Add DisCoCirc tutorial on babi task #42
Conversation
dimkart
changed the title
|
@ouissal-moumou Are there two separate notebooks? Can we just merge everything into one? If not, we will need a main index file that joins the two parts. @neiljdo FYI |
|
@dimkart it seems the two notebooks can be merged into one to improve the flow of the tutorial - the second notebook is just the training step + result visualization. Doing so also makes this tutorial similar to the existing ones. Also, the tutorial might benefit from some refactoring - I can work on this incrementally. |
|
@neiljdo Great, and if you think you can merge the two notebooks let's do this. |
There was a problem hiding this comment.
Overall good start on the tutorials, but there's a few places that could do with simplification to make the tutorial clearer. Some general comments to keep in mind:
- try to keep your explanatory sentences shorter. some are very verbose, which can make it harder to follow what you mean.
- in general it is more 'pythonic' to use list/dict comprehensions where you can rather than for loops. (at least in my opinion) this will typically make your code easier to follow too. I've tried to suggest the simplifications in places I think it would help, but feel free to disagree!
| "metadata": {}, | ||
| "source": [ | ||
| "\n", | ||
| "# Tutorial: babI6 Training and Preprocessing in Python\n", |
There was a problem hiding this comment.
Note, the correct capitalisation is bAbI6
| "# Tutorial: babI6 Training and Preprocessing in Python\n", | |
| "# Tutorial: bAbI6 Training and Preprocessing in Python\n", |
| "source": [ | ||
| "Before we delve into the code, we first highlight two new features of the new parser that will be used in this tutorial: the sandwich functor and foliated frames. \n", | ||
| "\n", | ||
| "In the previous versions of the parser, the semantic functor, while providing quantum implementations for boxes, wires and states, did not specify the quantum implementation of frames. The sandwich functor addresses this issue by introducing a novel construction that breaks down a frame into a sequence of boxes with the frame's contents. Now that we have these different frames, we can decide whether we want every layer in these frames to be assigned their operator or have the same operator for all the layers (different parameters assigned to the layers as opposed to having all the layers having the same parameter). For more detail on this, we recommend reading the paper explaining the theory behind the new parser [here](in_the_making)." |
There was a problem hiding this comment.
maybe break this block up and show the decomposition of a simple example sentence, or link to the other discocirc-basics tutorial.
| " List[str]]:\n", | ||
| " \"\"\"\n", | ||
| " reads the .txt file at `path`\n", | ||
| " returns 3 lists of equal length\n", |
There was a problem hiding this comment.
| " returns 3 lists of equal length\n", | |
| " returns 4 lists of equal length\n", |
| " processed_texts_list = []\n", | ||
| " for text in texts:\n", | ||
| " processed_text = \"\"\n", | ||
| " for sentence in text:\n", | ||
| " processed_text += sentence + \". \"\n", | ||
| " \n", | ||
| " processed_texts_list.append(processed_text)\n", | ||
| " \n", |
There was a problem hiding this comment.
| " processed_texts_list = []\n", | |
| " for text in texts:\n", | |
| " processed_text = \"\"\n", | |
| " for sentence in text:\n", | |
| " processed_text += sentence + \". \"\n", | |
| " \n", | |
| " processed_texts_list.append(processed_text)\n", | |
| " \n", | |
| " processed_texts_list = [\n", | |
| " \". \".join(text)\n", | |
| " for text in texts\n", | |
| " ]\n", |
simplify
| " (text, question, answer, text_length)\n", | ||
| " for text, question, answer, text_length in zip(texts, questions, answers, text_length)\n", | ||
| " if len(text) <= TEXT_LENGTH\n", |
There was a problem hiding this comment.
| " (text, question, answer, text_length)\n", | |
| " for text, question, answer, text_length in zip(texts, questions, answers, text_length)\n", | |
| " if len(text) <= TEXT_LENGTH\n", | |
| " (text, question, answer, text_length)\n", | |
| " for text, question, answer, text_length in zip(texts, questions, answers, text_length)\n", | |
| " if len(text) <= TEXT_LENGTH\n", |
| " quest_mid_layer = Id(qubit) if (qid1 == 0 or qid2 == 0) else Discard()\n", | ||
| " \n", | ||
| " \n", | ||
| " for k in range(1, len(text_circuit.cod)):\n", | ||
| " if k == qid1 or k == qid2:\n", | ||
| " quest_mid_layer = quest_mid_layer @ Id(qubit)\n", | ||
| " else:\n", | ||
| " quest_mid_layer = quest_mid_layer @ Discard()\n", |
There was a problem hiding this comment.
To me, this feels clearer, but up to you.
quest_mid_layer = Id().tensor(*[
Discard() if k in [qid1, qid2] else Id(qubit)
for k in range(len(text_circuit.cod))
])| "# Add the 'measure' field to each item\n", | ||
| "for key, value in babi6_datadict.items():\n", | ||
| " temp = -1 if value['answer'] == 0 else 1\n", | ||
| " value['measure'] = temp * value['text_length']\n", | ||
| "\n", | ||
| "# Group items by absolute value of measure\n", | ||
| "abs_value_groups = defaultdict(list)\n", | ||
| "for key, value in babi6_datadict.items():\n", | ||
| " abs_value = abs(value['measure'])\n", | ||
| " abs_value_groups[abs_value].append((key, value))\n", | ||
| "\n", |
There was a problem hiding this comment.
I think its probably clearer to avoid reference to the measure key here and keep it in terms of the text length/ans class explicitly, since abs(measure) == text length and measure > 0 == (ans == 'yes').
You can even avoid translating the answers into numbers as per the cell above and ignore that step.
| "id": "99b9ac0c", | ||
| "metadata": {}, | ||
| "source": [ | ||
| "We also modify the yes and no answers and replace their representations by 1s and 0s, i.e. by \"[1, 0]\"s and \"[0, 1]\"s respectively." |
There was a problem hiding this comment.
| "We also modify the yes and no answers and replace their representations by 1s and 0s, i.e. by \"[1, 0]\"s and \"[0, 1]\"s respectively." | |
| "We also modify the yes and no answers and replace their representations with a one-hot encoding over the possible assertions, i.e. by \"[1, 0]\"s and \"[0, 1]\"s respectively." |
| "metadata": {}, | ||
| "source": [ | ||
| "## 8. Training the Circuits and Tests\n", | ||
| "Now that we have the data ready, we proceed with the training as usual, except that, we have to deal with pairs of circuits instead of single circuits, which will be accommodated by overriding the forward() method in the model." |
There was a problem hiding this comment.
Should maybe also add here that the final output of the model is going to be a vector we can interpret as a probability distribution over the possible answers (in this case [yes, no]). This should give better context to the next cell where you update the answer encodings.
| "metadata": {}, | ||
| "outputs": [], | ||
| "source": [ | ||
| "all_circuits = []\n", |
There was a problem hiding this comment.
simplify
all_circuits = [
circuit
for circuit_tuple in training_circuits + val_circuits + test_circuits
for circuit in circuit_tuple
]
This adds a tutorial showing how to use a DisCoCirc model for solving the babi6 task.
This is the first version of the tutorial. Ignore any errors that might stem from previous unsuccessful runs of the experiments.