#!/usr/bin/env python
# encoding: utf-8
import torch
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable
import bob.core
logger = bob.core.log.setup("bob.learn.pytorch")
import time
import os
import numpy
[docs]class CNNTrainer(object):
"""
Class to train a CNN
Attributes
----------
network: :py:class:`torch.nn.Module`
The network to train
batch_size: int
The size of your minibatch
use_gpu: bool
If you would like to use the gpu
verbosity_level: int
The level of verbosity output to stdout
"""
def __init__(self, network, batch_size=64, use_gpu=False, verbosity_level=2, num_classes=2):
""" Init function
Parameters
----------
network: :py:class:`torch.nn.Module`
The network to train
batch_size: int
The size of your minibatch
use_gpu: bool
If you would like to use the gpu
verbosity_level: int
The level of verbosity output to stdout
num_classes: int
The number of classes
"""
self.network = network
self.num_classes = num_classes
self.batch_size = batch_size
self.use_gpu = use_gpu
self.criterion = nn.CrossEntropyLoss()
if self.use_gpu:
self.network.cuda()
bob.core.log.set_verbosity_level(logger, verbosity_level)
[docs] def load_and_initialize_model(self, model_filename):
""" Loads and initialize a model
Parameters
----------
model_filename: str
"""
try:
cp = torch.load(model_filename)
logger.info("model {} loaded".format(model_filename))
except RuntimeError:
# pre-trained model was probably saved using nn.DataParallel ...
cp = torch.load(model_filename, map_location='cpu')
logger.info("model {} loaded on CPU".format(model_filename))
if 'state_dict' in cp:
from collections import OrderedDict
new_state_dict = OrderedDict()
for k, v in cp['state_dict'].items():
name = k[7:]
new_state_dict[name] = v
cp['state_dict'] = new_state_dict
logger.info("state_dict modified")
###########################################################################################################
### for each defined architecture, get the output size in pre-trained model, and change it if necessary ###
# LightCNN
if isinstance(self.network, bob.learn.pytorch.architectures.LightCNN.LightCNN9) \
or isinstance(self.network, bob.learn.pytorch.architectures.LightCNN.LightCNN29) \
or isinstance(self.network, bob.learn.pytorch.architectures.LightCNN.LightCNN29v2):
last_layer_weight = 'fc2.weight'
last_layer_bias = 'fc2.bias'
num_classes_pretrained = cp['state_dict'][last_layer_weight].shape[0]
if num_classes_pretrained == self.num_classes:
self.network.load_state_dict(cp['state_dict'])
else:
var = 1.0 / (cp['state_dict'][last_layer_weight].shape[0])
np_weights = numpy.random.normal(loc=0.0, scale=var, size=((self.num_classes+1), cp['state_dict'][last_layer_weight].shape[1]))
cp['state_dict'][last_layer_weight] = torch.from_numpy(np_weights)
if not (isinstance(self.network, bob.learn.pytorch.architectures.LightCNN.LightCNN29v2)):
cp['state_dict'][last_layer_bias] = torch.zeros(((self.num_classes+1),))
self.network.load_state_dict(cp['state_dict'], strict=True)
logger.info("state_dict loaded for {} with {} classes".format(type(self.network), self.num_classes))
# CNN8
if isinstance(self.network, bob.learn.pytorch.architectures.CNN8):
num_classes_pretrained = cp['state_dict']['classifier.weight'].shape[0]
if num_classes_pretrained == self.num_classes:
self.network.load_state_dict(cp['state_dict'])
else:
var = 1.0 / (cp['state_dict']['classifier.weight'].shape[0])
np_weights = numpy.random.normal(loc=0.0, scale=var, size=((self.num_classes+1), cp['state_dict']['classifier.weight'].shape[1]))
cp['state_dict']['classifier.weight'] = torch.from_numpy(np_weights)
cp['state_dict']['classifier.bias'] = torch.zeros(((self.num_classes+1),))
#self.network.load_state_dict(cp['state_dict'], strict=False)
self.network.load_state_dict(cp['state_dict'], strict=True)
logger.info("state_dict loaded for {} with {} classes".format(type(self.network), self.num_classes))
# CASIANet
if isinstance(self.network, bob.learn.pytorch.architectures.CASIANet):
num_classes_pretrained = cp['state_dict']['classifier.weight'].shape[0]
if num_classes_pretrained == self.num_classes:
self.network.load_state_dict(cp['state_dict'])
else:
var = 1.0 / (cp['state_dict']['classifier.weight'].shape[0])
np_weights = numpy.random.normal(loc=0.0, scale=var, size=((self.num_classes+1), cp['state_dict']['classifier.weight'].shape[1]))
cp['state_dict']['classifier.weight'] = torch.from_numpy(np_weights)
cp['state_dict']['classifier.bias'] = torch.zeros(((self.num_classes+1),))
#self.network.load_state_dict(cp['state_dict'], strict=False)
self.network.load_state_dict(cp['state_dict'], strict=True)
logger.info("state_dict loaded for {} with {} classes".format(type(self.network), self.num_classes))
###########################################################################################################
start_epoch = 0
start_iter = 0
losses = []
if 'epoch' in cp.keys():
start_epoch = cp['epoch']
if 'iteration' in cp.keys():
start_iter = cp['iteration']
if 'losses' in cp.keys():
losses = cp['epoch']
return start_epoch, start_iter, losses
[docs] def save_model(self, output_dir, epoch=0, iteration=0, losses=None):
"""Save the trained network
Parameters
----------
output_dir: str
The directory to write the models to
epoch: int
the current epoch
iteration: int
the current (last) iteration
losses: list(float)
The list of losses since the beginning of training
"""
saved_filename = 'model_{}_{}.pth'.format(epoch, iteration)
saved_path = os.path.join(output_dir, saved_filename)
logger.info('Saving model to {}'.format(saved_path))
cp = {'epoch': epoch,
'iteration': iteration,
'loss': losses,
'state_dict': self.network.cpu().state_dict()
}
torch.save(cp, saved_path)
# moved the model back to GPU if needed
if self.use_gpu :
self.network.cuda()
[docs] def train(self, dataloader, n_epochs=20, learning_rate=0.01, output_dir='out', model=None):
"""Performs the training.
Parameters
----------
dataloader: :py:class:`torch.utils.data.DataLoader`
The dataloader for your data
n_epochs: int
The number of epochs you would like to train for
learning_rate: float
The learning rate for SGD optimizer.
output_dir: str
The directory where you would like to save models
"""
# if model exists, load it
if model is not None:
start_epoch, start_iter, losses = self.load_and_initialize_model(model)
if start_epoch != 0:
logger.info('Previous network was trained up to epoch {}, iteration {}'.format(start_epoch, start_iter))
if losses:
logger.info('Last loss = {}'.format(losses[-1]))
else:
logger.info('Starting training / fine-tuning from pre-trained model')
else:
start_epoch = 0
start_iter = 0
losses = []
logger.info('Starting training from scratch')
# setup optimizer
optimizer = optim.SGD(self.network.parameters(), learning_rate, momentum = 0.9, weight_decay = 0.0005)
# let's go
for epoch in range(start_epoch, n_epochs):
for i, data in enumerate(dataloader, 0):
if i >= start_iter:
start = time.time()
images = data['image']
labels = data['label']
batch_size = len(images)
if self.use_gpu:
images = images.cuda()
labels = labels.cuda()
imagesv = Variable(images)
labelsv = Variable(labels)
output, _ = self.network(imagesv)
loss = self.criterion(output, labelsv)
optimizer.zero_grad()
loss.backward()
optimizer.step()
end = time.time()
logger.info("[{}/{}][{}/{}] => Loss = {} (time spent: {})".format(epoch, n_epochs, i, len(dataloader), loss.item(), (end-start)))
losses.append(loss.item())
# do stuff - like saving models
logger.info("EPOCH {} DONE".format(epoch+1))
self.save_model(output_dir, epoch=(epoch+1), iteration=0, losses=losses)