Support exporting fpn (facebookresearch#372)

Summary:
Based on orionr's work

 - [x] Solve [the problem about GenerateProposals](facebookresearch#334 (comment))
 - [x] Use the existing [ResizeNearest](https://github.com/caffe2/caffe2/blob/master/caffe2/operators/resize_op.cc#L57) layer instead of UpsampleNearest. ResizeNearest has cpu implementation and neon optimization
 - [x] Make it work (with pytorch/pytorch#7091)

With this PR, FPN is supported in cooperation with pytorch/pytorch#7091. I have verified that it works on `e2e_faster_rcnn_R-50-FPN_1x.yaml`
Pull Request resolved: facebookresearch#372

Reviewed By: newstzpz

Differential Revision: D9213242

Pulled By: rbgirshick

fbshipit-source-id: 8fc7b77e6cbf08adaafd760505dd760df59bfd79

Author: daquexian

tools/convert_pkl_to_pb.py

@@ -50,6 +50,7 @@
 from detectron.utils.logging import setup_logging
 from detectron.utils.model_convert_utils import convert_op_in_proto
 from detectron.utils.model_convert_utils import op_filter
+import detectron.utils.blob as blob_utils
 import detectron.core.test_engine as test_engine
 import detectron.utils.c2 as c2_utils
 import detectron.utils.model_convert_utils as mutils
@@ -124,10 +125,41 @@ def unscope_name(name):
 
 
 def reset_names(names):
-    for i in range(0, len(names)):
+    for i in range(len(names)):
         names[i] = unscope_name(names[i])
 
 
+def convert_collect_and_distribute(
+    op, blobs,
+    roi_canonical_scale,
+    roi_canonical_level,
+    roi_max_level,
+    roi_min_level,
+    rpn_max_level,
+    rpn_min_level,
+    rpn_post_nms_topN,
+):
+    print('Converting CollectAndDistributeFpnRpnProposals'
+          ' Python -> C++:\n{}'.format(op))
+    assert op.name.startswith('CollectAndDistributeFpnRpnProposalsOp'), \
+        'Not valid CollectAndDistributeFpnRpnProposalsOp'
+
+    inputs = [x for x in op.input]
+    ret = core.CreateOperator(
+        'CollectAndDistributeFpnRpnProposals',
+        inputs,
+        list(op.output),
+        roi_canonical_scale=roi_canonical_scale,
+        roi_canonical_level=roi_canonical_level,
+        roi_max_level=roi_max_level,
+        roi_min_level=roi_min_level,
+        rpn_max_level=rpn_max_level,
+        rpn_min_level=rpn_min_level,
+        rpn_post_nms_topN=rpn_post_nms_topN,
+    )
+    return ret
+
+
 def convert_gen_proposals(
     op, blobs,
     rpn_pre_nms_topN,
@@ -136,19 +168,22 @@ def convert_gen_proposals(
     rpn_min_size,
 ):
     print('Converting GenerateProposals Python -> C++:\n{}'.format(op))
-    assert op.name.startswith("GenerateProposalsOp"), "Not valid GenerateProposalsOp"
+    assert op.name.startswith('GenerateProposalsOp'), 'Not valid GenerateProposalsOp'
 
-    spatial_scale = mutils.get_op_arg_valf(op, "spatial_scale", None)
+    spatial_scale = mutils.get_op_arg_valf(op, 'spatial_scale', None)
     assert spatial_scale is not None
 
+    lvl = int(op.input[0][-1]) if op.input[0][-1].isdigit() else None
+
     inputs = [x for x in op.input]
-    anchor_name = "anchor"
+    anchor_name = 'anchor{}'.format(lvl) if lvl else 'anchor'
     inputs.append(anchor_name)
-    blobs[anchor_name] = get_anchors(spatial_scale)
+    anchor_sizes = (cfg.FPN.RPN_ANCHOR_START_SIZE * 2.**(lvl - cfg.FPN.RPN_MIN_LEVEL),) if lvl else cfg.RPN.SIZES
+    blobs[anchor_name] = get_anchors(spatial_scale, anchor_sizes)
     print('anchors {}'.format(blobs[anchor_name]))
 
     ret = core.CreateOperator(
-        "GenerateProposals",
+        'GenerateProposals',
         inputs,
         list(op.output),
         spatial_scale=spatial_scale,
@@ -158,14 +193,13 @@ def convert_gen_proposals(
         min_size=rpn_min_size,
         correct_transform_coords=True,
     )
-
     return ret, anchor_name
 
 
-def get_anchors(spatial_scale):
+def get_anchors(spatial_scale, anchor_sizes):
     anchors = generate_anchors.generate_anchors(
         stride=1. / spatial_scale,
-        sizes=cfg.RPN.SIZES,
+        sizes=anchor_sizes,
         aspect_ratios=cfg.RPN.ASPECT_RATIOS).astype(np.float32)
     return anchors
 
@@ -188,36 +222,78 @@ def convert_op_name(op):
         reset_names(op.output)
         return [op]
 
-    @op_filter(type="Python", inputs=['rpn_cls_probs', 'rpn_bbox_pred', 'im_info'])
-    def convert_gen_proposal(op_in):
-        gen_proposals_op, ext_input = convert_gen_proposals(
-            op_in, blobs,
-            rpn_min_size=float(cfg.TEST.RPN_MIN_SIZE),
-            rpn_post_nms_topN=cfg.TEST.RPN_POST_NMS_TOP_N,
-            rpn_pre_nms_topN=cfg.TEST.RPN_PRE_NMS_TOP_N,
-            rpn_nms_thresh=cfg.TEST.RPN_NMS_THRESH,
-        )
-        net.external_input.extend([ext_input])
-        return [gen_proposals_op]
+    @op_filter(type='Python')
+    def convert_python(op):
+        if op.name.startswith('GenerateProposalsOp'):
+            gen_proposals_op, ext_input = convert_gen_proposals(
+                op, blobs,
+                rpn_min_size=float(cfg.TEST.RPN_MIN_SIZE),
+                rpn_post_nms_topN=cfg.TEST.RPN_POST_NMS_TOP_N,
+                rpn_pre_nms_topN=cfg.TEST.RPN_PRE_NMS_TOP_N,
+                rpn_nms_thresh=cfg.TEST.RPN_NMS_THRESH,
+            )
+            net.external_input.extend([ext_input])
+            return [gen_proposals_op]
+        elif op.name.startswith('CollectAndDistributeFpnRpnProposalsOp'):
+            collect_dist_op = convert_collect_and_distribute(
+                op, blobs,
+                roi_canonical_scale=cfg.FPN.ROI_CANONICAL_SCALE,
+                roi_canonical_level=cfg.FPN.ROI_CANONICAL_LEVEL,
+                roi_max_level=cfg.FPN.ROI_MAX_LEVEL,
+                roi_min_level=cfg.FPN.ROI_MIN_LEVEL,
+                rpn_max_level=cfg.FPN.RPN_MAX_LEVEL,
+                rpn_min_level=cfg.FPN.RPN_MIN_LEVEL,
+                rpn_post_nms_topN=cfg.TEST.RPN_POST_NMS_TOP_N,
+            )
+            return [collect_dist_op]
+        else:
+            raise ValueError('Failed to convert Python op {}'.format(
+                op.name))
+
+    # Only convert UpsampleNearest to ResizeNearest when converting to pb so that the existing models is unchanged
+    # https://github.com/facebookresearch/Detectron/pull/372#issuecomment-410248561
+    @op_filter(type='UpsampleNearest')
+    def convert_upsample_nearest(op):
+        for arg in op.arg:
+            if arg.name == 'scale':
+                scale = arg.i
+                break
+        else:
+            raise KeyError('No attribute "scale" in UpsampleNearest op')
+        resize_nearest_op = core.CreateOperator('ResizeNearest',
+                                                list(op.input),
+                                                list(op.output),
+                                                name=op.name,
+                                                width_scale=float(scale),
+                                                height_scale=float(scale))
+        return resize_nearest_op
 
-    @op_filter(input_has='rois')
+    @op_filter()
     def convert_rpn_rois(op):
-        for j in range(0, len(op.input)):
+        for j in range(len(op.input)):
             if op.input[j] == 'rois':
                 print('Converting op {} input name: rois -> rpn_rois:\n{}'.format(
                     op.type, op))
                 op.input[j] = 'rpn_rois'
+        for j in range(len(op.output)):
+            if op.output[j] == 'rois':
+                print('Converting op {} output name: rois -> rpn_rois:\n{}'.format(
+                    op.type, op))
+                op.output[j] = 'rpn_rois'
         return [op]
 
     @op_filter(type_in=['StopGradient', 'Alias'])
     def convert_remove_op(op):
         print('Removing op {}:\n{}'.format(op.type, op))
         return []
 
+    # We want to apply to all operators, including converted
+    # so run separately
+    convert_op_in_proto(net, convert_remove_op)
+    convert_op_in_proto(net, convert_upsample_nearest)
+    convert_op_in_proto(net, convert_python)
     convert_op_in_proto(net, convert_op_name)
-    convert_op_in_proto(net, [
-        convert_gen_proposal, convert_rpn_rois, convert_remove_op
-    ])
+    convert_op_in_proto(net, convert_rpn_rois)
 
     reset_names(net.external_input)
     reset_names(net.external_output)
@@ -272,6 +348,7 @@ def convert_model_gpu(args, net, init_net):
     cdo_cpu = mutils.get_device_option_cpu()
 
     CPU_OPS = [
+        ["CollectAndDistributeFpnRpnProposals", None],
         ["GenerateProposals", None],
         ["BBoxTransform", None],
         ["BoxWithNMSLimit", None],
@@ -424,10 +501,8 @@ def _prepare_blobs(
     im = cv2.resize(im, None, None, fx=im_scale, fy=im_scale,
                     interpolation=cv2.INTER_LINEAR)
 
-    blob = np.zeros([1, im.shape[0], im.shape[1], 3], dtype=np.float32)
-    blob[0, :, :, :] = im
-    channel_swap = (0, 3, 1, 2)  # swap channel to (k, c, h, w)
-    blob = blob.transpose(channel_swap)
+    # Reuse code in blob_utils and fit FPN
+    blob = blob_utils.im_list_to_blob([im])
 
     blobs = {}
     blobs['data'] = blob
@@ -462,7 +537,7 @@ def run_model_pb(args, net, init_net, im, check_blobs):
         )
 
     try:
-        workspace.RunNet(net.Proto().name)
+        workspace.RunNet(net)
         scores = workspace.FetchBlob('score_nms')
         classids = workspace.FetchBlob('class_nms')
         boxes = workspace.FetchBlob('bbox_nms')
@@ -520,13 +595,16 @@ def main():
         merge_cfg_from_list(args.opts)
     cfg.NUM_GPUS = 1
     assert_and_infer_cfg()
-    logger.info('Conerting model with config:')
+    logger.info('Converting model with config:')
     logger.info(pprint.pformat(cfg))
 
-    assert not cfg.MODEL.KEYPOINTS_ON, "Keypoint model not supported."
-    assert not cfg.MODEL.MASK_ON, "Mask model not supported."
-    assert not cfg.FPN.FPN_ON, "FPN not supported."
-    assert not cfg.RETINANET.RETINANET_ON, "RetinaNet model not supported."
+    # script will stop when it can't find an operator rather
+    # than stopping based on these flags
+    #
+    # assert not cfg.MODEL.KEYPOINTS_ON, "Keypoint model not supported."
+    # assert not cfg.MODEL.MASK_ON, "Mask model not supported."
+    # assert not cfg.FPN.FPN_ON, "FPN not supported."
+    # assert not cfg.RETINANET.RETINANET_ON, "RetinaNet model not supported."
 
     # load model from cfg
     model, blobs = load_model(args)