qubvel/efficientnet

Implementation of EfficientNet model. Keras and TensorFlow Keras.

 

efficientnet's Issues

error when creating model

hi,

Trying to train this. when running :

from efficientnet import EfficientNetB0
import keras
from keras.layers import Activation, Dense, Multiply, Input, Flatten, Dropout, Conv2D, BatchNormalization


inp_mask = Input(shape=(512,512,3))
bn_inp = inp_mask
pretrain_model_mask = EfficientNetB0(
input_shape=(512, 512, 3),  #SWITCH
include_top=False,
weights='imagenet',
#        dropout=0.5,
#        pooling='max'
)(bn_inp)

I get the following

Traceback (most recent call last):
  File "t.py", line 12, in <module>
    weights='imagenet',
  File "/home/m/.local/lib/python2.7/site-packages/efficientnet/model.py", line 380, in EfficientNetB0
    weights=weights, classes=classes)
  File "/home/m/.local/lib/python2.7/site-packages/efficientnet/model.py", line 358, in _get_model_by_name
    model = EfficientNet(input_shape, block_agrs_list, global_params, include_top=include_top)
  File "/home/m/.local/lib/python2.7/site-packages/efficientnet/model.py", line 282, in EfficientNet
    x = MBConvBlock(block_args, global_params)(x)
  File "/home/m/.local/lib/python2.7/site-packages/efficientnet/model.py", line 232, in block
    x = DropConnect(global_params.drop_connect_rate)(x)
  File "/home/m/.local/lib/python2.7/site-packages/efficientnet/layers.py", line 16, in __init__
    super().__init__()
TypeError: super() takes at least 1 argument (0 given)

Weight Conversion

Could you provide some details about how you converted the weights?

The official checkpoints for EfficientNet-B4 and -B5 have recently been released and I tried to convert them, but my B4 model only achieves 79.7% top-1 accuracy instead of 82.6%.

How to init for "channels_first"?

K.set_image_data_format('channels_first')

Hello, I'm wondering how to correctly use this library in a channels first setup. I already have my generator outputting (batches, channels, x, y) and tried initializing efficientnet with input shape (channels, x, y) but it doesn't seem to work. Any help is appreciated!

Cannot use with tf.keras

Consistently getting errors, either No module named 'efficientnet.tfkeras' or AttributeError: 'tuple' object has no attribute 'layer'. Using with standalone Keras, model compiles ok, however, I want to use tf.data.
I'm using tensorflow-gpu==1.14

error from ''from efficientnet import EfficientNetB0''

while running the code

from efficientnet import EfficientNetB0

I get error DLL load failed: The specified module could not be found.

tracebook
from efficientnet import EfficientNetB0
Traceback (most recent call last):

File "", line 1, in
from efficientnet import EfficientNetB0

File "C:\Users\Student\Anaconda2\lib\site-packages\efficientnet_init_.py", line 18, in
from .preprocessing import center_crop_and_resize, preprocess_input

File "C:\Users\Student\Anaconda2\lib\site-packages\efficientnet\preprocessing.py", line 16, in
from skimage.transform import resize

File "C:\Users\Student\Anaconda2\lib\site-packages\skimage_init_.py", line 135, in
from .data import data_dir

File "C:\Users\Student\Anaconda2\lib\site-packages\skimage\data_init_.py", line 16, in
from ._binary_blobs import binary_blobs
File "C:\Users\Student\Anaconda2\lib\site-packages\skimage\data_binary_blobs.py", line 2, in
from ..filters import gaussian

File "C:\Users\Student\Anaconda2\lib\site-packages\skimage\filters_init_.py", line 1, in
from .lpi_filter import inverse, wiener, LPIFilter2D

File "C:\Users\Student\Anaconda2\lib\site-packages\skimage\filters\lpi_filter.py", line 7, in
from scipy.fftpack import ifftshift

File "C:\Users\Student\Anaconda2\lib\site-packages\scipy\fftpack_init_.py", line 99, in
from .basic import *

File "C:\Users\Student\Anaconda2\lib\site-packages\scipy\fftpack\basic.py", line 12, in
**from . import _fftpack**

ImportError: DLL load failed: The specified module could not be found.

Warnings when creating model with tf.keras

Hi,

I am using v1.0.0b3 with tf.keras and I get the following warning multiple times when creating an EfficientNet model:

WARNING:tensorflow:Entity <bound method Dropout.call of <efficientnet.model.get_dropout..FixedDropout object at 0x0000017000529710>> could not be transformed and will be executed as-is. Please
report this to the AutgoGraph team. When filing the bug, set the verbosity to 10 (on Linux, export AUTOGRAPH_VERBOSITY=10) and attach the full output. Cause: converting <bound method Dropout.call of
<efficientnet.model.get_dropout..FixedDropout object at 0x0000017000529710>>: AssertionError: Bad argument number for Name: 3, expecting 4

Is this a known issue that I can safely ignore, or do you know if it could affect whether the dropout layer works correctly?

Best Regards
John

New weights available

Weights for B6 and B7 trained with auto-augment are available. Maybe worth adding?
https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet

Undocumented but needed function argument 'backend'

I'm now getting the error message

AttributeError: 'NoneType' object has no attribute 'image_data_format'

on this line. It appears that a backend argument needs to be provided when creating EfficientNet through any of the EfficientNetB* functions, but this is an undocumented argument as far as I can see.

What value should be provided for it?

is it possible to convert models trained by earlier version to last version?

I have a few models that I am not that keen on retraining....

Thank you!

ValueError: Unknown activation function:swish

when I try to load the trained model, it give me "ValueError: Unknown activation function:swish", do you know how to solve that? thanks a lot!

ImportError: cannot import name 'load_model'.

Hi, qubvel. I piped install and trying to import 'load_model, but it got ImportError: cannot import name 'load_model' message.

I do recheck the whole*.py files but I still can't find any function named 'load_model'. Maybe I should pip installed fom git+https://github.com/qubvel/efficientnet instead of pip install efficientnet directly.

Really appreciate your help.

inference errors when loading models due to backend.backend()

I cannot load models with the v1.0.0b1 version of your package. In model.py of your package you are calling backend.backend(): https://github.com/qubvel/efficientnet/blob/master/efficientnet/model.py#L246 .

I cannot see anywhere where this is supposed to be automatically initialized to the user's backend. The same file contains backend = None: https://github.com/qubvel/efficientnet/blob/master/efficientnet/model.py#L43

When trying to load a model this leads to the following error message:

Found 3662 validated image filenames.
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-13-3f84b7149375> in <module>
      8                       model_dir = "../",
      9                       folds = 1,
---> 10                       batch_size = 16)
     11 ("")

<ipython-input-12-eda66ccb6232> in pred_cv_test(df, model_dir, folds, nr_samples, batch_size)
     31                                     batch_size = batch_size)
     32 
---> 33         model = get_model(model_dir = model_dir, fold_nr = fold)
     34         preds = model.predict_generator(
     35             generator = test_generator,

<ipython-input-12-eda66ccb6232> in get_model(model_dir, fold_nr)
      9     best_epoch_model = load_model(
     10         model_dir + "effnet_reg_fold" + str(fold_nr) + ".h5",
---> 11         custom_objects = {"root_mse": root_mse})
     12     return(best_epoch_model)
     13 

~\Anaconda3\envs\r-tensorflow\lib\site-packages\tensorflow\python\keras\saving\save.py in load_model(filepath, custom_objects, compile)
    144       h5py is not None and (
    145           isinstance(filepath, h5py.File) or h5py.is_hdf5(filepath))):
--> 146     return hdf5_format.load_model_from_hdf5(filepath, custom_objects, compile)
    147 
    148   if isinstance(filepath, six.string_types):

~\Anaconda3\envs\r-tensorflow\lib\site-packages\tensorflow\python\keras\saving\hdf5_format.py in load_model_from_hdf5(filepath, custom_objects, compile)
    210     model_config = json.loads(model_config.decode('utf-8'))
    211     model = model_config_lib.model_from_config(model_config,
--> 212                                                custom_objects=custom_objects)
    213 
    214     # set weights

~\Anaconda3\envs\r-tensorflow\lib\site-packages\tensorflow\python\keras\saving\model_config.py in model_from_config(config, custom_objects)
     53                     '`Sequential.from_config(config)`?')
     54   from tensorflow.python.keras.layers import deserialize  # pylint: disable=g-import-not-at-top
---> 55   return deserialize(config, custom_objects=custom_objects)
     56 
     57 

~\Anaconda3\envs\r-tensorflow\lib\site-packages\tensorflow\python\keras\layers\serialization.py in deserialize(config, custom_objects)
     87       module_objects=globs,
     88       custom_objects=custom_objects,
---> 89       printable_module_name='layer')

~\Anaconda3\envs\r-tensorflow\lib\site-packages\tensorflow\python\keras\utils\generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)
    190             custom_objects=dict(
    191                 list(_GLOBAL_CUSTOM_OBJECTS.items()) +
--> 192                 list(custom_objects.items())))
    193       with CustomObjectScope(custom_objects):
    194         return cls.from_config(cls_config)

~\Anaconda3\envs\r-tensorflow\lib\site-packages\tensorflow\python\keras\engine\network.py in from_config(cls, config, custom_objects)
   1129         if layer in unprocessed_nodes:
   1130           for node_data in unprocessed_nodes.pop(layer):
-> 1131             process_node(layer, node_data)
   1132 
   1133     name = config.get('name')

~\Anaconda3\envs\r-tensorflow\lib\site-packages\tensorflow\python\keras\engine\network.py in process_node(layer, node_data)
   1085         flat_input_tensors = nest.flatten(input_tensors)
   1086         if len(flat_input_tensors) == 1:
-> 1087           layer(flat_input_tensors[0], **kwargs)
   1088         else:
   1089           layer(input_tensors, **kwargs)

~\Anaconda3\envs\r-tensorflow\lib\site-packages\tensorflow\python\keras\engine\base_layer.py in __call__(self, inputs, *args, **kwargs)
    632                     outputs = base_layer_utils.mark_as_return(outputs, acd)
    633                 else:
--> 634                   outputs = call_fn(inputs, *args, **kwargs)
    635 
    636             except TypeError as e:

~\Anaconda3\envs\r-tensorflow\lib\site-packages\tensorflow\python\keras\layers\core.py in call(self, inputs)
    358 
    359   def call(self, inputs):
--> 360     return self.activation(inputs)
    361 
    362   def compute_output_shape(self, input_shape):

~\Anaconda3\envs\r-tensorflow\lib\site-packages\efficientnet\model.py in swish(x)
    134     """
    135 
--> 136     if backend.backend() == 'tensorflow':
    137         try:
    138             # The native TF implementation has a more

AttributeError: 'NoneType' object has no attribute 'backend'

I can get it to work if I hardcode my backend by editing model.py.

In your inference notebook you are not loading a saved model from disk. Have you guys tested and successfully loaded in a model from disk with the new version?

Hyperparameters

Hi, I'm trying to train imagenet with EfficientnetB3. Would you please be kind enough to mention the correct combination of hyperparameters.

AttributeError: module 'tensorflow.python.keras.api._v1.keras.backend' has no attribute 'is_keras_tensor'

runners.py", line 80, in train
  model = model_fn(inputs, problem.output_spec())
File "/opt/venvs/tflow3.6/lib/python3.6/site-packages/gin/config.py", line 1032, in wrapper
  utils.augment_exception_message_and_reraise(e, err_str)
File "/opt/venvs/tflow3.6/lib/python3.6/site-packages/gin/utils.py", line 49, in augment_exception_message_and_reraise
  six.raise_from(proxy.with_traceback(exception.__traceback__), None)
File "<string>", line 3, in raise_from
File "/opt/venvs/tflow3.6/lib/python3.6/site-packages/gin/config.py", line 1009, in wrapper
  return fn(*new_args, **new_kwargs)
File "/opt/venvs/tflow3.6/lib/python3.6/site-packages/ml_glaucoma/models/efficientnet.py", line 45, in efficient_net
  input_tensor=inputs, **kwargs).outputs
File "/opt/venvs/tflow3.6/lib/python3.6/site-packages/efficientnet/__init__.py", line 57, in wrapper
  return func(*args, **kwargs)
File "/opt/venvs/tflow3.6/lib/python3.6/site-packages/efficientnet/model.py", line 548, in EfficientNetB4
  **kwargs)
File "/opt/venvs/tflow3.6/lib/python3.6/site-packages/efficientnet/model.py", line 369, in EfficientNet
  if not backend.is_keras_tensor(input_tensor):
File "/opt/venvs/tflow3.6/lib/python3.6/site-packages/tensorflow/python/util/deprecation_wrapper.py", line 106, in __getattr__
  attr = getattr(self._dw_wrapped_module, name)
Sep 08 03:15:55 ip-172-31-25-230 python[3136]: AttributeError: module 'tensorflow.python.keras.api._v1.keras.backend' has no attribute 'is_keras_tensor'
In call to configurable 'efficient_net' (<function efficient_net at 0x7feef701ce18>)

Batch size dramatically drop when training EfficientNetB4 comparing with InceptionV3

[Environment] GeForce RTX 2080 11GB
The max batch size dramatically drop to 8 in training EfficientNetB4 with 1 GPU, while the bacth size is 64 in training InceptionV3.

• EfficientNetB4 vs InceptionV3
• Params 19M vs 24M
• Input size 380 vs 299
• Max Batch size 8 vs 64

Thanks for any help.

when saving the model structure gets nameerror no drop_connect_rate

model_json = model.to_json()

with open("testing.json", "w") as json_file:
json_file.write(model_json)

---

NameError Traceback (most recent call last)
in ()
----> 1 model_json = model.to_json()
2
3
4 with open("testing.json", "w") as json_file:
5 json_file.write(model_json)

3 frames
/usr/local/lib/python3.6/dist-packages/efficientnet/layers.py in get_config(self)
35 def get_config(self):
36 config = super().get_config()
---> 37 config['drop_connect_rate'] = drop_connect_rate
38
39

NameError: name 'drop_connect_rate' is not defined

Is it suitable for feature extraction?

Hi,
I wanna use the pretrained weights to perform feature extraction. What would be the best way to go about this?

Is there any possible way to convert model to tflite?

Currently, I am loading the model and than trying to convert it to tflite. I've used just EffNet from 0.0.4 and keras/tfkeras from pre version, but issues occurred

ValueError: Unknown activation function:swish

Possibly wrong strides?

Hi, thanks for your Keras implementation! I'm currently looking through the model definition to solidify my understanding of the paper. I've noticed something odd in the parametrisation of the blocks: https://github.com/qubvel/efficientnet/blob/master/efficientnet/params.py#L170 -
's22' means a stride of [2,2] which means downsampling both image dimensions by half while 's11' means a stride of [1,1] which means no downsampling (this is obvious to you I guess). Your order of strides is: s11 s22 s22 s22 s11 s22 s11 but according to Table 1 in the paper, I think it should be s11 s22 s22 s1 s22 s22 s11. Do you agree or am I wrong somewhere?

version conflict while using efficientnet code

@qubvel what should be the version of tensorflow and keras so that there is no version conflict. This is not working in tensorflow-gpu==1.14 as well tensorflow-gpu==2.0.0-b1.
Thanks in advance.

Loading a model: "Unknown initializer: conv_kernel_initializer"

I trained an EfficientNetB3.
But trying to load the model with keras.load_model(modelfile)
I get this error:

File "...python3.6/site-packages/keras/utils/generic_utils.py", line 138, in deserialize_keras_object
    ': ' + class_name)
ValueError: Unknown initializer: conv_kernel_initializer

What do I miss?

transfer learning - retrain EfficientNetB7 RTX 2070 out of memory

config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.7
tf.keras.backend.set_session(tf.Session(config=config))

model = efn.EfficientNetB7()
model.summary()

# create new output layer
output_layer = Dense(5, activation='sigmoid', name="retrain_output")(model.get_layer('top_dropout').output)
new_model = Model(model.input, output=output_layer)
new_model.summary()
# lock previous weights

for i, l in enumerate(new_model.layers):
    if i < 228:
        l.trainable = False
# lock probs weights

new_model.compile(loss='mean_squared_error', optimizer='adam')

batch_size = 5
samples_per_epoch = 30
epochs = 20

# generate train data
train_datagen = ImageDataGenerator(
    shear_range=0.2,
    zoom_range=0.2,
    horizontal_flip=True,
    validation_split=0)

train_generator = train_datagen.flow_from_directory(
    train_data_input_folder,
    target_size=(input_dim, input_dim),
    batch_size=batch_size,
    class_mode='categorical',
    seed=2019,
    subset='training')

validation_generator = train_datagen.flow_from_directory(
    validation_data_input_folder,
    target_size=(input_dim, input_dim),
    batch_size=batch_size,
    class_mode='categorical',
    seed=2019,
    subset='validation')

new_model.fit_generator(
    train_generator,
    samples_per_epoch=samples_per_epoch,
    epochs=epochs,
    validation_steps=20,
    validation_data=validation_generator,
    nb_worker=24)

new_model.save(model_output_path)

2019-11-17 08:52:52.903583: I tensorflow/stream_executor/dso_loader.cc:152] successfully opened CUDA library libcublas.so.10.0 locally
....
...
2019-11-17 08:53:24.713020: I tensorflow/core/common_runtime/bfc_allocator.cc:641] 110 Chunks of size 27724800 totalling 2.84GiB
2019-11-17 08:53:24.713024: I tensorflow/core/common_runtime/bfc_allocator.cc:641] 6 Chunks of size 38814720 totalling 222.10MiB
2019-11-17 08:53:24.713027: I tensorflow/core/common_runtime/bfc_allocator.cc:641] 23 Chunks of size 54000128 totalling 1.16GiB
2019-11-17 08:53:24.713031: I tensorflow/core/common_runtime/bfc_allocator.cc:641] 1 Chunks of size 73760000 totalling 70.34MiB
2019-11-17 08:53:24.713034: I tensorflow/core/common_runtime/bfc_allocator.cc:645] Sum Total of in-use chunks: 5.45GiB
2019-11-17 08:53:24.713040: I tensorflow/core/common_runtime/bfc_allocator.cc:647] Stats:
Limit: 5856749158
InUse: 5848048896
MaxInUse: 5848061440
NumAllocs: 6140
MaxAllocSize: 3259170816

2019-11-17 08:53:24.713214: W tensorflow/core/common_runtime/bfc_allocator.cc:271] ****************************************************************************************************
2019-11-17 08:53:24.713232: W tensorflow/core/framework/op_kernel.cc:1401] OP_REQUIRES failed at cwise_ops_common.cc:70 : Resource exhausted: OOM when allocating tensor with shape[5,1344,38,38] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc
Traceback (most recent call last):
File "/home/naort/Desktop/deep-learning-data-preparation-tools/EfficientNet-Transfer-Learning-Boiler-Plate/model_retrain.py", line 76, in
nb_worker=24)
File "/usr/local/lib/python3.6/dist-packages/keras/legacy/interfaces.py", line 91, in wrapper
return func(*args, **kwargs)
File "/usr/local/lib/python3.6/dist-packages/keras/engine/training.py", line 1732, in fit_generator
initial_epoch=initial_epoch)
File "/usr/local/lib/python3.6/dist-packages/keras/engine/training_generator.py", line 220, in fit_generator
reset_metrics=False)
File "/usr/local/lib/python3.6/dist-packages/keras/engine/training.py", line 1514, in train_on_batch
outputs = self.train_function(ins)
File "/home/naort/.local/lib/python3.6/site-packages/tensorflow/python/keras/backend.py", line 3076, in call
run_metadata=self.run_metadata)
File "/home/naort/.local/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1439, in call
run_metadata_ptr)
File "/home/naort/.local/lib/python3.6/site-packages/tensorflow/python/framework/errors_impl.py", line 528, in exit
c_api.TF_GetCode(self.status.status))
tensorflow.python.framework.errors_impl.ResourceExhaustedError: OOM when allocating tensor with shape[5,1344,38,38] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc
[[{{node training/Adam/gradients/AddN_387-0-TransposeNHWCToNCHW-LayoutOptimizer}}]]
Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.

[[{{node Mean}}]]
Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.

Failed to convert TF to Keras (No module named 'condconv')

Hi @qubvel,
thank you so much for sharing this awesome tool!

While I was running ./scripts/convert_from_tf_to_keras.sh, attempting to convert the TF model into Keras, I encountered the following error:

==============================================
= .......................................... =
= Setting up the installation environment... =
= .......................................... =
==============================================
==========================================
= ...................................... =
= Converting the checkpoints to Keras... =
= ...................................... =
==========================================
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
~ ............................. ~
~ Converting efficientnet-b0... ~
~ ............................. ~
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Using TensorFlow backend.
Traceback (most recent call last):
  File "/Users/yunseong/repos/ai/qubvel-efficientnet/scripts/load_efficientnet.py", line 155, in <module>
    import eval_ckpt_main
  File "tf_src/eval_ckpt_main.py", line 31, in <module>
    import efficientnet_builder
  File "tf_src/efficientnet_builder.py", line 28, in <module>
    import efficientnet_model
  File "tf_src/efficientnet_model.py", line 37, in <module>
    from condconv import condconv_layers
ModuleNotFoundError: No module named 'condconv'

Do you have any clue how I can fix this issue?

Incomplete logic in handling input shape

In efficientnet/models.py :

efficientnet/efficientnet/model.py

Lines 359 to 388 in 035df1a

	if input_shape is None:
	input_shape = (default_input_shape, default_input_shape, 3)
	model = EfficientNet(
	input_shape,
	block_agrs_list,
	global_params,
	input_tensor=input_tensor,
	include_top=include_top,
	pooling=pooling,
	)
	model.name = model_name
	if weights:
	if not include_top:
	weights_name = model_name + "-notop"
	else:
	weights_name = model_name
	weights = IMAGENET_WEIGHTS[weights_name]
	weights_path = get_file(
	weights["name"],
	weights["url"],
	cache_subdir="models",
	md5_hash=weights["md5"],
	)
	model.load_weights(weights_path)
	return model

You fall back to the default if input_shape is None. However when we use the efficientnet backbone from segmentation models:
https://github.com/qubvel/segmentation_models/blob/ef6350388a0e774b8e602268c8f6a90ac0ed3a29/segmentation_models/unet/model.py#L14-L26

You can say that by default the input shape is (None, None, 3) and therefore the if statement shown above evaluates to False and we don't fall back to the default input. I recommend replacing with
if input_shape is None or None in input_shape:...

[tf 2.0] efficientnetB7 - Large dropout rate

Thank you so much for your great work on this and the segmentation_models!

Wenn I use the efficientnetB7 with tensorflow 2.0 (2.0.0-beta1) it prints:

Large dropout rate: 0.5625 (>0.5). In TensorFlow 2.x, dropout() uses dropout rate instead of keep_prob. Please ensure that this is intended

Is this already handled?

Ability to use tensorflow.keras instead of regular keras?

Could we choose tf.keras instead of keras? I tried saving the model and loading it via tf.keras load_model to convert it to tf.keras, but it gives me the same error as #2.

efficientnet similar mnasnet?

Looks like the structure of the model efficientnet is quite similar to mnasnet?

license of pretrained weights?

are the weights released with this repository also under APACHE-2.0 license?

use for object detection

Would it be possible to "easily" use this model as a backbone for object detection / instances segmentation as many other popular architectures like ResNet50? Thx

Problem in converting model to TFLite

I was trying to convert my EfficientNet model to TFLite:

converter = tf.lite.TFLiteConverter.from_keras_model_file('effNetB3.h5',custom_objects={'swish': Swish(swish)})
tfmodel = converter.convert()
open ("effNetB3.tflite" , "wb") .write(tfmodel)

It gave me the following error:
ValueError: Unknown layer: FixedDropout

tensorflow 2.0.0 doesn't work in efficientnet

this is the traceback

AttributeError Traceback (most recent call last)
in
1 import efficientnet
2
----> 3 model = efficientnet.EfficientNetB3(weights=None)

~/anaconda3/envs/py36/lib/python3.7/site-packages/efficientnet/model.py in EfficientNetB3(include_top, input_shape, input_tensor, weights, classes, pooling)
441 weights=weights,
442 classes=classes,
--> 443 pooling=pooling,
444 )
445

~/anaconda3/envs/py36/lib/python3.7/site-packages/efficientnet/model.py in _get_model_by_name(model_name, input_shape, input_tensor, include_top, weights, classes, pooling)
367 input_tensor=input_tensor,
368 include_top=include_top,
--> 369 pooling=pooling,
370 )
371

~/anaconda3/envs/py36/lib/python3.7/site-packages/efficientnet/model.py in EfficientNet(input_shape, block_args_list, global_params, input_tensor, include_top, pooling)
214 # Stem part
215 if input_tensor is None:
--> 216 inputs = KL.Input(shape=input_shape)
217 else:
218 if not K.is_keras_tensor(input_tensor):

~/anaconda3/envs/py36/lib/python3.7/site-packages/keras/engine/input_layer.py in Input(shape, batch_shape, name, dtype, sparse, tensor)
176 name=name, dtype=dtype,
177 sparse=sparse,
--> 178 input_tensor=tensor)
179 # Return tensor including _keras_shape and _keras_history.
180 # Note that in this case train_output and test_output are the same pointer.

~/anaconda3/envs/py36/lib/python3.7/site-packages/keras/legacy/interfaces.py in wrapper(*args, **kwargs)
89 warnings.warn('Update your ' + object_name + ' call to the ' +
90 'Keras 2 API: ' + signature, stacklevel=2)
---> 91 return func(*args, **kwargs)
92 wrapper._original_function = func
93 return wrapper

~/anaconda3/envs/py36/lib/python3.7/site-packages/keras/engine/input_layer.py in init(self, input_shape, batch_size, batch_input_shape, dtype, input_tensor, sparse, name)
37 if not name:
38 prefix = 'input'
---> 39 name = prefix + '_' + str(K.get_uid(prefix))
40 super(InputLayer, self).init(dtype=dtype, name=name)
41

~/anaconda3/envs/py36/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py in get_uid(prefix)
72 """
73 global _GRAPH_UID_DICTS
---> 74 graph = tf.get_default_graph()
75 if graph not in _GRAPH_UID_DICTS:
76 _GRAPH_UID_DICTS[graph] = defaultdict(int)

AttributeError: module 'tensorflow' has no attribute 'get_default_graph'

maybe lets make an update to requirements.txt?

Update preprocess_input() to support batches of images

Currently there's an assert on the input dimension to preprocess_input() that implies this function expects a single image; can you make this a bit more flexible, to allow for batches of images in the same way that the usual keras_applications.imagenet_utils.preprocess_input() function allows for a 3D or 4D input array? Thanks.

Importing keras backend fails when using preprocess_input

I'm trying to use efficientnet.model.preprocess_input as ImageDataGenerator(preprocessing_function=preprocess_input, ...). However, getting following error:

~/.virtualenvs/venv1/lib/python3.7/site-packages/keras_applications/imagenet_utils.py in preprocess_input(x, data_format, mode, **kwargs)
    184 
    185     if data_format is None:
--> 186         data_format = backend.image_data_format()
    187     if data_format not in {'channels_first', 'channels_last'}:
    188         raise ValueError('Unknown data_format ' + str(data_format))

AttributeError: 'NoneType' object has no attribute 'image_data_format'

Symptoms look very similar to keras-team/keras-applications#54 and it seems that solution can be also similar to keras-team/keras-applications#54 (comment)_ , i.e. replacing keras_applications with keras.applications here:

efficientnet/efficientnet/model.py

Line 39 in 91366f5

from keras_applications.imagenet_utils import preprocess_input as _preprocess_input

how can i retrian the efficientnet model？

please give me a demo, thanks!

Feature: merging keras and tf.keras frameworks

Hi @Callidior
I am going to adapt my segmentation-models lib for tf.keras, so I also need efficientnet adapted for tf.keras.
I saw your implementation of Efficientnet for keras-applications. Very nice work, but it is too long to wait until keras will merge it and make a new release.
It would be nice if you make a PR with same code to this repo, or allow to copy it.

Can this model be trained as well?

Or only inference?

Missing requirement.txt

I would like to know if there is a 'requirement.txt' for setting up efficient net using setup.py.

use drop_out=0.2 in the inference mode, isn't wrong way?

Hi
In the efficient-net model used dropout layer in the final prediction layer and also used in the Conv layers with rate 0.2 for training mode, Why in the inference mode didn't set rate of these layers to 1.0, and these layers don't operate in the inference mode?

EfficientNet version 1.0.0 contains lower accuracy than 0.0.4

With the included example/inference_example.ipynb, I am seeing lower prediction with version 1.0.0 as opposed to version 0.0.4.

Using the below code for testing both versions (with the exception of import efficientnet vs import efficientnet.keras/tfkeras) produce different prediction results:

!pip install efficientnet==0.0.4
!wget https://github.com/qubvel/efficientnet/raw/master/misc/panda.jpg
from keras.applications.imagenet_utils import decode_predictions
import numpy as np
import matplotlib.pyplot as plt
from efficientnet import EfficientNetB0
from efficientnet import center_crop_and_resize, preprocess_input

model = EfficientNetB0(weights='imagenet')
image = plt.imread('panda.jpg')

image_size = model.input_shape[1]
x = center_crop_and_resize(image, image_size=image_size)
x = preprocess_input(x)
x = np.expand_dims(x, 0)

y = model.predict(x)
decode_predictions(y)

Output from version 0.0.4:
[[('n02510455', 'giant_panda', 0.8347928),
('n02134084', 'ice_bear', 0.015602051),
('n02509815', 'lesser_panda', 0.0045535257),
('n02133161', 'American_black_bear', 0.002471913),
('n02132136', 'brown_bear', 0.0020707587)]]

Output from version 1.0.0:
[[('n02510455', 'giant_panda', 0.75878686),
('n02134084', 'ice_bear', 0.008354747),
('n02132136', 'brown_bear', 0.0072072297),
('n02509815', 'lesser_panda', 0.0041302275),
('n02120079', 'Arctic_fox', 0.0040210797)]]

At its current state (version 1.0.0), I am also seeing lower top-1 score training on CIFAR10 on EfficientNet B0 as opposed to the same on ResNet50, thinking that the issue may be caused by outdated/bad weights as shown above.

As backbone for detection task

i had a problem when using your efficiententB3 as backbone for object detection task(previous ResNet50).
while i use ResNet50,batchsize per GPU is 16. But using EfficientNetB3 batchsize per GPU is no more than 12. I also print(model.summary()) to show num of parameters.ResNet50 is 23,000,000+, EfficientNetB3 is 10,000,000+.
So,obviously,EfficientNetB3's parameters is about half the size of the ResNet50.However,i can only set batchsize to 8 or 12.This make me confused.
My GPU is Tesla P100 -16G mem

how to train a new model using my own dataset?

preprocess

I found that preprocess_input in official github contain random crop for training. and crop center for eval. However your code has just center crop for eval.
Is it necessary to do same as original?. or it is just an augmentation process. Now, i'm finetuned all layers but i just use only MEAN RGB and MEAN STD. Good val Acc and test Acc

E tensorflow/core/grappler/clusters/utils.cc:87] Failed to get device properties, error code: 30

HI, I tried to initialized the model with the following code:
import efficientnet.keras as efn
model = efn.EfficientNetB0(weights='imagenet')

but it returned E tensorflow/core/grappler/clusters/utils.cc:87] Failed to get device properties, error code: 30

Failed to install on Ubuntu 16.04.6

Hi,
I ran pip install -U git+https://github.com/qubvel/efficientnet
and received the following:

Collecting git+https://github.com/qubvel/efficientnet
Cloning https://github.com/qubvel/efficientnet to /tmp/pip-k06rtpj3-build
Collecting keras_applications<=1.0.8,>=1.0.7 (from efficientnet==1.0.0b3)
Downloading https://files.pythonhosted.org/packages/71/e3/19762fdfc62877ae9102edf6342d71b28fbfd9dea3d2f96a882ce099b03f/Keras_Applications-1.0.8-py3-none-any.whl (50kB)
100% |████████████████████████████████| 51kB 11.2MB/s
Collecting scikit-image (from efficientnet==1.0.0b3)
Downloading https://files.pythonhosted.org/packages/00/ba/9688afba6ab3714875a8d22284ce0362086e47f3cda60243e5d4de5e5943/scikit_image-0.15.0-cp35-cp35m-manylinux1_x86_64.whl (26.2MB)
100% |████████████████████████████████| 26.2MB 57kB/s
Collecting numpy>=1.9.1 (from keras_applications<=1.0.8,>=1.0.7->efficientnet==1.0.0b3)
Downloading https://files.pythonhosted.org/packages/d4/64/7619774f0bd8ef364d46a5df8eb1bc78784cd787324b9624f6793e72f787/numpy-1.17.1-cp35-cp35m-manylinux1_x86_64.whl (20.1MB)
100% |████████████████████████████████| 20.2MB 77kB/s
Requirement already up-to-date: h5py in ./.local/lib/python3.5/site-packages (from keras_applications<=1.0.8,>=1.0.7->efficientnet==1.0.0b3)
Collecting PyWavelets>=0.4.0 (from scikit-image->efficientnet==1.0.0b3)
Downloading https://files.pythonhosted.org/packages/6a/3f/d92fbd9b65f879197620f404c1603bfeaa54560743bb9fb3a6b48b66a707/PyWavelets-1.0.3-cp35-cp35m-manylinux1_x86_64.whl (4.4MB)
100% |████████████████████████████████| 4.4MB 373kB/s
Collecting pillow>=4.3.0 (from scikit-image->efficientnet==1.0.0b3)
Downloading https://files.pythonhosted.org/packages/d6/98/0d360dbc087933679398d73187a503533ec0547ba4ffd2115365605559cc/Pillow-6.1.0-cp35-cp35m-manylinux1_x86_64.whl (2.1MB)
100% |████████████████████████████████| 2.1MB 802kB/s
Requirement already up-to-date: imageio>=2.0.1 in ./.local/lib/python3.5/site-packages (from scikit-image->efficientnet==1.0.0b3)
Collecting matplotlib!=3.0.0,>=2.0.0 (from scikit-image->efficientnet==1.0.0b3)
Downloading https://files.pythonhosted.org/packages/12/d1/7b12cd79c791348cb0c78ce6e7d16bd72992f13c9f1e8e43d2725a6d8adf/matplotlib-3.1.1.tar.gz (37.8MB)
100% |████████████████████████████████| 37.8MB 42kB/s
Complete output from command python setup.py egg_info:

Beginning with Matplotlib 3.1, Python 3.6 or above is required.

This may be due to an out of date pip.

Make sure you have pip >= 9.0.1.


----------------------------------------

Command "python setup.py egg_info" failed with error code 1 in /tmp/pip-build-l3aisvf_/matplotlib/
You are using pip version 8.1.2, however version 19.2.3 is available.
You should consider upgrading via the 'pip install --upgrade pip' command.

Does anyone know the reason?

how to change efficientnet model?

@qubvel I want to change the downsampling factor from 32 to 16 of last layer keeping everything else same? how can I change it? Thanks in advance! :)

Please use `rate` instead of `keep_prob`. Rate should be set to `rate = 1 - keep_prob`.

ImportError: cannot import name '_validate_lengths'

ImportError: cannot import name '_validate_lengths'

This solves the problem, but it will be good to add requirements.txt

pip install -U scikit-image

python -c "import skimage; print(skimage.__version__)"
0.15.0

ValueError: Unknown initializer: EfficientConv2DKernelInitializer

I trained the model on my custom dataset, and now when I am trying to import the .h5 file, it is throwing me this error

model = load_model('model/model_weights_efficient.h5')
Traceback (most recent call last):
File "", line 1, in
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/saving/save.py", line 146, in load_model
return hdf5_format.load_model_from_hdf5(filepath, custom_objects, compile)
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/saving/hdf5_format.py", line 212, in load_model_from_hdf5
custom_objects=custom_objects)
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/saving/model_config.py", line 55, in model_from_config
return deserialize(config, custom_objects=custom_objects)
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/layers/serialization.py", line 89, in deserialize
printable_module_name='layer')
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/utils/generic_utils.py", line 192, in deserialize_keras_object
list(custom_objects.items())))
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/engine/network.py", line 1121, in from_config
process_layer(layer_data)
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/engine/network.py", line 1105, in process_layer
layer = deserialize_layer(layer_data, custom_objects=custom_objects)
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/layers/serialization.py", line 89, in deserialize
printable_module_name='layer')
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/utils/generic_utils.py", line 194, in deserialize_keras_object
return cls.from_config(cls_config)
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/engine/base_layer.py", line 446, in from_config
return cls(**config)
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/layers/convolutional.py", line 477, in init
kernel_initializer=initializers.get(kernel_initializer),
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/initializers.py", line 194, in get
return deserialize(identifier)
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/initializers.py", line 186, in deserialize
printable_module_name='initializer')
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/utils/generic_utils.py", line 181, in deserialize_keras_object
config, module_objects, custom_objects, printable_module_name)
File "/home/eleven/ccd/tf12/lib/python3.6/site-packages/tensorflow/python/keras/utils/generic_utils.py", line 166, in class_and_config_for_serialized_keras_object
raise ValueError('Unknown ' + printable_module_name + ': ' + class_name)
ValueError: Unknown initializer: EfficientConv2DKernelInitializer

Any way around?

load_model('path_to_model') not working

While trying to load model using load_model(), I am getting this error. ValueError: Unknown initializer: EfficientConv2DKernelInitializer

I followed the steps written on your readme.

Smaller model then EfficientNetB0

What are recommendations for params for creating smaller then EfficientNetB0 models for 128x128 input and about 1kk params?

# Base EfficientNetB0()
# Total params: 5,330,564
# Trainable params: 5,288,548
# Non-trainable params: 42,016