def validate(model, device, data_loader, labels_dict, logger):
logger.info("Validating model")
y_arr = np.zeros(shape=(len(data_loader), len(labels_dict)), dtype=np.float64)
y_pred_arr = np.zeros(shape=(len(data_loader), len(labels_dict)), dtype=np.float64)
print_freq = int(0.1*(len(data_loader)))
for i, sample in enumerate(data_loader):
if i % print_freq == 0:
logger.info("image num {:04d}...".format(i))
inputs = sample["image"].to(device)
labels = sample["label"].to(device)
outputs = model(inputs)
sftmax = Softmax()(outputs)
# sftmax[sftmax >= tau] = 1.0
# sftmax[sftmax < tau] = 0.0
y_arr[i] = labels.cpu().detach().numpy()[0]
y_pred_arr[i] = sftmax.cpu().detach().numpy()[0]
for i, label in enumerate(labels_dict.keys()):
class_y_pred = y_pred_arr[:, i:i + 1].reshape(y_pred_arr.shape[0])
class_y = y_arr[:, i:i + 1].reshape(y_arr.shape[0])
roc_auc = roc_auc_score(class_y, class_y_pred)
logger.info(" Class: '{}', roc auc: {:.4f}".format(label, roc_auc))
def recognizeImage():
gpu_available = torch.cuda.is_available()
base64string = request.get_json(force=True)['base64']
image = get_image(base64string)
if image.shape[1:] != (3, 32, 32):
return jsonify(error="only RGB image 32x32 accepted",
message="failure!")
model = get_model()
# breakpoint()
# get sample outputs
# breakpoint()
output_tensor = model(image.cuda())
output_tensor = Softmax(dim=1)(output_tensor)
prob_pred_tensor, pred_tensor = torch.max(output_tensor, 1)
output = np.squeeze(
output_tensor.detach().numpy()) if not gpu_available else np.squeeze(
output_tensor.cpu().detach().numpy())
prob_pred = np.squeeze(prob_pred_tensor.detach().numpy()
) if not gpu_available else np.squeeze(
prob_pred_tensor.cpu().detach().numpy())
pred = np.squeeze(
pred_tensor.detach().numpy()) if not gpu_available else np.squeeze(
pred_tensor.cpu().detach().numpy())
classes = [
'airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog',
'horse', 'ship', 'truck'
]
pred_dict = {
'predicted_class': str(classes[pred]),
'prob_predicted_class': str(prob_pred)
}
for i, prob in enumerate(output):
pred_dict[classes[i]] = str(prob)
return jsonify(prediction=pred_dict, message="success!")