merged

.gitignore

@@ -3,3 +3,4 @@
 *.pyc
 exp/
 data/**/images
+src/models/old

src/main.lua

@@ -1,32 +1,39 @@
 require 'paths'
 paths.dofile('ref.lua')     -- Parse command line input and do global variable initialization
-paths.dofile('data.lua')    -- Set up data processing
 paths.dofile('model.lua')   -- Read in network model
 paths.dofile('train.lua')   -- Load up training/testing functions
 
+-- Set up data loader
 torch.setnumthreads(1)
-local Dataloader = require 'dataloader'
-loader = Dataloader.create(opt)
+local Dataloader = paths.dofile('util/dataloader.lua')
+loader = Dataloader.create(opt, dataset, ref)
 
-isFinished = false -- Finish early if validation accuracy plateaus, can be adjusted with opt.threshold
+-- Initialize logs
+ref.log = {}
+ref.log.train = Logger(paths.concat(opt.save, 'train.log'), opt.continue)
+ref.log.valid = Logger(paths.concat(opt.save, 'valid.log'), opt.continue)
 
 -- Main training loop
 for i=1,opt.nEpochs do
-    train()
-    valid()
-    collectgarbage()
+    print("==> Starting epoch: " .. epoch .. "/" .. (opt.nEpochs + opt.epochNumber - 1))
+    if opt.trainIters > 0 then train() end
+    if opt.validIters > 0 then valid() end
     epoch = epoch + 1
-    if isFinished then break end
+    collectgarbage()
 end
 
--- Update options/reference for last epoch
+-- Update reference for last epoch
 opt.lastEpoch = epoch - 1
-torch.save(opt.save .. '/options.t7', opt)
-
--- Generate final predictions on validation set
-if opt.finalPredictions == 1 then predict() end
 
 -- Save model
 model:clearState()
-torch.save(paths.concat(opt.save,'final_model.t7'), model)
+torch.save(paths.concat(opt.save,'options.t7'), opt)
 torch.save(paths.concat(opt.save,'optimState.t7'), optimState)
+torch.save(paths.concat(opt.save,'final_model.t7'), model)
+
+-- Generate final predictions on validation set
+if opt.finalPredictions then
+	ref.log = {}
+	loader.test = Dataloader(opt, dataset, ref, 'test')
+	predict()
+end

src/misc/convert_annot.py

@@ -3,9 +3,18 @@
 import sys
 import mpii
 
-keys = ['index','person','imgname','center','scale','part','visible','normalize','torsoangle']
+keys = ['index','person','imgname','center','scale','part','visible','normalize','torsoangle','multi','istrain']
 annot = {k:[] for k in keys}
-dotrain = True
+
+# Set up index reference for multiperson training
+multiRef = np.zeros(mpii.nimages)
+trainRef = mpii.annot['img_train'][0][0][0]
+allIdxs = np.arange(0,trainRef.shape[0])
+with h5py.File('../../data/mpii/annot/multi-idxs.h5','r') as f:
+    mTrain = f['train'][:] - 1
+    mTest = f['test'][:] - 1
+    multiRef[allIdxs[trainRef == 1][mTrain]] = 1
+    multiRef[allIdxs[trainRef == 0][mTest]] = 1
 
 # Get image filenames
 imgnameRef = mpii.annot['annolist'][0][0][0]['image'][:]
@@ -14,37 +23,44 @@
     print "\r",idx,
     sys.stdout.flush()
 
-    if mpii.istrain(idx) == dotrain:
-        for person in xrange(mpii.numpeople(idx)):
-            c,s = mpii.location(idx,person)
-            if not c[0] == -1:
-                # Adjust center/scale slightly to avoid cropping limbs
-                # (in hindsight this should have been done in the Torch code...)
-                c[1] += 15 * s
-                s *= 1.25
+    for person in xrange(mpii.numpeople(idx)):
+        c,s = mpii.location(idx,person)
+        if not c[0] == -1:
+            # Add info to annotation list
+            annot['index'] += [idx]
+            annot['person'] += [person]
+            imgname = np.zeros(16)
+            refname = str(imgnameRef[idx][0][0][0][0])
+            for i in range(len(refname)): imgname[i] = ord(refname[i])
+            annot['imgname'] += [imgname]
+            annot['center'] += [c]
+            annot['scale'] += [s]
+            annot['multi'] += [multiRef[idx]]
 
+            if mpii.istrain(idx) == True:
                 # Part annotations and visibility
                 coords = np.zeros((16,2))
                 vis = np.zeros(16)
                 for part in xrange(16):
                    coords[part],vis[part] = mpii.partinfo(idx,person,part)
-
-                # Add info to annotation list
-                annot['index'] += [idx]
-                annot['person'] += [person]
-                annot['imgname'] += [str(imgnameRef[idx][0][0][0][0])]
-                annot['center'] += [c]
-                annot['scale'] += [s]
                 annot['part'] += [coords]
                 annot['visible'] += [vis]
                 annot['normalize'] += [mpii.normalization(idx,person)]
                 annot['torsoangle'] += [mpii.torsoangle(idx,person)]
+                annot['istrain'] += [1]
+            else:
+                annot['part'] += [-np.ones((16,2))]
+                annot['visible'] += [np.zeros(16)]
+                annot['normalize'] += [1]
+                annot['torsoangle'] += [0]
+                if trainRef[idx] == 0:  # Test image
+                    annot['istrain'] += [0]
+                else:   # Training image (something missing in annot)
+                    annot['istrain'] += [2]
 
 print ""
 
-with h5py.File('mpii-annot-train.h5','w') as f:
+with h5py.File('mpii-annot.h5','w') as f:
     f.attrs['name'] = 'mpii'
     for k in keys:
-        if not k == 'imgname': annot[k] = np.array(annot[k])
-        f[k] = annot[k]
-
+        f[k] = np.array(annot[k])