progamergov / pytorch-old-tensorflow-models

Models from Tensorflow and Keras converted to PyTorch

pytorch-old-tensorflow-models

This repository contains the following models converted from Tensorflow to PyTorch:

• Inception5h

• Keras Inceptionv3

pytorch-old-tensorflow-models's Issues

Question

Hey,
first I want to thank you for creating the Repo, it was very helpful so far for me.
Now I started looking into the exact output of the torch model and compared it with the TensorFlow model, however, they don't seem to match.
Could you please explain how you converted the TensorFlow model to Pytorch and that the proper input should be?

This is how I use it on the torch side

    model = Inception5h()
    fileurl = "https://github.com/ProGamerGov/pytorch-old-tensorflow-models/raw/master/inception5h.pth"
    model.load_state_dict(load_state_dict_from_url(fileurl))

    np.random.seed(0)
    input_image = np.float32(np.random.randn(224, 224, 3))
    torch_input = transforms.ToTensor()(input_image).unsqueeze_(0)
    output = model(torch_input)
    output_np = output.detach().numpy()
    np.save("torch_output.npy", output_np)

and on the tensorflow side:

      graph = tf.Graph()
    sess = tf.InteractiveSession(graph=graph)

    t_input = tf.placeholder(np.float32, name="our_input")  # define the input tensor
    t_preprocessed = tf.expand_dims(t_input, 0)

    with gfile.FastGFile("./tensorflow_inception_graph.pb", "rb") as f:
        graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())
    tf.import_graph_def(graph_def, {"input": t_preprocessed})

    def T(layer_name):
        return graph.get_tensor_by_name("import/%s:0" % layer_name)

    np.random.seed(0)
    input_image = np.float32(np.random.randn(224, 224, 3))
    softmax = T("softmax2")
    output = sess.run(softmax, {t_input: input_image})
    np.save("tf_output.npy", output)

If I compare the resulting output it is not identical, is this expected or am I missing something?

Thanks for your feedback!