yoheikikuta / senet-keras Goto Github PK

View Code? Open in Web Editor NEW

136.0 5.0 54.0 102 KB

Naive implementation of SENet in Keras

License: MIT License

Jupyter Notebook 28.53% Shell 1.09% Python 66.04% Dockerfile 4.34%

senet-keras's Introduction

SENet (Keras implementation)

New information

We provide a trained SEResNeXt model (training data: cifar10)
Google drive
You can try this model in evaluate-cifar10.ipynb.

Naive implementation of SENet models in Keras.

Transplanting https://github.com/taki0112/SENet-Tensorflow to Keras.
Only SE-ResNext at this stage.

Prerequisites

nvidia-docker environment

Environment constuction

Build a docker image (on the root directory of the repository)
```
$ docker build -t [tag name] -f docker/Dockerfile .
```

Create a container using the image

$ nvidia-docker run -it -v $PWD:/work [tag name]

Train a model

Train a model with cifar10 data.

(in the container) $ pwd
/work
(in the container) $ python train-cifar10.py

Note that this script is written in an insufficient way; use data generator in consideration of expansion to general image data). The training speed is slow. On a p3.2xlarge instance, it takes about 1.5 days.

Evaluate the model

Launch a jupyter notebook.

(in the container) $ bash launch_notebook.sh

Execute evaluate-cifar10.ipynb notebook.

Results

Accuracy plot of train/val.
Loss plot of train/val.
Accuracy for the test data.
92.38%

senet-keras's People

Contributors

Stargazers

Watchers

senet-keras's Issues

SE-resnet50

hello, have you SE-resnet50 in keras ?

model load error

I train senet model like this:

senet =SEResNeXt.SEResNeXt(224,12).model
adam = tensorflow.keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, decay=1e-6)
senet.compile(loss='categorical_crossentropy',optimizer=adam, metrics=['accuracy'])
filepath="model_val/weights-improvement-{epoch:02d}-{acc:.2f}.hdf5"
checkpoint = ModelCheckpoint(filepath, monitor='acc', verbose=1, save_best_only=True, mode='max')
senet.fit(X, Y ,epochs=200, batch_size=2,callbacks=[checkpoint ])

When I try to load trained model, i got this :

model = load_model('model_val/weights-improvement-32-0.95.hdf5') File "/home/tensorflow04/anaconda3/lib/python3.6/site-packages/keras/models.py", line 240, in load_model model = model_from_config(model_config, custom_objects=custom_objects) File "/home/tensorflow04/anaconda3/lib/python3.6/site-packages/keras/models.py", line 314, in model_from_config return layer_module.deserialize(config, custom_objects=custom_objects) File "/home/tensorflow04/anaconda3/lib/python3.6/site-packages/keras/layers/__init__.py", line 55, in deserialize printable_module_name='layer') File "/home/tensorflow04/anaconda3/lib/python3.6/site-packages/keras/utils/generic_utils.py", line 139, in deserialize_keras_object list(custom_objects.items()))) File "/home/tensorflow04/anaconda3/lib/python3.6/site-packages/keras/engine/topology.py", line 2490, in from_config process_layer(layer_data) File "/home/tensorflow04/anaconda3/lib/python3.6/site-packages/keras/engine/topology.py", line 2476, in process_layer custom_objects=custom_objects) File "/home/tensorflow04/anaconda3/lib/python3.6/site-packages/keras/layers/__init__.py", line 55, in deserialize printable_module_name='layer') File "/home/tensorflow04/anaconda3/lib/python3.6/site-packages/keras/utils/generic_utils.py", line 139, in deserialize_keras_object list(custom_objects.items()))) File "/home/tensorflow04/anaconda3/lib/python3.6/site-packages/keras/layers/core.py", line 696, in from_config printable_module_name='function in Lambda layer') File "/home/tensorflow04/anaconda3/lib/python3.6/site-packages/keras/utils/generic_utils.py", line 156, in deserialize_keras_object fn = module_objects.get(function_name) AttributeError: 'NoneType' object has no attribute 'get'

Can it support Tensorflow 2.0 with Windows10?

ValueError: Only tensors of same shape can be merged by layermultiply_1 Got input shapes: [(None, 96, 96, 64), (None, 1, 1, 64)]

The role of the "channel"

This project is very useful to me, thank you very much. But can you tell me what is the role of the "channel" in line 136 of the SEResNeXt.py file? The object "channel" appears only once in the entire file.

help

在吗？代码运行出错 ValueError: An operation has None for gradient. Please make sure that all of your ops have a gradient defined (i.e. are differentiable). Common ops without gradient: K.argmax, K.round, K.eval.

Can you provide a pre-training model?

imagenet or coco，thank you very much

batch_size=?

Excuse me, my computer can only run when batch_size=1, what is your batch_size set to?

Can SENet be converted to encoder-decoder architecture?

Hello, Mr. Kikuta,
Thank you for sharing your code of SENet based on Keras.
I'm curious if this SENet can be converted to encoder-decoder architecture with skip connections?
Just like U-Net which can do the segmentation work?
Thank you.

Cannot make it parallel

I create my model like this:
model = SEResNeXt(size=256, num_classes=4).model
When I try to run the model on several GPUs by
parallel_model = multi_gpu_model(model, gpus=nb_gpu)
I got this error:
Concatenate layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, None, None, 32), (None, 128, 128, 64), (None, None, None, 32)]

python = 3.4.3
keras = 2.0.9