资源算法caffemodel2pytorch

caffemodel2pytorch

2019-10-09 | |  70 |   0 |   0

This converter can be useful for porting Caffe code and layers to PyTorch. Features:

  • dump caffemodel weights to hdf5, npy, pt and json formats

  • load Caffe models and use them from PyTorch

  • mock PyCaffe API to allow for smooth porting of Caffe-using code (drop-in script for OICR for changing backend in train/eval to PyTorch is below):

    • Net, Blob, SGDSolver

  • wrapping Caffe's Python layers (see the OICR example)

  • example of ROI pooling in PyTorch without manual CUDA code compilation (see the OICR example)

The layer support isn't as complete as in https://github.com/marvis/pytorch-caffe. Currently it supports the following Caffe layers:

  • convolution (num_output, kernel_size, stride, pad, dilation; constant and gaussian weight/bias fillers)

  • inner_product (num_output; constant and gaussian weight/bias fillers)

  • max / avg pooling (kernel_size, stride, pad)

  • relu

  • dropout (dropout_ratio)

  • eltwise (prod, sum, max)

  • softmax (axis)

  • local response norm (local_size, alpha, beta)

Dependencies: protobuf with Python bindings, including protoc binary in PATH.

PRs to enable other layers or layer params are very welcome (see the definition of the modules dictionary in the code)!

License is MIT.

Dump weights to PT or HDF5

# prototxt and caffemodel from https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-readme-md# dumps to PT by default to VGG_ILSVRC_16_layers.caffemodel.ptpython -m caffemodel2pytorch VGG_ILSVRC_16_layers.caffemodel# dumps to HDF5 converted.h5python -m caffemodel2pytorch VGG_ILSVRC_16_layers.caffemodel -o converted.h5
# load dumped VGG16 in PyTorchimport collections, torch, torchvision, numpy, h5py
model = torchvision.models.vgg16()
model.features = torch.nn.Sequential(collections.OrderedDict(zip(['conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1', 'conv2_1', 'relu2_1', 'conv2_2', 'relu2_2', 'pool2', 'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2', 'conv3_3', 'relu3_3', 'pool3', 'conv4_1', 'relu4_1', 'conv4_2', 'relu4_2', 'conv4_3', 'relu4_3', 'pool4', 'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'conv5_3', 'relu5_3', 'pool5'], model.features)))
model.classifier = torch.nn.Sequential(collections.OrderedDict(zip(['fc6', 'relu6', 'drop6', 'fc7', 'relu7', 'drop7', 'fc8'], model.classifier)))

state_dict = h5py.File('converted.h5', 'r') # torch.load('VGG_ILSVRC_16_layers.caffemodel.pt')model.load_state_dict({l : torch.from_numpy(numpy.array(v)).view_as(p) for k, v in state_dict.items() for l, p in model.named_parameters() if k in l})

Run Caffe models using PyTorch as backend

import torchimport caffemodel2pytorch

model = caffemodel2pytorch.Net(	prototxt = 'VGG_ILSVRC_16_layers_deploy.prototxt',	weights = 'VGG_ILSVRC_16_layers.caffemodel',	caffe_proto = 'https://raw.githubusercontent.com/BVLC/caffe/master/src/caffe/proto/caffe.proto')
model.cuda()
model.eval()
torch.set_grad_enabled(False)# make sure to have right procedure of image normalization and channel reorderingimage = torch.autograd.Variable(torch.Tensor(8, 3, 224, 224).cuda())# outputs dict of PyTorch Variables# in this example the dict contains the only key "prob"#output_dict = model(data = image)# you can remove unneeded layers:del model.probdel model.fc8# a single input variable is interpreted as an input blob named "data"# in this example the dict contains the only key "fc7"output_dict = model(image)

Imitate pycaffe interface to help in porting

import numpy as npimport caffemodel2pytorch as caffe

caffe.set_mode_gpu()
caffe.set_device(0)# === LOADING AND USING THE NET IN EVAL MODE ===net = caffe.Net('VGG_ILSVRC_16_layers_deploy.prototxt', caffe.TEST, weights = 'VGG_ILSVRC_16_layers.caffemodel')# outputs a dict of NumPy arrays, data layer is sidesteppedblobs_out = net.forward(data = np.zeros((8, 3, 224, 224), dtype = np.float32))# access the last layerlayer = net.layers[-1]# converts and provides the output as NumPy arraynumpy_array = net.blobs['conv1_1'].data# access the loss weightsloss_weights = net.blob_loss_weights# === BASIC OPTIMIZER ===# this example uses paths from https://github.com/ppengtang/oicr# create an SGD solver, loads the net in train mode# it knows about base_lr, weight_decay, momentum, lr_mult, decay_mult, iter_size, lr policy step, step_size, gamma# it finds train.prototxt from the solver.prototxt's train_net or net parameterssolver = caffe.SGDSolver('oicr/models/VGG16/solver.prototxt')# load pretrained weightssolver.net.copy_from('oicr/data/imagenet_models/VGG16.v2.caffemodel')# runs one iteration of forward, backward, optimization; returns a float loss value# data layer must be registered or inputs must be provided as keyword argumentsloss = solver.step(1)

Drop-in script for OICR enabling PyTorch as backend for eval and training

Place caffe_pytorch_oicr.py and caffemodel2pytorch.py in the root oicr directory. To use the PyTorch backend in testing and in training, put a line import caffe_pytorch_oicr at the very top (before import _init_paths) in tools/test_net.py and tools/train_net.py respectively. It requires PyTorch and CuPy (for on-the-fly CUDA kernel compilation).

# caffe_pytorch_oicr.pyimport collectionsimport torchimport torch.nn.functional as Fimport cupyimport caffemodel2pytorch

caffemodel2pytorch.initialize('./caffe-oicr/src/caffe/proto/caffe.proto') # needs to be called explicitly for these porting scenarios to enable caffe.proto.caffe_pb2 variablecaffemodel2pytorch.set_mode_gpu()
caffemodel2pytorch.modules['GlobalSumPooling'] = lambda param: lambda pred: pred.sum(dim = 0, keepdim = True)
caffemodel2pytorch.modules['MulticlassCrossEntropyLoss'] = lambda param: lambda pred, labels, eps = 1e-6: F.binary_cross_entropy(pred.clamp(eps, 1 - eps), labels)
caffemodel2pytorch.modules['data'] = lambda param: __import__('roi_data_layer.layer').layer.RoIDataLayer() # wrapping a PyCaffe layercaffemodel2pytorch.modules['OICRLayer'] = lambda param: OICRLayer # wrapping a PyTorch functioncaffemodel2pytorch.modules['WeightedSoftmaxWithLoss'] = lambda param: WeightedSoftmaxWithLoss
caffemodel2pytorch.modules['ReLU'] = lambda param: torch.nn.ReLU(inplace = True) # wrapping a PyTorch modulecaffemodel2pytorch.modules['ROIPooling'] = lambda param: lambda input, rois: RoiPooling(param['pooled_h'], param['pooled_w'], param['spatial_scale'])(input, rois) # wrapping a PyTorch autograd functiondef WeightedSoftmaxWithLoss(prob, labels_ic, cls_loss_weights, eps = 1e-12):
	loss = -cls_loss_weights * F.log_softmax(prob, dim = -1).gather(-1, labels_ic.long().unsqueeze(-1)).squeeze(-1)
	valid_sum = cls_loss_weights.gt(eps).float().sum()	return loss.sum() / (loss.numel() if valid_sum == 0 else valid_sum)def OICRLayer(boxes, cls_prob, im_labels, cfg_TRAIN_FG_THRESH = 0.5):
    cls_prob = (cls_prob if cls_prob.size(-1) == im_labels.size(-1) else cls_prob[..., 1:]).clone()
    boxes = boxes[..., 1:]
    gt_boxes, gt_classes, gt_scores = [], [], []    for i in im_labels.eq(1).nonzero()[:, 1]:
        max_index = int(cls_prob[:, i].max(dim = 0)[1])
        gt_boxes.append(boxes[max_index])
        gt_classes.append(int(i) + 1)
        gt_scores.append(float(cls_prob[max_index, i]))
        cls_prob[max_index] = 0
    max_overlaps, gt_assignment = overlap(boxes, torch.stack(gt_boxes)).max(dim = 1)    return gt_assignment.new(gt_classes)[gt_assignment] * (max_overlaps > cfg_TRAIN_FG_THRESH).type_as(gt_assignment), max_overlaps.new(gt_scores)[gt_assignment]class RoiPooling(torch.autograd.Function):	CUDA_NUM_THREADS = 1024
	GET_BLOCKS = staticmethod(lambda N: (N + RoiPooling.CUDA_NUM_THREADS - 1) // RoiPooling.CUDA_NUM_THREADS)
	Stream = collections.namedtuple('Stream', ['ptr'])

	kernel_forward = b'''	#define FLT_MAX 340282346638528859811704183484516925440.0f	typedef float Dtype;	#define CUDA_KERNEL_LOOP(i, n) for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); i += blockDim.x * gridDim.x)	extern "C"	__global__ void ROIPoolForward(const int nthreads, const Dtype* bottom_data,		const Dtype spatial_scale, const int channels, const int height,		const int width, const int pooled_height, const int pooled_width,		const Dtype* bottom_rois, Dtype* top_data, int* argmax_data) {	  CUDA_KERNEL_LOOP(index, nthreads) { 		// (n, c, ph, pw) is an element in the pooled output		int pw = index % pooled_width;		int ph = (index / pooled_width) % pooled_height;		int c = (index / pooled_width / pooled_height) % channels;		int n = index / pooled_width / pooled_height / channels;		bottom_rois += n * 5;		int roi_batch_ind = bottom_rois[0];		int roi_start_w = round(bottom_rois[1] * spatial_scale);		int roi_start_h = round(bottom_rois[2] * spatial_scale);		int roi_end_w = round(bottom_rois[3] * spatial_scale);		int roi_end_h = round(bottom_rois[4] * spatial_scale);		// Force malformed ROIs to be 1x1		int roi_width = max(roi_end_w - roi_start_w + 1, 1);		int roi_height = max(roi_end_h - roi_start_h + 1, 1);		Dtype bin_size_h = static_cast<Dtype>(roi_height)						   / static_cast<Dtype>(pooled_height);		Dtype bin_size_w = static_cast<Dtype>(roi_width)						   / static_cast<Dtype>(pooled_width);		int hstart = static_cast<int>(floor(static_cast<Dtype>(ph)											* bin_size_h));		int wstart = static_cast<int>(floor(static_cast<Dtype>(pw)											* bin_size_w));		int hend = static_cast<int>(ceil(static_cast<Dtype>(ph + 1)										 * bin_size_h));		int wend = static_cast<int>(ceil(static_cast<Dtype>(pw + 1)										 * bin_size_w));		// Add roi offsets and clip to input boundaries		hstart = min(max(hstart + roi_start_h, 0), height);		hend = min(max(hend + roi_start_h, 0), height);		wstart = min(max(wstart + roi_start_w, 0), width);		wend = min(max(wend + roi_start_w, 0), width);		bool is_empty = (hend <= hstart) || (wend <= wstart);		// Define an empty pooling region to be zero		Dtype maxval = is_empty ? 0 : -FLT_MAX;		// If nothing is pooled, argmax = -1 causes nothing to be backprop'd		int maxidx = -1;		bottom_data += (roi_batch_ind * channels + c) * height * width;		for (int h = hstart; h < hend; ++h) {		  for (int w = wstart; w < wend; ++w) {			int bottom_index = h * width + w;			if (bottom_data[bottom_index] > maxval) {			  maxval = bottom_data[bottom_index];			  maxidx = bottom_index;			}		  }		}		top_data[index] = maxval;		argmax_data[index] = maxidx;	  }	}	'''

	kernel_backward = b'''	typedef float Dtype;	#define CUDA_KERNEL_LOOP(i, n) for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); i += blockDim.x * gridDim.x)	extern "C"	__global__ void ROIPoolBackward(const int nthreads, const Dtype* top_diff,		const int* argmax_data, const int num_rois, const Dtype spatial_scale,		const int channels, const int height, const int width,		const int pooled_height, const int pooled_width, Dtype* bottom_diff,		const Dtype* bottom_rois) {	  CUDA_KERNEL_LOOP(index, nthreads) {		// (n, c, h, w) coords in bottom data		int w = index % width;		int h = (index / width) % height;		int c = (index / width / height) % channels;		int n = index / width / height / channels;		Dtype gradient = 0;		// Accumulate gradient over all ROIs that pooled this element		for (int roi_n = 0; roi_n < num_rois; ++roi_n) {		  const Dtype* offset_bottom_rois = bottom_rois + roi_n * 5;		  int roi_batch_ind = offset_bottom_rois[0];		  // Skip if ROI's batch index doesn't match n		  if (n != roi_batch_ind) {			continue;		  }		  int roi_start_w = round(offset_bottom_rois[1] * spatial_scale);		  int roi_start_h = round(offset_bottom_rois[2] * spatial_scale);		  int roi_end_w = round(offset_bottom_rois[3] * spatial_scale);		  int roi_end_h = round(offset_bottom_rois[4] * spatial_scale);		  // Skip if ROI doesn't include (h, w)		  const bool in_roi = (w >= roi_start_w && w <= roi_end_w &&							   h >= roi_start_h && h <= roi_end_h);		  if (!in_roi) {			continue;		  }		  int offset = (roi_n * channels + c) * pooled_height * pooled_width;		  const Dtype* offset_top_diff = top_diff + offset;		  const int* offset_argmax_data = argmax_data + offset;		  // Compute feasible set of pooled units that could have pooled		  // this bottom unit		  // Force malformed ROIs to be 1x1		  int roi_width = max(roi_end_w - roi_start_w + 1, 1);		  int roi_height = max(roi_end_h - roi_start_h + 1, 1);		  Dtype bin_size_h = static_cast<Dtype>(roi_height)							 / static_cast<Dtype>(pooled_height);		  Dtype bin_size_w = static_cast<Dtype>(roi_width)							 / static_cast<Dtype>(pooled_width);		  int phstart = floor(static_cast<Dtype>(h - roi_start_h) / bin_size_h);		  int phend = ceil(static_cast<Dtype>(h - roi_start_h + 1) / bin_size_h);		  int pwstart = floor(static_cast<Dtype>(w - roi_start_w) / bin_size_w);		  int pwend = ceil(static_cast<Dtype>(w - roi_start_w + 1) / bin_size_w);		  phstart = min(max(phstart, 0), pooled_height);		  phend = min(max(phend, 0), pooled_height);		  pwstart = min(max(pwstart, 0), pooled_width);		  pwend = min(max(pwend, 0), pooled_width);		  for (int ph = phstart; ph < phend; ++ph) {			for (int pw = pwstart; pw < pwend; ++pw) {			  if (offset_argmax_data[ph * pooled_width + pw] == (h * width + w)) {				gradient += offset_top_diff[ph * pooled_width + pw];			  }			}		  }		}		bottom_diff[index] = gradient;	  }	}	'''
	cupy_init = cupy.array([])
	compiled_forward = cupy.cuda.compiler.compile_with_cache(kernel_forward).get_function('ROIPoolForward')
	compiled_backward = cupy.cuda.compiler.compile_with_cache(kernel_backward).get_function('ROIPoolBackward')	def __init__(self, pooled_height, pooled_width, spatial_scale):		self.pooled_height = pooled_height		self.pooled_width = pooled_width		self.spatial_scale = spatial_scale	def forward(self, images, rois):
		output = torch.cuda.FloatTensor(len(rois), images.size(1) * self.pooled_height * self.pooled_width)		self.argmax = torch.cuda.IntTensor(output.size()).fill_(-1)		self.input_size = images.size()		self.save_for_backward(rois)
		RoiPooling.compiled_forward(grid = (RoiPooling.GET_BLOCKS(output.numel()), 1, 1), block = (RoiPooling.CUDA_NUM_THREADS, 1, 1), args=[
			output.numel(), images.data_ptr(), cupy.float32(self.spatial_scale), self.input_size[-3], self.input_size[-2], self.input_size[-1],			self.pooled_height, self.pooled_width, rois.data_ptr(), output.data_ptr(), self.argmax.data_ptr()
			  ], stream=RoiPooling.Stream(ptr=torch.cuda.current_stream().cuda_stream))		return output	def backward(self, grad_output):
		rois, = self.saved_tensors
		grad_input = torch.cuda.FloatTensor(*self.input_size).zero_()
		RoiPooling.compiled_backward(grid = (RoiPooling.GET_BLOCKS(grad_input.numel()), 1, 1), block = (RoiPooling.CUDA_NUM_THREADS, 1, 1), args=[
			grad_input.numel(), grad_output.data_ptr(), self.argmax.data_ptr(), len(rois), cupy.float32(self.spatial_scale), self.input_size[-3],			self.input_size[-2], self.input_size[-1], self.pooled_height, self.pooled_width, grad_input.data_ptr(), rois.data_ptr()
			  ], stream=RoiPooling.Stream(ptr=torch.cuda.current_stream().cuda_stream))		return grad_input, None
		def overlap(box1, box2):
    b1, b2 = box1.t().contiguous(), box2.t().contiguous()
    xx1 = torch.max(b1[0].unsqueeze(1), b2[0].unsqueeze(0))
    yy1 = torch.max(b1[1].unsqueeze(1), b2[1].unsqueeze(0))
    xx2 = torch.min(b1[2].unsqueeze(1), b2[2].unsqueeze(0))
    yy2 = torch.min(b1[3].unsqueeze(1), b2[3].unsqueeze(0))
    inter = area(x1 = xx1, y1 = yy1, x2 = xx2, y2 = yy2)    return inter / (area(b1.t()).unsqueeze(1) + area(b2.t()).unsqueeze(0) - inter)def area(boxes = None, x1 = None, y1 = None, x2 = None, y2 = None):    return (boxes[..., 3] - boxes[..., 1] + 1) * (boxes[..., 2] - boxes[..., 0] + 1) if boxes is not None else (x2 - x1 + 1).clamp(min = 0) * (y2 - y1 + 1).clamp(min = 0)

Note: I've also had to replace utils/bbox.pyx by utils/cython_bbox.pyx and utils/nms.pyx by utils/cython_nms.pyx in lib/setup.py to deal with some setup.py issues.

上一篇:neural-assembly-compiler

下一篇:extension-cpp

用户评价
全部评价

热门资源

  • Keras-ResNeXt

    Keras ResNeXt Implementation of ResNeXt models...

  • seetafaceJNI

    项目介绍 基于中科院seetaface2进行封装的JAVA...

  • spark-corenlp

    This package wraps Stanford CoreNLP annotators ...

  • capsnet-with-caps...

    CapsNet with capsule-wise convolution Project ...

  • inferno-boilerplate

    This is a very basic boilerplate example for pe...