fix coco training issues

huangzsdy · Nov 8, 2016 · 91741f2 · 91741f2
1 parent 4e35411
commit 91741f2
Show file tree

Hide file tree

Showing 10 changed files with 37 additions and 31 deletions.
diff --git a/.gitignore b/.gitignore
@@ -1,4 +1,5 @@
 .ipynb_checkpoints
+.idea
 lib/build
 lib/pycocotools/_mask.c
 lib/pycocotools/_mask.so

diff --git a/lib/rpn/generate.py b/lib/rpn/generate.py
@@ -124,7 +124,7 @@ def imdb_proposals(net, imdb):
 def imdb_rpn_compute_stats(net, imdb, anchor_scales=(8,16,32),
                            feature_stride=16):
     raw_anchors = generate_anchors(scales=np.array(anchor_scales))
-
+    print raw_anchors.shape
     sums = 0
     squred_sums = 0
     counts = 0
@@ -146,10 +146,10 @@ def imdb_rpn_compute_stats(net, imdb, anchor_scales=(8,16,32),
         map_w[i] = width
         map_h[i] = height
 
-    for i in xrange(imdb.num_images):
+    for i in xrange(len(roidb)):
         if not i % 5000:
             print 'computing %d/%d' % (i, imdb.num_images)
-        im = cv2.imread(imdb.image_path_at(i))
+        im = cv2.imread(roidb[i]['image'])
         im_data, im_info = _get_image_blob(im)
         gt_boxes = roidb[i]['boxes']
         gt_boxes = gt_boxes * im_info[0, 2]
@@ -208,4 +208,6 @@ def imdb_rpn_compute_stats(net, imdb, anchor_scales=(8,16,32),
 
     means = sums / counts
     stds = np.sqrt(squred_sums / counts - means ** 2)
+    print means
+    print stds
     return means, stds
diff --git a/models/coco/ResNet-101/rfcn_alt_opt_5step_ohem/rpn_test.pt b/models/coco/ResNet-101/rfcn_alt_opt_5step_ohem/rpn_test.pt
@@ -6292,7 +6292,7 @@ layer {
   param { lr_mult: 1.0 }
   param { lr_mult: 2.0 }
   convolution_param {
-    num_output: 18   # 2(bg/fg) * 9(anchors)
+    num_output: 24   # 2(bg/fg) * 12(anchors)
     kernel_size: 1 pad: 0 stride: 1
     weight_filler { type: "gaussian" std: 0.01 }
     bias_filler { type: "constant" value: 0 }
@@ -6306,7 +6306,7 @@ layer {
   param { lr_mult: 1.0 }
   param { lr_mult: 2.0 }
   convolution_param {
-    num_output: 36   # 4 * 9(anchors)
+    num_output: 48   # 4 * 12(anchors)
     kernel_size: 1 pad: 0 stride: 1
     weight_filler { type: "gaussian" std: 0.01 }
     bias_filler { type: "constant" value: 0 }
@@ -6334,7 +6334,7 @@ layer {
   type: 'Reshape'
   bottom: 'rpn_cls_prob'
   top: 'rpn_cls_prob_reshape'
-  reshape_param { shape { dim: 0 dim: 18 dim: -1 dim: 0 } }
+  reshape_param { shape { dim: 0 dim: 24 dim: -1 dim: 0 } }
 }
 layer {
   name: 'proposal'

diff --git a/models/coco/ResNet-101/rfcn_alt_opt_5step_ohem/stage1_rpn_train.pt b/models/coco/ResNet-101/rfcn_alt_opt_5step_ohem/stage1_rpn_train.pt
@@ -6292,7 +6292,7 @@ layer {
   param { lr_mult: 1.0 }
   param { lr_mult: 2.0 }
   convolution_param {
-    num_output: 18   # 2(bg/fg) * 9(anchors)
+    num_output: 24   # 2(bg/fg) * 12(anchors)
     kernel_size: 1 pad: 0 stride: 1
     weight_filler { type: "gaussian" std: 0.01 }
     bias_filler { type: "constant" value: 0 }
@@ -6306,7 +6306,7 @@ layer {
   param { lr_mult: 1.0 }
   param { lr_mult: 2.0 }
   convolution_param {
-    num_output: 36   # 4 * 9(anchors)
+    num_output: 48   # 4 * 12(anchors)
     kernel_size: 1 pad: 0 stride: 1
     weight_filler { type: "gaussian" std: 0.01 }
     bias_filler { type: "constant" value: 0 }
@@ -6333,7 +6333,7 @@ layer {
   python_param {
     module: 'rpn.anchor_target_layer'
     layer: 'AnchorTargetLayer'
-    param_str: "'feat_stride': 16"
+    param_str: "'feat_stride': 16 \n'scales': !!python/tuple [4, 8, 16, 32]"
   }
 }
 layer {

diff --git a/models/coco/ResNet-101/rfcn_alt_opt_5step_ohem/stage2_rpn_train.pt b/models/coco/ResNet-101/rfcn_alt_opt_5step_ohem/stage2_rpn_train.pt
@@ -6292,7 +6292,7 @@ layer {
   param { lr_mult: 1.0 }
   param { lr_mult: 2.0 }
   convolution_param {
-    num_output: 18   # 2(bg/fg) * 9(anchors)
+    num_output: 24   # 2(bg/fg) * 12(anchors)
     kernel_size: 1 pad: 0 stride: 1
     weight_filler { type: "gaussian" std: 0.01 }
     bias_filler { type: "constant" value: 0 }
@@ -6306,7 +6306,7 @@ layer {
   param { lr_mult: 1.0 }
   param { lr_mult: 2.0 }
   convolution_param {
-    num_output: 36   # 4 * 9(anchors)
+    num_output: 48   # 4 * 12(anchors)
     kernel_size: 1 pad: 0 stride: 1
     weight_filler { type: "gaussian" std: 0.01 }
     bias_filler { type: "constant" value: 0 }
@@ -6333,7 +6333,7 @@ layer {
   python_param {
     module: 'rpn.anchor_target_layer'
     layer: 'AnchorTargetLayer'
-    param_str: "'feat_stride': 16"
+    param_str: "'feat_stride': 16 \n'scales': !!python/tuple [4, 8, 16, 32]"
   }
 }
 layer {

diff --git a/models/coco/ResNet-101/rfcn_alt_opt_5step_ohem/stage3_rpn_train.pt b/models/coco/ResNet-101/rfcn_alt_opt_5step_ohem/stage3_rpn_train.pt
@@ -6292,7 +6292,7 @@ layer {
   param { lr_mult: 1.0 }
   param { lr_mult: 2.0 }
   convolution_param {
-    num_output: 18   # 2(bg/fg) * 9(anchors)
+    num_output: 24   # 2(bg/fg) * 12(anchors)
     kernel_size: 1 pad: 0 stride: 1
     weight_filler { type: "gaussian" std: 0.01 }
     bias_filler { type: "constant" value: 0 }
@@ -6306,7 +6306,7 @@ layer {
   param { lr_mult: 1.0 }
   param { lr_mult: 2.0 }
   convolution_param {
-    num_output: 36   # 4 * 9(anchors)
+    num_output: 48   # 4 * 12(anchors)
     kernel_size: 1 pad: 0 stride: 1
     weight_filler { type: "gaussian" std: 0.01 }
     bias_filler { type: "constant" value: 0 }
@@ -6333,7 +6333,7 @@ layer {
   python_param {
     module: 'rpn.anchor_target_layer'
     layer: 'AnchorTargetLayer'
-    param_str: "'feat_stride': 16"
+    param_str: "'feat_stride': 16 \n'scales': !!python/tuple [4, 8, 16, 32]"
   }
 }
 layer {

diff --git a/models/coco/ResNet-101/rfcn_end2end/test_agnostic.prototxt b/models/coco/ResNet-101/rfcn_end2end/test_agnostic.prototxt
@@ -6954,7 +6954,7 @@ layer {
   param { lr_mult: 1.0 decay_mult: 1.0 }
   param { lr_mult: 2.0 decay_mult: 0 }
   convolution_param {
-    num_output: 18   # 2(bg/fg) * 9(anchors)
+    num_output: 24   # 2(bg/fg) * 12(anchors)
     kernel_size: 1 pad: 0 stride: 1
     weight_filler { type: "gaussian" std: 0.01 }
     bias_filler { type: "constant" value: 0 }
@@ -6968,7 +6968,7 @@ layer {
   param { lr_mult: 1.0 decay_mult: 1.0 }
   param { lr_mult: 2.0 decay_mult: 0 }
   convolution_param {
-    num_output: 36   # 4 * 9(anchors)
+    num_output: 48   # 4 * 12(anchors)
     kernel_size: 1 pad: 0 stride: 1
     weight_filler { type: "gaussian" std: 0.01 }
     bias_filler { type: "constant" value: 0 }
@@ -6995,7 +6995,7 @@ layer {
   type: 'Reshape'
   bottom: 'rpn_cls_prob'
   top: 'rpn_cls_prob_reshape'
-  reshape_param { shape { dim: 0 dim: 18 dim: -1 dim: 0 } }
+  reshape_param { shape { dim: 0 dim: 24 dim: -1 dim: 0 } }
 }
 layer {
   name: 'proposal'
@@ -7007,7 +7007,7 @@ layer {
   python_param {
     module: 'rpn.proposal_layer'
     layer: 'ProposalLayer'
-    param_str: "'feat_stride': 16"
+    param_str: "'feat_stride': 16 \n'scales': !!python/tuple [4, 8, 16, 32]"
   }
 }
 

diff --git a/models/coco/ResNet-101/rfcn_end2end/train_agnostic.prototxt b/models/coco/ResNet-101/rfcn_end2end/train_agnostic.prototxt
@@ -6952,7 +6952,7 @@ layer {
   param { lr_mult: 1.0 }
   param { lr_mult: 2.0 }
   convolution_param {
-    num_output: 18   # 2(bg/fg) * 9(anchors)
+    num_output: 24   # 2(bg/fg) * 12(anchors)
     kernel_size: 1 pad: 0 stride: 1
     weight_filler { type: "gaussian" std: 0.01 }
     bias_filler { type: "constant" value: 0 }
@@ -6967,7 +6967,7 @@ layer {
   param { lr_mult: 1.0 }
   param { lr_mult: 2.0 }
   convolution_param {
-    num_output: 36   # 4 * 9(anchors)
+    num_output: 48   # 4 * 12(anchors)
     kernel_size: 1 pad: 0 stride: 1
     weight_filler { type: "gaussian" std: 0.01 }
     bias_filler { type: "constant" value: 0 }
@@ -6996,7 +6996,7 @@ layer {
   python_param {
     module: 'rpn.anchor_target_layer'
     layer: 'AnchorTargetLayer'
-    param_str: "'feat_stride': 16"
+    param_str: "'feat_stride': 16 \n'scales': !!python/tuple [4, 8, 16, 32]"
   }
 }
 
@@ -7041,7 +7041,7 @@ layer {
   type: 'Reshape'
   bottom: 'rpn_cls_prob'
   top: 'rpn_cls_prob_reshape'
-  reshape_param { shape { dim: 0 dim: 18 dim: -1 dim: 0 } }
+  reshape_param { shape { dim: 0 dim: 24 dim: -1 dim: 0 } }
 }
 
 layer {
@@ -7055,7 +7055,7 @@ layer {
   python_param {
     module: 'rpn.proposal_layer'
     layer: 'ProposalLayer'
-    param_str: "'feat_stride': 16"
+    param_str: "'feat_stride': 16 \n'scales': !!python/tuple [4, 8, 16, 32]"
   }
 }
 

diff --git a/models/coco/ResNet-101/rfcn_end2end/train_agnostic_ohem.prototxt b/models/coco/ResNet-101/rfcn_end2end/train_agnostic_ohem.prototxt
@@ -6953,7 +6953,7 @@ layer {
   param { lr_mult: 1.0 }
   param { lr_mult: 2.0 }
   convolution_param {
-    num_output: 18   # 2(bg/fg) * 9(anchors)
+    num_output: 24   # 2(bg/fg) * 12(anchors)
     kernel_size: 1 pad: 0 stride: 1
     weight_filler { type: "gaussian" std: 0.01 }
     bias_filler { type: "constant" value: 0 }
@@ -6968,7 +6968,7 @@ layer {
   param { lr_mult: 1.0 }
   param { lr_mult: 2.0 }
   convolution_param {
-    num_output: 36   # 4 * 9(anchors)
+    num_output: 48   # 4 * 12(anchors)
     kernel_size: 1 pad: 0 stride: 1
     weight_filler { type: "gaussian" std: 0.01 }
     bias_filler { type: "constant" value: 0 }
@@ -6997,7 +6997,7 @@ layer {
   python_param {
     module: 'rpn.anchor_target_layer'
     layer: 'AnchorTargetLayer'
-    param_str: "'feat_stride': 16"
+    param_str: "'feat_stride': 16 \n'scales': !!python/tuple [4, 8, 16, 32]"
   }
 }
 
@@ -7042,7 +7042,7 @@ layer {
   type: 'Reshape'
   bottom: 'rpn_cls_prob'
   top: 'rpn_cls_prob_reshape'
-  reshape_param { shape { dim: 0 dim: 18 dim: -1 dim: 0 } }
+  reshape_param { shape { dim: 0 dim: 24 dim: -1 dim: 0 } }
 }
 
 layer {
@@ -7056,7 +7056,7 @@ layer {
   python_param {
     module: 'rpn.proposal_layer'
     layer: 'ProposalLayer'
-    param_str: "'feat_stride': 16"
+    param_str: "'feat_stride': 16 \n'scales': !!python/tuple [4, 8, 16, 32]"
   }
 }
 

diff --git a/tools/train_rfcn_alt_opt_5stage.py b/tools/train_rfcn_alt_opt_5stage.py
@@ -84,7 +84,7 @@ def get_solvers(imdb_name, net_name, model_name):
                    [net_name, model_name, 'stage2_rpn_solver360k480k.pt'],
                    [net_name, model_name, 'stage2_rfcn_ohem_solver360k480k.pt'],
                    [net_name, model_name, 'stage3_rpn_solver360k480k.pt']]
-        solvers = [os.path.join(cfg.MODELS_DIR, *s) for s in solvers]
+        solvers = [os.path.join('.', 'models', 'coco', *s) for s in solvers]
         # Iterations for each training stage
         max_iters = [480000, 480000, 480000, 480000, 480000]        
     else:
@@ -257,7 +257,10 @@ def rpn_compute_stats(queue=None, imdb_name=None, cfg=None, rpn_test_prototxt=No
         rpn_net = caffe.Net(rpn_test_prototxt, caffe.TEST)
         # Generate proposals on the imdb
         print 'start computing means/stds, it may take several minutes...'
-        means, stds = imdb_rpn_compute_stats(rpn_net, imdb)
+        if imdb_name.startswith('coco'):
+            means, stds = imdb_rpn_compute_stats(rpn_net, imdb, anchor_scales=(4, 8, 16, 32))
+        else:
+            means, stds = imdb_rpn_compute_stats(rpn_net, imdb, anchor_scales=(8, 16, 32))
         np.save(mean_file, means)
         np.save(std_file, stds)
     queue.put({'means': means, 'stds': stds})