def train(model_id, model_path, data_paths_path, feature_path, nb_epoch, batch_size, load_epoch):
start_time = time.clock()
t_la = [[], []]
t_l = [[], []]
t_a = [[], []]
v_l = [[], []]
v_a = [[], []]
fig = None
model = models.get_model_from_id(model_id)
if model is None:
return
model_path = model_path + model_id
# Load log
if not os.path.exists(model_path):
os.makedirs(model_path)
log = open(model_path + '/log.txt', "a")
log.write('\n\n\nTraining initialised: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now()))
if load_epoch == 0:
print 'Training model from scratch...'
log.write('\nTraining model from scratch...')
else:
if load_epoch < 0 or load_epoch is None: # get latest
for i in range(100, -1, -1):
if os.path.isfile(model_path + '/' + model_id + '_' + str(i) + '.h5'):
load_epoch = i
break
if load_epoch is None:
load_epoch = 0
if load_epoch == 0:
log.write('\nTraining model from scratch...')
else:
print 'Loading past model to train from:'
print model_path + '/' + model_id + '_' + str(load_epoch) + '.h5'
log.write('\nLoading past model to train from:')
log.write('\n' + model_path + '/' + model_id + '_' + str(load_epoch) + '.h5')
[t_l, v_l, v_a] = np.load(model_path + '/training_stats_' + str(load_epoch) + '.npy')
model.load_weights(model_path + '/' + model_id + '_' + str(load_epoch) + '.h5')
model = models.compile_model(model_id, model)
for e in range(load_epoch + 1, nb_epoch+1):
print(
"--------------------------------------------\nepoch %d\n--------------------------------------------" % e)
log.write(
"\n--------------------------------------------\nepoch %d\n--------------------------------------------" % e)
# get data
with open(data_paths_path + 'train_paths_equalised.txt') as f:
all_paths = f.readlines()
random.shuffle(all_paths) # randomise order every epoch!!
all_paths = [line.split() for line in all_paths] # split so x and y split
X_batch = []
Y_batch = []
sum_loss = 0
past = 0
count = 0
inner_count = 0
start_time_inner = time.clock()
for path in all_paths:
count += 1
x, y = models.load_input(model_id, feature_path, path)
X_batch.append(x)
Y_batch.append(y)
if (count % batch_size == 0) or (count == len(all_paths)):
# print 'B'
if count == len(all_paths):
inner_count + 1
Y_batch = np.squeeze(Y_batch)
loss, acc = model.train_on_batch(X_batch, Y_batch)
sum_loss += loss
inner_count += 1
# clear batch
X_batch = []
Y_batch = []
if (int((float(count) / len(all_paths)) * 100) > past) or (count == len(all_paths)):
tr = (len(all_paths) - count) / ((count) / (time.clock() - start_time_inner))
trt = ((nb_epoch - e + 1) * len(all_paths) - count) / (
((e - 1) * len(all_paths) + count) / (time.clock() - start_time))
print '(%d) [%.5f] Image: %d / %d; Epoch TR: %02d:%02d:%02d; Total TR: %02d:%02d:%02d;' % (
past, sum_loss / inner_count, count, len(all_paths), int((tr / 60) / 60), int((tr / 60) % 60),
int(tr % 60),
int((trt / 60) / 60), int((trt / 60) % 60), int(trt % 60))
log.close()
log = open(model_path + '/log.txt', "a")
log.write('\n(%d) [%.5f] Image: %d / %d; Epoch TR: %02d:%02d:%02d; Total TR: %02d:%02d:%02d;' % (
past, sum_loss / inner_count, count, len(all_paths), int((tr / 60) / 60), int((tr / 60) % 60),
int(tr % 60),
int((trt / 60) / 60), int((trt / 60) % 60), int(trt % 60)))
t_l[0].append((e - 1) + past * .01)
t_l[1].append(sum_loss / inner_count)
# graph it
if fig:
plt.close()
fig, ax1 = plt.subplots()
ax1.plot(t_l[0], t_l[1], 'g-')
ax1.plot(v_l[0], v_l[1], 'b-')
ax1.set_ylim(bottom=0)
ax2 = ax1.twinx()
ax2.plot(v_a[0], v_a[1], 'r-')
ax2.set_ylim(top=1)
# plt.plot(t_l[0], t_l[1])
# plt.plot(v_l[0],v_l[1])
# plt.plot(v_a[0],v_a[1])
# plt.show(block=False)
past += 10
sum_loss = 0
inner_count = 0
# if past > 0:
# break
print '--------------------------------------------'
print 'Validation results:'
log.write('\n--------------------------------------------')
log.write('\nValidation results:\n')
with open(data_paths_path + 'val_paths_equalised.txt') as f:
all_val_paths = f.readlines()
random.shuffle(all_val_paths) # randomise order every epoch!!
all_val_paths = [line.split() for line in all_val_paths] # split so x and y split
X_val = []
Y_val = []
count = 0
past = 0
val_metrics = []
for path in all_val_paths:
count += 1
x, y = models.load_input(model_id, feature_path, path)
X_val.append(x)
Y_val.append(y)
if (count % batch_size == 0) or (count == len(all_paths)):
# test
Y_val = np.squeeze(Y_val)
val_metrics.append(model.test_on_batch(X_val, Y_val))
# clear batch
X_val = []
Y_val = []
if int((float(count) / len(all_val_paths)) * 100) > past:
print('.'),
log.write('.')
past += 10
print '\n'
val_results = np.average(val_metrics, axis=0)
print val_results
log.write('\n' + str(val_results))
v_l[0].append(e)
v_l[1].append(val_results[0])
v_a[0].append(e)
v_a[1].append(val_results[1])
if e % 1 == 0:
if not os.path.exists(model_path):
os.makedirs(model_path)
model.save_weights(model_path + '/' + model_id + '_' + str(e) + '.h5', overwrite=True)
# fig.savefig(model_path + '/training.png')
fig.savefig(model_path + '/training.pdf')
np.save(model_path + '/training_stats_' + str(e) + '.npy', [t_l, v_l, v_a])
tt = time.clock() - start_time
print 'Total Time Taken: %02d:%02d:%02d;' % (int((tt / 60) / 60), int((tt / 60) % 60), int(tt % 60))
log.write('\n\nTotal Time Taken: %02d:%02d:%02d;' % (int((tt / 60) / 60), int((tt / 60) % 60), int(tt % 60)))
return model
def predict(model_id, model_path, data_paths_path, feature_path, split, batch_size=None, load_epoch=None, layers=['pred'], save_path=None, equalised=False):
start_time = time.clock()
output_classes = 7
model = models.get_model_from_id(model_id)
if model is None:
return
# Load log
model_path = model_path + model_id
if not os.path.exists(model_path):
os.makedirs(model_path)
if load_epoch is not None:
print 'Loading model: ' + model_path + '/' + model_id + '_' + str(load_epoch) + '.h5'
model.load_weights(model_path + '/' + model_id + '_' + str(load_epoch) + '.h5')
else:
print 'ERROR: Need load_epoch number to load'
return
# model = models.compile_model(model_id, model) # dont need to compile on prediction
# get data
if equalised:
with open(data_paths_path + split + '_paths_equalised.txt') as f:
all_paths = f.readlines()
else:
with open(data_paths_path + split + '_paths.txt') as f:
all_paths = f.readlines()
# all_paths = all_paths[:500]
all_paths = [line.split() for line in all_paths] # split so x and y split
for layer_name in layers:
# model, output_classes = models.get_model_from_id(model_id)
model = Model(input=model.input, output=model.get_layer(layer_name).output)
X_batch = []
Y_batch = []
Y_gt = None
Y_pred = None
past = 0
count = 0
inner_count = 0
for path in all_paths:
# print path
count += 1
cor_path = DRIVE + path[0][path[0].find('/DATASETS/')+1:]
if path[0] != cor_path:
# print 'Paths in .txt files seem incorrect'
# print 'Changed from: '+path[0]
# print 'Changed to: '+ cor_path
path[0] = cor_path
x, y = models.load_input(model_id, feature_path, path)
X_batch.append(x)
Y_batch.append(y)
if batch_size is not None:
if (count % batch_size == 0) or (count == len(all_paths)):
# Y_batch = np.eye(output_classes)[Y_batch]
# train
if Y_gt is None:
Y_pred = model.predict_on_batch(np.array(X_batch))
Y_gt = Y_batch
else:
Y_pred = np.append(Y_pred, model.predict_on_batch(np.array(X_batch)),axis=0)
Y_gt = np.append(Y_gt,Y_batch,axis=0)
inner_count += 1
# clear batches
X_batch, Y_batch = [], []
else:
Y_pred = model.predict(x)
Y_gt = np.eye(output_classes)[y]
if int((float(count) / len(all_paths)) * 100) > past:
tr = (len(all_paths) - count) / ((count) / (time.clock() - start_time))
print '(%d) Image: %d / %d; TR: %02d:%02d:%02d;' % (past, count, len(all_paths), int((tr / 60) / 60),
int((tr / 60) % 60), int(tr % 60))
past += 5
# if batch_size is None:
# Y_pred = model.predict_proba(X_batch, batch_size=32)
# save predictions to file
all_array = {}
if save_path is not None:
print '\nSaving ....'
save_path += model_id + '_' + str(load_epoch)+'/'
print save_path
for p in range(len(Y_gt)):
path = all_paths[p]
# write out to npy files
image_name = path[0].split('/')[len(path[0].split('/')) - 1]
if not os.path.exists(save_path + layer_name + '/npy/ind/'):
os.makedirs(save_path + layer_name + '/npy/ind/')
np.save(save_path + layer_name + '/npy/ind/' + image_name[:-4] + '.npy', np.squeeze(Y_pred[p]))
all_array[image_name[:-4]] = [Y_gt[p], Y_pred[p]]
# np.save(save_path + layer_name + '/npy/'+split+'.npy', all_array)
return Y_gt, Y_pred
