loss = compute_loss(b)
losses.append(loss)
mean_valid_loss = np.sqrt(np.mean(losses))
print " mean validation loss (RMSE):\t\t%.6f" % mean_valid_loss
losses_valid.append(mean_valid_loss)
now = time.time()
time_since_start = now - start_time
time_since_prev = now - prev_time
prev_time = now
est_time_left = time_since_start * (float(NUM_CHUNKS -
(e + 1)) / float(e + 1))
eta = datetime.now() + timedelta(seconds=est_time_left)
eta_str = eta.strftime("%c")
print " %s since start (%.2f s)" % (load_data.hms(time_since_start),
time_since_prev)
print " estimated %s to go (ETA: %s)" % (load_data.hms(est_time_left),
eta_str)
print
del chunk_data, xs_chunk, x_chunk, y_chunk, xs_valid, x_valid # memory cleanup
print "Compute predictions on validation set for analysis in batches"
predictions_list = []
for b in xrange(num_batches_valid):
# if b % 1000 == 0:
# print " batch %d/%d" % (b + 1, num_batches_valid)
predictions = compute_output(b)
predictions_list.append(predictions)
predictions = compute_output(b)
predictions_chunk_list.append(predictions)
predictions_chunk = np.vstack(predictions_chunk_list)
predictions_chunk = predictions_chunk[:
chunk_length] # cut off zeros / padding
print(" compute average over transforms")
predictions_chunk_avg = predictions_chunk.reshape(
-1, len(augmentation_transforms), 37).mean(1)
predictions_list.append(predictions_chunk_avg)
time_since_start = time.time() - start_time
print(" %s since start" % load_data.hms(time_since_start))
all_predictions = np.vstack(predictions_list)
print("Write predictions to %s" % target_path_valid)
load_data.save_gz(target_path_valid, all_predictions)
print("Evaluate")
rmse_valid = analysis['losses_valid'][-1]
rmse_augmented = np.sqrt(np.mean((y_valid - all_predictions)**2))
print(" MSE (last iteration):\t%.6f" % rmse_valid)
print(" MSE (augmented):\t%.6f" % rmse_augmented)
if DO_TEST:
print()
print("TEST SET")
if b % 1000 == 0:
print " batch %d/%d" % (b + 1, num_batches_chunk)
predictions = compute_output(b)
predictions_chunk_list.append(predictions)
predictions_chunk = np.vstack(predictions_chunk_list)
predictions_chunk = predictions_chunk[:chunk_length] # cut off zeros / padding
print " compute average over transforms"
predictions_chunk_avg = predictions_chunk.reshape(-1, len(augmentation_transforms), 37).mean(1)
predictions_list.append(predictions_chunk_avg)
time_since_start = time.time() - start_time
print " %s since start" % load_data.hms(time_since_start)
all_predictions = np.vstack(predictions_list)
print "Write predictions to %s" % target_path_valid
load_data.save_gz(target_path_valid, all_predictions)
print "Evaluate"
rmse_valid = analysis['losses_valid'][-1]
rmse_augmented = np.sqrt(np.mean((y_valid - all_predictions)**2))
print " MSE (last iteration):\t%.6f" % rmse_valid
print " MSE (augmented):\t%.6f" % rmse_augmented
def main(
model="resnet",
mode=params.mode,
num_chunks=params.num_chunks,
chunk_size=params.chunk_size,
input_sizes=params.input_sizes,
batch_size=params.batch_size,
nbands=params.nbands,
model_name=params.model_name,
):
#create a model (neural network)
multi_model = call_model(params, model=model)
print("Model loaded: {}".format(model), flush=True)
if mode == "train":
#create a csv logger that will store the history of the .fit function into a .csv file
with open(params.full_path_of_history, 'w',
newline='') as history_file:
writer = csv.writer(history_file)
writer.writerow(["chunk", "loss", "binary_accuracy"])
loss = optimizers.Adam(lr=params.learning_rate)
multi_model.compile(
optimizer=loss,
loss="binary_crossentropy",
metrics=[metrics.binary_accuracy],
)
if nbands == 3:
augmented_data_gen_pos = ra.realtime_augmented_data_gen_pos_col(
params=params,
num_chunks=params.num_chunks,
chunk_size=chunk_size,
target_sizes=input_sizes,
augmentation_params=params.default_augmentation_params,
)
augmented_data_gen_neg = ra.realtime_augmented_data_gen_neg_col(
params=params,
num_chunks=params.num_chunks,
chunk_size=chunk_size,
target_sizes=input_sizes,
augmentation_params=params.default_augmentation_params,
)
else:
augmented_data_gen_pos = ra.realtime_augmented_data_gen_pos(
params=params,
range_min=params.range_min,
range_max=params.range_max,
num_chunks=params.num_chunks,
chunk_size=chunk_size,
target_sizes=input_sizes,
normalize=params.normalize,
resize=params.resize,
augmentation_params=params.default_augmentation_params,
)
augmented_data_gen_neg = ra.realtime_augmented_data_gen_neg(
params=params,
num_chunks=params.num_chunks,
chunk_size=chunk_size,
target_sizes=input_sizes,
normalize=params.normalize,
resize=params.resize,
augmentation_params=params.default_augmentation_params,
)
train_gen_neg = load_data.buffered_gen_mp(
augmented_data_gen_neg, buffer_size=params.buffer_size)
train_gen_pos = load_data.buffered_gen_mp(
augmented_data_gen_pos, buffer_size=params.buffer_size)
loss_per_chunk = []
bin_acc_per_chunk = []
actual_begin_time = time.time()
try:
for chunk in range(params.num_chunks):
start_time = time.time()
chunk_data_pos, chunk_length = next(train_gen_pos)
y_train_pos = chunk_data_pos.pop()
X_train_pos = chunk_data_pos
chunk_data_neg, _ = next(train_gen_neg)
y_train_neg = chunk_data_neg.pop()
X_train_neg = chunk_data_neg
if False: #just to view some images positive or negative
imgs = chunk_data_pos[0]
imgs = np.squeeze(imgs)
for img in imgs:
plt.imshow(img / 255.0)
plt.show()
X_train = np.concatenate((X_train_pos[0], X_train_neg[0]))
y_train = np.concatenate((y_train_pos, y_train_neg))
y_train = y_train.astype(np.int32)
y_train = np.expand_dims(y_train, axis=1)
batches = 0
start_chunk_processing_time = time.time()
for batch in iterate_minibatches(X_train,
y_train,
batch_size,
shuffle=True):
X_batch, y_batch = batch
history = multi_model.fit(
X_batch / 255.0 - params.avg_img, y_batch)
batches += 1
print("Chunck neural net time: {0:.3f} seconds".format(
time.time() - start_chunk_processing_time),
flush=True)
#write results to csv for later use
writer.writerow([
str(chunk),
str(history.history["loss"][0]),
str(history.history["binary_accuracy"][0])
])
#store loss and accuracy in list
loss_per_chunk.append(history.history["loss"][0])
bin_acc_per_chunk.append(
history.history["binary_accuracy"][0])
# plot loss and accuracy on interval
if chunk % params.chunk_plot_interval == 0:
save_loss_and_acc_figure(loss_per_chunk,
bin_acc_per_chunk, params)
#empty the train data
X_train = None
y_train = None
print("Chunck {}/{} has been trained".format(
chunk + 1, params.num_chunks),
flush=True)
except KeyboardInterrupt:
multi_model.save_weights(params.full_path_of_weights)
print("interrupted by KEYBOARD!", flush=True)
print("saved weights to: {}".format(
params.full_path_of_weights),
flush=True)
end_time = time.time()
multi_model.save_weights(params.full_path_of_weights)
print("\nSaved weights to: {}".format(params.full_path_of_weights),
flush=True)
print("\nSaved results to: {}".format(params.full_path_of_history),
flush=True)
final_time = end_time - actual_begin_time
print("\nTotal time employed ",
load_data.hms(final_time),
flush=True)
if mode == "predict":
if nbands == 3:
augmented_data_gen_test_fixed = ra.realtime_fixed_augmented_data_test_col(
params=params,
target_sizes=input_sizes) # ,normalize=normalize)
else:
augmented_data_gen_test_fixed = ra.realtime_fixed_augmented_data_test(
params=params, target_sizes=input_sizes)
#load a trained model
multi_model.load_weights(params.full_path_predict_weights)
predictions = []
test_batches = 0
if params.augm_pred == True: #seems to be a boolean to control whether you want the test data to be augmented or not when performing a prediction on test data.
start_time = time.time()
for e, (chunk_data_test, chunk_length_test
) in enumerate(augmented_data_gen_test_fixed):
X_test = chunk_data_test
X_test = X_test[0]
X_test = X_test / 255.0 - params.avg_img
pred1 = multi_model.predict(X_test)
pred2 = multi_model.predict(
np.array([np.flipud(image) for image in X_test]))
pred3 = multi_model.predict(
np.array([np.fliplr(np.flipud(image))
for image in X_test]))
pred4 = multi_model.predict(
np.array([np.fliplr(image) for image in X_test]))
preds = np.mean([pred1, pred2, pred3, pred4], axis=0)
preds = preds.tolist()
predictions = predictions + preds
print("done with predict chunk: {}".format(e), flush=True)
else:
for e, (chunk_data_test, chunk_length_test
) in enumerate(augmented_data_gen_test_fixed):
X_test = chunk_data_test
X_test = X_test[0]
X_test = X_test / 255.0 - params.avg_img
pred1 = multi_model.predict(X_test)
preds = pred1.tolist()
predictions = predictions + preds
with open("pred_" + params.model_name + ".pkl", "wb") as f:
pickle.dump([[ra.test_data], [predictions]], f,
pickle.HIGHEST_PROTOCOL)
objects = []
with (open("pred_" + params.model_name + ".pkl", "rb")) as openfile:
while True:
try:
objects.append(pickle.load(openfile))
except EOFError:
break
f = open("pred_my_model.csv", "w")
x = str(objects[0])
f.write(x)
f.write("\n")
f.close()
losses.append(loss)
mean_valid_loss = np.sqrt(np.mean(losses))
print " mean validation loss (RMSE):\t\t%.6f" % mean_valid_loss
losses_valid.append(mean_valid_loss)
layers.dump_params(l6, e=e)
now = time.time()
time_since_start = now - start_time
time_since_prev = now - prev_time
prev_time = now
est_time_left = time_since_start * (float(NUM_CHUNKS - (e + 1)) / float(e + 1))
eta = datetime.now() + timedelta(seconds=est_time_left)
eta_str = eta.strftime("%c")
print " %s since start (%.2f s)" % (load_data.hms(time_since_start), time_since_prev)
print " estimated %s to go (ETA: %s)" % (load_data.hms(est_time_left), eta_str)
print
del chunk_data, xs_chunk, x_chunk, y_chunk, xs_valid, x_valid # memory cleanup
print "Compute predictions on validation set for analysis in batches"
predictions_list = []
for b in xrange(num_batches_valid):
# if b % 1000 == 0:
# print " batch %d/%d" % (b + 1, num_batches_valid)
predictions = compute_output(b)
predictions_list.append(predictions)
losses.append(loss)
mean_valid_loss = np.sqrt(np.mean(losses))
print(" mean validation loss (RMSE):\t\t%.6f" % mean_valid_loss)
losses_valid.append(mean_valid_loss)
now = time.time()
time_since_start = now - start_time
time_since_prev = now - prev_time
prev_time = now
est_time_left = time_since_start * (float(NUM_CHUNKS -
(e + 1)) / float(e + 1))
eta = datetime.now() + timedelta(seconds=est_time_left)
eta_str = eta.strftime("%c")
print(" %s since start (%.2f s)" %
(load_data.hms(time_since_start), time_since_prev))
print(" estimated %s to go (ETA: %s)" %
(load_data.hms(est_time_left), eta_str))
print()
del chunk_data, xs_chunk, x_chunk, y_chunk, xs_valid, x_valid # memory cleanup
print("Compute predictions on validation set for analysis in batches")
predictions_list = []
for b in range(num_batches_valid):
# if b % 1000 == 0:
# print " batch %d/%d" % (b + 1, num_batches_valid)
predictions = compute_output(b)
predictions_list.append(predictions)
losses.append(loss)
mean_valid_loss = np.sqrt(np.mean(losses))
print(" mean validation loss (RMSE):\t\t%.6f" % mean_valid_loss)
losses_valid.append(mean_valid_loss)
layers.dump_params(l6, e=e)
now = time.time()
time_since_start = now - start_time
time_since_prev = now - prev_time
prev_time = now
est_time_left = time_since_start * (float(NUM_CHUNKS - (e + 1)) / float(e + 1))
eta = datetime.now() + timedelta(seconds=est_time_left)
eta_str = eta.strftime("%c")
print(" %s since start (%.2f s)" % (load_data.hms(time_since_start), time_since_prev))
print(" estimated %s to go (ETA: %s)" % (load_data.hms(est_time_left), eta_str))
print()
del chunk_data, xs_chunk, x_chunk, y_chunk, xs_valid, x_valid # memory cleanup
print("Compute predictions on validation set for analysis in batches")
predictions_list = []
for b in range(num_batches_valid):
# if b % 1000 == 0:
# print " batch %d/%d" % (b + 1, num_batches_valid)
predictions = compute_output(b)
predictions_list.append(predictions)
