models_dict.py

import torch
import models
import params
from utils import NLL_loss
#

# Model Hyperparameters 
MODELS = {
    'MLP': {
        'name': 'MLP',
        'ref':models.MLP,
        'disc':'',
        'tag': '',
        'hyperparams':{
            'hidden dim':512,
            'tv only': False,
            'task': params.DUAL,
            'curriculum loss': False,
            'curriculum seq': False,
            'curriculum virtual':False,
        },
        'optimizer': torch.optim.Adam,
        'lc loss function': torch.nn.CrossEntropyLoss,
        'ttlc loss function': torch.nn.MSELoss,
        'data type': 'state',
        'state type': '',
    },
    'VLSTM': {
        'name': 'VLSTM',
        'ref':models.VanillaLSTM,
        'disc':'',
        'tag': '',
        'hyperparams':{
            'layer number': 1,
            'tv only': False,
            'hidden dim':512,
            'task': params.REGRESSION,
            'curriculum loss': False,
            'curriculum seq': False,
            'curriculum virtual':False,
        },
        'optimizer': torch.optim.Adam,
        'lc loss function': torch.nn.CrossEntropyLoss,
        'ttlc loss function': torch.nn.MSELoss,
        'data type': 'state',
        'state type': '',
    },
    'VGRU': {
        'name': 'VGRU',
        'ref':models.VanillaGRU,
        'disc':'',
        'tag': '',
        'hyperparams':{
            'layer number': 1,
            'tv only': False,
            'hidden dim':512,
            'task': params.REGRESSION,
            'curriculum loss': False,
            'curriculum seq': False,
            'curriculum virtual':False,
        },
        'optimizer': torch.optim.Adam,
        'lc loss function': torch.nn.CrossEntropyLoss,
        'ttlc loss function': torch.nn.MSELoss,
        'data type': 'state',
        'state type': '',
    },
    'VCNN':{
        'name': 'VCNN',
        'ref':models.VanillaCNN,
        'disc':'',
        'tag': '',
        'hyperparams':{
            'kernel size': 3,
            'channel number':32,
            'merge channels': True,
            'task': params.DUAL,
            'curriculum loss': False,
            'curriculum seq': False,
            'curriculum virtual':False,
            # 'probabilistic model':False,  # False True
            # 'LSTM model':True,
            'model type': 'CNN-LSTM-v2' # Choices = ['Resnet_LSTM', 'pretrainedResnetModel',
                                                    #  'pretrained_denseNet',
                                                    #  'probabilistic' , 'CNN-LSTM-v3', 'CNN-LSTM-v1',
                                                    #    'CNN-LSTM-v2', 'CNN-Linear']
        },
        'optimizer': torch.optim.Adam,
        'lc loss function': torch.nn.CrossEntropyLoss,
        'ttlc loss function': torch.nn.MSELoss(), #NLL_loss, # torch.nn.MSELoss(),  NLL_loss,
        # 'ttlc loss function': NLL_loss, #NLL_loss, # torch.nn.MSELoss(),  NLL_loss,
        'data type': 'image',
        'state type': '',
    },
    'REGIONATTCNN3':{
        'name': 'REGIONATTCNN3',
        'ref':models.ATTCNN3,
        'disc':'attention weights for quad-regions',
        'tag': '',
        'hyperparams':{
            'kernel size': 3,
            'channel number':16,
            'merge channels': True,
            'task': params.DUAL,
            'curriculum loss': False,
            'curriculum seq': False,
            'curriculum virtual':False,
        },
        'optimizer': torch.optim.Adam,
        'lc loss function': torch.nn.CrossEntropyLoss,
        'ttlc loss function': torch.nn.MSELoss,
        'data type': 'image',
        'state type': '',
    },
}