I can't download the NASNetMobile model with the include_top=False option
Closed this issue · 9 comments
System information.
Google Colab
Python 3.10.12
tensorflow 2.12.0
keras 2.12.0
Describe the problem.
I can't download the NASNetMobile model with the 'include_top=False' option.
NASNetMobile(weights='imagenet', include_top=True) works fine, but NASNetMobile(weights='imagenet', include_top=False) generates an Error:
ValueError: Layer count mismatch when loading weights from file. Model expected 382 layers, found 388 saved layers.
- Do you want to contribute a PR? - no
Standalone code to reproduce the issue.
from tensorflow.keras.applications.nasnet import NASNetMobile
model = NASNetMobile(weights='imagenet', include_top=False)
https://colab.research.google.com/drive/11apqiS93G4ZjRDrzQUU4MAc0UXkWj3Cx?usp=sharing
Source code / logs.
Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/nasnet/NASNet-mobile-no-top.h5
19993432/19993432 [==============================] - 0s 0us/step
ValueError Traceback (most recent call last)
in <cell line: 1>()
----> 1 model = NASNetMobile(weights='imagenet', include_top=False) #error
3 frames
/usr/local/lib/python3.10/dist-packages/keras/applications/nasnet.py in NASNetMobile(input_shape, include_top, weights, input_tensor, pooling, classes, classifier_activation)
425 backend that does not support separable convolutions.
426 """
--> 427 return NASNet(
428 input_shape,
429 penultimate_filters=1056,
/usr/local/lib/python3.10/dist-packages/keras/applications/nasnet.py in NASNet(input_shape, penultimate_filters, num_blocks, stem_block_filters, skip_reduction, filter_multiplier, include_top, weights, input_tensor, pooling, classes, default_size, classifier_activation)
323 file_hash="1ed92395b5b598bdda52abe5c0dbfd63",
324 )
--> 325 model.load_weights(weights_path)
326 elif default_size == 331: # large version
327 if include_top:
/usr/local/lib/python3.10/dist-packages/keras/utils/traceback_utils.py in error_handler(*args, **kwargs)
68 # To get the full stack trace, call:
69 # tf.debugging.disable_traceback_filtering()
---> 70 raise e.with_traceback(filtered_tb) from None
71 finally:
72 del filtered_tb
/usr/local/lib/python3.10/dist-packages/keras/saving/legacy/hdf5_format.py in load_weights_from_hdf5_group(f, model)
806 layer_names = filtered_layer_names
807 if len(layer_names) != len(filtered_layers):
--> 808 raise ValueError(
809 "Layer count mismatch when loading weights from file. "
810 f"Model expected {len(filtered_layers)} layers, found "
ValueError: Layer count mismatch when loading weights from file. Model expected 382 layers, found 388 saved layers.
Hi. I would like to contribute a fix for this.
The problem is arising because of input_shape is not specified when it is include_top=False. I guess you can add a check for that.
Hi @Frightera,
So should we pass the default (example value) if no input_shape is provided when include_top=False ?
Hi @Frightera Is this related to the input size issue? Isnt this a number of model layers issue?
So should we pass the default (example value) if no input_shape is provided when include_top=False ?
Yes, that's right. If you specify the input_shape when it is include_top=False, it should fix it.
Hi @Frightera Is this related to the input size issue? Isnt this a number of model layers issue?
As far as I remember, NasNet implementation was little bit different, so that error is little bit misleading. There should be a check to force the user the specify the input shape in that case.
Shall I add the check for this?
@jkoniecznyy,
I tried to execute the code by providing the input_shape with include_top=False, and it was executed without any error/issue. Kindly find the gist of it here.
model = NASNetMobile(weights='imagenet', include_top=False, input_shape=(224,224,3))
keras-team/keras#9812
keras-team/keras#9865
keras-team/keras#9891
Thank you!
Oh, I didn't expect that to be about the input_shape, thank you very much for solving my problem.
I think it would be helpful to standardize the behavior of all models in the absence of a specific input_shape. Because currently Mobilenet shows a warning, Xception, DenseNet and EfficientNet just load, and NASNetMobile shows an error related to the number of layers (even on version 2.13.0.)
Below you can find my gist:
https://colab.research.google.com/drive/1UZoMLrNp_y_2sjfkXUHlVzWxC42KCnQ5?usp=sharing
Yea I will write to pull to fix an default value if no input shape is provided.