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
#---------------------------------------------
def evalModelX(model, num_labels, postfix, main_folder_path, eval_method, gpu0, useGPU,
patch_size = 70, eval_metric = 'iou', test_augm = False, extra_patch = 30):
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()
_, gt, out, _ = predict(os.path.join(main_folder_path, img_str), model, num_labels, postfix,
main_folder_path, eval_method, gpu0, useGPU, patch_size=patch_size,
test_augm = test_augm, extra_patch = extra_patch)
curr_eval = METRICS.metricEval(eval_metric, out, gt, num_labels)
v+=curr_eval
return v / len(img_list)
def testPredict(img, model, num_labels, 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, useGPU, gpu0,
stride = stride, patch_size = patch_size,
test_augm = test_augm, extra_patch = extra_patch)
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, postfix, main_folder_path, eval_method, gpu0, useGPU,
stride = 50, patch_size = 70, test_augm = True, extra_patch = 30):
#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))
img = img.transpose((3,0,1,2))
img = img[np.newaxis, :]
gt = gt.transpose((3,0,1,2))
if eval_method == 0:
if useGPU:
out_v = model(Variable(torch.from_numpy(img).float(),volatile = True).cuda(gpu0))
else:
out_v = model(Variable(torch.from_numpy(img).float(),volatile = True))
out = out_v.data[0].cpu().numpy()
#FIX?
del out_v
out_v = Variable(torch.from_numpy(np.array([1])).float())
out_v = Variable(torch.from_numpy(np.array([1])).float())
elif eval_method == 1:
out = predictByPatches(img, model, num_labels, useGPU, gpu0, stride = stride, patch_size = patch_size, test_augm = test_augm, extra_patch = extra_patch)
out = out.squeeze()
#take argmax to get predictions
out = np.argmax(out, axis = 0)
#remove batch and label dimension
img = img.squeeze()
out = out.squeeze()
gt = gt.squeeze()
return img, gt, out, affine
def predictByPatches(img, model, num_labels, useGPU, gpu0, patch_size = 70, test_augm = False, stride = 50, extra_pad = 0, extra_patch = 30):
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)) % stride
dim2_pad = (stride - ((dim2-p_size) % stride)) % stride
dim3_pad = (stride - ((dim3-p_size) % stride)) % 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
out_shape = (img.shape[0], num_labels, img.shape[2], img.shape[3], img.shape[4])
out_total = np.zeros(out_shape)
out_counter = np.zeros(out_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)]
out_np = getPatchPrediction(img_patch, model, useGPU, gpu0, extra_pad = extra_pad, test_augm = test_augm)
out_np = removePatchOffset(out_np, i_l, i_r, j_l, j_r, k_l, k_r)
out_total[:,:, i:i+p_size,j:j+p_size,k:k+p_size] += out_np
out_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
out_np = getPatchPrediction(img_patch, model, useGPU, gpu0, extra_pad = extra_pad, test_augm = test_augm)
out_total[:, :, i:i+p_size, j:j+p_size, k:k+p_size] += out_np
out_counter[:, :, i:i+p_size, j:j+p_size, k:k+p_size] += 1
out_total = out_total / out_counter
#remove padding from predictions
nb, c, i_size, j_size, k_size = out_total.shape
out_total = out_total[:, :, x_1_off:i_size-x_2_off, y_1_off:j_size-y_2_off, z_1_off:k_size-z_2_off]
return out_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 = 0, 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
out_np_total = None
for i in range(num_augm):
img_patch_cp = np.copy(img_patch)
#AUGMENT IMAGE
if test_augm and i != 0:
pass
#apply augmentation
rot_x, rot_y, rot_z = AUG.getRotationVal([10,10,10])
zoom_val = AUG.getScalingVal(0.8, 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]
#MAKE PREDICTION
if useGPU:
out = model(Variable(torch.from_numpy(img_patch_cp).float(),volatile = True).cuda(gpu0))
else:
out = model(Variable(torch.from_numpy(img_patch_cp).float(),volatile = True))
out_np = out.data[0].cpu().numpy()
#output is (1 x 3 x dim1 x dim2 x dim3)
out_np = out_np[np.newaxis,:]
if test_augm and i != 0:
temp = np.copy(out_np)
out_np = None
#reverse augmentation on predictions
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(temp.shape[1]):
r = AUG.applyRotation([temp[:,j:j+1,:,:,:]], [-rot_x, -rot_y, -rot_z], [3])[0]
r = AUG.applyScale(r, [rev_zoom_i,rev_zoom_j,rev_zoom_k], [3])[0]
if not isinstance(out_np, np.ndarray):
out_np = np.zeros([1, temp.shape[1], r.shape[2], r.shape[3], r.shape[4]])
out_np[:, j,:,:,:] = r
out_np = numpySoftmax(out_np, 1)
if not isinstance(out_np_total, np.ndarray):
if pd == 0:
out_np_total = out_np
else:
out_np_total = out_np[:,:,pd:-pd, pd:-pd, pd:-pd]
else:
if pd ==0:
out_np_total += out_np
else:
out_np_total += out_np[:,:,pd:-pd, pd:-pd, pd:-pd]
return out_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)
