torchmerge 0.1.16540964

utils

Merge mutiple models or embedding into a single one very easily
in Pytorch.





## Usage

https://colab.research.google.com/drive/1vOFxEcLQdgCxCCJCkp-mxouTQ1f8F5FX?usp=sharing



### Example Short:

import os, random, numpy as np, pandas as pd ;from box import Box
from copy import deepcopy
import copy, collections
import torch
import torch.nn as nn
import torchvision


#############################################################################################
def test3d():    
   from box import Box ; from copy import deepcopy
   from torch.utils.data import DataLoader, TensorDataset
   

   ARG = Box({
       'MODE'   : 'mode1',
       'DATASET': {},
       'MODEL_INFO' : {},
   })
   PARAMS = {}


   ##################################################################
   if ARG.MODE == 'mode1':
       ARG.MODEL_INFO.TYPE = 'dataonly' 
       train_config                           = Box({})
       train_config.LR                        = 0.001
       train_config.SEED                      = 42
       train_config.DEVICE                    = 'cpu'
       train_config.BATCH_SIZE                = 64
       train_config.EPOCHS                    = 1
       train_config.EARLY_STOPPING_THLD       = 10
       train_config.VALID_FREQ                = 1
       train_config.SAVE_FILENAME             = './model.pt'
       train_config.TRAIN_RATIO               = 0.7
       train_config.VAL_RATIO                 = 0.2
       train_config.TEST_RATIO                = 0.1


   ####################################################################
   def load_DataFrame():
       return None


   def test_dataset_f_mnist(samples=100):
       from sklearn.model_selection import train_test_split
       from torchvision import transforms, datasets
       # Generate the transformations
       train_list_transforms = [transforms.ToTensor(),transforms.Lambda(lambda x: x.repeat(3, 1, 1))]

       dataset1 = datasets.FashionMNIST(root="data",train=True,
                                        transform=transforms.Compose(train_list_transforms),download=True,)
       
       #sampling the requred no. of samples from dataset 
       dataset1 = torch.utils.data.Subset(dataset1, np.arange(samples))
       X,Y    = [],  []
       for data, targets in dataset1:
           X.append(data)
           Y.append(targets)

       #Converting list to tensor format
       X,y = torch.stack(X),torch.Tensor(Y)

       train_r, test_r, val_r  = train_config.TRAIN_RATIO, train_config.TEST_RATIO,train_config.VAL_RATIO
       train_X, test_X, train_y, test_y = train_test_split(X,  y,  test_size=1 - train_r)
       valid_X, test_X, valid_y, test_y = train_test_split(test_X, test_y, test_size= test_r / (test_r + val_r))
       return (train_X, train_y, valid_X, valid_y, test_X , test_y)


   def prepro_dataset(self,df:pd.DataFrame=None):
       train_X ,train_y,valid_X ,valid_y,test_X, test_y = test_dataset_f_mnist(samples=100)
       return train_X ,train_y,valid_X ,valid_y,test_X,test_y


   
   ### modelA  ########################################################
   from torchvision import  models
   model_ft = models.resnet18(pretrained=True)
   embA_dim = model_ft.fc.in_features  ###

   ARG.modelA               = {}   
   ARG.modelA.name          = 'resnet18'
   ARG.modelA.nn_model      = model_ft
   ARG.modelA.layer_emb_id  = 'fc'
   ARG.modelA.architect     = [ embA_dim]  ### head s
   modelA = me.model_create(ARG.modelA)
   


   ### modelB  ########################################################
   from torchvision import  models
   model_ft = models.resnet50(pretrained=True)
   embB_dim = int(model_ft.fc.in_features)

   ARG.modelB               = {}   
   ARG.modelB.name          = 'resnet50'
   ARG.modelB.nn_model      = model_ft
   ARG.modelB.layer_emb_id  = 'fc'
   ARG.modelB.architect     = [embB_dim ]   ### head size
   modelB = me.model_create(ARG.modelB )




   ### merge_model  ###################################################
   ### EXPLICIT DEPENDENCY  
   ARG.merge_model           = {}
   ARG.merge_model.name      = 'modelmerge1'

   ARG.merge_model.architect                  = {}
   ARG.merge_model.architect.input_dim        =  embA_dim + embB_dim 

   ARG.merge_model.architect.merge_type       = 'cat'
   ARG.merge_model.architect.merge_layers_dim = [1024, 768]  ### Common embedding is 768
   ARG.merge_model.architect.merge_custom     = None


   ### Custom head
   ARG.merge_model.architect.head_layers_dim  = [ 128, 1]    ### Specific task    
   ARG.merge_model.architect.head_custom      = None
 
 
   ARG.merge_model.dataset       = {}
   ARG.merge_model.dataset.dirin = "/"
   ARG.merge_model.dataset.coly = 'ytarget'
   ARG.merge_model.train_config  = train_config


   model = me.MergeModel_create(ARG, model_create_list= [modelA, modelB ] )
   model.build()



   #### Run Model   ###################################################
   model.training(load_DataFrame, prepro_dataset) 

   model.save_weight('ztmp/model_x5.pt')
   model.load_weights('ztmp/model_x5.pt')
   inputs = torch.randn((train_config.BATCH_SIZE,3,28,28)).to(model.device)
   outputs = model.predict(inputs)
   print(outputs)