#!/usr/bin/env python
# coding: utf-8

# # Image Recognition Using Keras-ResNet50
# 
# ## About this project
# We're going to use a pretrained model included in Keras to do image recognition.
# Keras comes with several image recognition models including
# * VGG (<a href="http://www.robots.ox.ac.uk/~vgg/" target="_blank">Visual Geometry Group at the University of Oxford</a>)
# * ResNet50 (MS research)
# * Inception-v3 (Google)
# * Xception (Keras)
# 
# We'll use the ResNet50 in this project. <a href="https://arxiv.org/pdf/1512.03385.pdf" target="_blank">This</a> paper describes ResNets.
# Check out this <a href="https://github.com/Kulbear/deep-learning-coursera/blob/master/Convolutional%20Neural%20Networks/Residual%20Networks%20-%20v1.ipynb" target="_blank">link</a> for details on the model and ResNets in general. 
# 
# We have a few images to pick from to see how well the image classifier performs 
# 
# 

# ## Import and set up model

# In[1]:


import numpy as np
# from tensorflow.keras import backend
from keras.preprocessing import image
from tensorflow.keras.applications import resnet50 # <model


#Image visualization
get_ipython().run_line_magic('pylab', 'inline')
import matplotlib.pyplot as plt
import matplotlib.image as mpimg



# In[2]:


#Now we load the imported model 
model =resnet50.ResNet50() #Create a new instance of the model 


# ### Details of the model layers

# After we load the model we can use `summary()` to look at the model's layers and data flow structure. 

# In[18]:


model.summary()


# ## Load image to predict

# ### Downsample the image

# In[15]:


#Load in the sample image we want to predict
imgloc='shark.jpg'

img =image.load_img(imgloc) #Load in the image

print(shape(img))#Size of image

#Look at loaded image
plt.figure()
plt.imshow(img) 
plt.title('Original')
plt.show()  # display it



#Because the image is too big we need to downsample it in order to 
#pass it to Keras. Keras needs imagesize to be matched to 
#the number of input nodes in the model in this case 224x224 pxls 

imgName = imgloc#'bay.jpg'
#Re- Load in the sample image we want to predict
img =image.load_img(imgName, target_size=(224,224)) #Load in the image and resize


print(shape(img))#Size of image

#Look at loaded image
plt.figure()
plt.imshow(img) 
plt.title('Downsampled')
plt.show()  # display it


# ### More Preprocessing steps

# In[16]:


#Convert image to array (flatten)
x=image.img_to_array(img)

#The neural network is actually expecting (a list) more than 1 image so we will trick it
#Add a 4th dimension to the array

x= np.expand_dims(x,axis=0)

#Normalize data to 0-1 instead of 0-255
x=resnet50.preprocess_input(x)


# ### Run neural network and make prediction

# In[17]:


#Now we run the normalized data through the network and predict
predictions =model.predict(x)

#We will get back a predictions object with 1000 element array. 
#Each element reps a probability that the input matches each of the 1000 objects 
#that the network was trained on.

#Here we use a function to tell us the names of the objects that 
#the network predicted. We only want to top 10 so we ask for ony 10

prediction_classes =resnet50.decode_predictions(predictions, top=10)

#Print out all the predictions
for imagenet_id, name, likelihood in prediction_classes[0]:
    print("-{}: {:2f} % likelihood".format(name, likelihood*100))
    
#Look at loaded image
plt.figure()
plt.imshow(img) 
plt.title('Downsampled')
plt.show()  # display it


# ## Put it all into a function!

# In[31]:


def imgClasser(locc):
    imgName = locc
    #Re- Load in the sample image we want to predict
    img =image.load_img(imgName, target_size=(224,224)) #Load in the image and resize
    
    #Convert image to array (flatten)
    x=image.img_to_array(img)

    #The neural network is actually expecting (a list) more than 1 image so we will trick it
    #Add a 4th dimension to the array

    x= np.expand_dims(x,axis=0)

    #Normalize data to 0-1 instead of 0-255
    x=resnet50.preprocess_input(x)
    
    
    #Now we run the normalized data through the network and predict
    predictions =model.predict(x)

    #We will get back a predictions object with 1000 element array. 
    #Each element reps a probability that the input matches each of the 1000 objects 
    #that the network was trained on.

    #Here we use a function to tell us the names of the objects that 
    #the network predicted. We only want to top 10 so we ask for ony 10

    prediction_classes =resnet50.decode_predictions(predictions, top=10)

    #Print out all the predictions
    for imagenet_id, name, likelihood in prediction_classes[0]:
        print("-{}: {:2f} % likelihood".format(name, likelihood*100))
    #Look at loaded image
    plt.figure()
    plt.imshow(img) 
    plt.title('Downsampled')
    plt.show()  # display it
    


# # Classify Image
# 
# Finally, we use the model to identify an object in a new image. We pass the image to the model and print the top 10 most likely objects that the model thinks it is. 

# In[30]:


imgClasser('burger.jpg')


# We see that the model is able to successfully identify the image as containing a **cheeseburger** with a **99.9% liklihood**!