timm/timm_image.py

# Copyright 2021 Zilliz. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import logging
import numpy
import os
from pathlib import Path
from typing import List, Union

import towhee
from towhee.operator.base import NNOperator, OperatorFlag
from towhee.types.arg import arg, to_image_color
from towhee import register
from towhee.types import Image

import torch
from torch import nn

from PIL import Image as PILImage

import timm
from timm.data.transforms_factory import create_transform
from timm.data import resolve_data_config
from timm.models.factory import create_model

import warnings

warnings.filterwarnings('ignore')
log = logging.getLogger('timm_op')


@register(output_schema=['vec'])
class TimmImage(NNOperator):
    """
    Pytorch image embedding operator that uses the Pytorch Image Model (timm) collection.
    Args:
        model_name (`str`):
            Which model to use for the embeddings.
        num_classes (`int = 1000`):
            Number of classes for classification.
        skip_preprocess (`bool = False`):
            Whether skip image transforms.
    """

    def __init__(self,
                 model_name: str = 'resnet50',
                 num_classes: int = 1000,
                 skip_preprocess: bool = False,
                 pretrained: bool = True,
                 device: str = None
                 ) -> None:
        super().__init__()
        if device is None:
            device = 'cuda' if torch.cuda.is_available() else 'cpu'
        self.device = device
        self.model_name = model_name
        self.model = create_model(self.model_name, pretrained=pretrained, num_classes=num_classes)
        self.model.eval()
        self.model.to(self.device)

        self.config = resolve_data_config({}, model=self.model)
        self.tfms = create_transform(**self.config)
        self.skip_tfms = skip_preprocess

    def __call__(self, data: Union[List[towhee._types.Image], towhee._types.Image]):
        if not isinstance(data, list):
            imgs = [data]
        else:
            imgs = data
        img_list = []
        for img in imgs:
            img = self.convert_img(img)
            img = img if self.skip_tfms else self.tfms(img)
            img_list.append(img)
        inputs = torch.stack(img_list)
        inputs = inputs.to(self.device)
        features = self.model.forward_features(inputs)
        if features.dim() == 4:
            global_pool = nn.AdaptiveAvgPool2d(1)
            features = global_pool(features)
        features = features.to('cpu').flatten(1)
        if isinstance(data, list):
            vecs = list(features.detach().numpy())
        else:
            vecs = features.squeeze(0).detach().numpy()
        return vecs

    @arg(1, to_image_color('RGB'))
    def convert_img(self, img: towhee._types.Image):
        img = PILImage.fromarray(img.astype('uint8'), 'RGB')
        return img

    def save_model(self, format: str = 'pytorch', path: str = 'default'):
        if path == 'default':
            path = str(Path(__file__).parent)
            path = os.path.join(path, 'saved', format)
            os.makedirs(path, exist_ok=True)
            name = self.model_name.replace('/', '-')
            path = os.path.join(path, name)
        dummy_input = torch.rand((1,) + self.config['input_size'])
        if format == 'pytorch':
            path = path + '.pt'
            torch.save(self.model, path)
        elif format == 'torchscript':
            path = path + '.pt'
            try:
                try:
                    jit_model = torch.jit.script(self.model)
                except Exception:
                    jit_model = torch.jit.trace(self.model, dummy_input, strict=False)
                torch.jit.save(jit_model, path)
            except Exception as e:
                log.error(f'Fail to save as torchscript: {e}.')
                raise RuntimeError(f'Fail to save as torchscript: {e}.')
        elif format == 'onnx':
            path = path + '.onnx'
            try:
                torch.onnx.export(self.model,
                                  dummy_input,
                                  path,
                                  input_names=['input_0'],
                                  output_names=['output_0'],
                                  opset_version=13,
                                  dynamic_axes={
                                      'input_0': {0: 'batch_size'},
                                      'output_0': {0: 'batch_size'}
                                  },
                                  do_constant_folding=True
                                  )
            except Exception as e:
                log.error(f'Fail to save as onnx: {e}.')
                raise RuntimeError(f'Fail to save as onnx: {e}.')
        # todo: elif format == 'tensorrt':
        else:
            log.error(f'Unsupported format "{format}".')

    @staticmethod
    def supported_model_names(format: str = None):
        full_list = timm.list_models(pretrained=True)
        full_list.sort()
        if format is None:
            model_list = full_list
        elif format == 'pytorch':
            to_remove = []
            assert set(to_remove).issubset(set(full_list))
            model_list = list(set(full_list) - set(to_remove))
        elif format == 'onnx':
            to_remove = []
            assert set(to_remove).issubset(set(full_list))
            model_list = list(set(full_list) - set(to_remove))
        # todo: elif format == 'torchscript':
        # todo: elif format == 'tensorrt'
        else:
            log.error(f'Invalid format "{format}". Currently supported formats: "pytorch".')
        return model_list

    def input_schema(self):
        return [(Image, (-1, -1, 3))]

    def output_schema(self):
        image = Image(numpy.random.randn(480, 480, 3), "RGB")
        ret = self(image)
        data_type = type(ret.reshape(-1)[0])
        return [(data_type, ret.shape)]