Re-train a neural network discarding its memory?

Question

r = s \[Function] i \[Function] ImageResize[i, s];
ab = ColorSeparate[#, "LAB"][[2 ;;]] &;
out = ImageData /@ ab[# // r@{56, 56}] &;
imgs = Import /@ FileNames["*.jpg", dir, 2]; (*dir = somewhere on my disk*)
data = Table[i -> out@i, {i, imgs}];

---

raw = NetModel[
  "Colorful Image Colorization Trained on ImageNet Competition Data",
  "UninitializedEvaluationNet"
];
nex = NetTrain[raw, data, TargetDevice -> "GPU"]

performs training but produces results that fail to colorize no matter what. (All experiences described are same with/without TargetDevice -> "GPU".)

The LAB A and B channels the above produces after training are always empty.

1. Using NetInitialize[raw] before training does not help.
2. NetModel["Colorful Image Colorization Trained on ImageNet Competition \ Data", "Properties"] has no
   1. TrainingNetwork
   2. UninitializedTrainingNetwork

---

raw = NetModel[
  "Colorful Image Colorization Trained on ImageNet Competition Data",
  "EvaluationNet"
];
nex = NetTrain[raw, data, TargetDevice -> "GPU"]

produces different results but it is not clear to me what this does.

1. Does the above continue training an already-trained EvaluationNetwork?
2. Is there a way to discard the memory of a neural network to have it start training from scratch?