coremltools
coremltools copied to clipboard
apple
RuntimeError: PyTorch convert function for op 'max_pool2d_with_indices' not implemented.
🐞Describe the bug
- I am trying to convert my existing pytorch model and I do not directly
use max_pool2d_with_indicesanywhere. However, I run into the following errorRuntimeError: PyTorch convert function for op 'max_pool2d_with_indices' not implemented.Here are some parts of my model.
class RegularBottleNeck(nn.Module):
def __init__(self,
input_ch,
output_ch,
projection_ratio=4,
regularizer_prob=0,
dilation=0,
assymmetric=False,
bias=False):
super(RegularBottleNeck, self).__init__()
reduced_depth = input_ch // projection_ratio
self.ext_branch_1 = nn.Sequential(
nn.Conv2d(in_channels=input_ch,
out_channels=reduced_depth,
kernel_size=1,
stride=1,
padding=0,
bias=bias),
nn.BatchNorm2d(reduced_depth),
nn.PReLU()
)
if dilation:
self.ext_branch_2 = nn.Sequential(
nn.Conv2d(in_channels=reduced_depth,
out_channels=reduced_depth,
kernel_size=3,
stride=1,
padding=dilation,
bias=bias,
dilation=dilation),
nn.BatchNorm2d(reduced_depth),
nn.PReLU()
)
elif assymmetric:
self.ext_branch_2 = nn.Sequential(
nn.Conv2d(in_channels=reduced_depth,
out_channels=reduced_depth,
kernel_size=(5, 1),
stride=1,
padding=(2, 0),
bias=bias),
nn.Conv2d(in_channels=reduced_depth,
out_channels=reduced_depth,
kernel_size=(1, 5),
stride=1,
padding=(0, 2),
bias=bias),
nn.BatchNorm2d(reduced_depth),
nn.PReLU()
)
else: # Regular Bottle Neck
self.ext_branch_2 = nn.Sequential(
nn.Conv2d(in_channels=reduced_depth,
out_channels=reduced_depth,
kernel_size=3,
stride=1,
padding=1,
bias=bias),
nn.BatchNorm2d(reduced_depth),
nn.PReLU()
)
self.ext_branch_3 = nn.Sequential(
nn.Conv2d(in_channels=reduced_depth,
out_channels=output_ch,
kernel_size=1,
stride=1,
padding=0,
bias=bias),
nn.BatchNorm2d(output_ch),
nn.PReLU()
)
self.regularizer = nn.Dropout2d(p=regularizer_prob)
self.prelu = nn.PReLU()
class DownSampleBottleNeck(nn.Module):
def __init__(self,
input_ch,
output_ch,
projection_ratio=4,
regularizer_prob=0,
bias=False):
super(DownSampleBottleNeck, self).__init__()
reduced_depth = input_ch // projection_ratio
self.ext_branch_1 = nn.Sequential(
nn.Conv2d(in_channels=input_ch,
out_channels=reduced_depth,
kernel_size=2,
stride=2,
padding=0,
bias=bias),
nn.BatchNorm2d(reduced_depth),
nn.PReLU()
)
self.ext_branch_2 = nn.Sequential(
nn.Conv2d(in_channels=reduced_depth,
out_channels=reduced_depth,
kernel_size=3,
stride=1,
padding=1,
bias=bias),
nn.BatchNorm2d(reduced_depth),
nn.PReLU()
)
self.ext_branch_3 = nn.Sequential(
nn.Conv2d(in_channels=reduced_depth,
out_channels=output_ch,
kernel_size=1,
stride=1,
padding=0,
bias=bias),
nn.BatchNorm2d(output_ch),
nn.PReLU()
)
self.max_pool = nn.MaxPool2d(kernel_size=2, stride=2, padding=0, return_indices=True)
self.regularizer = nn.Dropout2d(p=regularizer_prob)
self.prelu = nn.PReLU()
class UpSampleBottleNeck(nn.Module):
def __init__(self,
input_ch,
output_ch,
projection_ratio=4,
regularizer_prob=0,
bias=False):
super(UpSampleBottleNeck, self).__init__()
reduced_depth = input_ch // projection_ratio
self.main_branch = nn.Sequential(
nn.Conv2d(in_channels=input_ch,
out_channels=output_ch,
kernel_size=3,
stride=1,
padding=1,
bias=bias),
nn.BatchNorm2d(output_ch)
)
self.ext_branch_1 = nn.Sequential(
nn.Conv2d(in_channels=input_ch,
out_channels=reduced_depth,
kernel_size=1,
stride=1,
padding=0,
bias=bias),
nn.BatchNorm2d(reduced_depth),
nn.PReLU()
)
self.ext_branch_2 = nn.Sequential(
nn.ConvTranspose2d(in_channels=reduced_depth,
out_channels=reduced_depth,
kernel_size=3,
stride=2,
padding=1,
output_padding=1,
bias=bias),
nn.BatchNorm2d(reduced_depth),
nn.PReLU()
)
self.ext_branch_3 = nn.Sequential(
nn.Conv2d(in_channels=reduced_depth,
out_channels=output_ch,
kernel_size=1,
stride=1,
padding=0,
bias=bias),
nn.BatchNorm2d(output_ch),
nn.PReLU()
)
self.max_unpool = nn.MaxUnpool2d(kernel_size=2, stride=2, padding=0)
self.regularizer = nn.Dropout2d(p=regularizer_prob)
self.prelu = nn.PReLU()
System environment
- coremltools version 4:
- python version 3.8.0:
- pytorch 1.6.0
Need this too. A lot of recent models use this operator. I tried to rewrite the computing of the indices directly on the model, but I didn't succeed. I've also posted a question on stack overflow on this specific issue
@kir486680 - which of the three models you shared is generating this error? How are you calling coremltools.convert?
Since we have not heard back here, I'm going to close this issue. If we ever get the requested information, I'll reopen.
I also face this issue, so I am adding comments here.
RuntimeError: PyTorch convert function for op 'max_pool2d_with_indices' not implemented.
How to reproduce:
- Download Resnet ONNX model
- Convert to PyTorch using onnx2pytorch.
import onnx
onnx_model = onnx.load(model_name)
from onnx2pytorch import ConvertModel
pytorch_model = ConvertModel(onnx_model)
import torch
torch.save(pytorch_model, 'resnet152-7.pt')
- Convert to TorchScript using Model Tracing
pytorch_model = torch.load('resnet152-7.pt')
# Example input, needed by jit tracer.
example = torch.rand(1, 3, 224, 224)
# # Generate TorchScript by tracing.
pytorch_model.eval()
traced = torch.jit.trace(pytorch_model, example)
- Convert to Core ML.
import coremltools as ct
model = ct.convert(traced, inputs=[ct.TensorType(shape=example.shape)])
@TimYao18 - can you give us a minimal example (standalone code we can just copy and paste)?
import onnx
from onnx2pytorch import ConvertModel
import torch
model_name = 'resnet152-v1-7.onnx'
onnx_model = onnx.load(model_name)
pytorch_model = ConvertModel(onnx_model)
# Example input, needed by jit tracer.
example = torch.rand(1, 3, 224, 224)
# # Generate TorchScript by tracing.
pytorch_model.eval()
traced = torch.jit.trace(pytorch_model, example)
import coremltools as ct
model = ct.convert(traced, inputs=[ct.TensorType(shape=example.shape)])
This isn't standalone or minimal. It uses two external packages and a full sized model.
Ideally, we'd have something that could be used as a unit test.