performance regression? caffe -> caffe2

[tdp2110]

I have an install of caffe and caffe2 on my desktop linux machine (specs at end of post). I have an NVIDIA GPU and I believe both builds (caffe and caffe2) are using the GPU and CUDNN (both projects were built from source). I compared performance of caffe and caffe2 running squeezenet and found significantly better performance in caffe (using GPU mode) than the new caffe2. Here are my scripts

Caffe on GPU

import caffe
import numpy as np
import time


net_file = '/home/tom/caffe_models/SqueezeNet/SqueezeNet_v1.0/deploy.prototxt'
caffemodel_file = '/home/tom/caffe_models/SqueezeNet/SqueezeNet_v1.0/squeezenet_v1.0.caffemodel'


height, width = 227, 227

caffe.set_mode_gpu()
caffe.set_device(0)

net = caffe.Net(net_file, caffemodel_file, caffe.TEST)
net.blobs['data'].reshape(1, 3, height, width)

times =[]
np.random.seed(123456789)
for _ in xrange(100):
    image = np.random.rand(1, 3, height, width)
    net.blobs['data'].data[...] = image

    t0 = time.time()
    net.forward()
    times.append(time.time() - t0)

print "min time %s, mean time %s, max time %s, time stdev %s" % (
    min(times), np.mean(times), max(times), np.std(times))

and output: min time 0.0024528503418, mean time 0.0026821231842, max time 0.0160629749298, time stdev 0.00134584111746

Caffe2

from caffe2.python import workspace
import numpy as np
import time

def get_predictor():
    with open("init_net.pb", "rb") as f:
        init_net = f.read()
    with open("predict_net.pb", "rb") as f:
        predict_net = f.read()

    return workspace.Predictor(init_net, predict_net)

predictor = get_predictor()

input_image_size = 227
height, width = input_image_size, input_image_size

times = []
np.random.seed(123456789)
for _ in xrange(100):
    image = np.random.rand(1, 3, height, width).astype(np.float32)
    arg = [image]

    t0 = time.time()
    results = predictor.run(arg)
    times.append(time.time() - t0)

print "min time %s, mean time %s, max time %s, time stdev %s" % (
    min(times), np.mean(times), max(times), np.std(times))

and output: min time 0.0587921142578, mean time 0.0777468800545, max time 0.131857872009, time stdev 0.0156148011639

(I saw similar performance between the caffe2 model zoo protocol buffers and those generated from a caffe model using caffe2.python.caffe_translator).

I'm pretty sure my caffe2 setup is using GPU because python -m caffe2.python.operator_test.relu_op_test runs OK and references "engine=CUDNN").

So the question is, is this expected? Am I doing something wrong? I've tried a few other models, and caffe+GPU seems to beat caffe2 every time.

---

Here are my machine specs: I'm running 64-bit Ubuntu 16.04, Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz x 8, 16 GB RAM. And my GPU is a GeForce GTX 1080 with NVIDIA-SMI 375.51 Driver Version: 375.51

[salexspb]

have you checked which engine are you using in Caffe2? You could verify that by looking into predict_net.Proto(). It should have "engine=CUDNN" for every op that you expect to run on CUDA

[tdp2110]

Thanks. So when I parse that string (predict_net) into a caffe2_pb2.NetDef protobuf and iterate over its opfields, none of them have CUDNN engine (in fact, they're all empty strings).

Does this mean that in order to run on GPU, the NetDef protobuf explicitly has to say so?

[tdp2110]

Hmm, this seems to be related to https://github.com/caffe2/caffe2/issues/323, which doesn't seem to be resolved yet.

[sergey-serebryakov]

@tdp2110 It seems your code, the way it's written, must run on CPU. You have to manually specify the appropriate device placement (at least, now).

[tdp2110]

Hi @sergey-serebryakov, thanks. I tried your def load_net_def from #323, with device_opts = core.DeviceOption(caffe2_pb2.CUDA, gpu_id) (having gpu_id = 0) on my predict_net in this example, and I get

Traceback (most recent call last):
  File "squeezenet/run_in_caffe2_simple.py", line 41, in <module>
    results = predictor.run(arg)
RuntimeError: [enforce fail at blob.h:76] IsType<T>(). wrong type for the Blob instance. Blob contains caffe2::Tensor<caffe2::CPUContext> while caller expects caffe2::Tensor<caffe2::CUDAContext> .
Offending Blob name: data.
Error from operator:
input: "data" input: "conv1_w" input: "conv1_b" output: "conv1" type: "Conv" arg { name: "stride" i: 2 } arg { name: "pad" i: 0 } arg { name: "kernel" i: 7 } device_option { device_type: 1 cuda_gpu_id: 0 }

If I use the same device_opts on both predict_net and init_net, I get

Traceback (most recent call last):
  File "squeezenet/run_in_caffe2_simple.py", line 39, in <module>
    results = predictor.run(arg)
RuntimeError: [enforce fail at predictor.cc:11] blob->template IsType<TensorCPU>(). Blob is not a CPU Tensor: data

I'll try to dig into the source to try and make sense of this, but maybe you already know.

[sergey-serebryakov]

@tdp2110 Do you have tensors that you feed into workspace? You can pass device_opts to FeedBlob method as well:

   workspace.FeedBlob(
        '%sdata' % (var_scope),
        np.random.randn(*shape).astype(np.float32),
        device_option=device_opt
    )

I think these are the only places (loading predict and init nets and feeding blobs) where I use device opts.

[tdp2110]

I wasn't explicitly pushing into the workspace with FeedBlob, I was using the Predictor interface, as in some of the tutorials

    image = np.random.rand(1, 3, height, width).astype(np.float32)

    results = predictor.run([image])

I'll try looking into FeedBlob.

[salexspb]

@bwasti , did you have a chance to have a look into Predictor interface and running on GPU ? You have been looking into this if I am not mistaken.

[salexspb]

@tdp2110 , could you try the manual approach for now? (without Predictor)

[tdp2110]

@salexspb thanks, I'll look into the lower-level approach for now.

[raininglixinyu]

@tdp2110 Hi ,i met the same problem when i try to use the gpu, such as blob->template IsType<TensorCPU>(). Blob is not a CPU Tensor: data did you find the reason that cause the problem?

[shahromil16]

@raininglixinyu Assign CPU and GPU separately for the different layers. Try something implemented by KeyKy here: https://github.com/KeyKy/caffe2/blob/master/caffe2/python/tutorials/Run_Alexnet_in_CPU_and_GPU_mode.ipynb

[tdp2110]

Hi @raininglixinyu , I haven't looked at this in a few weeks, and caffe2 is moving pretty fast, but the last time I was working on this I believe my problem was basically that the Predictor interface didn't seem to expose a GPU mode. @salexspb is referring to a lower-level API for constructing nets which does allow you to run on GPU, but I haven't set up an example using it. If I do get a working example for this, I'll be happy to post it. That error in the blob setting you're getting is presumably an incompatibility between the types the net holds and the type you're trying to stuff into it.