onnx-coreml
onnx-coreml copied to clipboard
onnx
non-classifier case in coremltools4.0+
❓Question
Is it possible to convert a pytorch model withtout classification layer into mlmodel? I don't know how to do it because in the example, I am asked to specify the label.
System Information
- If applicable
Hi, have you ever encountered this situation?
One day ago, I convert my pytorch model to onnx. It works well. But now I use your tool to do the conversion directly. It seems not good.
This is the complete error.
TypeError: Input strides has type <class 'coremltools.converters.mil.mil.types.type_tensor.tensor.
Hi, have you ever encountered this situation? [image: image] https://user-images.githubusercontent.com/30463291/86051128-d90cd880-ba4c-11ea-88aa-c3038dafa5d6.png One day ago, I convert my pytorch model to onnx. It works well. But now I use your tool to do the conversion directly. It seems not good.
This is the complete error. TypeError: Input strides has type <class 'coremltools.converters.mil.mil.types.type_tensor.tensor..tensor'> not compatible with expected type IntTensorInputType
it works well until the 7-th part [image: image] https://user-images.githubusercontent.com/30463291/86051706-be872f00-ba4d-11ea-9354-b908148d270e.png
On Mon, 29 Jun 2020 at 21:03, Aseem Wadhwa [email protected] wrote:
Just don't provide the classifier_config argument to the ct.convert call, as shown here https://coremltools.readme.io/docs/unified-conversion-api#conversion-from-pytorch
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/onnx/onnx-coreml/issues/581#issuecomment-651329773, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHINKO6DHUZADM3JUWXVDR3RZDXRPANCNFSM4OLRJ5WA .
@huangguanbin can you please provide the pytorch model, which can reproduce this error? Otherwise its hard to debug the error.
no problem. I will do it right now. Give me a second.
On Mon, Jun 29, 2020 at 11:23 PM Aseem Wadhwa [email protected] wrote:
@huangguanbin https://github.com/huangguanbin can you please provide the pytorch model, which can reproduce this error? Otherwise its hard to debug the error.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/onnx/onnx-coreml/issues/581#issuecomment-651401396, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHINKO26V5E5WAGMSM5RQ23RZEH63ANCNFSM4OLRJ5WA .
https://colab.research.google.com/drive/1hfv_Qbl6ZHDbXq_HB9bo3tV89KM7Skoq?usp=sharing This is the notebook where you can reproduce the bug. And you can download the pytorch model by clicking the shared link. It contains a zip file. https://drive.google.com/file/d/1etvQPPGlr4YxBEsCgOeduicErYN18BLg/view?usp=sharing All you need to do is just go through the notebook. Make sure the zip file is in the /content folder. The checkpoint(aka the pytorch model itself) is in the /AlignedReID-master/util/log if you need to access it.
On Mon, Jun 29, 2020 at 11:24 PM 黄冠斌 [email protected] wrote:
no problem. I will do it right now. Give me a second.
On Mon, Jun 29, 2020 at 11:23 PM Aseem Wadhwa [email protected] wrote:
@huangguanbin https://github.com/huangguanbin can you please provide the pytorch model, which can reproduce this error? Otherwise its hard to debug the error.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/onnx/onnx-coreml/issues/581#issuecomment-651401396, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHINKO26V5E5WAGMSM5RQ23RZEH63ANCNFSM4OLRJ5WA .
If you encountered any issue to run my code, plz let me know ASAP. I will reply to you on time.
Is the zip accessible "?
On Mon, 29 Jun 2020 at 23:52, Aseem Wadhwa [email protected] wrote:
The colab is not accessible...
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/onnx/onnx-coreml/issues/581#issuecomment-651409968, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHINKOZJCTOK4TH6Z3O5AVLRZELKBANCNFSM4OLRJ5WA .
https://colab.research.google.com/drive/1hfv_Qbl6ZHDbXq_HB9bo3tV89KM7Skoq?usp=sharing try this one.
On Mon, Jun 29, 2020 at 11:52 PM 黄冠斌 [email protected] wrote:
Is the zip accessible "?
On Mon, 29 Jun 2020 at 23:52, Aseem Wadhwa [email protected] wrote:
The colab is not accessible...
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/onnx/onnx-coreml/issues/581#issuecomment-651409968, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHINKOZJCTOK4TH6Z3O5AVLRZELKBANCNFSM4OLRJ5WA .
if it is not accessible again, copy the code below onto colab.
!unzip AlignedReID-master.zip !pip install coremltools==4.0b1
%cd AlignedReID-master/
import torch from util.FeatureExtractor import FeatureExtractor from torchvision import transforms from IPython import embed import models from scipy.spatial.distance import cosine, euclidean from util.utils import * from sklearn.preprocessing import normalize
def pool2d(tensor, type= 'max'): sz = tensor.size() if type == 'max': x = torch.nn.functional.max_pool2d(tensor, kernel_size=(int(sz[2]/8 ), sz[3])) if type == 'mean': x = torch.nn.functional.mean_pool2d(tensor, kernel_size=(int(sz[2]/8 ), sz[3])) x = x[0].cpu().data.numpy() x = np.transpose(x,(2,1,0))[0] return x
if name == 'main': os.environ['CUDA_VISIBLE_DEVICES'] = "0" use_gpu = torch.cuda.is_available() if torch.cuda.is_available(): map_location = lambda storage, loc: storage.cuda() else: map_location = 'cpu'
model = models.init_model(name='resnet50', num_classes=751, loss={ 'softmax', 'metric'}, use_gpu=use_gpu,aligned=True)
checkpoint = torch.load( "/content/AlignedReID-master/util/log/checkpoint_ep300.pth.tar" , map_location=map_location) model.load_state_dict(checkpoint['state_dict'])
model.eval()
Trace with random data
example_input = torch.rand(1, 3, 256,128) traced_model = torch.jit.trace(model, example_input)
traced_model.save("reid.pt") import coremltools as ct
Convert the saved PyTorch model to Core ML
mlmodel = ct.convert("reid.pt", inputs=[ct.TensorType(shape=(1, 3, 256, 128))])
On Mon, Jun 29, 2020 at 11:54 PM 黄冠斌 [email protected] wrote:
https://colab.research.google.com/drive/1hfv_Qbl6ZHDbXq_HB9bo3tV89KM7Skoq?usp=sharing try this one.
On Mon, Jun 29, 2020 at 11:52 PM 黄冠斌 [email protected] wrote:
Is the zip accessible "?
On Mon, 29 Jun 2020 at 23:52, Aseem Wadhwa [email protected] wrote:
The colab is not accessible...
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/onnx/onnx-coreml/issues/581#issuecomment-651409968, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHINKOZJCTOK4TH6Z3O5AVLRZELKBANCNFSM4OLRJ5WA .
I changed the link of zip file by mistake. If you want to download it again. use this one. https://drive.google.com/file/d/1ufh5kndXMdWfKPRT85dDtEwhZ1MXnPg7/view?usp=sharing
I would appreciate it if you can also check how precise the conversion is. Because some say conversion leads to loss of precision. That means given same input matrix but get different output matrix. THANK YOU!!
On Mon, Jun 29, 2020 at 11:55 PM 黄冠斌 [email protected] wrote:
if it is not accessible again, copy the code below onto colab.
!unzip AlignedReID-master.zip !pip install coremltools==4.0b1
%cd AlignedReID-master/
import torch from util.FeatureExtractor import FeatureExtractor from torchvision import transforms from IPython import embed import models from scipy.spatial.distance import cosine, euclidean from util.utils import * from sklearn.preprocessing import normalize
def pool2d(tensor, type= 'max'): sz = tensor.size() if type == 'max': x = torch.nn.functional.max_pool2d(tensor, kernel_size=(int(sz[2]/8 ), sz[3])) if type == 'mean': x = torch.nn.functional.mean_pool2d(tensor, kernel_size=(int(sz[2]/8 ), sz[3])) x = x[0].cpu().data.numpy() x = np.transpose(x,(2,1,0))[0] return x
if name == 'main': os.environ['CUDA_VISIBLE_DEVICES'] = "0" use_gpu = torch.cuda.is_available() if torch.cuda.is_available(): map_location = lambda storage, loc: storage.cuda() else: map_location = 'cpu'
model = models.init_model(name='resnet50', num_classes=751, loss={ 'softmax', 'metric'}, use_gpu=use_gpu,aligned=True)
checkpoint = torch.load( "/content/AlignedReID-master/util/log/checkpoint_ep300.pth.tar" , map_location=map_location) model.load_state_dict(checkpoint['state_dict'])
model.eval()
Trace with random data
example_input = torch.rand(1, 3, 256,128) traced_model = torch.jit.trace(model, example_input)
traced_model.save("reid.pt") import coremltools as ct
Convert the saved PyTorch model to Core ML
mlmodel = ct.convert("reid.pt", inputs=[ct.TensorType(shape=(1, 3, 256, 128))])
On Mon, Jun 29, 2020 at 11:54 PM 黄冠斌 [email protected] wrote:
https://colab.research.google.com/drive/1hfv_Qbl6ZHDbXq_HB9bo3tV89KM7Skoq?usp=sharing try this one.
On Mon, Jun 29, 2020 at 11:52 PM 黄冠斌 [email protected] wrote:
Is the zip accessible "?
On Mon, 29 Jun 2020 at 23:52, Aseem Wadhwa [email protected] wrote:
The colab is not accessible...
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/onnx/onnx-coreml/issues/581#issuecomment-651409968, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHINKOZJCTOK4TH6Z3O5AVLRZELKBANCNFSM4OLRJ5WA .
Any progress on this?
On Tue, 30 Jun 2020 at 01:13, 黄冠斌 [email protected] wrote:
I would appreciate it if you can also check how precise the conversion is. Because some say conversion leads to loss of precision. That means given same input matrix but get different output matrix. THANK YOU!!
On Mon, Jun 29, 2020 at 11:55 PM 黄冠斌 [email protected] wrote:
if it is not accessible again, copy the code below onto colab.
!unzip AlignedReID-master.zip !pip install coremltools==4.0b1
%cd AlignedReID-master/
import torch from util.FeatureExtractor import FeatureExtractor from torchvision import transforms from IPython import embed import models from scipy.spatial.distance import cosine, euclidean from util.utils import * from sklearn.preprocessing import normalize
def pool2d(tensor, type= 'max'): sz = tensor.size() if type == 'max': x = torch.nn.functional.max_pool2d(tensor, kernel_size=(int(sz[2]/8 ), sz[3])) if type == 'mean': x = torch.nn.functional.mean_pool2d(tensor, kernel_size=(int(sz[2]/ 8), sz[3])) x = x[0].cpu().data.numpy() x = np.transpose(x,(2,1,0))[0] return x
if name == 'main': os.environ['CUDA_VISIBLE_DEVICES'] = "0" use_gpu = torch.cuda.is_available() if torch.cuda.is_available(): map_location = lambda storage, loc: storage.cuda() else: map_location = 'cpu'
model = models.init_model(name='resnet50', num_classes=751, loss={ 'softmax', 'metric'}, use_gpu=use_gpu,aligned=True)
checkpoint = torch.load( "/content/AlignedReID-master/util/log/checkpoint_ep300.pth.tar" , map_location=map_location) model.load_state_dict(checkpoint['state_dict'])
model.eval()
Trace with random data
example_input = torch.rand(1, 3, 256,128) traced_model = torch.jit.trace(model, example_input)
traced_model.save("reid.pt") import coremltools as ct
Convert the saved PyTorch model to Core ML
mlmodel = ct.convert("reid.pt", inputs=[ct.TensorType(shape=(1, 3, 256, 128))])
On Mon, Jun 29, 2020 at 11:54 PM 黄冠斌 [email protected] wrote:
https://colab.research.google.com/drive/1hfv_Qbl6ZHDbXq_HB9bo3tV89KM7Skoq?usp=sharing try this one.
On Mon, Jun 29, 2020 at 11:52 PM 黄冠斌 [email protected] wrote:
Is the zip accessible "?
On Mon, 29 Jun 2020 at 23:52, Aseem Wadhwa [email protected] wrote:
The colab is not accessible...
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/onnx/onnx-coreml/issues/581#issuecomment-651409968, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHINKOZJCTOK4TH6Z3O5AVLRZELKBANCNFSM4OLRJ5WA .
I had the same issue (TypeError: Input strides has type <class 'coremltools.converters.mil.mil.types.type_tensor.tensor..tensor'> not compatible with expected type IntTensorInputType). I am wondering if it is the issue related compatibility between coreML tensor and torch Tensor?
i had same issue:
TypeError: Input strides has type <class 'coremltools.converters.mil.mil.types.type_tensor.tensor.
Seems like the error message for type mismatch of input tensors is not very descriptive. @quangphap208 @JialeHu Which models are you hitting this error on? Since type mismatch can happen in any model for different reasons. Some of the them might be bugs but others might be genuine errors. Therefore, this needs to investigated on a model basis.
@huangguanbin Can you please re-convert your model after incorporating this following change?
Replace this line
strides = inputs[2]
in https://github.com/apple/coremltools/blob/e87f1d0d6ca1dc9cd812f14f777ead0bba70bf43/coremltools/converters/mil/frontend/torch/ops.py#L1421
with following code snippet
strides = inputs[2]
if strides.op.op_type == "const" and (not strides.val):
strides = mb.const(val=kernel_sizes.val, name=strides.name)
Since this issue is not related to ONNX. For PyTorch related issues in the future, please use coremltools repository https://github.com/apple/coremltools.
@huangguanbin This PR https://github.com/apple/coremltools/pull/769 fixes the reported issue. Please verify if it solves your problem.
@huangguanbin This PR apple/coremltools#769 fixes the reported issue. Please verify if it solves your problem.
This solves the issue for me. Thanks a lot.
strides = inputs[2] if strides.op.op_type == "const" and (not strides.val): strides = mb.const(val=kernel_sizes.val, name=strides.name)
I tried what you said, but it is not working yet.
I checked the latest notification. But I'm not sure what i should modify.