{ "cells": [ { "cell_type": "code", "execution_count": 4, "metadata": {}, "outputs": [], "source": [ "# David Ouyang 12/5/2019\n", "\n", "# Notebook which:\n", "# 1. Downloads weights\n", "# 2. Initializes model and imports weights\n", "# 3. Performs test time evaluation of videos (already preprocessed with ConvertDICOMToAVI.ipynb)\n", "\n", "import re\n", "import os, os.path\n", "from os.path import splitext\n", "import pydicom as dicom\n", "import numpy as np\n", "from pydicom.uid import UID, generate_uid\n", "import shutil\n", "from multiprocessing import dummy as multiprocessing\n", "import time\n", "import subprocess\n", "import datetime\n", "from datetime import date\n", "import sys\n", "import cv2\n", "import matplotlib.pyplot as plt\n", "import sys\n", "from shutil import copy\n", "import math\n", "import torch\n", "import torchvision\n", "\n", "sys.path.append(\"..\")\n", "import echonet\n", "\n", "import wget \n", "\n", "#destinationFolder = \"/Users/davidouyang/Dropbox/Echo Research/CodeBase/Output\"\n", "destinationFolder = \"C:\\\\Users\\\\Windows\\\\Dropbox\\\\Echo Research\\\\CodeBase\\\\Output\"\n", "#videosFolder = \"/Users/davidouyang/Dropbox/Echo Research/CodeBase/a4c-video-dir\"\n", "videosFolder = \"C:\\\\Users\\\\Windows\\\\Dropbox\\\\Echo Research\\\\CodeBase\\\\a4c-video-dir\"\n", "#DestinationForWeights = \"/Users/davidouyang/Dropbox/Echo Research/CodeBase/EchoNetDynamic-Weights\"\n", "DestinationForWeights = \"C:\\\\Users\\\\Windows\\\\Dropbox\\\\Echo Research\\\\CodeBase\\\\EchoNetDynamic-Weights\"" ] }, { "cell_type": "code", "execution_count": 5, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "The weights are at C:\\Users\\Windows\\Dropbox\\Echo Research\\CodeBase\\EchoNetDynamic-Weights\n", "Segmentation Weights already present\n", "EF Weights already present\n" ] } ], "source": [ "# Download model weights\n", "\n", "if os.path.exists(DestinationForWeights):\n", " print(\"The weights are at\", DestinationForWeights)\n", "else:\n", " print(\"Creating folder at \", DestinationForWeights, \" to store weights\")\n", " os.mkdir(DestinationForWeights)\n", " \n", "segmentationWeightsURL = 'https://github.com/douyang/EchoNetDynamic/releases/download/v1.0.0/deeplabv3_resnet50_random.pt'\n", "ejectionFractionWeightsURL = 'https://github.com/douyang/EchoNetDynamic/releases/download/v1.0.0/r2plus1d_18_32_2_pretrained.pt'\n", "\n", "\n", "if not os.path.exists(os.path.join(DestinationForWeights, os.path.basename(segmentationWeightsURL))):\n", " print(\"Downloading Segmentation Weights, \", segmentationWeightsURL,\" to \",os.path.join(DestinationForWeights,os.path.basename(segmentationWeightsURL)))\n", " filename = wget.download(segmentationWeightsURL, out = DestinationForWeights)\n", "else:\n", " print(\"Segmentation Weights already present\")\n", " \n", "if not os.path.exists(os.path.join(DestinationForWeights, os.path.basename(ejectionFractionWeightsURL))):\n", " print(\"Downloading EF Weights, \", ejectionFractionWeightsURL,\" to \",os.path.join(DestinationForWeights,os.path.basename(ejectionFractionWeightsURL)))\n", " filename = wget.download(ejectionFractionWeightsURL, out = DestinationForWeights)\n", "else:\n", " print(\"EF Weights already present\")\n", " \n" ] }, { "cell_type": "code", "execution_count": 6, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "loading weights from C:\\Users\\Windows\\Dropbox\\Echo Research\\CodeBase\\EchoNetDynamic-Weights\\r2plus1d_18_32_2_pretrained\n", "cuda is available, original weights\n", "external_test ['0X1A05DFFFCAFB253B.avi', '0X1A0A263B22CCD966.avi', '0X1A2A76BDB5B98BED.avi', '0X1A2C60147AF9FDAE.avi', '0X1A2E9496910EFF5B.avi', '0X1A3D565B371DC573.avi', '0X1A3E7BF1DFB132FB.avi', '0X1A5FAE3F9D37794E.avi', '0X1A6ACFE7B286DAFC.avi', '0X1A8D85542DBE8204.avi', '23_Apical_4_chamber_view.dcm.avi', '62_Apical_4_chamber_view.dcm.avi', '64_Apical_4_chamber_view.dcm.avi']\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████████████████████████████████████████████████████████████████████████████| 10/10 [00:10<00:00, 1.00s/it]\n", "100%|████████████████████████████████████████████████████████| 13/13 [00:29<00:00, 2.26s/it, 3122.29 (3440.26) / 0.00]\n" ] } ], "source": [ "# Initialize and Run EF model\n", "\n", "frames = 32\n", "period = 1 #2\n", "batch_size = 20\n", "model = torchvision.models.video.r2plus1d_18(pretrained=False)\n", "model.fc = torch.nn.Linear(model.fc.in_features, 1)\n", "\n", "\n", "\n", "print(\"loading weights from \", os.path.join(DestinationForWeights, \"r2plus1d_18_32_2_pretrained\"))\n", "\n", "if torch.cuda.is_available():\n", " print(\"cuda is available, original weights\")\n", " device = torch.device(\"cuda\")\n", " model = torch.nn.DataParallel(model)\n", " model.to(device)\n", " checkpoint = torch.load(os.path.join(DestinationForWeights, os.path.basename(ejectionFractionWeightsURL)))\n", " model.load_state_dict(checkpoint['state_dict'])\n", "else:\n", " print(\"cuda is not available, cpu weights\")\n", " device = torch.device(\"cpu\")\n", " checkpoint = torch.load(os.path.join(DestinationForWeights, os.path.basename(ejectionFractionWeightsURL)), map_location = \"cpu\")\n", " state_dict_cpu = {k[7:]: v for (k, v) in checkpoint['state_dict'].items()}\n", " model.load_state_dict(state_dict_cpu)\n", "\n", "\n", "# try some random weights: final_r2+1d_model_regression_EF_sgd_skip1_32frames.pth.tar\n", "# scp ouyangd@arthur2:~/Echo-Tracing-Analysis/final_r2+1d_model_regression_EF_sgd_skip1_32frames.pth.tar \"C:\\Users\\Windows\\Dropbox\\Echo Research\\CodeBase\\EchoNetDynamic-Weights\"\n", "#Weights = \"final_r2+1d_model_regression_EF_sgd_skip1_32frames.pth.tar\"\n", "\n", "\n", "output = os.path.join(destinationFolder, \"cedars_ef_output.csv\")\n", "\n", "ds = echonet.datasets.Echo(split = \"external_test\", external_test_location = videosFolder, crops=\"all\")\n", "print(ds.split, ds.fnames)\n", "\n", "mean, std = echonet.utils.get_mean_and_std(ds)\n", "\n", "kwargs = {\"target_type\": \"EF\",\n", " \"mean\": mean,\n", " \"std\": std,\n", " \"length\": frames,\n", " \"period\": period,\n", " }\n", "\n", "ds = echonet.datasets.Echo(split = \"external_test\", external_test_location = videosFolder, **kwargs, crops=\"all\")\n", "\n", "test_dataloader = torch.utils.data.DataLoader(ds, batch_size = 1, num_workers = 5, shuffle = True, pin_memory=(device.type == \"cuda\"))\n", "loss, yhat, y = echonet.utils.video.run_epoch(model, test_dataloader, \"test\", None, device, save_all=True, blocks=25)\n", "\n", "with open(output, \"w\") as g:\n", " for (filename, pred) in zip(ds.fnames, yhat):\n", " for (i,p) in enumerate(pred):\n", " g.write(\"{},{},{:.4f}\\n\".format(filename, i, p))\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Initialize and Run Segmentation model\n", "\n", "torch.cuda.empty_cache()\n", "\n", "\n", "videosFolder = \"C:\\\\Users\\\\Windows\\\\Dropbox\\\\Echo Research\\\\CodeBase\\\\View Classification\\\\AppearsA4c\\\\Resized2\"\n", "\n", "def collate_fn(x):\n", " x, f = zip(*x)\n", " i = list(map(lambda t: t.shape[1], x))\n", " x = torch.as_tensor(np.swapaxes(np.concatenate(x, 1), 0, 1))\n", " return x, f, i\n", "\n", "dataloader = torch.utils.data.DataLoader(echonet.datasets.Echo(split=\"external_test\", external_test_location = videosFolder, target_type=[\"Filename\"], length=None, period=1, mean=mean, std=std),\n", " batch_size=10, num_workers=0, shuffle=False, pin_memory=(device.type == \"cuda\"), collate_fn=collate_fn)\n", "if not all([os.path.isfile(os.path.join(destinationFolder, \"labels\", os.path.splitext(f)[0] + \".npy\")) for f in dataloader.dataset.fnames]):\n", " # Save segmentations for all frames\n", " # Only run if missing files\n", "\n", " pathlib.Path(os.path.join(destinationFolder, \"labels\")).mkdir(parents=True, exist_ok=True)\n", " block = 1024\n", " model.eval()\n", "\n", " with torch.no_grad():\n", " for (x, f, i) in tqdm.tqdm(dataloader):\n", " x = x.to(device)\n", " y = np.concatenate([model(x[i:(i + block), :, :, :])[\"out\"].detach().cpu().numpy() for i in range(0, x.shape[0], block)]).astype(np.float16)\n", " start = 0\n", " for (filename, offset) in zip(f, i):\n", " np.save(os.path.join(destinationFolder, \"labels\", os.path.splitext(filename)[0]), y[start:(start + offset), 0, :, :])\n", " start += offset\n", " \n", "dataloader = torch.utils.data.DataLoader(echonet.datasets.Echo(split=\"external_test\", external_test_location = videosFolder, target_type=[\"Filename\"], length=None, period=1, segmentation=os.path.join(destinationFolder, \"labels\")),\n", " batch_size=1, num_workers=8, shuffle=False, pin_memory=False)\n", "if not all(os.path.isfile(os.path.join(destinationFolder, \"videos\", f)) for f in dataloader.dataset.fnames):\n", " pathlib.Path(os.path.join(destinationFolder, \"videos\")).mkdir(parents=True, exist_ok=True)\n", " pathlib.Path(os.path.join(destinationFolder, \"size\")).mkdir(parents=True, exist_ok=True)\n", " echonet.utils.latexify()\n", " with open(os.path.join(destinationFolder, \"size.csv\"), \"w\") as g:\n", " g.write(\"Filename,Frame,Size,ComputerSmall\\n\")\n", " for (x, filename) in tqdm.tqdm(dataloader):\n", " x = x.numpy()\n", " for i in range(len(filename)):\n", " img = x[i, :, :, :, :].copy()\n", " logit = img[2, :, :, :].copy()\n", " img[1, :, :, :] = img[0, :, :, :]\n", " img[2, :, :, :] = img[0, :, :, :]\n", " img = np.concatenate((img, img), 3)\n", " img[0, :, :, 112:] = np.maximum(255. * (logit > 0), img[0, :, :, 112:])\n", "\n", " img = np.concatenate((img, np.zeros_like(img)), 2)\n", " size = (logit > 0).sum(2).sum(1)\n", " try:\n", " trim_min = sorted(size)[round(len(size) ** 0.05)]\n", " except:\n", " import code; code.interact(local=dict(globals(), **locals()))\n", " trim_max = sorted(size)[round(len(size) ** 0.95)]\n", " trim_range = trim_max - trim_min\n", " peaks = set(scipy.signal.find_peaks(-size, distance=20, prominence=(0.50 * trim_range))[0])\n", " for (x, y) in enumerate(size):\n", " g.write(\"{},{},{},{}\\n\".format(filename[0], x, y, 1 if x in peaks else 0))\n", " fig = plt.figure(figsize=(size.shape[0] / 50 * 1.5, 3))\n", " plt.scatter(np.arange(size.shape[0]) / 50, size, s=1)\n", " ylim = plt.ylim()\n", " for p in peaks:\n", " plt.plot(np.array([p, p]) / 50, ylim, linewidth=1)\n", " plt.ylim(ylim)\n", " plt.title(os.path.splitext(filename[i])[0])\n", " plt.xlabel(\"Seconds\")\n", " plt.ylabel(\"Size (pixels)\")\n", " plt.tight_layout()\n", " plt.savefig(os.path.join(destinationFolder, \"size\", os.path.splitext(filename[i])[0] + \".pdf\"))\n", " plt.close(fig)\n", " size -= size.min()\n", " size = size / size.max()\n", " size = 1 - size\n", " for (x, y) in enumerate(size):\n", " img[:, :, int(round(115 + 100 * y)), int(round(x / len(size) * 200 + 10))] = 255.\n", " interval = np.array([-3, -2, -1, 0, 1, 2, 3])\n", " for a in interval:\n", " for b in interval:\n", " img[:, x, a + int(round(115 + 100 * y)), b + int(round(x / len(size) * 200 + 10))] = 255.\n", " if x in peaks:\n", " img[:, :, 200:225, b + int(round(x / len(size) * 200 + 10))] = 255.\n", " echonet.utils.savevideo(os.path.join(destinationFolder, \"videos\", filename[i]), img.astype(np.uint8), 50) " ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.7.4" } }, "nbformat": 4, "nbformat_minor": 2 }