santhalakshminarayana

def conv_block(X,filters,block):
    # resiudal block with dilated convolutions
    # add skip connection at last after doing convoluion operation to input X
    
    b = 'block_'+str(block)+'_'
    f1,f2,f3 = filters
    X_skip = X
    # block_a
    X = Convolution2D(filters=f1,kernel_size=(1,1),dilation_rate=(1,1),
                      padding='same',kernel_initializer='he_normal',name=b+'a')(X)

santhalakshminarayana

import argparse
import cv2
import numpy as np

def normalize_kernel(kernel, k_width, k_height, scaling_factor = 1.0):
    '''Zero-summing normalize kernel'''
    
    K_EPS = 1.0e-12
    # positive and negative sum of kernel values
    pos_range, neg_range = 0, 0

santhalakshminarayana

def get_batches_x(tot_seq, batch_size):
    ind = np.random.permutation(tot_seq).tolist()
    i = 0
    for i in range(0, tot_seq, batch_size):
        batch_ids = ind[i:i+batch_size]
        yield X[batch_ids], Y[batch_ids]
        
class Quote_Generator(nn.Module):
    def __init__(self, embed_size, hidden_size, vocab_len):
        super(Quote_Generator, self).__init__()

santhalakshminarayana

# Prepare Train Data
x_train=[]
y_train=[]
person_rep=dict()
person_folders=os.listdir(path+'/Images_crop/')
for i,person in enumerate(person_folders):
  person_rep[i]=person
  image_names=os.listdir('Images_crop/'+person+'/')
  for image_name in image_names:
    img=load_img(path+'/Images_crop/'+person+'/'+image_name,target_size=(224,224))

santhalakshminarayana

def predict(x, temp):
    probs = F.softmax(x / temp, dim = 0)
    probs = np.squeeze(probs.detach().cpu().numpy())
    ind = np.random.choice(vocab_len, 1, p = probs)
    return ind[0]
 
generated_text = ['there','is','no','one','love']
curr_len = 0
embeds = []
is_end = word_to_int[';']

santhalakshminarayana

window = 5 # max_seq_length
sequences, next_words = [], []
for quote in quotes:
    words = quote.split(' ')
    for i in range(0,len(words) - window + 1):
        sequences.append(words[i:i+window])
        if (i + window) < len(words):
            next_words.append(words[i+window])
        else:
            next_words.append(';')

santhalakshminarayana

tsne = TSNE(n_components=2, perplexity=40, n_iter=300,metric='cosine')
n_points = 100
emb_tsne = tsne.fit_transform(emb[:n_points, :])

labels = list(word_to_int.keys())[:n_points]
x = emb_tsne[:,0]
y = emb_tsne[:,1]
plt.figure(figsize=(16, 16)) 
for i in range(n_points):
    plt.scatter(x[i],y[i])

santhalakshminarayana

def get_batch(batch_size):
    ind = np.random.permutation(occs.size).tolist()
    i = 0
    for i in range(0, tot_pairs, batch_size):
        batch_ids = ind[i:i+batch_size]
        yield p1[batch_ids], p2[batch_ids], occs[batch_ids]

device = None
if torch.cuda.is_available():
    device = torch.device("cuda:0")

santhalakshminarayana

def f_x(x, x_max, alpha):
    x = (x/x_max)**alpha
    return torch.min(x, torch.ones_like(x)).to(device)

def weight_mse(w_x, x, log_x):
    loss = w_x * F.mse_loss(x, log_x, reduction='none')
    return torch.mean(loss).to(device)
  
def glove_train(glove):
    epochs = 100

santhalakshminarayana

tot_pairs = len(co_occ_matrix)
p1, p2,occs = list(), list(), list()
for i in co_occ_matrix.keys():
    p1.append(i[0])
    p2.append(i[1])
    occs.append(co_occ_matrix[i])
p1 = np.array(p1)
p2 = np.array(p2)
occs = np.array(occs)