[theano-users] VGG16_Places_365 and VGG16_Hybrid_Places

ephi5757 via theano-users

2018-04-26 19:26:43 UTC

In the code to test the hybrid ImageNet+Places CNN, change the line: model
= VGG16_Hubrid_1365(weights='places', include_top=False) to model =
VGG16_Hubrid_1365(weights='places', include_top=True). Then
preds.shape=(1365L) and the code is debugged.
Note however, that there is a known issue with the pre-trained weights so
that the top-5 class results for different images are practically identical
and also the top-5 confidence levels are very low, i.e., 2-3%. Refer
to the GitHub page https://github.com/GKalliatakis/Keras-VGG16-places365
<https://www.google.com/url?q=https%3A%2F%2Fgithub.com%2FGKalliatakis%2FKeras-VGG16-places365&sa=D&sntz=1&usg=AFQjCNGBJUu-M4S4av5I6ghg4TfyoM_wHg> under
"issues" for updates.
====================================================================.

Post by ephi5757 via theano-users
. I am testing version of a python script (see below) that I found at
https://github.com/GKalliatakis/Keras-VGG16-places365
<https://www.google.com/url?q=https%3A%2F%2Fgithub.com%2FGKalliatakis%2FKeras-VGG16-places365&sa=D&sntz=1&usg=AFQjCNGBJUu-M4S4av5I6ghg4TfyoM_wHg>
to predict the VGG places 365 CNN classes and probabilities for several
test images.

1) I found that the image variable "x" has dimension 4 and

shape (1L, 224L, 224L, 3L).

2) I found that the "preds" variable has dimension 1 and

shape (365L,) so that np.sort(preds)[::-1][0:5] looked something like [
0.03043453 0.02591384 0.02299733 0.01979745 0.01885794].

*** Are these the predicted probabilities? if so, they are

quite small. please advise.

3) I also found that "top_preds" variable has dimension 1

and shape (5L,) so that np.argsort(preds)[::-1][0:5] looked something like
[236 99 21 317 77].

*** These appear to be the top-5 predicted classes, wherein

the SCENE CATEGORIES are: museum/indoor; coffee_shop; art_studio;
staircase; and campus.

HOWEVER, my image was image_path='tank_desert' and when I

change the test image to 'tank_forest' I get similar classes, i.e., [99
236 21 77 362].

*** I suspect this test script for the VGG16 Places 365

model is hardwired for the jpg image 'restaurant.jpg' as shown in the
example on the github web page for the Places 365 CNN .

*** Note that when I test the attached 'restaurant.jpg' the

top-5 classes are [99, 236, 21, 77, 193] corresponding to the categories
coffee_shop, museum/indoor, art_studio, campus, inn/outdoor.

.
Is there a way to change a configuration file and/or a part

of a .py code to allow for new and different test images?

Any comments or suggestions that you may offer would be

helpful.

Best Regards,
Arnold
PS: I downloaded the categories list locally to my notebook computer

because I was having trouble accessing the file online as the code ran.

.
PSS: to follow up... I modified the script for VGG places 365 CNN to

test the VGG16_hybrid_places_1365 CNN (see further below) and found that
now the variable "x" has dimension 4 and shape 4 (1L, 224L, 224L, 3L). I
also found that "preds" have dimension 3 and shape (7L, 7L, 512L) and
that "top_preds" have shape (5, 7L, 512L) so that I get a very large
array of top-5 class predictions.

***Why are the prediction arrays of a different shape in this case? I

expected "preds" to have shape (1365L) and top_preds again to have shape
(5L).

Again, Any comments or suggestions that you may offer would be helpful.
Best,
Arnold
====================================================================
#script for VGG places 365 CNN
import keras
import numpy as np
import os
from VGG16_places_365 import VGG16_Places365 from keras.preprocessing
import image from places_utils import preprocess_input model =
VGG16_Places365(weights='places')
img_path = r'C:\Users\atunick\VGG16_hybrid_places_1365\tank_desert.jpg'
#img_path = 'tank_desert.jpg'
img = image.load_img(img_path, target_size=(224, 224)) x =
image.img_to_array(img) x = np.expand_dims(x, axis=0) x =
preprocess_input(x) print x.ndim, x.shape, 'x'
predictions_to_return = 5
preds = model.predict(x)[0]
print 'preds', preds.ndim, preds.shape, np.sort(preds)[::-1][0:5]
top_preds = np.argsort(preds)[::-1][0:predictions_to_return]
print 'top_preds', top_preds.ndim, top_preds.shape, top_preds # load
the class label file_name =
r'C:\Users\atunick\VGG16_hybrid_places_1365\categories_places365.txt'
# synset_url = 'Caution-

https://raw.githubusercontent.com/CSAILVision/places365/master/categories_places365.txt
'

# os.system('wget ' + synset_url)
classes = list()
classes.append(line.strip().split(' ')[0][3:]) classes =
tuple(classes) #print classes print('--SCENE CATEGORIES:') # output
#print top_preds[i]
print(classes[top_preds[i]])
print 'completed'
=============================================
#script to test the VGG16_hybrid_places_1365 CNN
import keras
import numpy as np
import os
from VGG16_hybrid_places_1365 import VGG16_Hubrid_1365
from keras.preprocessing import image
from places_utils import preprocess_input
model = VGG16_Hubrid_1365(weights='places', include_top=False)
img_path = r'C:\Users\atunick\VGG16_hybrid_places_1365\tank_desert.jpg'
#img_path = 'tank_desert.jpg'
img = image.load_img(img_path, target_size=(224, 224))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
print x.ndim, x.shape
predictions_to_return = 5
preds = model.predict(x)[0]
print 'preds', preds.ndim, preds.shape, np.sort(preds)[::-1][0:5]
top_preds = np.argsort(preds)[::-1][0:predictions_to_return]
print'top_preds', top_preds.ndim, top_preds.shape, top_preds
print ' '
# load the class label
file_name =

r'C:\Users\atunick\VGG16_hybrid_places_1365\categories_hybrid1365.txt'

# synset_url = 'Caution-

https://raw.githubusercontent.com/CSAILVision/places365/master/categories_hybrid1365.txt
'

# os.system('wget ' + synset_url)
classes = list()
classes.append(line.strip().split(' ')[0][3:])
classes = tuple(classes)
print('--SCENE CATEGORIES:')
# output the prediction
print(classes[top_preds[i]])
print 'completed'

--
---
You received this message because you are subscribed to the Google Groups "theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to theano-users+***@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.