Training Resnet50 error

[jiangzihanict]

I am using latest keras and keras-application to train a resnet50 model, but found following error. I have been google for a long time but haven't found a solution.

The related code: input_shape=(64,64,3) model = keras.applications.resnet50.ResNet50(include_top=False, weights=None, input_shape=input_shape, classes=2) train_gen = ImageDataGenerator(preprocessing_function=keras.applications.resnet50.preprocess_input) train_iter = train_gen.flow(tr_x, tr_y, batch_size=args.batch_size) val_gen = ImageDataGenerator(preprocessing_function=keras.applications.resnet50.preprocess_input) val_iter = val_gen.flow(val_x, val_y,batch_size=args.val_batch_size) model.fit_generator(train_iter, steps_per_epoch=train_batches // hvd.size(), callbacks=callbacks, epochs=args.epochs, verbose=1, workers=4, validation_data=val_iter, validation_steps=3 * val_batches // hvd.size())

Error: ValueError: Error when checking target: expected activation_49 to have 4 dimensions, but got array with shape (32, 2) I am using latest keras and keras-application to train a resnet50 model, but found following error. I have been google for a long time but haven't found a solution.

The related code: input_shape=(64,64,3) model = keras.applications.resnet50.ResNet50(include_top=False, weights=None, input_shape=input_shape, classes=2) train_gen = ImageDataGenerator(preprocessing_function=keras.applications.resnet50.preprocess_input) train_iter = train_gen.flow(tr_x, tr_y, batch_size=args.batch_size) val_gen = ImageDataGenerator(preprocessing_function=keras.applications.resnet50.preprocess_input) val_iter = val_gen.flow(val_x, val_y,batch_size=args.val_batch_size) model.fit_generator(train_iter, steps_per_epoch=train_batches // hvd.size(), callbacks=callbacks, epochs=args.epochs, verbose=1, workers=4, validation_data=val_iter, validation_steps=3 * val_batches // hvd.size())

Error: ValueError: Error when checking target: expected activation_49 to have 4 dimensions, but got array with shape (32, 2)

Do I need to resize the image shape from (64,64) to (224,224)?

Thanks!

[taehoonlee]

@jiangzihanict, If you want to perform a classification task, the include_top should be True.

[jiangzihanict]

@taehoonlee My image size is (64,64). if the include_top is true, the image size only to be (224,224)

[taehoonlee]

@jiangzihanict, It is possible to call model = keras.applications.resnet50.ResNet50(include_top=True, weights=None, input_shape=(64, 64, 3), classes=2). Please try pip install -U keras keras-applications.