Not compatible with current pix2pix and its models exported

[skabbit]

If you try your own trained models, you'll get the error:
tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder' with dtype string and shape [1]

Current pix2pix (mode=export) use base64 PNG as an input value. It is easy to fix webcam-pix2pix-tensorflow in predictor.py, patch starting from line 77:

-        img = np.expand_dims(img, 0) if len(img.shape) < 4 else img
-        model_out = self.sess.run( [self.output_op], { self.input_op: img })[0]

+        pil_img = Image.fromarray(img, mode="RGB")
+        buff = BytesIO()
+        pil_img.save(buff, format="PNG")
+        img_base64 = base64.urlsafe_b64encode(buff.getvalue())
+        model_out = self.sess.run( [self.output_op], { "Placeholder:0": [img_base64] })[0]

[skabbit]

Oops. Re-tested this solution, it works somehow another way.
Resulted image is not really connected to the source, no readable shapes or anything.
I've rollback the "export" mode of pix2pix to an older version, and it works well.
Will discover the things further.