import sys
import os
import evalMetrics as METRICS
import PP
import numpy as np
import torch
import torch.nn as nn
from torch.autograd import Variable
import augmentations as AUG
#---------------------------------------------
#Evaluation functions for PrivCNNs
#---------------------------------------------
def evalModelX(model, num_labels, num_labels2, postfix, main_folder_path, eval_method, gpu0, useGPU,
patch_size = 70, eval_metric = 'iou', test_augm = False, extra_patch = 30, priv_eval = True):
eval_list = main_folder_path + 'val' + postfix + '.txt'
img_list = open(eval_list).readlines()
v = 0
v_priv = 0
for img_str in img_list:
img_str = img_str.rstrip()
_, gt1, out1, gt2, out2, _ = predict(os.path.join(main_folder_path, img_str), model, num_labels, num_labels2,
postfix, main_folder_path, eval_method, gpu0, useGPU, patch_size=patch_size,
test_augm = test_augm, extra_patch = extra_patch, priv_eval = priv_eval)
v += METRICS.metricEval(eval_metric, out2, gt2, num_labels)
v_priv += METRICS.metricEval(eval_metric, out1, gt1, num_labels2)
return v_priv / len(img_list), v / len(img_list)
def testPredict(img, model, num_labels, num_labels2, eval_method, gpu0, useGPU, stride = 50, patch_size = 70, test_augm = True, extra_patch = 30, get_soft = False):
if eval_method == 0:
if useGPU:
_, out = model(Variable(torch.from_numpy(img).float(),volatile = True).cuda(gpu0))
else:
_, out = model(Variable(torch.from_numpy(img).float(),volatile = True))
out = out.data[0].cpu().numpy()
elif eval_method == 1:
_, out = predictByPatches(img, model, num_labels, num_labels2, useGPU, gpu0,
stride = stride, patch_size = patch_size,
test_augm = test_augm, extra_patch = extra_patch, priv_eval = False)
out = out.squeeze()
if get_soft:
return out
#take argmax to get predictions
out = np.argmax(out, axis = 0)
#remove batch and label dimension
out = out.squeeze()
return out
#returns the image as numpy, the ground truth and the prediction given model and input path
#affine = True, returns the affine transformation from loading the scan
def predict(img_path, model, num_labels, num_labels2, postfix, main_folder_path, eval_method, gpu0, useGPU,
stride = 50, patch_size = 70, test_augm = True, extra_patch = 30, priv_eval = True):
#read image
img = PP.numpyFromScan(img_path)
#read wmh
gt_path = img_path.replace('slices', 'gt_slices').replace('FLAIR', 'wmh').replace('/pre','')
gt, affine = PP.numpyFromScan(gt_path, get_affine = True, makebin = (num_labels == 2))
gif_path = img_path.replace('scans', 'gifs').replace('FLAIR', 'parcellation').replace('/pre','')
gif = PP.numpyFromScan(gif_path)
img = img.transpose((3,0,1,2))
img = img[np.newaxis, :]
gt = gt.transpose((3,0,1,2))
gif = gif.transpose((3,0,1,2))
if eval_method == 0:
if useGPU:
out1_v, out2_v = model(Variable(torch.from_numpy(img).float(),volatile=True).cuda(gpu0))
else:
out1_v, out2_v = model(Variable(torch.from_numpy(img).float(),volatile=True))
out1 = out1_v.data[0].cpu().numpy()
out2 = out2_v.data[0].cpu().numpy()
del out1_v, out2_v
elif eval_method == 1:
out1, out2 = predictByPatches(img, model, num_labels, num_labels2, useGPU, gpu0,
stride = stride, test_augm = test_augm, patch_size = patch_size,
extra_patch = extra_patch, priv_eval = priv_eval)
out1 = out1.squeeze()
out1 = np.argmax(out1, axis = 0)
out1 = out1.squeeze()
out2 = out2.squeeze()
out2 = np.argmax(out2, axis = 0)
out2 = out2.squeeze()
#remove batch and label dimension
img = img.squeeze()
return img, gif, out1, gt, out2, affine
def predictByPatches(img, model, num_labels, num_labels2, useGPU, gpu0, patch_size = 70, test_augm = False, stride = 50, extra_pad = 0, extra_patch = 30, priv_eval = True):
batch_num, num_channels, dim1, dim2, dim3 = img.shape
p_size = patch_size
#add padding to each dim s.t. % stride = 0
dim1_pad = stride - ((dim1-p_size) % stride)
dim2_pad = stride - ((dim2-p_size) % stride)
dim3_pad = stride - ((dim3-p_size) % stride)
x_1_off, x_2_off = int(round(dim1_pad/2.0)), dim1_pad//2
y_1_off, y_2_off = int(round(dim2_pad/2.0)), dim2_pad//2
z_1_off, z_2_off = int(round(dim3_pad/2.0)), dim3_pad//2
img = np.lib.pad(img, ((0,0),(0,0), (x_1_off, x_2_off), (y_1_off, y_2_off), (z_1_off, z_2_off)), mode='minimum')
_, _, padded_dim1, padded_dim2, padded_dim3 = img.shape
out2_shape = (img.shape[0], num_labels, img.shape[2], img.shape[3], img.shape[4])
out1_shape = (img.shape[0], num_labels2, img.shape[2], img.shape[3], img.shape[4])
out1_total = np.zeros(out1_shape, dtype=np.float16)
out1_counter = np.zeros(out1_shape, dtype=np.int8)
out2_total = np.zeros(out2_shape)
out2_counter = np.zeros(out2_shape)
extra_p = extra_patch / 2
for i in range(0, padded_dim1 - p_size + 1, stride):
for j in range(0, padded_dim2 - p_size + 1, stride):
for k in range(0, padded_dim3 - p_size + 1, stride):
if extra_p != 0:
i_l, i_r = getExtraPatchOffsets(i, 0, padded_dim1 - p_size, extra_p)
j_l, j_r = getExtraPatchOffsets(j, 0, padded_dim2 - p_size, extra_p)
k_l, k_r = getExtraPatchOffsets(k, 0, padded_dim3 - p_size, extra_p)
img_patch = img[:,:, (i-i_l):(i+p_size+i_r),(j-j_l):(j+p_size+j_r),(k-k_l):(k+p_size+k_r)]
out1_np, out2_np = getPatchPrediction(img_patch, model, useGPU, gpu0, extra_pad = extra_pad, test_augm = test_augm)
out1_np = removePatchOffset(out1_np, i_l, i_r, j_l, j_r, k_l, k_r)
out2_np = removePatchOffset(out2_np, i_l, i_r, j_l, j_r, k_l, k_r)
if priv_eval:
out1_total[:,:, i:i+p_size,j:j+p_size,k:k+p_size] += out1_np
out1_counter[:, :, i:i+p_size, j:j+p_size, k:k+p_size] += 1
out2_total[:,:, i:i+p_size,j:j+p_size,k:k+p_size] += out2_np
out2_counter[:, :, i:i+p_size, j:j+p_size, k:k+p_size] += 1
else:
img_patch = img[:, :, i:i+p_size, j:j+p_size, k:k+p_size]
#make a prediction on this image patch, adding extra padding during prediction and augmenting
#the result is of the same shape and size as the original img patch
out1_np, out2_np = getPatchPrediction(img_patch, model, useGPU, gpu0, extra_pad = extra_pad, test_augm = test_augm)
#too memory intensive
if priv_eval:
out1_total[:, :, i:i+p_size, j:j+p_size, k:k+p_size] += out1_np.astype(np.float16)
out1_counter[:, :, i:i+p_size, j:j+p_size, k:k+p_size] += 1
out2_total[:, :, i:i+p_size, j:j+p_size, k:k+p_size] += out2_np
out2_counter[:, :, i:i+p_size, j:j+p_size, k:k+p_size] += 1
if priv_eval:
out1_total = out1_total / out1_counter
out2_total = out2_total / out2_counter
#remove padding from predictions
out1_total = out1_total[:, :, x_1_off:-x_2_off, y_1_off:-y_2_off, z_1_off:-z_2_off]
out2_total = out2_total[:, :, x_1_off:-x_2_off, y_1_off:-y_2_off, z_1_off:-z_2_off]
return out1_total, out2_total
def getExtraPatchOffsets(v, low_bound, upper_bound, extra_p):
v_left = 0
v_right = 0
if v - extra_p > low_bound:
v_left = extra_p
if v + extra_p < upper_bound:
v_right = extra_p
return v_left, v_right
#list of tuple [(i_l, i_r), (j_l, j_r)]
def removePatchOffset(np_arr, i_l, i_r, j_l, j_r, k_l, k_r):
bn, c, s_i, s_j, s_k = np_arr.shape
return np_arr[:,:,(i_l):(s_i-i_r), (j_l):(s_j-j_r), (k_l):(s_k-k_r)]
def getPatchPrediction(img_patch, model, useGPU, gpu0, extra_pad = 10, test_augm = False):
pd = extra_pad/2
padding = ((0,0), (0,0), (pd, pd), (pd, pd), (pd,pd))
img_patch = np.pad(img_patch, padding, 'constant')
num_augm = 1
if test_augm:
num_augm = 3
out1_np_total = None
out2_np_total = None
for i in range(num_augm):
img_patch_cp = np.copy(img_patch)
if test_augm and i != 0:
#apply augmentation
rot_x, rot_y, rot_z = AUG.getRotationVal([10,10,10])
zoom_val = AUG.getScalingVal(0.9, 1.1)
img_patch_cp = AUG.applyScale([img_patch_cp], zoom_val, [3])[0]
img_patch_cp = AUG.applyRotation([img_patch_cp], [rot_x, rot_y, rot_z], [3])[0]
if useGPU:
out1, out2 = model(Variable(torch.from_numpy(img_patch_cp).float(),volatile=True).cuda(gpu0))
else:
out1, out2 = model(Variable(torch.from_numpy(img_patch_cp).float(),volatile=True))
out1_np = out1.data[0].cpu().numpy()
out2_np = out2.data[0].cpu().numpy()
del out1, out2
#output is (1 x 3 x dim1 x dim2 x dim3)
out1_np = out1_np[np.newaxis,:]
out2_np = out2_np[np.newaxis,:]
if test_augm and i != 0:
temp2 = np.copy(out2_np)
out2_np = None
rev_zoom_i = float(img_patch.shape[2]) / img_patch_cp.shape[2]
rev_zoom_j = float(img_patch.shape[3]) / img_patch_cp.shape[3]
rev_zoom_k = float(img_patch.shape[4]) / img_patch_cp.shape[4]
for j in range(temp2.shape[1]):
r2 = AUG.applyRotation([temp2[:,j:j+1,:,:,:]], [-rot_x, -rot_y, -rot_z], [3])[0]
r2 = AUG.applyScale([r2], [rev_zoom_i,rev_zoom_j,rev_zoom_k], [3])[0]
if not isinstance(out2_np, np.ndarray):
out2_np = np.zeros([1, temp2.shape[1], r2.shape[2], r2.shape[3], r2.shape[4]])
out2_np[:, j,:,:,:] = r2
out2_np = numpySoftmax(out2_np, 1)
nb, c, n_i, n_j, n_k = out2_np.shape
if not isinstance(out1_np_total, np.ndarray):
out1_np_total = out1_np[:,:,(pd):(n_i-pd),(pd):(n_j-pd),(pd):(n_k-pd)]
out2_np_total = out2_np[:,:,(pd):(n_i-pd),(pd):(n_j-pd),(pd):(n_k-pd)]
else:
out2_np_total += out2_np[:,:,(pd):(n_i-pd),(pd):(n_j-pd),(pd):(n_k-pd)]
return (out1_np_total), (out2_np_total / num_augm)
def numpySoftmax(x, axis_):
e_x = np.exp(x - np.max(x))
return e_x / (e_x.sum(axis=axis_) + 0.00001)
