', input_tensor=Input(shape=(224,224,3)))

If you are using Keras ImageDataGenerator, the input shape will be automatically inferred.

from deep-learning-models.

Regarding the error:

input size must be at least 197x197,

Had the same issue, but disappeared after upgrading keras to 2.2.4 (with backend tensorflow 1.12.0 in case it matters).

from deep-learning-models.

@MrXu
hi, i also use the resnet50, and i want set input_size=(48,48,3) but it show me an error.

valueerror input size must be at least 197x197,

i check the doc, it show me ,

width and height should be no smaller than 32. E.g. (200, 200, 3) would be one valid value.,

It's so weird.

from deep-learning-models.

You should use the functional API. So for example try:

  if K.image_dim_ordering() == 'tf':
    inp = Input(shape=(224, 224, 3), name='input_image')
else:
    inp = Input(shape=(3, 224, 224), name='input_image')


main_model = Resnet50(include_top=False)(inp)

main_out = Dropout(0.5)(main_model)
main_out = Dense(512, activation='relu', name='fcc_0')(main_out)
main_out = Dropout(0.5)(main_model)
main_out = Dense(output_classes, activation='softmax', name='class_id')(main_out)

model = Model(input=inp, output=artist_out)

from deep-learning-models.

@AntreasAntoniou Thanks, but I need to load weights for the Sequential model. Also, when I try to use main_model = Resnet50(include_top=False)(inp), it complains Exception: You are attempting to share a same BatchNormalization layer across different data flows.

from deep-learning-models.