Challenge: Working with tfrecords

[johnny-mueller]

I have a problem with the input. I have built a model based on the Yolov3 structure that uses tfrecords as input. Currently I have the requirement that I cannot pass a numpy array. Naive I tried to use tfrecords as input but did not get any useful results.

I get the following error message:

ValueError: Graph disconnected: cannot obtain value for tensor Tensor("input_7:0", shape=(None, 26, 26, 256), dtype=float32) at layer "input_7". The following previous layers 
were accessed without issue: []

Here is the corresponding code I used for this. The loading and preprocessing is based on the preprocessing before training the model

train_dataset = dataset.load_fake_dataset()
train_dataset = dataset.load_tfrecord_dataset(dataset_path, class_names_path, image_size)
train_dataset = train_dataset.batch(batch_size)
train_dataset = train_dataset.map(lambda x, y: (dataset.transform_images(x, image_size), dataset.transform_targets(y, anchors, anchor_masks, image_size)))

explainer = tf_explain.core.GradCAM()
grid = explainer.explain((train_dataset, None), model, class_index=20)
explainer.save(grid, ".", "grad_cam.png")

Is there a workaround or approach how I can approach this?

[RaphaelMeudec]

@johnny-mueller tf-explain does not support (yet) tf.data inputs, only numpy array. So a workaround is to use .take on your dataset and pass it to the explainer as numpy arrays. Supporting tf.data.Dataset is on the roadmap, will comment on this issue once it's done!

[johnny-mueller]

@RaphaelMeudec Thanks for the workaround. I just implemented it and it seems to work. Nevertheless I currently have problems with the input layer.

ValueError: Input 0 of layer conv2d is incompatible with the layer: expected ndim=4, found ndim=5. Full shape received: [1, 1, 416, 416, 3]