Detectron: Train Faster R-CNN with different backbone

[marinaKollmitz]

Thank you very much for this great project! Could you please help me training Faster R-CNN with another backbone architecture, for example GoggleNet. I used the pickle_caffe_blobs.py script to convert the pretrained bvlc_googlenet.caffemodel to a .pkl model as mentioned at issue 77. But how can I train with it? Do I need to setup a .py file for the model like the VGG16.py in lib/modeling? Or is there a way to generate this file as well? Thank you for your help!

[youngwanLEE]

+1

[filipetrocadoferreira]

+2

[marinaKollmitz]

I now implemented the conv body for the network in python and used the converted .pkl file as pretrained weights. To convert the .caffemodel weights to the pickle file, I only kept the conv body part of the network in the deploy.prototxt file. Here is my python implementation, in case anyone is interested. It is a smaller version of the GoogLeNet, called GoogLeNet-xxs, described here. The original network has more inception parts:

def add_googlenet_xxs_conv5_body(model):

    model.Conv('data', 'conv1-7x7_s2', 3, 64, 7, pad=3, stride=2)
    model.Relu('conv1-7x7_s2', 'conv1-7x7_s2')
    model.MaxPool('conv1-7x7_s2', 'pool1-3x3_s2', kernel=3, pad=0, stride=2)     
    model.LRN('pool1-3x3_s2', 'pool1-norm1', size=5, alpha=0.0001, beta=0.75)

    #stop gradient?
    #model.StopGradient('pool1-norm1', 'pool1-norm1')
    model.Conv('pool1-norm1', 'conv2-3x3_reduce', 64, 64, 1, pad=0, stride=1) #input size=?
    model.Relu('conv2-3x3_reduce', 'conv2-3x3_reduce')
    model.Conv('conv2-3x3_reduce', 'conv2-3x3', 64, 192, 3, pad=1, stride=1)
    model.Relu('conv2-3x3', 'conv2-3x3')
    model.LRN('conv2-3x3', 'conv2-norm2', size=5, alpha=0.0001, beta=0.75)
    model.MaxPool('conv2-norm2', 'pool2-3x3_s2', kernel=3, stride=2, pad=0)

    #inception 3a
    model.Conv('pool2-3x3_s2', 'inception_3a-1x1', 192, 64, 1, pad=0, stride=1)
    model.Relu('inception_3a-1x1', 'inception_3a-1x1')
    model.Conv('pool2-3x3_s2', 'inception_3a-3x3_reduce', 192, 96, 1, pad=0, stride=1)
    model.Relu('inception_3a-3x3_reduce', 'inception_3a-3x3_reduce')
    model.Conv('inception_3a-3x3_reduce', 'inception_3a-3x3', 96, 128, 3, pad=1, stride=1)
    model.Relu('inception_3a-3x3', 'inception_3a-3x3')
    model.Conv('pool2-3x3_s2', 'inception_3a-5x5_reduce', 192, 16, 1, pad=0, stride=1)
    model.Relu('inception_3a-5x5_reduce', 'inception_3a-5x5_reduce')
    model.Conv('inception_3a-5x5_reduce', 'inception_3a-5x5', 16, 32, 5, pad=2, stride=1)
    model.Relu('inception_3a-5x5', 'inception_3a-5x5')
    model.MaxPool('pool2-3x3_s2', 'inception_3a-pool', kernel=3, stride=1, pad=1)
    model.Conv('inception_3a-pool', 'inception_3a-pool_proj', 192, 32, 1, stride=1, pad=0)
    model.Relu('inception_3a-pool_proj', 'inception_3a-pool_proj')
    model.Concat(['inception_3a-1x1','inception_3a-3x3', 'inception_3a-5x5', 'inception_3a-pool_proj'], 'inception_3a-output')

    #inception 3b
    model.Conv('inception_3a-output', 'inception_3b-1x1', 256, 128, 1, pad=0, stride=1)
    model.Relu('inception_3b-1x1', 'inception_3b-1x1')
    model.Conv('inception_3a-output', 'inception_3b-3x3_reduce', 256, 128, 1, pad=0, stride=1)
    model.Relu('inception_3b-3x3_reduce', 'inception_3b-3x3_reduce')
    model.Conv('inception_3b-3x3_reduce', 'inception_3b-3x3', 128, 192, 3, pad=1, stride=1)
    model.Relu('inception_3b-3x3', 'inception_3b-3x3')
    model.Conv('inception_3a-output', 'inception_3b-5x5_reduce', 256, 32, 1, pad=0, stride=1)
    model.Relu('inception_3b-5x5_reduce', 'inception_3b-5x5_reduce')
    model.Conv('inception_3b-5x5_reduce', 'inception_3b-5x5', 32, 96, 5, pad=2, stride=1)
    model.Relu('inception_3b-5x5', 'inception_3b-5x5')
    model.MaxPool('inception_3a-output', 'inception_3b-pool', kernel=3, stride=1, pad=1)
    model.Conv('inception_3b-pool', 'inception_3b-pool_proj', 256, 64, 1, pad=0, stride=1)
    model.Relu('inception_3b-pool_proj', 'inception_3b-pool_proj')
    model.Concat(['inception_3b-1x1', 'inception_3b-3x3', 'inception_3b-5x5', 'inception_3b-pool_proj'], 'inception_3b-output')

    model.MaxPool('inception_3b-output', 'pool3-3x3_s2', kernel=3, stride=2, pad=0)

    #inception 4a
    model.Conv('pool3-3x3_s2', 'inception_4a-1x1', 480, 192, 1, pad=0, stride=1)
    model.Relu('inception_4a-1x1', 'inception_4a-1x1')
    model.Conv('pool3-3x3_s2', 'inception_4a-3x3_reduce', 480, 96, 1, pad=0, stride=1)
    model.Relu('inception_4a-3x3_reduce', 'inception_4a-3x3_reduce')
    model.Conv('inception_4a-3x3_reduce', 'inception_4a-3x3', 96, 208, 3, pad=1, stride=1)
    model.Relu('inception_4a-3x3', 'inception_4a-3x3')
    model.Conv('pool3-3x3_s2', 'inception_4a-5x5_reduce', 480, 16, 1, pad=0, stride=1)
    model.Relu('inception_4a-5x5_reduce', 'inception_4a-5x5_reduce')
    model.Conv('inception_4a-5x5_reduce', 'inception_4a-5x5', 16, 48, 5, pad=2, stride=1)
    model.Relu('inception_4a-5x5', 'inception_4a-5x5')
    model.MaxPool('pool3-3x3_s2', 'inception_4a-pool', kernel=3, stride=1, pad=1)
    model.Conv('inception_4a-pool', 'inception_4a-pool_proj', 480, 64, 1, pool=0, stride=1)
    model.Relu('inception_4a-pool_proj', 'inception_4a-pool_proj')
    blob_out = model.Concat(['inception_4a-1x1', 'inception_4a-3x3', 'inception_4a-5x5', 'inception_4a-pool_proj'], 'inception_4a-output')

    return blob_out, 512, 1. / 16.

I am still curious if there is a way to load a network without implementing the network structure in python. Since issue #339 targets this question a little more in detail, I will close this issue here.

[filipetrocadoferreira]

where can we get the weights?

[marinaKollmitz]

Short answer: You can download my generated pretrained weights for the GoogLeNet-xxs here

Longer answer, how I generated the file: In general, what I did is download the .caffemodel and deploy.prototxt from the caffe model zoo here. Then I deleted everything from the deploy.txt except the convolution part I want to train with. Next I fed it to the detectron/tools/pickle_caffe_blobs.py script.

Some details: I had to rename the layers in the implementation, because detectron strips off the layer names before the last / when saving the blobs to model_final.pkl. I replaced all "/"s with "-"s in the layer names. As a consequence, I also had to rename the blobs in the pretrained.pkl file. I changed one part in the pickle_caffe_blobs.py script to rename the blobs when saving to file:

def normalize_googlenet_name(name):
    name = name.replace('/', '-')
    return name

def pickle_weights(out_file_name, weights):
    blobs = {
        normalize_googlenet_name(blob.name): utils.Caffe2TensorToNumpyArray(blob)
        for blob in weights.protos
    }
    with open(out_file_name, 'w') as f:
        pickle.dump(blobs, f, protocol=pickle.HIGHEST_PROTOCOL)
    print('Wrote blobs:')
    print(sorted(blobs.keys()))

I will upload everything to my fork soon, in the meantime, I hope it helps.

[filipetrocadoferreira]

Thank you so much. Good Work. I'll give a try.

Edit: But it would be nice to use directly net definition from '.pb' files to retrain/finetune. This question is general to caffe2 (not detectron)

[rohitbhio]

Hi @marinaKollmitz ,

Awesome work. Thanks for defining CONV body method. Have you also defined a similar method for Fully connected layers?

Does the method defining fully connected layer scheme has similar structure to that of VGG?

Thanks again

[marinaKollmitz]

I just copied the VGG fully connected layers.

[rohitbhio]

@marinaKollmitz : Thanks for confirmation. I did use VGG FC layers.

[srikanth-kilaru]

Problem solved by following recommendation at https://github.com/facebookresearch/Detectron/issues/32 to USE_NCCL True

[srikanth-kilaru]

After training the model, I get the following errors during test.
command:
python2 tools/test_net.py --cfg configs/getting_started/ml349_2gpu_e2e_faster_rcnn
_Inception_ResNetv2.yaml --multi-gpu-testing TEST.WEIGHTS /tmp/detectron-output/train/coco_2014_train/generalized_

rcnn/model_final.pkl NUM_GPUS 2

Errors

==================
INFO net.py: 88: conv1-7x7_s2_w not found
INFO net.py: 88: conv1-7x7_s2_b not found
INFO net.py: 88: conv2-3x3_reduce_w not found
INFO net.py: 88: conv2-3x3_reduce_b not found
INFO net.py: 88: conv2-3x3_w not found
INFO net.py: 88: conv2-3x3_b not found
INFO net.py: 88: inception_3a-1x1_w not found
INFO net.py: 88: inception_3a-1x1_b not found
INFO net.py: 88: inception_3a-3x3_reduce_w not found
INFO net.py: 88: inception_3a-3x3_reduce_b not found
INFO net.py: 88: inception_3a-3x3_w not found
INFO net.py: 88: inception_3a-3x3_b not found
INFO net.py: 88: inception_3a-5x5_reduce_w not found
INFO net.py: 88: inception_3a-5x5_reduce_b not found
INFO net.py: 88: inception_3a-5x5_w not found
INFO net.py: 88: inception_3a-5x5_b not found
INFO net.py: 88: inception_3a-pool_proj_w not found
INFO net.py: 88: inception_3a-pool_proj_b not found
INFO net.py: 88: inception_3b-1x1_w not found
INFO net.py: 88: inception_3b-1x1_b not found
INFO net.py: 88: inception_3b-3x3_reduce_w not found
INFO net.py: 88: inception_3b-3x3_reduce_b not found
INFO net.py: 88: inception_3b-3x3_w not found
INFO net.py: 88: inception_3b-3x3_b not found
INFO net.py: 88: inception_3b-5x5_reduce_w not found
INFO net.py: 88: inception_3b-5x5_reduce_b not found
INFO net.py: 88: inception_3b-5x5_w not found
INFO net.py: 88: inception_3b-5x5_b not found
INFO net.py: 88: inception_3b-pool_proj_w not found
INFO net.py: 88: inception_3b-pool_proj_b not found
INFO net.py: 88: inception_4a-1x1_w not found
INFO net.py: 88: inception_4a-1x1_b not found
INFO net.py: 88: inception_4a-3x3_reduce_w not found
INFO net.py: 88: inception_4a-3x3_reduce_b not found
INFO net.py: 88: inception_4a-3x3_w not found
INFO net.py: 88: inception_4a-3x3_b not found
INFO net.py: 88: inception_4a-5x5_reduce_w not found
INFO net.py: 88: inception_4a-5x5_reduce_b not found
INFO net.py: 88: inception_4a-5x5_w not found
INFO net.py: 88: inception_4a-5x5_b not found
INFO net.py: 88: inception_4a-pool_proj_w not found
INFO net.py: 88: inception_4a-pool_proj_b not found
INFO net.py: 88: conv_rpn_w not found
INFO net.py: 88: conv_rpn_b not found
INFO net.py: 88: rpn_cls_logits_w not found
INFO net.py: 88: rpn_cls_logits_b not found
INFO net.py: 88: rpn_bbox_pred_w not found
INFO net.py: 88: rpn_bbox_pred_b not found
INFO net.py: 88: head_conv1_w not found
INFO net.py: 88: head_conv1_gn_s not found
INFO net.py: 88: head_conv1_gn_b not found
INFO net.py: 88: head_conv2_w not found
INFO net.py: 88: head_conv2_gn_s not found
INFO net.py: 88: head_conv2_gn_b not found
INFO net.py: 88: head_conv3_w not found
INFO net.py: 88: head_conv3_gn_s not found
INFO net.py: 88: head_conv3_gn_b not found
INFO net.py: 88: head_conv4_w not found
INFO net.py: 88: head_conv4_gn_s not found
INFO net.py: 88: head_conv4_gn_b not found
INFO net.py: 88: _mask_fcn1_w not found
INFO net.py: 88: _mask_fcn1_gn_s not found
INFO net.py: 88: _mask_fcn1_gn_b not found
INFO net.py: 88: _mask_fcn2_w not found
INFO net.py: 88: _mask_fcn2_gn_s not found
INFO net.py: 88: _mask_fcn2_gn_b not found
INFO net.py: 88: _mask_fcn3_w not found
INFO net.py: 88: _mask_fcn3_gn_s not found
INFO net.py: 88: _mask_fcn3_gn_b not found
INFO net.py: 88: _mask_fcn4_w not found
INFO net.py: 88: _mask_fcn4_gn_s not found
INFO net.py: 88: _mask_fcn4_gn_b not found
INFO net.py: 88: conv5_mask_w not found
INFO net.py: 88: conv5_mask_b not found
INFO net.py: 88: mask_fcn_logits_w not found
INFO net.py: 88: mask_fcn_logits_b not found
I0529 11:23:29.991972 25591 net_dag_utils.cc:102] Operator graph pruning prior to chain compute took: 5.1028e-05 secs
I0529 11:23:29.992153 25591 net_dag.cc:46] Number of parallel execution chains 25 Number of operators = 78
I0529 11:23:29.995748 25591 net_dag_utils.cc:102] Operator graph pruning prior to chain compute took: 3.4177e-05 secs
I0529 11:23:29.995859 25591 net_dag.cc:46] Number of parallel execution chains 18 Number of operators = 53
I0529 11:23:29.997187 25591 net_dag_utils.cc:102] Operator graph pruning prior to chain compute took: 1.3739e-05 secs
I0529 11:23:29.997264 25591 net_dag.cc:46] Number of parallel execution chains 1 Number of operators = 17
E0529 11:23:54.638962 26563 net_dag.cc:195] Exception from operator chain starting at '' (type 'Conv'): caffe2::EnforceNotMet: [enforce fail at blob.h:84] IsType(). wrong type for the Blob instance. Blob contains nullptr (uninitialized) while caller expects caffe2::Tensor .
Offending Blob name: gpu_0/conv1-7x7_s2_w.
Error from operator:
input: "gpu_0/data" input: "gpu_0/conv1-7x7_s2_w" input: "gpu_0/conv1-7x7_s2_b" output: "gpu_0/conv1-7x7_s2" name: "" type: "Conv" arg { name: "kernel" i: 7 } arg { name: "exhaustive_search" i: 0 } arg { name: "pad" i: 3 } arg { name: "order" s: "NCHW" } arg { name: "stride" i: 2 } device_option { device_type: 1 cuda_gpu_id: 0 } engine: "CUDNN"
WARNING workspace.py: 185: Original python traceback for operator 0 in network generalized_rcnn in exception above (most recent call last):
WARNING workspace.py: 190: File "/home/srikilaru/detectron/tools/test_net.py", line 116, in
WARNING workspace.py: 190: File "/home/srikilaru/detectron/detectron/core/test_engine.py", line 128, in run_inference
WARNING workspace.py: 190: File "/home/srikilaru/detectron/detectron/core/test_engine.py", line 125, in result_getter
WARNING workspace.py: 190: File "/home/srikilaru/detectron/detectron/core/test_engine.py", line 235, in test_net
WARNING workspace.py: 190: File "/home/srikilaru/detectron/detectron/core/test_engine.py", line 328, in initialize_model_from_cfg
WARNING workspace.py: 190: File "/home/srikilaru/detectron/detectron/modeling/model_builder.py", line 124, in create
WARNING workspace.py: 190: File "/home/srikilaru/detectron/detectron/modeling/model_builder.py", line 89, in generalized_rcnn
WARNING workspace.py: 190: File "/home/srikilaru/detectron/detectron/modeling/model_builder.py", line 230, in build_generic_detection_model
WARNING workspace.py: 190: File "/home/srikilaru/detectron/detectron/modeling/optimizer.py", line 54, in build_data_parallel_model
WARNING workspace.py: 190: File "/home/srikilaru/detectron/detectron/modeling/model_builder.py", line 170, in _single_gpu_build_func
WARNING workspace.py: 190: File "/home/srikilaru/detectron/detectron/modeling/Inception_ResNetv2.py", line 27, in add_inception_resnetv2_xxs_conv5_body
WARNING workspace.py: 190: File "/home/srikilaru/pytorch/build/caffe2/python/cnn.py", line 97, in Conv
WARNING workspace.py: 190: File "/home/srikilaru/pytorch/build/caffe2/python/brew.py", line 107, in scope_wrapper
WARNING workspace.py: 190: File "/home/srikilaru/pytorch/build/caffe2/python/helpers/conv.py", line 186, in conv
WARNING workspace.py: 190: File "/home/srikilaru/pytorch/build/caffe2/python/helpers/conv.py", line 139, in _Conv
INFO net.py: 88: inception_3b-pool_proj_b not found
Base
Traceback (most recent call last):
File "/home/srikilaru/detectron/tools/test_net.py", line 116, in
check_expected_results=True,
File "/home/srikilaru/detectron/detectron/core/test_engine.py", line 128, in run_inference
all_results = result_getter()
File "/home/srikilaru/detectron/detectron/core/test_engine.py", line 125, in result_getter
gpu_id=gpu_id
File "/home/srikilaru/detectron/detectron/core/test_engine.py", line 258, in test_net
model, im, box_proposals, timers
File "/home/srikilaru/detectron/detectron/core/test.py", line 66, in im_detect_all
model, im, cfg.TEST.SCALE, cfg.TEST.MAX_SIZE, boxes=box_proposals
File "/home/srikilaru/detectron/detectron/core/test.py", line 158, in im_detect_bbox
workspace.RunNet(model.net.Proto().name)
File "/home/srikilaru/pytorch/build/caffe2/python/workspace.py", line 217, in RunNet
StringifyNetName(name), num_iter, allow_fail,
File "/home/srikilaru/pytorch/build/caffe2/python/workspace.py", line 178, in CallWithExceptionIntercept
return func(args, *kwargs)
RuntimeError: [enforce fail at blob.h:84] IsType(). wrong type for the Blob instance. Blob contains nullptr (uni
nitialized) while caller expects caffe2::Tensor .
Offending Blob name: gpu_0/conv1-7x7_s2_w.
Error from operator:
input: "gpu_0/data" input: "gpu_0/conv1-7x7_s2_w" input: "gpu_0/conv1-7x7_s2_b" output: "gpu_0/conv1-7x7_s2" name:
"" type: "Conv" arg { name: "kernel" i: 7 } arg { name: "exhaustive_search" i: 0 } arg { name: "pad" i: 3 } arg { n
ame: "order" s: "NCHW" } arg { name: "stride" i: 2 } device_option { device_type: 1 cuda_gpu_id: 0 } engine: "CUDNN
"
Traceback (most recent call last):
File "tools/test_net.py", line 116, in
check_expected_results=True,
File "/home/srikilaru/detectron/detectron/core/test_engine.py", line 128, in run_inference
all_results = result_getter()
File "/home/srikilaru/detectron/detectron/core/test_engine.py", line 108, in result_getter
multi_gpu=multi_gpu_testing
File "/home/srikilaru/detectron/detectron/core/test_engine.py", line 155, in test_net_on_dataset
weights_file, dataset_name, proposal_file, num_images, output_dir
File "/home/srikilaru/detectron/detectron/core/test_engine.py", line 188, in multi_gpu_test_net_on_dataset
'detection', num_images, binary, output_dir, opts
File "/home/srikilaru/detectron/detectron/utils/subprocess.py", line 95, in process_in_parallel
log_subprocess_output(i, p, output_dir, tag, start, end)
File "/home/srikilaru/detectron/detectron/utils/subprocess.py", line 133, in log_subprocess_output
assert ret == 0, 'Range subprocess failed (exit code: {})'.format(ret)
AssertionError: Range subprocess failed (exit code: 1)

[johannathiemich]

Could you please explain how I can make use of that .pkl file you provided?
Just loading the .pkl file without changing anything in the config file does not seem to work.

I tried adding the code from your post (add_googlenet_xxs_conv5_body(model)) to the FPN.py file and referenced the method in the corresponding config file when using your provided .pkl file.
But I got an error saying that the blob returned by the add_google_net method does not have attribute "len" therefore an assertion failed. It seems to expect a list of tensors but since the last layer is a Concat layer, only one tensor is returned I think.

So how did you make all of this work? Which changes do I have to make?