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=0770bf3825f26d4397522b49f4f661be9ab2535e;hb=HEAD;hpb=e6757e3b043bf0ade8ad71335f72c7159865f218

diff --git a/cnn-svrt.py b/cnn-svrt.py
index 0770bf3..a6b9cab 100755
--- a/cnn-svrt.py
+++ b/cnn-svrt.py
@@ -19,29 +19,35 @@
 #  General Public License for more details.
 #
 #  You should have received a copy of the GNU General Public License
-#  along with selector.  If not, see <http://www.gnu.org/licenses/>.
+#  along with svrt.  If not, see <http://www.gnu.org/licenses/>.
 
 import time
 import argparse
 import math
+
 import distutils.util
+import re
+import signal
 
 from colorama import Fore, Back, Style
 
 # Pytorch
 
 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
 
-import vignette_set
+import svrtset
 
 ######################################################################
 
@@ -56,6 +62,13 @@ parser.add_argument('--nb_train_samples',
 parser.add_argument('--nb_test_samples',
                     type = int, default = 10000)
 
+parser.add_argument('--nb_validation_samples',
+                    type = int, default = 10000)
+
+parser.add_argument('--validation_error_threshold',
+                    type = float, default = 0.0,
+                    help = 'Early training termination criterion')
+
 parser.add_argument('--nb_epochs',
                     type = int, default = 50)
 
@@ -65,31 +78,47 @@ parser.add_argument('--batch_size',
 parser.add_argument('--log_file',
                     type = str, default = 'default.log')
 
+parser.add_argument('--nb_exemplar_vignettes',
+                    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',
                     help = 'Should we compute the test errors of loaded models')
 
+parser.add_argument('--problems',
+                    type = str, default = '1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23',
+                    help = 'What problems to process')
+
 args = parser.parse_args()
 
 ######################################################################
 
-log_file = open(args.log_file, 'w')
+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()
 
 print(Fore.RED + 'Logging into ' + args.log_file + Style.RESET_ALL)
 
 # Log and prints the string, with a time stamp. Does not log the
 # remark
+
 def log_string(s, remark = ''):
-    global pred_log_t
+    global pred_log_t, last_tag_t
 
     t = time.time()
 
@@ -100,10 +129,31 @@ def log_string(s, remark = ''):
 
     pred_log_t = t
 
-    log_file.write('[' + time.ctime() + '] ' + elapsed + ' ' + s + '\n')
+    if t > last_tag_t + 3600:
+        last_tag_t = t
+        print(Fore.RED + time.ctime() + Style.RESET_ALL)
+
+    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)
 
 ######################################################################
 
@@ -122,6 +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)
@@ -129,7 +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))
@@ -145,6 +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)
@@ -155,7 +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)
@@ -190,7 +243,132 @@ class AfrozeDeepNet(nn.Module):
 
 ######################################################################
 
-def train_model(model, train_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)
+        output = model.forward(Variable(input))
+        wta_prediction = output.data.max(1)[1].view(-1)
+
+        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, model_filename, train_set, validation_set, nb_epochs_done = 0):
     batch_size = args.batch_size
     criterion = nn.CrossEntropyLoss()
 
@@ -201,7 +379,7 @@ def train_model(model, train_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)
@@ -212,25 +390,26 @@ def train_model(model, train_set):
             loss.backward()
             optimizer.step()
         dt = (time.time() - start_t) / (e + 1)
+
         log_string('train_loss {:d} {:f}'.format(e + 1, acc_loss),
                    ' [ETA ' + time.ctime(time.time() + dt * (args.nb_epochs - e)) + ']')
 
-    return model
+        torch.save([ model.state_dict(), e + 1 ], model_filename)
 
-######################################################################
+        if validation_set is not None:
+            nb_validation_errors = nb_errors(model, validation_set)
 
-def nb_errors(model, data_set):
-    ne = 0
-    for b in range(0, data_set.nb_batches):
-        input, target = data_set.get_batch(b)
-        output = model.forward(Variable(input))
-        wta_prediction = output.data.max(1)[1].view(-1)
+            log_string('validation_error {:.02f}% {:d} {:d}'.format(
+                100 * nb_validation_errors / validation_set.nb_samples,
+                nb_validation_errors,
+                validation_set.nb_samples)
+            )
 
-        for i in range(0, data_set.batch_size):
-            if wta_prediction[i] != target[i]:
-                ne = ne + 1
+            if nb_validation_errors / validation_set.nb_samples <= args.validation_error_threshold:
+                log_string('below validation_error_threshold')
+                break
 
-    return ne
+    return model
 
 ######################################################################
 
@@ -247,6 +426,37 @@ def int_to_suffix(n):
     else:
         return str(n)
 
+class vignette_logger():
+    def __init__(self, delay_min = 60):
+        self.start_t = time.time()
+        self.last_t = self.start_t
+        self.delay_min = delay_min
+
+    def __call__(self, n, m):
+        t = time.time()
+        if t > self.last_t + self.delay_min:
+            dt = (t - self.start_t) / m
+            log_string('sample_generation {:d} / {:d}'.format(
+                m,
+                n), ' [ETA ' + time.ctime(time.time() + dt * (n - m)) + ']'
+            )
+            self.last_t = t
+
+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):
+        b = n[k] // data_set.batch_size
+        m = n[k] % data_set.batch_size
+        i, t = data_set.get_batch(b)
+        i = i[m].float()
+        i.sub_(i.min())
+        i.div_(i.max())
+        if k == 0: patchwork = Tensor(nb, 1, i.size(1), i.size(2))
+        patchwork[k].copy_(i)
+
+    torchvision.utils.save_image(patchwork, name)
+
 ######################################################################
 
 if args.nb_train_samples%args.batch_size > 0 or args.nb_test_samples%args.batch_size > 0:
@@ -255,25 +465,35 @@ if args.nb_train_samples%args.batch_size > 0 or args.nb_test_samples%args.batch_
 
 if args.compress_vignettes:
     log_string('using_compressed_vignettes')
-    VignetteSet = vignette_set.CompressedVignetteSet
+    VignetteSet = svrtset.CompressedVignetteSet
 else:
     log_string('using_uncompressed_vignettes')
-    VignetteSet = vignette_set.VignetteSet
+    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 range(1, 24):
+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()
@@ -282,17 +502,18 @@ for problem_number in range(1, 24):
     ##################################################
     # 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)
 
@@ -300,14 +521,29 @@ for problem_number in range(1, 24):
 
         train_set = VignetteSet(problem_number,
                                 args.nb_train_samples, args.batch_size,
-                                cuda = torch.cuda.is_available())
+                                cuda = torch.cuda.is_available(),
+                                logger = vignette_logger())
 
         log_string('data_generation {:0.2f} samples / s'.format(
             train_set.nb_samples / (time.time() - t))
         )
 
-        train_model(model, train_set)
-        torch.save(model.state_dict(), model_filename)
+        if args.nb_exemplar_vignettes > 0:
+            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,
+                                         args.nb_validation_samples, args.batch_size,
+                                         cuda = torch.cuda.is_available(),
+                                         logger = vignette_logger())
+        else:
+            validation_set = None
+
+        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)
@@ -322,7 +558,7 @@ for problem_number in range(1, 24):
     ##################################################
     # 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()
 
@@ -330,11 +566,8 @@ for problem_number in range(1, 24):
                                args.nb_test_samples, args.batch_size,
                                cuda = torch.cuda.is_available())
 
-        log_string('data_generation {:0.2f} samples / s'.format(
-            test_set.nb_samples / (time.time() - t))
-        )
-
-        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,