[ { "path": ".gitignore", "content": "*.pyc\ncaffe_converter\n.idea\n" }, { "path": "README.md", "content": "# MTCNN_face_detection_and_alignment\n\n## About\n\n This is a python/mxnet implementation of [Zhang](https://kpzhang93.github.io/)'s work ****. it's fast and accurate, see [link](https://github.com/kpzhang93/MTCNN_face_detection_alignment). \n\n It should have **almost** the same output with the original work, for mxnet fans and those can't afford matlab :)\n\n[中文blog](https://pangyupo.github.io/2016/10/22/mxnet-mtcnn/)\n\n## Requirement\t \n\n- opencv \n\n ​\tI use cv2 for image io and resize(much faster than skimage), the input image's channel is acutally BGR\n\n- mxnet \n\n ​\t**please update to the newest version, we need 'full' mode in Pooling operation**\n\nOnly tested on Linux and Mac\n\n## Test\n\nrun:\n\n ``python main.py`` \n\nyou can change `ctx` to `mx.gpu(0)` for faster detection\n\n--- update 20161028 ---\n\nby setting ``num_worker=4`` ``accurate_landmark=False`` we can reduce the detection time by 1/4-1/3, the bboxes are still the same, but we skip the last landmark fine-tune stage( mtcnn_v1 ). \n\n--- update 20161207 ---\n\nadd function `extract_face_chips`, examples:\n\n![1](http://7vikw0.com1.z0.glb.clouddn.com/chip_0.png)\n![2](http://7vikw0.com1.z0.glb.clouddn.com/chip_3.png)\n![3](http://7vikw0.com1.z0.glb.clouddn.com/chip_2.png)\n![4](http://7vikw0.com1.z0.glb.clouddn.com/chip_1.png)\n\nsee `mtcnn_detector.py` for the details about the parameters. this function use [dlib](http://dlib.net/)'s align strategy, which works well on profile images :) \n## Results\n\n![big4](http://7xsc78.com1.z0.glb.clouddn.com/face_mtcnn.png)\n\n\n\n## License\n\nMIT LICENSE\n\n\n\n## Reference\n\nK. Zhang and Z. Zhang and Z. Li and Y. Qiao Joint, Face Detection and Alignment Using Multitask Cascaded Convolutional Networks, IEEE Signal Processing Letters\n" }, { "path": "helper.py", "content": "# coding: utf-8\n# YuanYang\nimport math\nimport cv2\nimport numpy as np\n\n\ndef nms(boxes, overlap_threshold, mode='Union'):\n \"\"\"\n non max suppression\n\n Parameters:\n ----------\n box: numpy array n x 5\n input bbox array\n overlap_threshold: float number\n threshold of overlap\n mode: float number\n how to compute overlap ratio, 'Union' or 'Min'\n Returns:\n -------\n index array of the selected bbox\n \"\"\"\n # if there are no boxes, return an empty list\n if len(boxes) == 0:\n return []\n\n # if the bounding boxes integers, convert them to floats\n if boxes.dtype.kind == \"i\":\n boxes = boxes.astype(\"float\")\n\n # initialize the list of picked indexes\n pick = []\n\n # grab the coordinates of the bounding boxes\n x1, y1, x2, y2, score = [boxes[:, i] for i in range(5)]\n\n area = (x2 - x1 + 1) * (y2 - y1 + 1)\n idxs = np.argsort(score)\n\n # keep looping while some indexes still remain in the indexes list\n while len(idxs) > 0:\n # grab the last index in the indexes list and add the index value to the list of picked indexes\n last = len(idxs) - 1\n i = idxs[last]\n pick.append(i)\n\n xx1 = np.maximum(x1[i], x1[idxs[:last]])\n yy1 = np.maximum(y1[i], y1[idxs[:last]])\n xx2 = np.minimum(x2[i], x2[idxs[:last]])\n yy2 = np.minimum(y2[i], y2[idxs[:last]])\n\n # compute the width and height of the bounding box\n w = np.maximum(0, xx2 - xx1 + 1)\n h = np.maximum(0, yy2 - yy1 + 1)\n\n inter = w * h\n if mode == 'Min':\n overlap = inter / np.minimum(area[i], area[idxs[:last]])\n else:\n overlap = inter / (area[i] + area[idxs[:last]] - inter)\n\n # delete all indexes from the index list that have\n idxs = np.delete(idxs, np.concatenate(([last],\n np.where(overlap > overlap_threshold)[0])))\n\n return pick\n\ndef adjust_input(in_data):\n \"\"\"\n adjust the input from (h, w, c) to ( 1, c, h, w) for network input\n\n Parameters:\n ----------\n in_data: numpy array of shape (h, w, c)\n input data\n Returns:\n -------\n out_data: numpy array of shape (1, c, h, w)\n reshaped array\n \"\"\"\n if in_data.dtype is not np.dtype('float32'):\n out_data = in_data.astype(np.float32)\n else:\n out_data = in_data\n\n out_data = out_data.transpose((2,0,1))\n out_data = np.expand_dims(out_data, 0)\n out_data = (out_data - 127.5)*0.0078125\n return out_data\n\ndef generate_bbox(map, reg, scale, threshold):\n \"\"\"\n generate bbox from feature map\n Parameters:\n ----------\n map: numpy array , n x m x 1\n detect score for each position\n reg: numpy array , n x m x 4\n bbox\n scale: float number\n scale of this detection\n threshold: float number\n detect threshold\n Returns:\n -------\n bbox array\n \"\"\"\n stride = 2\n cellsize = 12\n\n t_index = np.where(map>threshold)\n\n # find nothing\n if t_index[0].size == 0:\n return np.array([])\n\n dx1, dy1, dx2, dy2 = [reg[0, i, t_index[0], t_index[1]] for i in range(4)]\n\n reg = np.array([dx1, dy1, dx2, dy2])\n score = map[t_index[0], t_index[1]]\n boundingbox = np.vstack([np.round((stride*t_index[1]+1)/scale),\n np.round((stride*t_index[0]+1)/scale),\n np.round((stride*t_index[1]+1+cellsize)/scale),\n np.round((stride*t_index[0]+1+cellsize)/scale),\n score,\n reg])\n\n return boundingbox.T\n\n\ndef detect_first_stage(img, net, scale, threshold):\n \"\"\"\n run PNet for first stage\n \n Parameters:\n ----------\n img: numpy array, bgr order\n input image\n scale: float number\n how much should the input image scale\n net: PNet\n worker\n Returns:\n -------\n total_boxes : bboxes\n \"\"\"\n height, width, _ = img.shape\n hs = int(math.ceil(height * scale))\n ws = int(math.ceil(width * scale))\n \n im_data = cv2.resize(img, (ws,hs))\n \n # adjust for the network input\n input_buf = adjust_input(im_data)\n output = net.predict(input_buf)\n boxes = generate_bbox(output[1][0,1,:,:], output[0], scale, threshold)\n\n if boxes.size == 0:\n return None\n\n # nms\n pick = nms(boxes[:,0:5], 0.5, mode='Union')\n boxes = boxes[pick]\n return boxes\n\ndef detect_first_stage_warpper( args ):\n return detect_first_stage(*args)\n" }, { "path": "main.py", "content": "# coding: utf-8\nimport mxnet as mx\nfrom mtcnn_detector import MtcnnDetector\nimport cv2\nimport os\nimport time\n\ndetector = MtcnnDetector(model_folder='model', ctx=mx.cpu(0), num_worker = 4 , accurate_landmark = False)\n\n\nimg = cv2.imread('test2.jpg')\n\n# run detector\nresults = detector.detect_face(img)\n\nif results is not None:\n\n total_boxes = results[0]\n points = results[1]\n \n # extract aligned face chips\n chips = detector.extract_image_chips(img, points, 144, 0.37)\n for i, chip in enumerate(chips):\n cv2.imshow('chip_'+str(i), chip)\n cv2.imwrite('chip_'+str(i)+'.png', chip)\n\n draw = img.copy()\n for b in total_boxes:\n cv2.rectangle(draw, (int(b[0]), int(b[1])), (int(b[2]), int(b[3])), (255, 255, 255))\n\n for p in points:\n for i in range(5):\n cv2.circle(draw, (p[i], p[i + 5]), 1, (0, 0, 255), 2)\n\n cv2.imshow(\"detection result\", draw)\n cv2.waitKey(0)\n\n# --------------\n# test on camera\n# --------------\n'''\ncamera = cv2.VideoCapture(0)\nwhile True:\n grab, frame = camera.read()\n img = cv2.resize(frame, (320,180))\n\n t1 = time.time()\n results = detector.detect_face(img)\n print 'time: ',time.time() - t1\n\n if results is None:\n continue\n\n total_boxes = results[0]\n points = results[1]\n\n draw = img.copy()\n for b in total_boxes:\n cv2.rectangle(draw, (int(b[0]), int(b[1])), (int(b[2]), int(b[3])), (255, 255, 255))\n\n for p in points:\n for i in range(5):\n cv2.circle(draw, (p[i], p[i + 5]), 1, (255, 0, 0), 2)\n cv2.imshow(\"detection result\", draw)\n cv2.waitKey(30)\n'''\n" }, { "path": "model/det1-symbol.json", "content": "{\n \"nodes\": [\n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"data\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"10\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv1\", \n \"inputs\": [[0, 0], [1, 0], [2, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu1_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu1\", \n \"inputs\": [[3, 0], [4, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(2,2)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool1\", \n \"inputs\": [[5, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"16\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv2\", \n \"inputs\": [[6, 0], [7, 0], [8, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu2_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu2\", \n \"inputs\": [[9, 0], [10, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"32\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv3\", \n \"inputs\": [[11, 0], [12, 0], [13, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu3_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu3\", \n \"inputs\": [[14, 0], [15, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv4_2_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv4_2_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(1,1)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"4\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv4_2\", \n \"inputs\": [[16, 0], [17, 0], [18, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv4_1_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv4_1_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(1,1)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"2\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv4_1\", \n \"inputs\": [[16, 0], [20, 0], [21, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"SoftmaxActivation\", \n \"param\": {\"mode\": \"channel\"}, \n \"name\": \"prob1\", \n \"inputs\": [[22, 0]], \n \"backward_source_id\": -1\n }\n ], \n \"arg_nodes\": [\n 0, \n 1, \n 2, \n 4, \n 7, \n 8, \n 10, \n 12, \n 13, \n 15, \n 17, \n 18, \n 20, \n 21\n ], \n \"heads\": [[19, 0], [23, 0]]\n}" }, { "path": "model/det1.prototxt", "content": "name: \"PNet\"\ninput: \"data\"\ninput_dim: 1\ninput_dim: 3\ninput_dim: 12\ninput_dim: 12\n\nlayer {\n name: \"conv1\"\n type: \"Convolution\"\n bottom: \"data\"\n top: \"conv1\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 0\n }\n convolution_param {\n num_output: 10\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n }\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"PReLU1\"\n type: \"PReLU\"\n bottom: \"conv1\"\n top: \"conv1\"\n}\nlayer {\n name: \"pool1\"\n type: \"Pooling\"\n bottom: \"conv1\"\n top: \"pool1\"\n pooling_param {\n pool: MAX\n kernel_size: 2\n stride: 2\n }\n}\n\nlayer {\n name: \"conv2\"\n type: \"Convolution\"\n bottom: \"pool1\"\n top: \"conv2\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 0\n }\n convolution_param {\n num_output: 16\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n }\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"PReLU2\"\n type: \"PReLU\"\n bottom: \"conv2\"\n top: \"conv2\"\n}\n\nlayer {\n name: \"conv3\"\n type: \"Convolution\"\n bottom: \"conv2\"\n top: \"conv3\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 0\n }\n convolution_param {\n num_output: 32\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n }\n bias_filler {\n\t type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"PReLU3\"\n type: \"PReLU\"\n bottom: \"conv3\"\n top: \"conv3\"\n}\n\n\nlayer {\n name: \"conv4-1\"\n type: \"Convolution\"\n bottom: \"conv3\"\n top: \"conv4-1\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 0\n }\n convolution_param {\n num_output: 2\n kernel_size: 1\n stride: 1\n weight_filler {\n type: \"xavier\"\n }\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\n\nlayer {\n name: \"conv4-2\"\n type: \"Convolution\"\n bottom: \"conv3\"\n top: \"conv4-2\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 0\n }\n convolution_param {\n num_output: 4\n kernel_size: 1\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"prob1\"\n type: \"Softmax\"\n bottom: \"conv4-1\"\n top: \"prob1\"\n}\n" }, { "path": "model/det2-symbol.json", "content": "{\n \"nodes\": [\n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"data\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"28\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv1\", \n \"inputs\": [[0, 0], [1, 0], [2, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu1_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu1\", \n \"inputs\": [[3, 0], [4, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(3,3)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool1\", \n \"inputs\": [[5, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"48\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv2\", \n \"inputs\": [[6, 0], [7, 0], [8, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu2_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu2\", \n \"inputs\": [[9, 0], [10, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(3,3)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool2\", \n \"inputs\": [[11, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(2,2)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"64\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv3\", \n \"inputs\": [[12, 0], [13, 0], [14, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu3_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu3\", \n \"inputs\": [[15, 0], [16, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv4_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv4_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"128\"\n }, \n \"name\": \"conv4\", \n \"inputs\": [[17, 0], [18, 0], [19, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu4_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu4\", \n \"inputs\": [[20, 0], [21, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv5_2_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv5_2_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"4\"\n }, \n \"name\": \"conv5_2\", \n \"inputs\": [[22, 0], [23, 0], [24, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv5_1_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv5_1_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"2\"\n }, \n \"name\": \"conv5_1\", \n \"inputs\": [[22, 0], [26, 0], [27, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prob1_label\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"SoftmaxOutput\", \n \"param\": {\n \"grad_scale\": \"1\", \n \"ignore_label\": \"-1\", \n \"multi_output\": \"False\", \n \"normalization\": \"null\", \n \"use_ignore\": \"False\"\n }, \n \"name\": \"prob1\", \n \"inputs\": [[28, 0], [29, 0]], \n \"backward_source_id\": -1\n }\n ], \n \"arg_nodes\": [\n 0, \n 1, \n 2, \n 4, \n 7, \n 8, \n 10, \n 13, \n 14, \n 16, \n 18, \n 19, \n 21, \n 23, \n 24, \n 26, \n 27, \n 29\n ], \n \"heads\": [[25, 0], [30, 0]]\n}" }, { "path": "model/det2.prototxt", "content": "name: \"RNet\"\ninput: \"data\"\ninput_dim: 1\ninput_dim: 3\ninput_dim: 24\ninput_dim: 24\n\n\n##########################\n######################\nlayer {\n name: \"conv1\"\n type: \"Convolution\"\n bottom: \"data\"\n top: \"conv1\"\n param {\n lr_mult: 0\n decay_mult: 0\n }\n param {\n lr_mult: 0\n decay_mult: 0\n }\n convolution_param {\n num_output: 28\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"prelu1\"\n type: \"PReLU\"\n bottom: \"conv1\"\n top: \"conv1\"\n propagate_down: true\n}\nlayer {\n name: \"pool1\"\n type: \"Pooling\"\n bottom: \"conv1\"\n top: \"pool1\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n}\n\nlayer {\n name: \"conv2\"\n type: \"Convolution\"\n bottom: \"pool1\"\n top: \"conv2\"\n param {\n lr_mult: 0\n decay_mult: 0\n }\n param {\n lr_mult: 0\n decay_mult: 0\n }\n convolution_param {\n num_output: 48\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"prelu2\"\n type: \"PReLU\"\n bottom: \"conv2\"\n top: \"conv2\"\n propagate_down: true\n}\nlayer {\n name: \"pool2\"\n type: \"Pooling\"\n bottom: \"conv2\"\n top: \"pool2\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n}\n####################################\n\n##################################\nlayer {\n name: \"conv3\"\n type: \"Convolution\"\n bottom: \"pool2\"\n top: \"conv3\"\n param {\n lr_mult: 0\n decay_mult: 0\n }\n param {\n lr_mult: 0\n decay_mult: 0\n }\n convolution_param {\n num_output: 64\n kernel_size: 2\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"prelu3\"\n type: \"PReLU\"\n bottom: \"conv3\"\n top: \"conv3\"\n propagate_down: true\n}\n###############################\n\n###############################\n\nlayer {\n name: \"conv4\"\n type: \"InnerProduct\"\n bottom: \"conv3\"\n top: \"conv4\"\n param {\n lr_mult: 0\n decay_mult: 0\n }\n param {\n lr_mult: 0\n decay_mult: 0\n }\n inner_product_param {\n num_output: 128\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"prelu4\"\n type: \"PReLU\"\n bottom: \"conv4\"\n top: \"conv4\"\n}\n\nlayer {\n name: \"conv5-1\"\n type: \"InnerProduct\"\n bottom: \"conv4\"\n top: \"conv5-1\"\n param {\n lr_mult: 0\n decay_mult: 0\n }\n param {\n lr_mult: 0\n decay_mult: 0\n }\n inner_product_param {\n num_output: 2\n #kernel_size: 1\n #stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"conv5-2\"\n type: \"InnerProduct\"\n bottom: \"conv4\"\n top: \"conv5-2\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n num_output: 4\n #kernel_size: 1\n #stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"prob1\"\n type: \"Softmax\"\n bottom: \"conv5-1\"\n top: \"prob1\"\n}" }, { "path": "model/det3-symbol.json", "content": "{\n \"nodes\": [\n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"data\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"32\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv1\", \n \"inputs\": [[0, 0], [1, 0], [2, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu1_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu1\", \n \"inputs\": [[3, 0], [4, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(3,3)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool1\", \n \"inputs\": [[5, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"64\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv2\", \n \"inputs\": [[6, 0], [7, 0], [8, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu2_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu2\", \n \"inputs\": [[9, 0], [10, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(3,3)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool2\", \n \"inputs\": [[11, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"64\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv3\", \n \"inputs\": [[12, 0], [13, 0], [14, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu3_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu3\", \n \"inputs\": [[15, 0], [16, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(2,2)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool3\", \n \"inputs\": [[17, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv4_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv4_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(2,2)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"128\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv4\", \n \"inputs\": [[18, 0], [19, 0], [20, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu4_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu4\", \n \"inputs\": [[21, 0], [22, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv5_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv5_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"256\"\n }, \n \"name\": \"conv5\", \n \"inputs\": [[23, 0], [24, 0], [25, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu5_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu5\", \n \"inputs\": [[26, 0], [27, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv6_3_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv6_3_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"10\"\n }, \n \"name\": \"conv6_3\", \n \"inputs\": [[28, 0], [29, 0], [30, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv6_2_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv6_2_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"4\"\n }, \n \"name\": \"conv6_2\", \n \"inputs\": [[28, 0], [32, 0], [33, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv6_1_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv6_1_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"2\"\n }, \n \"name\": \"conv6_1\", \n \"inputs\": [[28, 0], [35, 0], [36, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prob1_label\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"SoftmaxOutput\", \n \"param\": {\n \"grad_scale\": \"1\", \n \"ignore_label\": \"-1\", \n \"multi_output\": \"False\", \n \"normalization\": \"null\", \n \"use_ignore\": \"False\"\n }, \n \"name\": \"prob1\", \n \"inputs\": [[37, 0], [38, 0]], \n \"backward_source_id\": -1\n }\n ], \n \"arg_nodes\": [\n 0, \n 1, \n 2, \n 4, \n 7, \n 8, \n 10, \n 13, \n 14, \n 16, \n 19, \n 20, \n 22, \n 24, \n 25, \n 27, \n 29, \n 30, \n 32, \n 33, \n 35, \n 36, \n 38\n ], \n \"heads\": [[31, 0], [34, 0], [39, 0]]\n}" }, { "path": "model/det3.prototxt", "content": "name: \"ONet\"\ninput: \"data\"\ninput_dim: 1\ninput_dim: 3\ninput_dim: 48\ninput_dim: 48\n##################################\nlayer {\n name: \"conv1\"\n type: \"Convolution\"\n bottom: \"data\"\n top: \"conv1\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 32\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"prelu1\"\n type: \"PReLU\"\n bottom: \"conv1\"\n top: \"conv1\"\n}\nlayer {\n name: \"pool1\"\n type: \"Pooling\"\n bottom: \"conv1\"\n top: \"pool1\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n}\nlayer {\n name: \"conv2\"\n type: \"Convolution\"\n bottom: \"pool1\"\n top: \"conv2\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 64\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\n\nlayer {\n name: \"prelu2\"\n type: \"PReLU\"\n bottom: \"conv2\"\n top: \"conv2\"\n}\nlayer {\n name: \"pool2\"\n type: \"Pooling\"\n bottom: \"conv2\"\n top: \"pool2\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n}\n\nlayer {\n name: \"conv3\"\n type: \"Convolution\"\n bottom: \"pool2\"\n top: \"conv3\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n\tnum_output: 64\n\tkernel_size: 3\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"prelu3\"\n type: \"PReLU\"\n bottom: \"conv3\"\n top: \"conv3\"\n}\nlayer {\n name: \"pool3\"\n type: \"Pooling\"\n bottom: \"conv3\"\n top: \"pool3\"\n pooling_param {\n pool: MAX\n kernel_size: 2\n stride: 2\n }\n}\nlayer {\n name: \"conv4\"\n type: \"Convolution\"\n bottom: \"pool3\"\n top: \"conv4\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n\tnum_output: 128\n\tkernel_size: 2\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"prelu4\"\n type: \"PReLU\"\n bottom: \"conv4\"\n top: \"conv4\"\n}\n\n\nlayer {\n name: \"conv5\"\n type: \"InnerProduct\"\n bottom: \"conv4\"\n top: \"conv5\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n\t#kernel_size: 3\n\tnum_output: 256\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\n\nlayer {\n name: \"drop5\"\n type: \"Dropout\"\n bottom: \"conv5\"\n top: \"conv5\"\n dropout_param {\n dropout_ratio: 0.25\n }\n}\nlayer {\n name: \"prelu5\"\n type: \"PReLU\"\n bottom: \"conv5\"\n top: \"conv5\"\n}\n\n\nlayer {\n name: \"conv6-1\"\n type: \"InnerProduct\"\n bottom: \"conv5\"\n top: \"conv6-1\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n #kernel_size: 1\n\tnum_output: 2\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"conv6-2\"\n type: \"InnerProduct\"\n bottom: \"conv5\"\n top: \"conv6-2\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n \t#kernel_size: 1\n\tnum_output: 4\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"conv6-3\"\n type: \"InnerProduct\"\n bottom: \"conv5\"\n top: \"conv6-3\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n \t#kernel_size: 1\n\tnum_output: 10\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\nlayer {\n name: \"prob1\"\n type: \"Softmax\"\n bottom: \"conv6-1\"\n top: \"prob1\"\n}\n" }, { "path": "model/det4-symbol.json", "content": "{\n \"nodes\": [\n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"data\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"SliceChannel\", \n \"param\": {\n \"axis\": \"1\", \n \"num_outputs\": \"5\", \n \"squeeze_axis\": \"False\"\n }, \n \"name\": \"slice\", \n \"inputs\": [[0, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_1_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_1_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"28\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv1_1\", \n \"inputs\": [[1, 0], [2, 0], [3, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu1_1_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu1_1\", \n \"inputs\": [[4, 0], [5, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(3,3)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool1_1\", \n \"inputs\": [[6, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_1_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_1_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"48\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv2_1\", \n \"inputs\": [[7, 0], [8, 0], [9, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu2_1_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu2_1\", \n \"inputs\": [[10, 0], [11, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(3,3)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool2_1\", \n \"inputs\": [[12, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_1_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_1_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(2,2)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"64\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv3_1\", \n \"inputs\": [[13, 0], [14, 0], [15, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu3_1_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu3_1\", \n \"inputs\": [[16, 0], [17, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_2_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_2_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"28\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv1_2\", \n \"inputs\": [[1, 1], [19, 0], [20, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu1_2_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu1_2\", \n \"inputs\": [[21, 0], [22, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(3,3)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool1_2\", \n \"inputs\": [[23, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_2_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_2_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"48\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv2_2\", \n \"inputs\": [[24, 0], [25, 0], [26, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu2_2_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu2_2\", \n \"inputs\": [[27, 0], [28, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(3,3)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool2_2\", \n \"inputs\": [[29, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_2_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_2_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(2,2)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"64\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv3_2\", \n \"inputs\": [[30, 0], [31, 0], [32, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu3_2_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu3_2\", \n \"inputs\": [[33, 0], [34, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_3_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_3_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"28\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv1_3\", \n \"inputs\": [[1, 2], [36, 0], [37, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu1_3_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu1_3\", \n \"inputs\": [[38, 0], [39, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(3,3)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool1_3\", \n \"inputs\": [[40, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_3_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_3_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"48\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv2_3\", \n \"inputs\": [[41, 0], [42, 0], [43, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu2_3_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu2_3\", \n \"inputs\": [[44, 0], [45, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(3,3)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool2_3\", \n \"inputs\": [[46, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_3_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_3_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(2,2)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"64\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv3_3\", \n \"inputs\": [[47, 0], [48, 0], [49, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu3_3_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu3_3\", \n \"inputs\": [[50, 0], [51, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_4_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_4_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"28\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv1_4\", \n \"inputs\": [[1, 3], [53, 0], [54, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu1_4_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu1_4\", \n \"inputs\": [[55, 0], [56, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(3,3)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool1_4\", \n \"inputs\": [[57, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_4_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_4_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"48\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv2_4\", \n \"inputs\": [[58, 0], [59, 0], [60, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu2_4_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu2_4\", \n \"inputs\": [[61, 0], [62, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(3,3)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool2_4\", \n \"inputs\": [[63, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_4_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_4_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(2,2)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"64\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv3_4\", \n \"inputs\": [[64, 0], [65, 0], [66, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu3_4_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu3_4\", \n \"inputs\": [[67, 0], [68, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_5_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv1_5_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"28\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv1_5\", \n \"inputs\": [[1, 4], [70, 0], [71, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu1_5_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu1_5\", \n \"inputs\": [[72, 0], [73, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(3,3)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool1_5\", \n \"inputs\": [[74, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_5_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv2_5_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(3,3)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"48\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv2_5\", \n \"inputs\": [[75, 0], [76, 0], [77, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu2_5_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu2_5\", \n \"inputs\": [[78, 0], [79, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Pooling\", \n \"param\": {\n \"global_pool\": \"False\", \n \"kernel\": \"(3,3)\", \n \"pad\": \"(0,0)\", \n \"pool_type\": \"max\", \n \"pooling_convention\": \"full\", \n \"stride\": \"(2,2)\"\n }, \n \"name\": \"pool2_5\", \n \"inputs\": [[80, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_5_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"conv3_5_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Convolution\", \n \"param\": {\n \"cudnn_off\": \"False\", \n \"cudnn_tune\": \"off\", \n \"dilate\": \"(1,1)\", \n \"kernel\": \"(2,2)\", \n \"no_bias\": \"False\", \n \"num_filter\": \"64\", \n \"num_group\": \"1\", \n \"pad\": \"(0,0)\", \n \"stride\": \"(1,1)\", \n \"workspace\": \"1024\"\n }, \n \"name\": \"conv3_5\", \n \"inputs\": [[81, 0], [82, 0], [83, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu3_5_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu3_5\", \n \"inputs\": [[84, 0], [85, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"Concat\", \n \"param\": {\n \"dim\": \"1\", \n \"num_args\": \"5\"\n }, \n \"name\": \"concat\", \n \"inputs\": [[18, 0], [35, 0], [52, 0], [69, 0], [86, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc4_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc4_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"256\"\n }, \n \"name\": \"fc4\", \n \"inputs\": [[87, 0], [88, 0], [89, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu4_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu4\", \n \"inputs\": [[90, 0], [91, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc4_1_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc4_1_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"64\"\n }, \n \"name\": \"fc4_1\", \n \"inputs\": [[92, 0], [93, 0], [94, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu4_1_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu4_1\", \n \"inputs\": [[95, 0], [96, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc5_1_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc5_1_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"2\"\n }, \n \"name\": \"fc5_1\", \n \"inputs\": [[97, 0], [98, 0], [99, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc4_2_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc4_2_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"64\"\n }, \n \"name\": \"fc4_2\", \n \"inputs\": [[92, 0], [101, 0], [102, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu4_2_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu4_2\", \n \"inputs\": [[103, 0], [104, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc5_2_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc5_2_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"2\"\n }, \n \"name\": \"fc5_2\", \n \"inputs\": [[105, 0], [106, 0], [107, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc4_3_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc4_3_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"64\"\n }, \n \"name\": \"fc4_3\", \n \"inputs\": [[92, 0], [109, 0], [110, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu4_3_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu4_3\", \n \"inputs\": [[111, 0], [112, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc5_3_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc5_3_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"2\"\n }, \n \"name\": \"fc5_3\", \n \"inputs\": [[113, 0], [114, 0], [115, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc4_4_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc4_4_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"64\"\n }, \n \"name\": \"fc4_4\", \n \"inputs\": [[92, 0], [117, 0], [118, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu4_4_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu4_4\", \n \"inputs\": [[119, 0], [120, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc5_4_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc5_4_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"2\"\n }, \n \"name\": \"fc5_4\", \n \"inputs\": [[121, 0], [122, 0], [123, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc4_5_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc4_5_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"64\"\n }, \n \"name\": \"fc4_5\", \n \"inputs\": [[92, 0], [125, 0], [126, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"prelu4_5_gamma\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"LeakyReLU\", \n \"param\": {\n \"act_type\": \"prelu\", \n \"lower_bound\": \"0.125\", \n \"slope\": \"0.25\", \n \"upper_bound\": \"0.334\"\n }, \n \"name\": \"prelu4_5\", \n \"inputs\": [[127, 0], [128, 0]], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc5_5_weight\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"null\", \n \"param\": {}, \n \"name\": \"fc5_5_bias\", \n \"inputs\": [], \n \"backward_source_id\": -1\n }, \n {\n \"op\": \"FullyConnected\", \n \"param\": {\n \"no_bias\": \"False\", \n \"num_hidden\": \"2\"\n }, \n \"name\": \"fc5_5\", \n \"inputs\": [[129, 0], [130, 0], [131, 0]], \n \"backward_source_id\": -1\n }\n ], \n \"arg_nodes\": [\n 0, \n 2, \n 3, \n 5, \n 8, \n 9, \n 11, \n 14, \n 15, \n 17, \n 19, \n 20, \n 22, \n 25, \n 26, \n 28, \n 31, \n 32, \n 34, \n 36, \n 37, \n 39, \n 42, \n 43, \n 45, \n 48, \n 49, \n 51, \n 53, \n 54, \n 56, \n 59, \n 60, \n 62, \n 65, \n 66, \n 68, \n 70, \n 71, \n 73, \n 76, \n 77, \n 79, \n 82, \n 83, \n 85, \n 88, \n 89, \n 91, \n 93, \n 94, \n 96, \n 98, \n 99, \n 101, \n 102, \n 104, \n 106, \n 107, \n 109, \n 110, \n 112, \n 114, \n 115, \n 117, \n 118, \n 120, \n 122, \n 123, \n 125, \n 126, \n 128, \n 130, \n 131\n ], \n \"heads\": [[100, 0], [108, 0], [116, 0], [124, 0], [132, 0]]\n}" }, { "path": "model/det4.prototxt", "content": "name: \"LNet\"\ninput: \"data\"\ninput_dim: 1\ninput_dim: 15\ninput_dim: 24\ninput_dim: 24\n\nlayer {\n name: \"slicer_data\"\n type: \"Slice\"\n bottom: \"data\"\n top: \"data241\"\n top: \"data242\"\n top: \"data243\"\n top: \"data244\"\n top: \"data245\"\n slice_param {\n axis: 1\n slice_point: 3\n slice_point: 6\n slice_point: 9\n slice_point: 12\n }\n}\nlayer {\n name: \"conv1_1\"\n type: \"Convolution\"\n bottom: \"data241\"\n top: \"conv1_1\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 28\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu1_1\"\n type: \"PReLU\"\n bottom: \"conv1_1\"\n top: \"conv1_1\"\n\n}\nlayer {\n name: \"pool1_1\"\n type: \"Pooling\"\n bottom: \"conv1_1\"\n top: \"pool1_1\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n}\n\nlayer {\n name: \"conv2_1\"\n type: \"Convolution\"\n bottom: \"pool1_1\"\n top: \"conv2_1\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 48\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu2_1\"\n type: \"PReLU\"\n bottom: \"conv2_1\"\n top: \"conv2_1\"\n}\nlayer {\n name: \"pool2_1\"\n type: \"Pooling\"\n bottom: \"conv2_1\"\n top: \"pool2_1\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n \n}\nlayer {\n name: \"conv3_1\"\n type: \"Convolution\"\n bottom: \"pool2_1\"\n top: \"conv3_1\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 64\n kernel_size: 2\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu3_1\"\n type: \"PReLU\"\n bottom: \"conv3_1\"\n top: \"conv3_1\"\n}\n##########################\nlayer {\n name: \"conv1_2\"\n type: \"Convolution\"\n bottom: \"data242\"\n top: \"conv1_2\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 28\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu1_2\"\n type: \"PReLU\"\n bottom: \"conv1_2\"\n top: \"conv1_2\"\n\n}\nlayer {\n name: \"pool1_2\"\n type: \"Pooling\"\n bottom: \"conv1_2\"\n top: \"pool1_2\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n}\n\nlayer {\n name: \"conv2_2\"\n type: \"Convolution\"\n bottom: \"pool1_2\"\n top: \"conv2_2\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 48\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu2_2\"\n type: \"PReLU\"\n bottom: \"conv2_2\"\n top: \"conv2_2\"\n}\nlayer {\n name: \"pool2_2\"\n type: \"Pooling\"\n bottom: \"conv2_2\"\n top: \"pool2_2\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n \n}\nlayer {\n name: \"conv3_2\"\n type: \"Convolution\"\n bottom: \"pool2_2\"\n top: \"conv3_2\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 64\n kernel_size: 2\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu3_2\"\n type: \"PReLU\"\n bottom: \"conv3_2\"\n top: \"conv3_2\"\n}\n##########################\n##########################\nlayer {\n name: \"conv1_3\"\n type: \"Convolution\"\n bottom: \"data243\"\n top: \"conv1_3\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 28\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu1_3\"\n type: \"PReLU\"\n bottom: \"conv1_3\"\n top: \"conv1_3\"\n\n}\nlayer {\n name: \"pool1_3\"\n type: \"Pooling\"\n bottom: \"conv1_3\"\n top: \"pool1_3\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n}\n\nlayer {\n name: \"conv2_3\"\n type: \"Convolution\"\n bottom: \"pool1_3\"\n top: \"conv2_3\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 48\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu2_3\"\n type: \"PReLU\"\n bottom: \"conv2_3\"\n top: \"conv2_3\"\n}\nlayer {\n name: \"pool2_3\"\n type: \"Pooling\"\n bottom: \"conv2_3\"\n top: \"pool2_3\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n \n}\nlayer {\n name: \"conv3_3\"\n type: \"Convolution\"\n bottom: \"pool2_3\"\n top: \"conv3_3\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 64\n kernel_size: 2\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu3_3\"\n type: \"PReLU\"\n bottom: \"conv3_3\"\n top: \"conv3_3\"\n}\n##########################\n##########################\nlayer {\n name: \"conv1_4\"\n type: \"Convolution\"\n bottom: \"data244\"\n top: \"conv1_4\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 28\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu1_4\"\n type: \"PReLU\"\n bottom: \"conv1_4\"\n top: \"conv1_4\"\n\n}\nlayer {\n name: \"pool1_4\"\n type: \"Pooling\"\n bottom: \"conv1_4\"\n top: \"pool1_4\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n}\n\nlayer {\n name: \"conv2_4\"\n type: \"Convolution\"\n bottom: \"pool1_4\"\n top: \"conv2_4\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 48\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu2_4\"\n type: \"PReLU\"\n bottom: \"conv2_4\"\n top: \"conv2_4\"\n}\nlayer {\n name: \"pool2_4\"\n type: \"Pooling\"\n bottom: \"conv2_4\"\n top: \"pool2_4\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n \n}\nlayer {\n name: \"conv3_4\"\n type: \"Convolution\"\n bottom: \"pool2_4\"\n top: \"conv3_4\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 64\n kernel_size: 2\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu3_4\"\n type: \"PReLU\"\n bottom: \"conv3_4\"\n top: \"conv3_4\"\n}\n##########################\n##########################\nlayer {\n name: \"conv1_5\"\n type: \"Convolution\"\n bottom: \"data245\"\n top: \"conv1_5\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 28\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu1_5\"\n type: \"PReLU\"\n bottom: \"conv1_5\"\n top: \"conv1_5\"\n\n}\nlayer {\n name: \"pool1_5\"\n type: \"Pooling\"\n bottom: \"conv1_5\"\n top: \"pool1_5\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n}\n\nlayer {\n name: \"conv2_5\"\n type: \"Convolution\"\n bottom: \"pool1_5\"\n top: \"conv2_5\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 48\n kernel_size: 3\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu2_5\"\n type: \"PReLU\"\n bottom: \"conv2_5\"\n top: \"conv2_5\"\n}\nlayer {\n name: \"pool2_5\"\n type: \"Pooling\"\n bottom: \"conv2_5\"\n top: \"pool2_5\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n \n}\nlayer {\n name: \"conv3_5\"\n type: \"Convolution\"\n bottom: \"pool2_5\"\n top: \"conv3_5\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n convolution_param {\n num_output: 64\n kernel_size: 2\n stride: 1\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu3_5\"\n type: \"PReLU\"\n bottom: \"conv3_5\"\n top: \"conv3_5\"\n}\n##########################\nlayer { \n\tname: \"concat\" \n\tbottom: \"conv3_1\" \n\tbottom: \"conv3_2\" \n\tbottom: \"conv3_3\" \n\tbottom: \"conv3_4\" \n\tbottom: \"conv3_5\" \n\ttop: \"conv3\" \n\ttype: \"Concat\" \n\tconcat_param { \n\t\taxis: 1 \n\t} \n}\n##########################\nlayer {\n name: \"fc4\"\n type: \"InnerProduct\"\n bottom: \"conv3\"\n top: \"fc4\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n num_output: 256\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu4\"\n type: \"PReLU\"\n bottom: \"fc4\"\n top: \"fc4\"\n}\n############################\nlayer {\n name: \"fc4_1\"\n type: \"InnerProduct\"\n bottom: \"fc4\"\n top: \"fc4_1\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n num_output: 64\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu4_1\"\n type: \"PReLU\"\n bottom: \"fc4_1\"\n top: \"fc4_1\"\n}\nlayer {\n name: \"fc5_1\"\n type: \"InnerProduct\"\n bottom: \"fc4_1\"\n top: \"fc5_1\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n num_output: 2\n weight_filler {\n type: \"xavier\"\n\t #type: \"constant\"\n #value: 0\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\n\n\n#########################\nlayer {\n name: \"fc4_2\"\n type: \"InnerProduct\"\n bottom: \"fc4\"\n top: \"fc4_2\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n num_output: 64\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu4_2\"\n type: \"PReLU\"\n bottom: \"fc4_2\"\n top: \"fc4_2\"\n}\nlayer {\n name: \"fc5_2\"\n type: \"InnerProduct\"\n bottom: \"fc4_2\"\n top: \"fc5_2\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n num_output: 2\n weight_filler {\n type: \"xavier\"\n\t #type: \"constant\"\n #value: 0\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\n\n#########################\nlayer {\n name: \"fc4_3\"\n type: \"InnerProduct\"\n bottom: \"fc4\"\n top: \"fc4_3\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n num_output: 64\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu4_3\"\n type: \"PReLU\"\n bottom: \"fc4_3\"\n top: \"fc4_3\"\n}\nlayer {\n name: \"fc5_3\"\n type: \"InnerProduct\"\n bottom: \"fc4_3\"\n top: \"fc5_3\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n num_output: 2\n weight_filler {\n type: \"xavier\"\n\t #type: \"constant\"\n #value: 0\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\n\n#########################\nlayer {\n name: \"fc4_4\"\n type: \"InnerProduct\"\n bottom: \"fc4\"\n top: \"fc4_4\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n num_output: 64\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu4_4\"\n type: \"PReLU\"\n bottom: \"fc4_4\"\n top: \"fc4_4\"\n}\nlayer {\n name: \"fc5_4\"\n type: \"InnerProduct\"\n bottom: \"fc4_4\"\n top: \"fc5_4\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n num_output: 2\n weight_filler {\n type: \"xavier\"\n\t #type: \"constant\"\n #value: 0\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\n\n#########################\nlayer {\n name: \"fc4_5\"\n type: \"InnerProduct\"\n bottom: \"fc4\"\n top: \"fc4_5\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n num_output: 64\n weight_filler {\n type: \"xavier\"\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n \n}\nlayer {\n name: \"prelu4_5\"\n type: \"PReLU\"\n bottom: \"fc4_5\"\n top: \"fc4_5\"\n}\nlayer {\n name: \"fc5_5\"\n type: \"InnerProduct\"\n bottom: \"fc4_5\"\n top: \"fc5_5\"\n param {\n lr_mult: 1\n decay_mult: 1\n }\n param {\n lr_mult: 2\n decay_mult: 1\n }\n inner_product_param {\n num_output: 2\n weight_filler {\n type: \"xavier\"\n\t #type: \"constant\"\n #value: 0\n\t}\n bias_filler {\n type: \"constant\"\n value: 0\n }\n }\n}\n\n#########################\n\n" }, { "path": "mtcnn_detector.py", "content": "# coding: utf-8\nimport os\nimport mxnet as mx\nimport numpy as np\nimport math\nimport cv2\nfrom multiprocessing import Pool\nfrom itertools import repeat\nfrom itertools import izip\nfrom helper import nms, adjust_input, generate_bbox, detect_first_stage_warpper\n\nclass MtcnnDetector(object):\n \"\"\"\n Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Neural Networks\n see https://github.com/kpzhang93/MTCNN_face_detection_alignment\n this is a mxnet version\n \"\"\"\n def __init__(self,\n model_folder='.',\n minsize = 20,\n threshold = [0.6, 0.7, 0.8],\n factor = 0.709,\n num_worker = 1,\n accurate_landmark = False,\n ctx=mx.cpu()):\n \"\"\"\n Initialize the detector\n\n Parameters:\n ----------\n model_folder : string\n path for the models\n minsize : float number\n minimal face to detect\n threshold : float number\n detect threshold for 3 stages\n factor: float number\n scale factor for image pyramid\n num_worker: int number\n number of processes we use for first stage\n accurate_landmark: bool\n use accurate landmark localization or not\n\n \"\"\"\n self.num_worker = num_worker\n self.accurate_landmark = accurate_landmark\n\n # load 4 models from folder\n models = ['det1', 'det2', 'det3','det4']\n models = [ os.path.join(model_folder, f) for f in models]\n \n self.PNets = []\n for i in range(num_worker):\n workner_net = mx.model.FeedForward.load(models[0], 1, ctx=ctx)\n self.PNets.append(workner_net)\n\n self.Pool = Pool(num_worker)\n\n self.RNet = mx.model.FeedForward.load(models[1], 1, ctx=ctx)\n self.ONet = mx.model.FeedForward.load(models[2], 1, ctx=ctx)\n self.LNet = mx.model.FeedForward.load(models[3], 1, ctx=ctx)\n\n self.minsize = float(minsize)\n self.factor = float(factor)\n self.threshold = threshold\n\n\n def convert_to_square(self, bbox):\n \"\"\"\n convert bbox to square\n\n Parameters:\n ----------\n bbox: numpy array , shape n x 5\n input bbox\n\n Returns:\n -------\n square bbox\n \"\"\"\n square_bbox = bbox.copy()\n\n h = bbox[:, 3] - bbox[:, 1] + 1\n w = bbox[:, 2] - bbox[:, 0] + 1\n max_side = np.maximum(h,w)\n square_bbox[:, 0] = bbox[:, 0] + w*0.5 - max_side*0.5\n square_bbox[:, 1] = bbox[:, 1] + h*0.5 - max_side*0.5\n square_bbox[:, 2] = square_bbox[:, 0] + max_side - 1\n square_bbox[:, 3] = square_bbox[:, 1] + max_side - 1\n return square_bbox\n\n def calibrate_box(self, bbox, reg):\n \"\"\"\n calibrate bboxes\n\n Parameters:\n ----------\n bbox: numpy array, shape n x 5\n input bboxes\n reg: numpy array, shape n x 4\n bboxex adjustment\n\n Returns:\n -------\n bboxes after refinement\n\n \"\"\"\n w = bbox[:, 2] - bbox[:, 0] + 1\n w = np.expand_dims(w, 1)\n h = bbox[:, 3] - bbox[:, 1] + 1\n h = np.expand_dims(h, 1)\n reg_m = np.hstack([w, h, w, h])\n aug = reg_m * reg\n bbox[:, 0:4] = bbox[:, 0:4] + aug\n return bbox\n\n \n def pad(self, bboxes, w, h):\n \"\"\"\n pad the the bboxes, alse restrict the size of it\n\n Parameters:\n ----------\n bboxes: numpy array, n x 5\n input bboxes\n w: float number\n width of the input image\n h: float number\n height of the input image\n Returns :\n ------s\n dy, dx : numpy array, n x 1\n start point of the bbox in target image\n edy, edx : numpy array, n x 1\n end point of the bbox in target image\n y, x : numpy array, n x 1\n start point of the bbox in original image\n ex, ex : numpy array, n x 1\n end point of the bbox in original image\n tmph, tmpw: numpy array, n x 1\n height and width of the bbox\n\n \"\"\"\n tmpw, tmph = bboxes[:, 2] - bboxes[:, 0] + 1, bboxes[:, 3] - bboxes[:, 1] + 1\n num_box = bboxes.shape[0]\n\n dx , dy= np.zeros((num_box, )), np.zeros((num_box, ))\n edx, edy = tmpw.copy()-1, tmph.copy()-1\n\n x, y, ex, ey = bboxes[:, 0], bboxes[:, 1], bboxes[:, 2], bboxes[:, 3]\n\n tmp_index = np.where(ex > w-1)\n edx[tmp_index] = tmpw[tmp_index] + w - 2 - ex[tmp_index]\n ex[tmp_index] = w - 1\n\n tmp_index = np.where(ey > h-1)\n edy[tmp_index] = tmph[tmp_index] + h - 2 - ey[tmp_index]\n ey[tmp_index] = h - 1\n\n tmp_index = np.where(x < 0)\n dx[tmp_index] = 0 - x[tmp_index]\n x[tmp_index] = 0\n\n tmp_index = np.where(y < 0)\n dy[tmp_index] = 0 - y[tmp_index]\n y[tmp_index] = 0\n\n return_list = [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph]\n return_list = [item.astype(np.int32) for item in return_list]\n\n return return_list\n\n def slice_index(self, number):\n \"\"\"\n slice the index into (n,n,m), m < n\n Parameters:\n ----------\n number: int number\n number\n \"\"\"\n def chunks(l, n):\n \"\"\"Yield successive n-sized chunks from l.\"\"\"\n for i in range(0, len(l), n):\n yield l[i:i + n]\n num_list = range(number)\n return list(chunks(num_list, self.num_worker))\n \n\n def detect_face(self, img):\n \"\"\"\n detect face over img\n Parameters:\n ----------\n img: numpy array, bgr order of shape (1, 3, n, m)\n input image\n Retures:\n -------\n bboxes: numpy array, n x 5 (x1,y2,x2,y2,score)\n bboxes\n points: numpy array, n x 10 (x1, x2 ... x5, y1, y2 ..y5)\n landmarks\n \"\"\"\n\n # check input\n MIN_DET_SIZE = 12\n\n if img is None:\n return None\n\n # only works for color image\n if len(img.shape) != 3:\n return None\n\n # detected boxes\n total_boxes = []\n\n height, width, _ = img.shape\n minl = min( height, width)\n\n # get all the valid scales\n scales = []\n m = MIN_DET_SIZE/self.minsize\n minl *= m\n factor_count = 0\n while minl > MIN_DET_SIZE:\n scales.append(m*self.factor**factor_count)\n minl *= self.factor\n factor_count += 1\n\n #############################################\n # first stage\n #############################################\n #for scale in scales:\n # return_boxes = self.detect_first_stage(img, scale, 0)\n # if return_boxes is not None:\n # total_boxes.append(return_boxes)\n \n sliced_index = self.slice_index(len(scales))\n total_boxes = []\n for batch in sliced_index:\n local_boxes = self.Pool.map( detect_first_stage_warpper, \\\n izip(repeat(img), self.PNets[:len(batch)], [scales[i] for i in batch], repeat(self.threshold[0])) )\n total_boxes.extend(local_boxes)\n \n # remove the Nones \n total_boxes = [ i for i in total_boxes if i is not None]\n\n if len(total_boxes) == 0:\n return None\n \n total_boxes = np.vstack(total_boxes)\n\n if total_boxes.size == 0:\n return None\n\n # merge the detection from first stage\n pick = nms(total_boxes[:, 0:5], 0.7, 'Union')\n total_boxes = total_boxes[pick]\n\n bbw = total_boxes[:, 2] - total_boxes[:, 0] + 1\n bbh = total_boxes[:, 3] - total_boxes[:, 1] + 1\n\n # refine the bboxes\n total_boxes = np.vstack([total_boxes[:, 0]+total_boxes[:, 5] * bbw,\n total_boxes[:, 1]+total_boxes[:, 6] * bbh,\n total_boxes[:, 2]+total_boxes[:, 7] * bbw,\n total_boxes[:, 3]+total_boxes[:, 8] * bbh,\n total_boxes[:, 4]\n ])\n\n total_boxes = total_boxes.T\n total_boxes = self.convert_to_square(total_boxes)\n total_boxes[:, 0:4] = np.round(total_boxes[:, 0:4])\n\n #############################################\n # second stage\n #############################################\n num_box = total_boxes.shape[0]\n\n # pad the bbox\n [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(total_boxes, width, height)\n # (3, 24, 24) is the input shape for RNet\n input_buf = np.zeros((num_box, 3, 24, 24), dtype=np.float32)\n\n for i in range(num_box):\n tmp = np.zeros((tmph[i], tmpw[i], 3), dtype=np.uint8)\n tmp[dy[i]:edy[i]+1, dx[i]:edx[i]+1, :] = img[y[i]:ey[i]+1, x[i]:ex[i]+1, :]\n input_buf[i, :, :, :] = adjust_input(cv2.resize(tmp, (24, 24)))\n\n output = self.RNet.predict(input_buf)\n\n # filter the total_boxes with threshold\n passed = np.where(output[1][:, 1] > self.threshold[1])\n total_boxes = total_boxes[passed]\n\n if total_boxes.size == 0:\n return None\n\n total_boxes[:, 4] = output[1][passed, 1].reshape((-1,))\n reg = output[0][passed]\n\n # nms\n pick = nms(total_boxes, 0.7, 'Union')\n total_boxes = total_boxes[pick]\n total_boxes = self.calibrate_box(total_boxes, reg[pick])\n total_boxes = self.convert_to_square(total_boxes)\n total_boxes[:, 0:4] = np.round(total_boxes[:, 0:4])\n\n #############################################\n # third stage\n #############################################\n num_box = total_boxes.shape[0]\n\n # pad the bbox\n [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(total_boxes, width, height)\n # (3, 48, 48) is the input shape for ONet\n input_buf = np.zeros((num_box, 3, 48, 48), dtype=np.float32)\n\n for i in range(num_box):\n tmp = np.zeros((tmph[i], tmpw[i], 3), dtype=np.float32)\n tmp[dy[i]:edy[i]+1, dx[i]:edx[i]+1, :] = img[y[i]:ey[i]+1, x[i]:ex[i]+1, :]\n input_buf[i, :, :, :] = adjust_input(cv2.resize(tmp, (48, 48)))\n\n output = self.ONet.predict(input_buf)\n\n # filter the total_boxes with threshold\n passed = np.where(output[2][:, 1] > self.threshold[2])\n total_boxes = total_boxes[passed]\n\n if total_boxes.size == 0:\n return None\n\n total_boxes[:, 4] = output[2][passed, 1].reshape((-1,))\n reg = output[1][passed]\n points = output[0][passed]\n\n # compute landmark points\n bbw = total_boxes[:, 2] - total_boxes[:, 0] + 1\n bbh = total_boxes[:, 3] - total_boxes[:, 1] + 1\n points[:, 0:5] = np.expand_dims(total_boxes[:, 0], 1) + np.expand_dims(bbw, 1) * points[:, 0:5]\n points[:, 5:10] = np.expand_dims(total_boxes[:, 1], 1) + np.expand_dims(bbh, 1) * points[:, 5:10]\n\n # nms\n total_boxes = self.calibrate_box(total_boxes, reg)\n pick = nms(total_boxes, 0.7, 'Min')\n total_boxes = total_boxes[pick]\n points = points[pick]\n \n if not self.accurate_landmark:\n return total_boxes, points\n\n #############################################\n # extended stage\n #############################################\n num_box = total_boxes.shape[0]\n patchw = np.maximum(total_boxes[:, 2]-total_boxes[:, 0]+1, total_boxes[:, 3]-total_boxes[:, 1]+1)\n patchw = np.round(patchw*0.25)\n\n # make it even\n patchw[np.where(np.mod(patchw,2) == 1)] += 1\n\n input_buf = np.zeros((num_box, 15, 24, 24), dtype=np.float32)\n for i in range(5):\n x, y = points[:, i], points[:, i+5]\n x, y = np.round(x-0.5*patchw), np.round(y-0.5*patchw)\n [dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph] = self.pad(np.vstack([x, y, x+patchw-1, y+patchw-1]).T,\n width,\n height)\n for j in range(num_box):\n tmpim = np.zeros((tmpw[j], tmpw[j], 3), dtype=np.float32)\n tmpim[dy[j]:edy[j]+1, dx[j]:edx[j]+1, :] = img[y[j]:ey[j]+1, x[j]:ex[j]+1, :]\n input_buf[j, i*3:i*3+3, :, :] = adjust_input(cv2.resize(tmpim, (24, 24)))\n\n output = self.LNet.predict(input_buf)\n\n pointx = np.zeros((num_box, 5))\n pointy = np.zeros((num_box, 5))\n\n for k in range(5):\n # do not make a large movement\n tmp_index = np.where(np.abs(output[k]-0.5) > 0.35)\n output[k][tmp_index[0]] = 0.5\n\n pointx[:, k] = np.round(points[:, k] - 0.5*patchw) + output[k][:, 0]*patchw\n pointy[:, k] = np.round(points[:, k+5] - 0.5*patchw) + output[k][:, 1]*patchw\n\n points = np.hstack([pointx, pointy])\n points = points.astype(np.int32)\n\n return total_boxes, points\n\n def list2colmatrix(self, pts_list):\n \"\"\"\n convert list to column matrix\n Parameters:\n ----------\n pts_list:\n input list\n Retures:\n -------\n colMat: \n\n \"\"\"\n assert len(pts_list) > 0\n colMat = []\n for i in range(len(pts_list)):\n colMat.append(pts_list[i][0])\n colMat.append(pts_list[i][1])\n colMat = np.matrix(colMat).transpose()\n return colMat\n\n def find_tfrom_between_shapes(self, from_shape, to_shape):\n \"\"\"\n find transform between shapes\n Parameters:\n ----------\n from_shape: \n to_shape: \n Retures:\n -------\n tran_m:\n tran_b:\n \"\"\"\n assert from_shape.shape[0] == to_shape.shape[0] and from_shape.shape[0] % 2 == 0\n\n sigma_from = 0.0\n sigma_to = 0.0\n cov = np.matrix([[0.0, 0.0], [0.0, 0.0]])\n\n # compute the mean and cov\n from_shape_points = from_shape.reshape(from_shape.shape[0]/2, 2)\n to_shape_points = to_shape.reshape(to_shape.shape[0]/2, 2)\n mean_from = from_shape_points.mean(axis=0)\n mean_to = to_shape_points.mean(axis=0)\n\n for i in range(from_shape_points.shape[0]):\n temp_dis = np.linalg.norm(from_shape_points[i] - mean_from)\n sigma_from += temp_dis * temp_dis\n temp_dis = np.linalg.norm(to_shape_points[i] - mean_to)\n sigma_to += temp_dis * temp_dis\n cov += (to_shape_points[i].transpose() - mean_to.transpose()) * (from_shape_points[i] - mean_from)\n\n sigma_from = sigma_from / to_shape_points.shape[0]\n sigma_to = sigma_to / to_shape_points.shape[0]\n cov = cov / to_shape_points.shape[0]\n\n # compute the affine matrix\n s = np.matrix([[1.0, 0.0], [0.0, 1.0]])\n u, d, vt = np.linalg.svd(cov)\n\n if np.linalg.det(cov) < 0:\n if d[1] < d[0]:\n s[1, 1] = -1\n else:\n s[0, 0] = -1\n r = u * s * vt\n c = 1.0\n if sigma_from != 0:\n c = 1.0 / sigma_from * np.trace(np.diag(d) * s)\n\n tran_b = mean_to.transpose() - c * r * mean_from.transpose()\n tran_m = c * r\n\n return tran_m, tran_b\n\n def extract_image_chips(self, img, points, desired_size=256, padding=0):\n \"\"\"\n crop and align face\n Parameters:\n ----------\n img: numpy array, bgr order of shape (1, 3, n, m)\n input image\n points: numpy array, n x 10 (x1, x2 ... x5, y1, y2 ..y5)\n desired_size: default 256\n padding: default 0\n Retures:\n -------\n crop_imgs: list, n\n cropped and aligned faces \n \"\"\"\n crop_imgs = []\n for p in points:\n shape =[]\n for k in range(len(p)/2):\n shape.append(p[k])\n shape.append(p[k+5])\n\n if padding > 0:\n padding = padding\n else:\n padding = 0\n # average positions of face points\n mean_face_shape_x = [0.224152, 0.75610125, 0.490127, 0.254149, 0.726104]\n mean_face_shape_y = [0.2119465, 0.2119465, 0.628106, 0.780233, 0.780233]\n\n from_points = []\n to_points = []\n\n for i in range(len(shape)/2):\n x = (padding + mean_face_shape_x[i]) / (2 * padding + 1) * desired_size\n y = (padding + mean_face_shape_y[i]) / (2 * padding + 1) * desired_size\n to_points.append([x, y])\n from_points.append([shape[2*i], shape[2*i+1]])\n\n # convert the points to Mat\n from_mat = self.list2colmatrix(from_points)\n to_mat = self.list2colmatrix(to_points)\n\n # compute the similar transfrom\n tran_m, tran_b = self.find_tfrom_between_shapes(from_mat, to_mat)\n\n probe_vec = np.matrix([1.0, 0.0]).transpose()\n probe_vec = tran_m * probe_vec\n\n scale = np.linalg.norm(probe_vec)\n angle = 180.0 / math.pi * math.atan2(probe_vec[1, 0], probe_vec[0, 0])\n\n from_center = [(shape[0]+shape[2])/2.0, (shape[1]+shape[3])/2.0]\n to_center = [0, 0]\n to_center[1] = desired_size * 0.4\n to_center[0] = desired_size * 0.5\n\n ex = to_center[0] - from_center[0]\n ey = to_center[1] - from_center[1]\n\n rot_mat = cv2.getRotationMatrix2D((from_center[0], from_center[1]), -1*angle, scale)\n rot_mat[0][2] += ex\n rot_mat[1][2] += ey\n\n chips = cv2.warpAffine(img, rot_mat, (desired_size, desired_size))\n crop_imgs.append(chips)\n\n return crop_imgs\n\n" } ]