ShairozS

import torch
import numpy as np
import torch.nn.functional as F

def form_pairs(inA, inB):
    
    '''
    Form pairs from two tensors of embeddings. It is assumed that the embeddings at corresponding batch positions are similar
    and all other batch positions are dissimilar 
    

ShairozS

class BertEmbedder(torch.nn.Module):
    
    
    def __init__(self, bert_model='bert-base-uncased'):
        
        '''
        Initialize a BeRT model and use it to create tensor sentence embeddings of length 768
        
        Made from the guide at: https://mccormickml.com/2019/05/14/BERT-word-embeddings-tutorial/
        

ShairozS

ShairozS

def form_triplets(inA, inB):
        
    '''
    Form triplets from two tensors of embeddings. It is assumed that the embeddings at corresponding batch positions are similar
    and all other batch positions are dissimilar 
    
    i.e inA[i] ~ inB[i] and inA[i] !~ inB[j] for all i =! j
    '''
    
    b, emb_size = inA.shape

ShairozS

def form_pairs(inA, inB):
    
    '''
    Form pairs from two tensors of embeddings. It is assumed that the embeddings at corresponding batch positions are similar
    and all other batch positions are dissimilar 
    
    i.e inA[i] ~ inB[i] and inA[i] !~ inB[j] for all i =! j
    '''
    

ShairozS

ShairozS

class ContrastiveLoss(torch.nn.Module):
    
    
    """
    Contrastive loss function.
    Based on: http://yann.lecun.com/exdb/publis/pdf/hadsell-chopra-lecun-06.pdf
    
    args:
        distance (function): A function that returns the distance between two tensors - should be a valid metric over R; default= L2 distance
        margin (scalar): The margin value between positive and negative class ; default=1.0

ShairozS

import torch
import numpy as np



class Trainer:

    ##########################
    #                        #
    #     Initialization     #

ShairozS

import torch
import numpy as np

class Trainer:

    ##########################
    #                        #
    #     Initialization     #
    #                        #
    ##########################

ShairozS

class LabeledContrastiveDataset(Dataset):
    """
    Take a folder containing sub-folders of images, where the sub-folder name is the image class, and generate
    pairs of images with the same class label for contrastive learning. This procedure fixes the batch size since
    every batch contains all classes and every batch element is a unique pairing of each class. 
    """

    def __init__(self, folder, transforms=None):
        
        labels = os.listdir(folder)