Сould you please clarify, It's ok to have pool size less than stride size?
eg nasnet.py line 620
p1 = AveragePooling2D((1, 1), strides=(2, 2), padding='valid', name='adjust_avg_pool_1_%s' % id)(p)

There is an error in nasnet.py. Line 533f are just a repetition of the lines before. In the original implementation, the input to the second branch is padded and then shifted by one. This can be done like this:

p2 = ZeroPadding2D(padding=((0, 1), (0, 1)))(x)
p2 = Cropping2D(cropping=((1, 0), (1, 0)))(p2)

Which is taken from here: https://github.com/johannesu/NASNet-keras/blob/master/nasnet.py

Hi, there. I encountered a problem: KeyError: "Unable to open object (object 'global_average_pooling2d_1' doesn't exist)".
Here my code snippet:

model_name = "nasnet_large_500_300"

net_model = NASNetLarge(input_shape=input_shape, weights='imagenet', include_top=False, pooling='avg')
# append classification layer
x = net_model.output
final_output = Dense(nb_classes, activation='sigmoid', name='fc11')(x)

model = Model(inputs=net_model.input, outputs=final_output)

Could someone help me with this?
UPD: As I found keras couldn't load weights from h5 file

there is no train script, how can i train this model rather than predict

Hi, thank you for fixing nasnet's low accuracy problem. But after I updated the code and weights follow your lead, I found that every time I predict on "elephant.jpg", it got a different score. Here is my prediction code and some results:
input_img = image.load_img('elephant.jpg', target_size=(224,224))
img = image.img_to_array(input_img)
img = np.expand_dims(img, axis=0)
img = preprocess_input(img)
x = tf.keras.applications.nasnet.NASNetMobile(include_top=True,dropout=0, weights='imagenet',
input_tensor=None, input_shape=(224,224,3),
)
preds = x.predict(img)
print('Predicted:', decode_predictions(preds, top=3)[0])

Predicted: [('n02504458', 'African_elephant', 0.6310287), ('n02504013', 'Indian_elephant', 0.16810806), ('n01871265', 'tusker', 0.060016993)]
Predicted: [('n02504458', 'African_elephant', 0.68221015), ('n02504013', 'Indian_elephant', 0.17037697), ('n01871265', 'tusker', 0.057558082)]

I would really appreciate it if you could help me with this problem.

Thanks in advance.

Would it be possible to extend the functionality beyond just static input sizes? This is useful when we are using this as a backbone to a fully convolutional model. When I create a model with spatial input dimensions as None then the model breaks during inference. because the spatial dimension of two layers being added is off by a factor of 2.

hello @titu1994 , thanks for your nice work! I want to know if there any tutorials to guide me how to train NASNet with my own dataset. Looking forward to your reply ：）

Due to some network issues, urlretrieve raises error for me when it tries to download model. My solution to that normally is to manually download the .h5 file and put it in the .keras/models file. However, when I did the same here for NASNet the error still persisted. I did some digging and realized that the filed downloaded is named 'NASNet_large_no_top.h5' with capitalizations, but in your nasnet.py you are passing in 'nasnet_large_no_top.h5' with all lower case to keras_utils.getfile(). I fixed the error by renaming the .h5 file accordingly. I know this is a small thing, but it is not immediately obvious and could be frustrating to deal with. Is it possible to change the filename to the appropriate lower case version?

Hi, I’m really sorry for this kind of spam but I really need your help.

The reason why I’m forwarding this message is that I’ve been working for a month on a NEW ACTIVATION FUNCTION that according to my experiments works better tan RELU, ELU, LEAKY RELU, SELU, PRELU, SWISH…
Or any other activation function.
I’ve been paying AWS GPUs for the experiments from my own money and working on this as a side project (I’m 17 and still at High School), but now I need your help in order to publish it on the internet for open access to everyone since Arxiv asks me for endorsements. I think the paper is ready for publishing and we could talk about it more extensively if you wished to.

Any help will be much appreciated.
Thanks in advance,
Eric Alcaide.

This is error which i observed during the execution,
ValueError: Error when checking target: expected predictions to have shape (None, 10) but got array with shape (48, 2)

I used this NASNet for my data which is having two classes. Can you guide me to resolve this issue

I tried to use NASNET you wrote, I find adjust_block function in nasnet.py has a bug:
BatchNormalization should have diff name, but it is not now. I changed these name to adjust_bn_1 and adjust_bn_2_ then pass

Kindly help me sort it out. I am using the model with the following parameters:
model = NASNet(classes=2,input_shape=(224, 224, 3), weights=None, penultimate_filters=4032,
nb_blocks=6,use_auxilary_branch=True, skip_reduction=False)

This model as it stands is not useful for custom datasets since there is no easy way to save weights with Model Checkpoint. The layer names are too long. Im hacking my way through trying to figure out a solution but the fact that weights are saveable (at least in Keras 2.1.3) should come with a disclaimer. I realize it is an HD5 issue but still extremely frustrating.

ModelCheckpoint is already set to save weights only and save best only.

I think line 646 in nasnet.py is missing the addition of x2 and x5.

Thanks for the amazing work
I am having issue regarding network learning. My NASNet model ain't learning. Training accuracy is improving but validation accuracy isn't changing and stuck at 0.4194.
Training data=600 imgs, testing data=62 imgs
Image shape= (224,224,3)
Epochs=10-15

Looks like this parameter not used
https://github.com/titu1994/Keras-NASNet/blob/master/nasnet.py#L117

Have you managed to train NASNet Cifar on the cifar10 dataset? If yes, which results have you obtained? Thanks in advance.

edited - thought this was related to include_top, but seems not

e.g.

from nasnet import NASNet
n = NASNet()

gives

ValueError: Operands could not be broadcast together with shapes (331, 331, 168) (331, 331, 96)

I guess this is the same as this, however I'm using tensorflow 1.6

Thanks for the extremely useful work! :)

I run the code on readme.md

from nasnet import NASNet
model = NASNet(input_shape=(331, 331, 3),
           penultimate_filters=4032,
           nb_blocks=6,
           stem_filters=96,
           skip_reduction=True,
           use_auxilary_branch=False,
           filters_multiplier=2,
           dropout=0.5,
           classes=1000)

And I got.

Traceback (most recent call last): File "<input>", line 12, in <module> File "/home/palm/PycharmProjects/lstm_test/nasnet.py", line 221, in NASNet x, p = _normal_A(x, p, filters, weight_decay, id='%d' % (i)) File "/home/palm/PycharmProjects/lstm_test/nasnet.py", line 665, in _normal_A x3 = add([x3, p], name='normal_add_3_%s' % id) File "/home/palm/miniconda2/lib/python2.7/site-packages/keras/layers/merge.py", line 455, in add return Add(**kwargs)(inputs) File "/home/palm/miniconda2/lib/python2.7/site-packages/keras/engine/topology.py", line 571, in __call__ self.build(input_shapes) File "/home/palm/miniconda2/lib/python2.7/site-packages/keras/layers/merge.py", line 84, in build output_shape = self._compute_elemwise_op_output_shape(output_shape, shape) File "/home/palm/miniconda2/lib/python2.7/site-packages/keras/layers/merge.py", line 55, in _compute_elemwise_op_output_shape str(shape1) + ' ' + str(shape2)) ValueError: Operands could not be broadcast together with shapes (331, 331, 168) (331, 331, 96)

hello @titu1994 , I am using your code to train my dataset, and i want to train it with a pretrain model that you provide in nasnet.py. But the problem is that my category is 361, and the pre-trained model is 1000, how do I modify it? Looking forward for your reply! : )