demo.py about keras-retinanet

[hgaiser]

Not a demo.py like the one in py-faster-rcnn, but you can look at this notebook for an example on how to visualize the results from Pascal VOC.

ps. if you want you can make a PR for one.

[kaishijeng]

I am not familiar with keras. Can you help me how to load a Keras-retinanet pretrained model in Keras. I use keras.load_model below and got an error, unknown layer 'UpsampleLike' import keras from keras.models import load_model import matplotlib.pyplot as plt import cv2 import os import numpy as np import time import tensorflow as tf model = load_model('snapshots/resnet50_voc_final.h5') Error: Using TensorFlow backend. /usr/local/lib/python2.7/dist-packages/google/protobuf/__init__.py:37: UserWarning: Module numpy was already imported from /usr/local/lib/python2.7/dist-packages/numpy/__init__.pyc, but /usr/lib/python2.7/dist-packages is being added to sys.path __import__('pkg_resources').declare_namespace(__name__) /usr/local/lib/python2.7/dist-packages/google/protobuf/__init__.py:37: UserWarning: Module six was already imported from /usr/local/lib/python2.7/dist-packages/six.pyc, but /usr/lib/python2.7/dist-packages is being added to sys.path __import__('pkg_resources').declare_namespace(__name__) Traceback (most recent call last): File "examples/test.py", line 12, in <module> model = load_model('snapshots/resnet50_voc_final.h5') File "/usr/local/lib/python2.7/dist-packages/keras/models.py", line 239, in load_model model = model_from_config(model_config, custom_objects=custom_objects) File "/usr/local/lib/python2.7/dist-packages/keras/models.py", line 313, in model_from_config return layer_module.deserialize(config, custom_objects=custom_objects) File "/usr/local/lib/python2.7/dist-packages/keras/layers/__init__.py", line 54, in deserialize printable_module_name='layer') File "/usr/local/lib/python2.7/dist-packages/keras/utils/generic_utils.py", line 139, in deserialize_keras_object list(custom_objects.items()))) File "/usr/local/lib/python2.7/dist-packages/keras/engine/topology.py", line 2487, in from_config process_layer(layer_data) File "/usr/local/lib/python2.7/dist-packages/keras/engine/topology.py", line 2473, in process_layer custom_objects=custom_objects) File "/usr/local/lib/python2.7/dist-packages/keras/layers/__init__.py", line 54, in deserialize printable_module_name='layer') File "/usr/local/lib/python2.7/dist-packages/keras/utils/generic_utils.py", line 133, in deserialize_keras_object ': ' + class_name) ValueError: Unknown layer: UpsampleLike

…

On Thu, Oct 12, 2017 at 6:23 AM, Hans Gaiser ***@***.***> wrote: Closed #21 <#21>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#21 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AMGg3kmo7yV8nGgM8hyM7SFin5EGk3Xvks5srhK0gaJpZM4P2haM> .

[hgaiser]

You need to let keras know where some of the custom objects are, by calling model = keras.models.load_model("model.h5", custom_objects=keras_retinanet.custom_objects).

However, as of writing this model loading is broken. The workaround is to create an identical model and call models.load_weights("model.h5", by_name=True). Or alternatively load the model definition first and then load weights. The rest can be seen in the notebook.

[kaishijeng]

It works with models.load_weights("model.h5", by_name=True).
Now I am facing a different issue which use mode. predict with one input image. My code is below:

print('Creating model, this may take a second...')
model = create_model()
model.load_weights("snapshots/resnet50_voc_final.h5", by_name=True)
image = cv2.imread(sys.argv[1])
image -= min(image.flatten())
image /= max(image.flatten())
image *= 255
image = image.astype(np.uint8)
start = time.time()
predictions, detections = model.predict(image)
print("processing time: ", time.time() - start)

im = image.copy()
predicted_labels = np.argmax(detections[0, :, 4:], axis=1)
scores = np.max(detections[0, :, 4:], axis=1)

for idx, (label, score) in enumerate(zip(predicted_labels, scores)):
     if label == 0:
        continue
     b = detections[0, idx, :4].astype(int)
     cv2.rectangle(im, (b[0], b[1]), (b[2], b[3]), (0, 255, 0), 3)
     caption = "{} {:.3f}".format(class_names[label], score)
     cv2.putText(im, caption, (b[0], b[1] - 10), cv2.FONT_HERSHEY_PLAIN, 1.5, (0, 0, 0), 3)
     cv2.putText(im, caption, (b[0], b[1] - 10), cv2.FONT_HERSHEY_PLAIN, 1.5, (255, 255, 255), 2)
plt.figure(figsize=(15, 15))
plt.axis('off')
plt.imshow(im)
plt.show()

But I got the following errors:

Traceback (most recent call last):
File "./examples/demo-pascal.py", line 65, in
predictions, detections = model.predict(image)
File "/usr/local/lib/python2.7/dist-packages/keras/engine/training.py", line 1695, in predict
check_batch_axis=False)
File "/usr/local/lib/python2.7/dist-packages/keras/engine/training.py", line 132, in _standardize_input_data
str(array.shape))
ValueError: Error when checking : expected input_1 to have 4 dimensions, but got array with shape (1288, 2048, 3)

[hgaiser]

You don't preprocess your image properly. It should firstly have 4 dimension ((batch, height, width, channels)). In addition the imagenet mean should be subtracted and the image should be rescaled by 1/255. For an example on how to preprocess an image you can look here : https://github.com/delftrobotics/keras-retinanet/blob/master/keras_retinanet/preprocessing/coco.py#L107

Perhaps I'll add a demo.py which follows this use case, but for now you could try replicating the preprocessing steps from the data generator.

PS. It's not so much that preprocessing has to be done this way, but it does have to be consistent with what you used during training.

[kaishijeng]

Do you still have a plan to add demo.py?

Thanks,

[hgaiser]

Not really no. If you make a nice one we can add it to the repository.