import os
import numpy as np
import matplotlib.pyplot as plt
from sklearn.metrics import jaccard_similarity_score
from data_reader import get_patches_dir
from model import get_unet
GOLD_IMG_4Classes_3 = ['fieldborders_7_4109150_merged', 'fieldborders_28_4109150_merged', 'wsb_9_4109150_merged']
GOLD_IMG_4Classes_4 = ['waterways_2285_4109150_merged', 'waterways_2633_4109150_merged',
'waterways_1285_4109150_merged', 'fieldborders_19_4109150_merged']
AMT_SMALL_VAL = 100
def predict_from_val(model, amt_pred, trs, config, x):
prd = np.zeros((amt_pred, config.ISZ, config.ISZ, config.NUM_CLASSES)).astype(np.float32)
tmp = model.predict(x, batch_size=1)
for i in range(amt_pred):
prd_i = tmp[i]
for c in range(config.NUM_CLASSES):
prd[i,:,:,c] = prd_i[:,:,c] > trs[c]
return prd
def display_pred(pred_res, true_masks, config, modelName, amt_pred, trs, min_pred_sum, output_folder):
os.makedirs(output_folder, exist_ok=True)
trs_str = "trs_none"
if trs != None:
trs_str = '_'.join([str(x) for x in trs])
trs_str = "trs_" + trs_str
#print("Saving predictions when np.sum(pred) >", min_pred_sum)
nothing_saved = True
for p in range(amt_pred):
for i in range(config.NUM_CLASSES):
pred = pred_res[p,:, :, i]
true_mask = true_masks[p, :, :, i]
sum_pred = np.sum(pred)
sum_mask = np.sum(true_mask)
if sum_pred > min_pred_sum :
#and sum_mask > min_pred_sum:
jk = jaccard_similarity_score(true_mask, pred)
#print("Calc jaccard", jk)
fn = os.path.join(output_folder,"{4}{0}_p{1}_cl{2}_{3}.png".format(modelName, p, i, trs_str, jk))
#print("Saving predictions with np.sum {0} to {1}".format(sum, fn))
plt.imsave(fn, pred, cmap='hot')
fn_tr= os.path.join(output_folder,"{4}{0}_p{1}_TRUE_cl{2}_{3}.png".format(modelName, p, i, trs_str, jk))
plt.imsave(fn_tr, true_mask, cmap='hot')
nothing_saved = False
if (nothing_saved):
print("All predictions did not satisfy: sum_pred > min_pred_sum, nothing saved. Min_pred_sum:", min_pred_sum)
def check_predict_folder(model, model_name, val_dir, predict_dir, config, loss_mode, amt_pred, verbose, imageNames =[]):
x_val, y_val = get_patches_dir(val_dir, config, shuffleOn=False, amt=amt_pred, verbose=verbose,
imageNames=imageNames)
for i in range(1,10):
tr = []
for j in range(config.NUM_CLASSES):
tr.append(i / 10.0)
check_predict_model(model, model_name, config, predic_cnt=amt_pred,
trs = tr, x=x_val, y=y_val, min_pred_sum = 10,
output_folder= predict_dir)
def check_predict_small_test(model, model_name, predict_dir, config, loss_mode):
verbose = False
amt_pred = config.AMT_SMALL_VAL
predict_dir = predict_dir + "_small_test"
val_dir = config.SMAL_VAL_DIR
return check_predict_folder(model, model_name, val_dir, predict_dir, config, loss_mode, amt_pred, verbose)
def check_predict_gold(model, model_name, predict_dir, config, loss_mode):
amt_pred = 3
verbose = False
imageNames = GOLD_IMG_4Classes_4
predict_dir = predict_dir + "_gold"
val_dir = config.VAL_DIR
return check_predict_folder(model, model_name, val_dir, predict_dir, config, loss_mode, amt_pred, verbose, imageNames= imageNames)
def check_predict_model(model, model_name, config, predic_cnt, trs, x, y, min_pred_sum, output_folder):
pred_res = predict_from_val(model, predic_cnt, trs, config, x)
display_pred(pred_res, y, config, modelName=model_name, amt_pred=predic_cnt, trs=trs,
min_pred_sum=min_pred_sum,
output_folder=output_folder)
def check_predict(model_name, weights_folder, config, loss_mode, predic_cnt, trs, x, y, min_pred_sum, output_folder):
model = get_unet(config, loss_mode)
model.load_weights(os.path.join(weights_folder, model_name))
check_predict_model(model, model_name, config, predic_cnt, trs, x, y, min_pred_sum, output_folder)
