Merge pull request #225 from facebookresearch/bugz7

Generative model with start features

light/world/souls/base_soul.py

@@ -190,7 +190,7 @@ def build_context(self, partner_name=None, quest_txt=None):
             txt += quest_txt
         txt += "\n"
         return txt
-
+        
     def build_dialog_context(self, quest_txt=None):
         # Initial context.
         txt = self.build_context(quest_txt)
@@ -212,7 +212,8 @@ def build_dialog_context(self, quest_txt=None):
                 # reset conversation when unsafe utterances are in the history
                 txt = ""
         dtxt = dtxt.lstrip(" ")
-        return txt + dtxt
+        final = txt + dtxt
+        return final
 
     @classmethod
     def load_generic_act_model(cls, generic_act_model_file):

light/world/souls/models/generative_heuristic_model_with_start_feature_soul.py

@@ -0,0 +1,54 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from light.world.souls.models.generative_heuristic_model_soul import GenerativeHeuristicModelSoul
+
+class GenerativeHeuristicModelWithStartFeatureSoul(GenerativeHeuristicModelSoul):
+
+    def add_startswith_tokens(self, context, dialogue_txt):
+        # extract partner name
+        partner_name = ""
+        if self.target_node._last_interaction_partner_id != None:
+            partner = self.world.oo_graph.get_node(
+                self.target_node._last_interaction_partner_id
+            )
+        if partner is not None:
+            partner_name = partner.get_prefix_view()
+        if len(dialogue_txt) < 3:
+            feature = "START " + partner_name
+        else:
+            feature = "CONTINUE " + partner_name
+        final = context + dialogue_txt + '\n' + feature
+        #print(final)
+        return final
+
+    def build_dialog_context(self, quest_txt=None):
+        # Initial context.
+        txt = self.build_context(quest_txt)
+        # Dialogue/interaction context.
+        dtxt = ""
+        agent = self.target_node
+        agent_id = agent.node_id
+        turn_id = None
+        for d in agent._last_interaction_history:
+            current_turn_id = d[0][0]
+            if turn_id == None or turn_id == current_turn_id:
+                dtxt += " " + d[1]
+            else:
+                dtxt = dtxt.lstrip(" ")
+                dtxt += "\n" + d[1]
+            turn_id = current_turn_id
+            is_safe = d[0][2]
+            if not is_safe:
+                # reset conversation when unsafe utterances are in the history
+                dtxt = self.build_context(quest_txt)
+        dtxt = dtxt.lstrip(" ")
+
+        # Add starting context features, can help the model.            
+        final = self.add_startswith_tokens(txt, dtxt)
+
+        return final
+

projects/dialog_and_act_trainer/SETUP.txt

@@ -35,6 +35,13 @@ py convert.py -t fromfile -ffdp  /checkpoint/light/projects/dialog_and_act_train
 
 etc.   (for all the files in  /checkpoint/light/projects/dialog_and_act_trainer/raw/ .. )
 
+#OR
+#
+# we make that final data we want with convo starter tokens (to try to preent starting convo as if in middle ):
+
+py convert_withstarter.py -t fromfile -ffdp  /checkpoint/light/projects/dialog_and_act_trainer/raw/light_train.txt  -of  /checkpoint/light/projects/dialog_and_act_trainer/light_withstarter_train.txt
+
+
 AND
 
 # for the type picker (dialog, act, emote) task:

projects/dialog_and_act_trainer/convert.py

@@ -29,6 +29,7 @@ def clean(msg):
     txt = msg["text"]
     res = []
     app = ""
+
     for t in txt.split("\n"):
         if t.startswith("_") and "_object_desc" not in t:
             if (

projects/dialog_and_act_trainer/convert_withstarter.py

@@ -0,0 +1,161 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+"""
+Convert a dataset into the ParlAI text format.
+
+## Examples
+
+```shell
+parlai convert_data_to_parlai_format -t babi:task1k:1 --outfile /tmp/dump
+```
+"""
+
+from parlai.core.params import ParlaiParser
+from parlai.agents.repeat_label.repeat_label import RepeatLabelAgent
+from parlai.core.worlds import create_task
+from parlai.utils.misc import msg_to_str, TimeLogger
+import parlai.utils.logging as logging
+from parlai.core.script import ParlaiScript, register_script
+import random
+import tempfile
+import copy
+
+
+def clean(msg):
+    msg = copy.deepcopy(msg)
+    txt = msg["text"]
+    res = []
+    app = ""
+    convo_has_started = False
+    for t in txt.split("\n"):
+        if "_self_say " in t:
+            convo_has_started = True
+        if "_partner_say " in t:
+            convo_has_started = True
+        if "_partner_name " in t:
+            partner_name = t.replace("_partner_name ", "")
+
+    first_convo_line = True
+    for t in txt.split("\n"):
+        if t.startswith("_") and "_object_desc" not in t:
+            if (
+                t.startswith("_self_act")
+                or t.startswith("_self_emote")
+                or t.startswith("_partner_act")
+                or t.startswith("_partner_emote")
+            ):
+                t = t.replace("_self_act ", "")
+                t = t.replace("_partner_act ", "")
+                t = t.replace("_self_emote ", "")
+                t = t.replace("_partner_emote ", "")
+                app = app + " *" + t + "*"
+            else:
+                #if ("_partner_say " in t) or ("_self_say" in t):
+                #    #import pdb; pdb.set_trace()
+                if ("_partner_say " in t) or ("_self_say" in t):
+                    first_convo_line = False
+                t = t.replace("_self_say ", "")
+                t = t.replace("_partner_say ", "")
+                res.append(t + app)
+                app = ""
+
+    if not convo_has_started:
+        res.append("START " + partner_name)
+    else:
+        res.append("CONTINUE " + partner_name)
+
+    msg.force_set("text", "\n".join(res))
+    #print(res)
+    return msg
+
+
+def dump_data(opt):
+    # create repeat label agent and assign it to the specified task
+    agent = RepeatLabelAgent(opt)
+    world = create_task(opt, agent)
+    opt.log()
+    ignorefields = opt.get("ignore_fields", "")
+    if opt["outfile"] is None:
+        outfile = tempfile.mkstemp(
+            prefix="{}_{}_".format(opt["task"], opt["datatype"]), suffix=".txt"
+        )[1]
+    else:
+        outfile = opt["outfile"]
+
+    if opt["num_examples"] == -1:
+        num_examples = world.num_examples()
+    else:
+        num_examples = opt["num_examples"]
+    log_timer = TimeLogger()
+
+    logging.debug("starting to convert...")
+    logging.info(f"saving output to {outfile}")
+    fw = open(outfile, "w")
+    for _ in range(num_examples):
+        world.parley()
+        acts = world.get_acts()
+        value = acts[0].get("labels", acts[0].pop("eval_labels", None))
+        acts[0].force_set("labels", value)
+
+        msg = clean(acts[0])
+
+        txt = msg_to_str(msg, ignore_fields=ignorefields)
+        fw.write(txt + "\n")
+        if acts[0].get("episode_done", False):
+            fw.write("\n")
+
+        if log_timer.time() > opt["log_every_n_secs"]:
+            text, _log = log_timer.log(world.total_parleys, world.num_examples())
+            logging.info(text)
+
+        if world.epoch_done():
+            logging.info("epoch done")
+            break
+    fw.close()
+
+
+def setup_args():
+    # Get command line arguments
+    parser = ParlaiParser(description="Dump a task to a standardized format")
+    parser.add_argument(
+        "-n",
+        "--num-examples",
+        default=-1,
+        type=int,
+        help="Total number of exs to convert, -1 to convert all examples",
+    )
+    parser.add_argument(
+        "-of",
+        "--outfile",
+        default=None,
+        type=str,
+        help="Output file where to save, by default will be created in tmp",
+    )
+    parser.add_argument(
+        "-if",
+        "--ignore-fields",
+        default="id",
+        type=str,
+        help="Ignore these fields from the message (returned with .act() )",
+    )
+    parser.add_argument("-ltim", "--log-every-n-secs", type=float, default=2)
+    parser.set_defaults(datatype="train:stream")
+    return parser
+
+
+@register_script("convert_to_parlai", hidden=True)
+class ConvertDataToParlaiFormat(ParlaiScript):
+    @classmethod
+    def setup_args(cls):
+        return setup_args()
+
+    def run(self):
+        return dump_data(self.opt)
+
+
+if __name__ == "__main__":
+    random.seed(42)
+    ConvertDataToParlaiFormat.main()

projects/dialog_and_act_trainer/train/gen_withstarter_sweep.py

@@ -0,0 +1,105 @@
+# Copyright (c) 2017-present, Facebook, Inc.
+# All rights reserved.
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree. An additional grant
+# of patent rights can be found in the PATENTS file in the same directory.
+
+
+from parlai_internal.projects.param_sweep_utils.param_sweep import run_grid
+
+SWEEP_NAME = "gen_withstarter"
+name_keys = {}
+
+grid = {
+    "-t": [
+        '"fromfile:fromfile_datapath=/checkpoint/light/projects/dialog_and_act_trainer/light_withstarter_train.txt,fromfile:fromfile_datapath=/checkpoint/light/projects/dialog_and_act_trainer/light_withstarter_wild_train.txt"'
+    ],
+    "-et": [
+        '"fromfile:fromfile_datapath=/checkpoint/light/projects/dialog_and_act_trainer/light_withstarter_wild_valid.txt"',
+    ],
+    "--max-train-time": [(60 * 60) * 36],
+    "--model": ["parlai_internal.projects.meena.unlikely.boringul:UnlikelihoodAgent"],
+    "--eval-skip-generation": [False],
+    "--validation-max-exs": [200],
+    "-vp": [1000],
+    "-vmt": ["boring_fails"],
+    "-vmm": ["min"],
+    "--save-after-valid": [True],
+    "-vtim": [(60 * 60) * 30],
+    "-stim": [(60) * 60 * 1],
+    "--inference": ["beam"],
+    "--beam_min_length": [20],
+    "--beam_context_block_ngram": [-1],
+    "--beam_block_ngram": [-1],
+    "--beam_size": [1],
+    "--ul-type": ["std0"],
+    "--seq-ul-nc": [4],
+    "--seq-ul-nl": [4],
+    "--seq-ul-ratio": [0.25],
+    "--train-boring-repeats": [True],
+    "--train-context-repeats": [False],
+    "--train-label-repeats": [False],
+    "--attention-dropout": [0.00],
+    "--batchsize": ["64"],
+    "--embedding-size": [2560],
+    "--ffn-size": [10240],
+    "--variant": ["prelayernorm"],
+    "--n-heads": [32],
+    "--n-positions": [128],
+    "--n-encoder-layers": [2],
+    "--n-decoder-layers": [24],
+    "--history-add-global-end-token": [
+        "end",  # hack to get newline delimiter
+    ],
+    "--dict-tokenizer": ["bytelevelbpe"],
+    "--dict-file": [
+        "/checkpoint/parlai/zoo/meena/20200319_meenav0data_tall_2.7B_adamoptimizer/20200319_13.3ppl_200kupdates/model.dict"
+    ],
+    "--dropout": [0.1],
+    "--fp16": [True],
+    "--init-model": [
+        # emily's good good
+        "/checkpoint/parlai/zoo/q_function/generative2.7B_bst_0331/model",
+    ],
+    "--label-truncate": [128],
+    "--lr-scheduler": ["reduceonplateau"],
+    "--lr-scheduler-patience": [3],
+    "--optimizer": ["adam"],
+    "--relu-dropout": [0.0],
+    "--activation": ["gelu"],
+    "--model-parallel": ["true"],
+    "--text-truncate": [128],
+    "--truncate": [128],
+    "--warmup_updates": [200],
+    "--fp16-impl": ["mem_efficient"],
+    "--update-freq": [4],
+    "--gradient-clip": [0.1],
+    "--skip-generation": [True],
+    "--log_every_n_secs": [10],
+    "-lr": [7e-6],
+}
+
+if __name__ == "__main__":
+    run_grid(
+        grid,
+        name_keys,
+        SWEEP_NAME,
+        partition="learnfair",
+        # partition='dev',
+        jobtime="24:00:00",
+        PARLAI_PATH="/private/home/jase/src/ParlAI/",
+        gpus=8,
+        nodes=1,
+        volta=True,
+        volta32=True,
+        saveroot="/checkpoint/light/projects/dialog_and_act_trainer/models/"
+        + SWEEP_NAME,
+        include_job_id=True,
+        create_model_file=True,
+        hashname=False,
+        # fixedname='model',
+        requeue=True,
+        mem_gb=400,
+        data_parallel=True,
+        copy_env=False,
+    )