X-Git-Url: https://www.fleuret.org/cgi-bin/gitweb/gitweb.cgi?p=pysvrt.git;a=blobdiff_plain;f=cnn-svrt.py;h=a6b9cabdb8e1ebbf2c961549bd0764d8c0466e05;hp=d6c7169e917e5c864decb77975d6d31c52197109;hb=HEAD;hpb=1b7eb64f1a3de3761ff887b4cfbc25a81a60b00e diff --git a/cnn-svrt.py b/cnn-svrt.py index d6c7169..a6b9cab 100755 --- a/cnn-svrt.py +++ b/cnn-svrt.py @@ -24,8 +24,10 @@ import time import argparse import math + import distutils.util import re +import signal from colorama import Fore, Back, Style @@ -35,10 +37,12 @@ import torch import torchvision from torch import optim +from torch import multiprocessing from torch import FloatTensor as Tensor from torch.autograd import Variable from torch import nn from torch.nn import functional as fn + from torchvision import datasets, transforms, utils # SVRT @@ -75,15 +79,18 @@ parser.add_argument('--log_file', type = str, default = 'default.log') parser.add_argument('--nb_exemplar_vignettes', - type = int, default = -1) + type = int, default = 32) parser.add_argument('--compress_vignettes', type = distutils.util.strtobool, default = 'True', help = 'Use lossless compression to reduce the memory footprint') -parser.add_argument('--deep_model', - type = distutils.util.strtobool, default = 'True', - help = 'Use Afroze\'s Alexnet-like deep model') +parser.add_argument('--save_test_mistakes', + type = distutils.util.strtobool, default = 'False') + +parser.add_argument('--model', + type = str, default = 'deepnet', + help = 'What model to use') parser.add_argument('--test_loaded_models', type = distutils.util.strtobool, default = 'False', @@ -98,6 +105,10 @@ args = parser.parse_args() ###################################################################### log_file = open(args.log_file, 'a') +log_file.write('\n') +log_file.write('@@@@@@@@@@@@@@@@@@@ ' + time.ctime() + ' @@@@@@@@@@@@@@@@@@@\n') +log_file.write('\n') + pred_log_t = None last_tag_t = time.time() @@ -125,7 +136,24 @@ def log_string(s, remark = ''): log_file.write(re.sub(' ', '_', time.ctime()) + ' ' + elapsed + ' ' + s + '\n') log_file.flush() - print(Fore.BLUE + time.ctime() + ' ' + Fore.GREEN + elapsed + Style.RESET_ALL + ' ' + s + Fore.CYAN + remark + Style.RESET_ALL) + print(Fore.BLUE + time.ctime() + ' ' + Fore.GREEN + elapsed \ + + Style.RESET_ALL + + ' ' \ + + s + Fore.CYAN + remark \ + + Style.RESET_ALL) + +###################################################################### + +def handler_sigint(signum, frame): + log_string('got sigint') + exit(0) + +def handler_sigterm(signum, frame): + log_string('got sigterm') + exit(0) + +signal.signal(signal.SIGINT, handler_sigint) +signal.signal(signal.SIGTERM, handler_sigterm) ###################################################################### @@ -144,6 +172,8 @@ def log_string(s, remark = ''): # -- full(84x2) -> 2 1 class AfrozeShallowNet(nn.Module): + name = 'shallownet' + def __init__(self): super(AfrozeShallowNet, self).__init__() self.conv1 = nn.Conv2d(1, 6, kernel_size=21) @@ -151,7 +181,6 @@ class AfrozeShallowNet(nn.Module): self.conv3 = nn.Conv2d(16, 120, kernel_size=18) self.fc1 = nn.Linear(120, 84) self.fc2 = nn.Linear(84, 2) - self.name = 'shallownet' def forward(self, x): x = fn.relu(fn.max_pool2d(self.conv1(x), kernel_size=2)) @@ -167,6 +196,9 @@ class AfrozeShallowNet(nn.Module): # Afroze's DeepNet class AfrozeDeepNet(nn.Module): + + name = 'deepnet' + def __init__(self): super(AfrozeDeepNet, self).__init__() self.conv1 = nn.Conv2d( 1, 32, kernel_size=7, stride=4, padding=3) @@ -177,7 +209,6 @@ class AfrozeDeepNet(nn.Module): self.fc1 = nn.Linear(1536, 256) self.fc2 = nn.Linear(256, 256) self.fc3 = nn.Linear(256, 2) - self.name = 'deepnet' def forward(self, x): x = self.conv1(x) @@ -212,7 +243,110 @@ class AfrozeDeepNet(nn.Module): ###################################################################### -def nb_errors(model, data_set): +class DeepNet2(nn.Module): + name = 'deepnet2' + + def __init__(self): + super(DeepNet2, self).__init__() + self.nb_channels = 512 + self.conv1 = nn.Conv2d( 1, 32, kernel_size=7, stride=4, padding=3) + self.conv2 = nn.Conv2d( 32, self.nb_channels, kernel_size=5, padding=2) + self.conv3 = nn.Conv2d(self.nb_channels, self.nb_channels, kernel_size=3, padding=1) + self.conv4 = nn.Conv2d(self.nb_channels, self.nb_channels, kernel_size=3, padding=1) + self.conv5 = nn.Conv2d(self.nb_channels, self.nb_channels, kernel_size=3, padding=1) + self.fc1 = nn.Linear(16 * self.nb_channels, 512) + self.fc2 = nn.Linear(512, 512) + self.fc3 = nn.Linear(512, 2) + + def forward(self, x): + x = self.conv1(x) + x = fn.max_pool2d(x, kernel_size=2) + x = fn.relu(x) + + x = self.conv2(x) + x = fn.max_pool2d(x, kernel_size=2) + x = fn.relu(x) + + x = self.conv3(x) + x = fn.relu(x) + + x = self.conv4(x) + x = fn.relu(x) + + x = self.conv5(x) + x = fn.max_pool2d(x, kernel_size=2) + x = fn.relu(x) + + x = x.view(-1, 16 * self.nb_channels) + + x = self.fc1(x) + x = fn.relu(x) + + x = self.fc2(x) + x = fn.relu(x) + + x = self.fc3(x) + + return x + +###################################################################### + +class DeepNet3(nn.Module): + name = 'deepnet3' + + def __init__(self): + super(DeepNet3, self).__init__() + self.conv1 = nn.Conv2d( 1, 32, kernel_size=7, stride=4, padding=3) + self.conv2 = nn.Conv2d( 32, 128, kernel_size=5, padding=2) + self.conv3 = nn.Conv2d(128, 128, kernel_size=3, padding=1) + self.conv4 = nn.Conv2d(128, 128, kernel_size=3, padding=1) + self.conv5 = nn.Conv2d(128, 128, kernel_size=3, padding=1) + self.conv6 = nn.Conv2d(128, 128, kernel_size=3, padding=1) + self.conv7 = nn.Conv2d(128, 128, kernel_size=3, padding=1) + self.fc1 = nn.Linear(2048, 256) + self.fc2 = nn.Linear(256, 256) + self.fc3 = nn.Linear(256, 2) + + def forward(self, x): + x = self.conv1(x) + x = fn.max_pool2d(x, kernel_size=2) + x = fn.relu(x) + + x = self.conv2(x) + x = fn.max_pool2d(x, kernel_size=2) + x = fn.relu(x) + + x = self.conv3(x) + x = fn.relu(x) + + x = self.conv4(x) + x = fn.relu(x) + + x = self.conv5(x) + x = fn.max_pool2d(x, kernel_size=2) + x = fn.relu(x) + + x = self.conv6(x) + x = fn.relu(x) + + x = self.conv7(x) + x = fn.relu(x) + + x = x.view(-1, 2048) + + x = self.fc1(x) + x = fn.relu(x) + + x = self.fc2(x) + x = fn.relu(x) + + x = self.fc3(x) + + return x + +###################################################################### + +def nb_errors(model, data_set, mistake_filename_pattern = None): ne = 0 for b in range(0, data_set.nb_batches): input, target = data_set.get_batch(b) @@ -222,12 +356,19 @@ def nb_errors(model, data_set): for i in range(0, data_set.batch_size): if wta_prediction[i] != target[i]: ne = ne + 1 - + if mistake_filename_pattern is not None: + img = input[i].clone() + img.sub_(img.min()) + img.div_(img.max()) + k = b * data_set.batch_size + i + filename = mistake_filename_pattern.format(k, target[i]) + torchvision.utils.save_image(img, filename) + print(Fore.RED + 'Wrote ' + filename + Style.RESET_ALL) return ne ###################################################################### -def train_model(model, train_set, validation_set): +def train_model(model, model_filename, train_set, validation_set, nb_epochs_done = 0): batch_size = args.batch_size criterion = nn.CrossEntropyLoss() @@ -238,7 +379,7 @@ def train_model(model, train_set, validation_set): start_t = time.time() - for e in range(0, args.nb_epochs): + for e in range(nb_epochs_done, args.nb_epochs): acc_loss = 0.0 for b in range(0, train_set.nb_batches): input, target = train_set.get_batch(b) @@ -253,6 +394,8 @@ def train_model(model, train_set, validation_set): log_string('train_loss {:d} {:f}'.format(e + 1, acc_loss), ' [ETA ' + time.ctime(time.time() + dt * (args.nb_epochs - e)) + ']') + torch.save([ model.state_dict(), e + 1 ], model_filename) + if validation_set is not None: nb_validation_errors = nb_errors(model, validation_set) @@ -299,7 +442,7 @@ class vignette_logger(): ) self.last_t = t -def save_examplar_vignettes(data_set, nb, name): +def save_exemplar_vignettes(data_set, nb, name): n = torch.randperm(data_set.nb_samples).narrow(0, 0, nb) for k in range(0, nb): @@ -320,8 +463,6 @@ if args.nb_train_samples%args.batch_size > 0 or args.nb_test_samples%args.batch_ print('The number of samples must be a multiple of the batch size.') raise -log_string('############### start ###############') - if args.compress_vignettes: log_string('using_compressed_vignettes') VignetteSet = svrtset.CompressedVignetteSet @@ -329,20 +470,30 @@ else: log_string('using_uncompressed_vignettes') VignetteSet = svrtset.VignetteSet +######################################## +model_class = None +for m in [ AfrozeShallowNet, AfrozeDeepNet, DeepNet2, DeepNet3 ]: + if args.model == m.name: + model_class = m + break +if model_class is None: + print('Unknown model ' + args.model) + raise + +log_string('using model class ' + m.name) +######################################## + for problem_number in map(int, args.problems.split(',')): log_string('############### problem ' + str(problem_number) + ' ###############') - if args.deep_model: - model = AfrozeDeepNet() - else: - model = AfrozeShallowNet() + model = model_class() if torch.cuda.is_available(): model.cuda() model_filename = model.name + '_pb:' + \ str(problem_number) + '_ns:' + \ - int_to_suffix(args.nb_train_samples) + '.param' + int_to_suffix(args.nb_train_samples) + '.pth' nb_parameters = 0 for p in model.parameters(): nb_parameters += p.numel() @@ -351,17 +502,18 @@ for problem_number in map(int, args.problems.split(',')): ################################################## # Tries to load the model - need_to_train = False try: - model.load_state_dict(torch.load(model_filename)) + model_state_dict, nb_epochs_done = torch.load(model_filename) + model.load_state_dict(model_state_dict) log_string('loaded_model ' + model_filename) except: - need_to_train = True + nb_epochs_done = 0 + ################################################## # Train if necessary - if need_to_train: + if nb_epochs_done < args.nb_epochs: log_string('training_model ' + model_filename) @@ -377,8 +529,8 @@ for problem_number in map(int, args.problems.split(',')): ) if args.nb_exemplar_vignettes > 0: - save_examplar_vignettes(train_set, args.nb_exemplar_vignettes, - 'examplar_{:d}.png'.format(problem_number)) + save_exemplar_vignettes(train_set, args.nb_exemplar_vignettes, + 'exemplar_{:d}.png'.format(problem_number)) if args.validation_error_threshold > 0.0: validation_set = VignetteSet(problem_number, @@ -388,8 +540,10 @@ for problem_number in map(int, args.problems.split(',')): else: validation_set = None - train_model(model, train_set, validation_set) - torch.save(model.state_dict(), model_filename) + train_model(model, model_filename, + train_set, validation_set, + nb_epochs_done = nb_epochs_done) + log_string('saved_model ' + model_filename) nb_train_errors = nb_errors(model, train_set) @@ -404,7 +558,7 @@ for problem_number in map(int, args.problems.split(',')): ################################################## # Test if necessary - if need_to_train or args.test_loaded_models: + if nb_epochs_done < args.nb_epochs or args.test_loaded_models: t = time.time() @@ -412,7 +566,8 @@ for problem_number in map(int, args.problems.split(',')): args.nb_test_samples, args.batch_size, cuda = torch.cuda.is_available()) - nb_test_errors = nb_errors(model, test_set) + nb_test_errors = nb_errors(model, test_set, + mistake_filename_pattern = 'mistake_{:06d}_{:d}.png') log_string('test_error {:d} {:.02f}% {:d} {:d}'.format( problem_number,