X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?p=pysvrt.git;a=blobdiff_plain;f=cnn-svrt.py;h=a6b9cabdb8e1ebbf2c961549bd0764d8c0466e05;hp=d5685f426f518233d710598ea5f1ece4c1e7ce68;hb=HEAD;hpb=664435944d9750efb805d9a2035f1d4f4c238a25

diff --git a/cnn-svrt.py b/cnn-svrt.py
index d5685f4..a6b9cab 100755
--- a/cnn-svrt.py
+++ b/cnn-svrt.py
@@ -1,117 +1,579 @@
-#!/usr/bin/env python-for-pytorch
+#!/usr/bin/env python
+
+#  svrt is the ``Synthetic Visual Reasoning Test'', an image
+#  generator for evaluating classification performance of machine
+#  learning systems, humans and primates.
+#
+#  Copyright (c) 2017 Idiap Research Institute, http://www.idiap.ch/
+#  Written by Francois Fleuret <francois.fleuret@idiap.ch>
+#
+#  This file is part of svrt.
+#
+#  svrt is free software: you can redistribute it and/or modify it
+#  under the terms of the GNU General Public License version 3 as
+#  published by the Free Software Foundation.
+#
+#  svrt is distributed in the hope that it will be useful, but
+#  WITHOUT ANY WARRANTY; without even the implied warranty of
+#  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+#  General Public License for more details.
+#
+#  You should have received a copy of the GNU General Public License
+#  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
 
-from _ext import svrt
+# SVRT
+
+import svrtset
+
+######################################################################
+
+parser = argparse.ArgumentParser(
+    description = "Convolutional networks for the SVRT. Written by Francois Fleuret, (C) Idiap research institute.",
+    formatter_class = argparse.ArgumentDefaultsHelpFormatter
+)
+
+parser.add_argument('--nb_train_samples',
+                    type = int, default = 100000)
+
+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)
+
+parser.add_argument('--batch_size',
+                    type = int, default = 100)
+
+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('--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, '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, last_tag_t
+
+    t = time.time()
+
+    if pred_log_t is None:
+        elapsed = 'start'
+    else:
+        elapsed = '+{:.02f}s'.format(t - pred_log_t)
+
+    pred_log_t = t
+
+    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)
 
 ######################################################################
-# The data
 
-def generate_set(p, n):
-    target = torch.LongTensor(n).bernoulli_(0.5)
-    input = svrt.generate_vignettes(p, target)
-    input = input.view(input.size(0), 1, input.size(1), input.size(2)).float()
-    return Variable(input), Variable(target)
+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)
 
 ######################################################################
 
-# 128x128 --conv(9)-> 120x120 --max(4)-> 30x30 --conv(6)-> 25x25 --max(5)-> 5x5
+# Afroze's ShallowNet
+
+#                       map size   nb. maps
+#                     ----------------------
+#    input                128x128    1
+# -- conv(21x21 x 6)   -> 108x108    6
+# -- max(2x2)          -> 54x54      6
+# -- conv(19x19 x 16)  -> 36x36      16
+# -- max(2x2)          -> 18x18      16
+# -- conv(18x18 x 120) -> 1x1        120
+# -- reshape           -> 120        1
+# -- full(120x84)      -> 84         1
+# -- full(84x2)        -> 2          1
+
+class AfrozeShallowNet(nn.Module):
+    name = 'shallownet'
 
-class Net(nn.Module):
     def __init__(self):
-        super(Net, self).__init__()
-        self.conv1 = nn.Conv2d(1, 10, kernel_size=9)
-        self.conv2 = nn.Conv2d(10, 20, kernel_size=6)
-        self.fc1 = nn.Linear(500, 100)
-        self.fc2 = nn.Linear(100, 2)
+        super(AfrozeShallowNet, self).__init__()
+        self.conv1 = nn.Conv2d(1, 6, kernel_size=21)
+        self.conv2 = nn.Conv2d(6, 16, kernel_size=19)
+        self.conv3 = nn.Conv2d(16, 120, kernel_size=18)
+        self.fc1 = nn.Linear(120, 84)
+        self.fc2 = nn.Linear(84, 2)
 
     def forward(self, x):
-        x = fn.relu(fn.max_pool2d(self.conv1(x), kernel_size=4, stride=4))
-        x = fn.relu(fn.max_pool2d(self.conv2(x), kernel_size=5, stride=5))
-        x = x.view(-1, 500)
+        x = fn.relu(fn.max_pool2d(self.conv1(x), kernel_size=2))
+        x = fn.relu(fn.max_pool2d(self.conv2(x), kernel_size=2))
+        x = fn.relu(self.conv3(x))
+        x = x.view(-1, 120)
         x = fn.relu(self.fc1(x))
         x = self.fc2(x)
         return x
 
-def train_model(train_input, train_target):
-    model, criterion = Net(), nn.CrossEntropyLoss()
+######################################################################
+
+# 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)
+        self.conv2 = nn.Conv2d( 32,  96, kernel_size=5, padding=2)
+        self.conv3 = nn.Conv2d( 96, 128, kernel_size=3, padding=1)
+        self.conv4 = nn.Conv2d(128, 128, kernel_size=3, padding=1)
+        self.conv5 = nn.Conv2d(128,  96, kernel_size=3, padding=1)
+        self.fc1 = nn.Linear(1536, 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 = x.view(-1, 1536)
+
+        x = self.fc1(x)
+        x = fn.relu(x)
+
+        x = self.fc2(x)
+        x = fn.relu(x)
+
+        x = self.fc3(x)
+
+        return x
+
+######################################################################
+
+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()
 
     if torch.cuda.is_available():
-        model.cuda()
         criterion.cuda()
 
-    nb_epochs = 25
-    optimizer, bs = optim.SGD(model.parameters(), lr = 1e-1), 100
+    optimizer = optim.SGD(model.parameters(), lr = 1e-2)
+
+    start_t = time.time()
 
-    for k in range(0, nb_epochs):
-        for b in range(0, nb_train_samples, bs):
-            output = model.forward(train_input.narrow(0, b, bs))
-            loss = criterion(output, train_target.narrow(0, b, bs))
+    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)
+            output = model.forward(Variable(input))
+            loss = criterion(output, Variable(target))
+            acc_loss = acc_loss + loss.data[0]
             model.zero_grad()
             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)) + ']')
+
+        torch.save([ model.state_dict(), e + 1 ], model_filename)
+
+        if validation_set is not None:
+            nb_validation_errors = nb_errors(model, validation_set)
+
+            log_string('validation_error {:.02f}% {:d} {:d}'.format(
+                100 * nb_validation_errors / validation_set.nb_samples,
+                nb_validation_errors,
+                validation_set.nb_samples)
+            )
+
+            if nb_validation_errors / validation_set.nb_samples <= args.validation_error_threshold:
+                log_string('below validation_error_threshold')
+                break
 
     return model
 
 ######################################################################
 
-def print_test_error(model, test_input, test_target):
-    bs = 100
-    nb_test_errors = 0
+for arg in vars(args):
+    log_string('argument ' + str(arg) + ' ' + str(getattr(args, arg)))
+
+######################################################################
+
+def int_to_suffix(n):
+    if n >= 1000000 and n%1000000 == 0:
+        return str(n//1000000) + 'M'
+    elif n >= 1000 and n%1000 == 0:
+        return str(n//1000) + 'K'
+    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
 
-    for b in range(0, nb_test_samples, bs):
-        output = model.forward(test_input.narrow(0, b, bs))
-        _, wta = torch.max(output.data, 1)
+    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
 
-        for i in range(0, bs):
-            if wta[i][0] != test_target.narrow(0, b, bs).data[i]:
-                nb_test_errors = nb_test_errors + 1
+def save_exemplar_vignettes(data_set, nb, name):
+    n = torch.randperm(data_set.nb_samples).narrow(0, 0, nb)
 
-    print('TEST_ERROR {:.02f}% ({:d}/{:d})'.format(
-        100 * nb_test_errors / nb_test_samples,
-        nb_test_errors,
-        nb_test_samples)
-    )
+    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)
 
 ######################################################################
 
-nb_train_samples = 100000
-nb_test_samples = 10000
+if args.nb_train_samples%args.batch_size > 0 or args.nb_test_samples%args.batch_size > 0:
+    print('The number of samples must be a multiple of the batch size.')
+    raise
 
-for p in range(1, 24):
-    print('-- PROBLEM #{:d} --'.format(p))
+if args.compress_vignettes:
+    log_string('using_compressed_vignettes')
+    VignetteSet = svrtset.CompressedVignetteSet
+else:
+    log_string('using_uncompressed_vignettes')
+    VignetteSet = svrtset.VignetteSet
 
-    t1 = time.time()
-    train_input, train_target = generate_set(p, nb_train_samples)
-    test_input, test_target = generate_set(p, nb_test_samples)
-    if torch.cuda.is_available():
-        train_input, train_target = train_input.cuda(), train_target.cuda()
-        test_input, test_target = test_input.cuda(), test_target.cuda()
+########################################
+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) + ' ###############')
+
+    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) + '.pth'
+
+    nb_parameters = 0
+    for p in model.parameters(): nb_parameters += p.numel()
+    log_string('nb_parameters {:d}'.format(nb_parameters))
+
+    ##################################################
+    # Tries to load the model
+
+    try:
+        model_state_dict, nb_epochs_done = torch.load(model_filename)
+        model.load_state_dict(model_state_dict)
+        log_string('loaded_model ' + model_filename)
+    except:
+        nb_epochs_done = 0
+
+
+    ##################################################
+    # Train if necessary
+
+    if nb_epochs_done < args.nb_epochs:
+
+        log_string('training_model ' + model_filename)
+
+        t = time.time()
+
+        train_set = VignetteSet(problem_number,
+                                args.nb_train_samples, args.batch_size,
+                                cuda = torch.cuda.is_available(),
+                                logger = vignette_logger())
+
+        log_string('data_generation {:0.2f} samples / s'.format(
+            train_set.nb_samples / (time.time() - t))
+        )
+
+        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)
+
+        log_string('train_error {:d} {:.02f}% {:d} {:d}'.format(
+            problem_number,
+            100 * nb_train_errors / train_set.nb_samples,
+            nb_train_errors,
+            train_set.nb_samples)
+        )
+
+    ##################################################
+    # Test if necessary
 
-    mu, std = train_input.data.mean(), train_input.data.std()
-    train_input.data.sub_(mu).div_(std)
-    test_input.data.sub_(mu).div_(std)
+    if nb_epochs_done < args.nb_epochs or args.test_loaded_models:
 
-    t2 = time.time()
-    print('[data generation {:.02f}s]'.format(t2 - t1))
-    model = train_model(train_input, train_target)
+        t = time.time()
 
-    t3 = time.time()
-    print('[train {:.02f}s]'.format(t3 - t2))
-    print_test_error(model, test_input, test_target)
+        test_set = VignetteSet(problem_number,
+                               args.nb_test_samples, args.batch_size,
+                               cuda = torch.cuda.is_available())
 
-    t4 = time.time()
+        nb_test_errors = nb_errors(model, test_set,
+                                   mistake_filename_pattern = 'mistake_{:06d}_{:d}.png')
 
-    print('[test {:.02f}s]'.format(t4 - t3))
-    print()
+        log_string('test_error {:d} {:.02f}% {:d} {:d}'.format(
+            problem_number,
+            100 * nb_test_errors / test_set.nb_samples,
+            nb_test_errors,
+            test_set.nb_samples)
+        )
 
 ######################################################################