def on_epoch_end(self, epoch, logs=None):
for x, y_true in self.data:
predictions = [output_encoder.decode(y, self.anchors, self.model)
for y in self.model(x)]
ground_truth = [y.to_tensor() for y in y_true]
self.mean_AP_metric.update_state(ground_truth, predictions)
logs['val_meanAP'] = self.mean_AP_metric.result().numpy()
self.mean_AP_metric.reset_state()
for run in range(1, config.NUM_RUNS + 1):
weights_dir = 'weights_naive_pasting_ablation_{}'.format(run)
model_name = '{}_{}_samples_{}'.format(architecture.lower(),
train_samples,
train_type)
print('\nGenerating results for {}'.format(model_name))
model_file = model_name + '.h5'
model_path = os.path.join(weights_dir, model_file)
model.load_weights(model_path)
meanAP_metric.reset_state()
for x, y_true in test_data:
ground_truth = [y.to_tensor() for y in y_true]
predictions = [
output_encoder.decode(y, anchors, model)
for y in model(x)
]
meanAP_metric.update_state(ground_truth, predictions)
test_meanAP = meanAP_metric.result().numpy()
results['run_{}'.format(run)].append(test_meanAP)
print('test meanAP:', test_meanAP)
del model
results = pd.DataFrame(results)
results.to_csv('results_naive_pasting_ablation.csv', index=False)
model_name = architecture.lower() + '_{}'.format(train_type)
print('\nGenerating results for {}'.format(model_name))
model_file = model_name + '.h5'
model_path = os.path.join(weights_dir, model_file)
if not os.path.exists(model_path):
raise Exception('Model weights at {} not found'.format(model_path))
model.load_weights(model_path)
meanAP_metric.reset_state()
for x, y_true in test_data:
ground_truth = [y.to_tensor() for y in y_true]
predictions = [
output_encoder.decode(y, anchors, model) for y in model(x)
]
meanAP_metric.update_state(ground_truth, predictions)
test_meanAP = meanAP_metric.result().numpy()
results['run_{}'.format(run)].append(test_meanAP)
print('test meanAP:', test_meanAP)
for train_samples in config.TRAIN_SAMPLES:
for train_type in ['from_scratch', 'finetuned']:
results['architecture'].append(architecture)
results['train_samples'].append(train_samples)
results['train_type'].append(train_type)
for run in range(1, config.NUM_RUNS + 1):
results['num_real_samples'].append(num_real_samples)
results['num_fake_samples'].append(num_fake_samples)
results['train_type'].append(train_type)
for run in range(1, config.NUM_RUNS+1):
weights_dir = 'weights_pretraining_ablation_{}'.format(run)
model_name = '{}_{}_{}x'.format(architecture.lower(), train_type,
prop_fake_samples)
print('\nGenerating results for {}'.format(model_name))
model_file = model_name + '.h5'
model_path = os.path.join(weights_dir, model_file)
model.load_weights(model_path)
meanAP_metric.reset_state()
for x, y_true in test_data:
ground_truth = [y.to_tensor() for y in y_true]
predictions = [output_encoder.decode(y, anchors, model)
for y in model(x)]
meanAP_metric.update_state(ground_truth, predictions)
test_meanAP = meanAP_metric.result().numpy()
results['run_{}'.format(run)].append(test_meanAP)
print('test meanAP:', test_meanAP)
del model
results = pd.DataFrame(results)
results.to_csv('results_pretraining_ablation.csv', index=False)