BIG Reorganize examples (#4213)

* Created using Colaboratory * [examples] reorganize files * remove run_tpu_glue.py as superseded by TPU support in Trainer * Bugfix: int, not tuple * move files around
2020-05-07 13:48:44 -04:00
parent cafa6a9e29
commit 0ae96ff8a7
65 changed files with 1355 additions and 1308 deletions
--- a/examples/text-generation/README.md
+++ b/examples/text-generation/README.md
@@ -0,0 +1,15 @@
+## Language generation
+
+Based on the script [`run_generation.py`](https://github.com/huggingface/transformers/blob/master/examples/run_generation.py).
+
+Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL, XLNet, CTRL.
+A similar script is used for our official demo [Write With Transfomer](https://transformer.huggingface.co), where you
+can try out the different models available in the library.
+
+Example usage:
+
+```bash
+python run_generation.py \
+    --model_type=gpt2 \
+    --model_name_or_path=gpt2
+```
--- a/examples/text-generation/pplm/README.md
+++ b/examples/text-generation/pplm/README.md
@@ -0,0 +1,54 @@
+# Plug and Play Language Models: a Simple Approach to Controlled Text Generation
+
+Authors: [Sumanth Dathathri](https://dathath.github.io/), [Andrea Madotto](https://andreamad8.github.io/), Janice Lan, Jane Hung, Eric Frank, [Piero Molino](https://w4nderlu.st/), [Jason Yosinski](http://yosinski.com/), and [Rosanne Liu](http://www.rosanneliu.com/)
+
+This folder contains the original code used to run the Plug and Play Language Model (PPLM).
+
+Paper link: https://arxiv.org/abs/1912.02164
+
+Blog link: https://eng.uber.com/pplm
+
+Please check out the repo under uber-research for more information: https://github.com/uber-research/PPLM
+
+
+## Setup
+
+```bash
+git clone https://github.com/huggingface/transformers && cd transformers
+pip install .
+pip install nltk torchtext # additional requirements.
+cd examples/pplm
+```
+
+## PPLM-BoW 
+
+### Example command for bag-of-words control
+
+```bash
+python run_pplm.py -B military --cond_text "The potato" --length 50 --gamma 1.5 --num_iterations 3 --num_samples 10 --stepsize 0.03 --window_length 5 --kl_scale 0.01 --gm_scale 0.99 --colorama --sample
+```
+
+### Tuning hyperparameters for bag-of-words control
+
+1. Increase `--stepsize` to intensify topic control, and decrease its value to soften the control. `--stepsize 0` recovers the original uncontrolled GPT-2 model. 
+
+2. If the language being generated is repetitive (For e.g. "science science experiment experiment"), there are several options to consider: </br>
+	a) Reduce the `--stepsize` </br>
+	b) Increase `--kl_scale` (the KL-loss coefficient) or decrease `--gm_scale` (the gm-scaling term) </br>
+	c) Add `--grad-length xx` where xx is an (integer <= length, e.g. `--grad-length 30`).</br>
+
+
+## PPLM-Discrim
+
+### Example command for discriminator based sentiment control
+
+```bash
+python run_pplm.py -D sentiment --class_label 2 --cond_text "My dog died" --length 50 --gamma 1.0 --num_iterations 10 --num_samples 10 --stepsize 0.04 --kl_scale 0.01 --gm_scale 0.95 --sample
+```
+
+### Tuning hyperparameters for discriminator control
+
+1. Increase `--stepsize` to intensify topic control, and decrease its value to soften the control. `--stepsize 0` recovers the original uncontrolled GPT-2 model. 
+
+2. Use `--class_label 3` for negative, and `--class_label 2` for positive
+
--- a/examples/text-generation/pplm/imgs/headfigure.png
+++ b/examples/text-generation/pplm/imgs/headfigure.png
--- a/examples/text-generation/pplm/imgs/wooly.png
+++ b/examples/text-generation/pplm/imgs/wooly.png
--- a/examples/text-generation/pplm/pplm_classification_head.py
+++ b/examples/text-generation/pplm/pplm_classification_head.py
@@ -0,0 +1,19 @@
+import torch
+
+
+class ClassificationHead(torch.nn.Module):
+    """Classification Head for  transformer encoders"""
+
+    def __init__(self, class_size, embed_size):
+        super().__init__()
+        self.class_size = class_size
+        self.embed_size = embed_size
+        # self.mlp1 = torch.nn.Linear(embed_size, embed_size)
+        # self.mlp2 = (torch.nn.Linear(embed_size, class_size))
+        self.mlp = torch.nn.Linear(embed_size, class_size)
+
+    def forward(self, hidden_state):
+        # hidden_state = F.relu(self.mlp1(hidden_state))
+        # hidden_state = self.mlp2(hidden_state)
+        logits = self.mlp(hidden_state)
+        return logits
--- a/examples/text-generation/pplm/run_pplm.py
+++ b/examples/text-generation/pplm/run_pplm.py
@@ -0,0 +1,794 @@
+#! /usr/bin/env python3
+# coding=utf-8
+
+# Copyright (c) 2019 Uber Technologies, Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Example command with bag of words:
+python examples/run_pplm.py -B space --cond_text "The president" --length 100 --gamma 1.5 --num_iterations 3 --num_samples 10 --stepsize 0.01 --window_length 5 --kl_scale 0.01 --gm_scale 0.95
+
+Example command with discriminator:
+python examples/run_pplm.py -D sentiment --class_label 3 --cond_text "The lake" --length 10 --gamma 1.0 --num_iterations 30 --num_samples 10 --stepsize 0.01 --kl_scale 0.01 --gm_scale 0.95
+"""
+
+import argparse
+import json
+from operator import add
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch.autograd import Variable
+from tqdm import trange
+
+from pplm_classification_head import ClassificationHead
+from transformers import GPT2Tokenizer
+from transformers.file_utils import cached_path
+from transformers.modeling_gpt2 import GPT2LMHeadModel
+
+
+PPLM_BOW = 1
+PPLM_DISCRIM = 2
+PPLM_BOW_DISCRIM = 3
+SMALL_CONST = 1e-15
+BIG_CONST = 1e10
+
+BAG_OF_WORDS_ARCHIVE_MAP = {
+    "legal": "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/legal.txt",
+    "military": "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/military.txt",
+    "politics": "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/politics.txt",
+    "religion": "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/religion.txt",
+    "science": "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/science.txt",
+    "space": "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/space.txt",
+    "technology": "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/technology.txt",
+}
+
+DISCRIMINATOR_MODELS_PARAMS = {
+    "clickbait": {
+        "url": "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/discriminators/clickbait_classifier_head.pt",
+        "class_size": 2,
+        "embed_size": 1024,
+        "class_vocab": {"non_clickbait": 0, "clickbait": 1},
+        "default_class": 1,
+        "pretrained_model": "gpt2-medium",
+    },
+    "sentiment": {
+        "url": "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/discriminators/SST_classifier_head.pt",
+        "class_size": 5,
+        "embed_size": 1024,
+        "class_vocab": {"very_positive": 2, "very_negative": 3},
+        "default_class": 3,
+        "pretrained_model": "gpt2-medium",
+    },
+}
+
+
+def to_var(x, requires_grad=False, volatile=False, device="cuda"):
+    if torch.cuda.is_available() and device == "cuda":
+        x = x.cuda()
+    elif device != "cuda":
+        x = x.to(device)
+    return Variable(x, requires_grad=requires_grad, volatile=volatile)
+
+
+def top_k_filter(logits, k, probs=False):
+    """
+    Masks everything but the k top entries as -infinity (1e10).
+    Used to mask logits such that e^-infinity -> 0 won't contribute to the
+    sum of the denominator.
+    """
+    if k == 0:
+        return logits
+    else:
+        values = torch.topk(logits, k)[0]
+        batch_mins = values[:, -1].view(-1, 1).expand_as(logits)
+        if probs:
+            return torch.where(logits < batch_mins, torch.ones_like(logits) * 0.0, logits)
+        return torch.where(logits < batch_mins, torch.ones_like(logits) * -BIG_CONST, logits)
+
+
+def perturb_past(
+    past,
+    model,
+    last,
+    unpert_past=None,
+    unpert_logits=None,
+    accumulated_hidden=None,
+    grad_norms=None,
+    stepsize=0.01,
+    one_hot_bows_vectors=None,
+    classifier=None,
+    class_label=None,
+    loss_type=0,
+    num_iterations=3,
+    horizon_length=1,
+    window_length=0,
+    decay=False,
+    gamma=1.5,
+    kl_scale=0.01,
+    device="cuda",
+):
+    # Generate inital perturbed past
+    grad_accumulator = [(np.zeros(p.shape).astype("float32")) for p in past]
+
+    if accumulated_hidden is None:
+        accumulated_hidden = 0
+
+    if decay:
+        decay_mask = torch.arange(0.0, 1.0 + SMALL_CONST, 1.0 / (window_length))[1:]
+    else:
+        decay_mask = 1.0
+
+    # TODO fix this comment (SUMANTH)
+    # Generate a mask is gradient perturbated is based on a past window
+    _, _, _, curr_length, _ = past[0].shape
+
+    if curr_length > window_length and window_length > 0:
+        ones_key_val_shape = tuple(past[0].shape[:-2]) + tuple([window_length]) + tuple(past[0].shape[-1:])
+
+        zeros_key_val_shape = (
+            tuple(past[0].shape[:-2]) + tuple([curr_length - window_length]) + tuple(past[0].shape[-1:])
+        )
+
+        ones_mask = torch.ones(ones_key_val_shape)
+        ones_mask = decay_mask * ones_mask.permute(0, 1, 2, 4, 3)
+        ones_mask = ones_mask.permute(0, 1, 2, 4, 3)
+
+        window_mask = torch.cat((ones_mask, torch.zeros(zeros_key_val_shape)), dim=-2).to(device)
+    else:
+        window_mask = torch.ones_like(past[0]).to(device)
+
+    # accumulate perturbations for num_iterations
+    loss_per_iter = []
+    new_accumulated_hidden = None
+    for i in range(num_iterations):
+        print("Iteration ", i + 1)
+        curr_perturbation = [
+            to_var(torch.from_numpy(p_), requires_grad=True, device=device) for p_ in grad_accumulator
+        ]
+
+        # Compute hidden using perturbed past
+        perturbed_past = list(map(add, past, curr_perturbation))
+        _, _, _, curr_length, _ = curr_perturbation[0].shape
+        all_logits, _, all_hidden = model(last, past=perturbed_past)
+        hidden = all_hidden[-1]
+        new_accumulated_hidden = accumulated_hidden + torch.sum(hidden, dim=1).detach()
+        # TODO: Check the layer-norm consistency of this with trained discriminator (Sumanth)
+        logits = all_logits[:, -1, :]
+        probs = F.softmax(logits, dim=-1)
+
+        loss = 0.0
+        loss_list = []
+        if loss_type == PPLM_BOW or loss_type == PPLM_BOW_DISCRIM:
+            for one_hot_bow in one_hot_bows_vectors:
+                bow_logits = torch.mm(probs, torch.t(one_hot_bow))
+                bow_loss = -torch.log(torch.sum(bow_logits))
+                loss += bow_loss
+                loss_list.append(bow_loss)
+            print(" pplm_bow_loss:", loss.data.cpu().numpy())
+
+        if loss_type == 2 or loss_type == 3:
+            ce_loss = torch.nn.CrossEntropyLoss()
+            # TODO why we need to do this assignment and not just using unpert_past? (Sumanth)
+            curr_unpert_past = unpert_past
+            curr_probs = torch.unsqueeze(probs, dim=1)
+            wte = model.resize_token_embeddings()
+            for _ in range(horizon_length):
+                inputs_embeds = torch.matmul(curr_probs, wte.weight.data)
+                _, curr_unpert_past, curr_all_hidden = model(past=curr_unpert_past, inputs_embeds=inputs_embeds)
+                curr_hidden = curr_all_hidden[-1]
+                new_accumulated_hidden = new_accumulated_hidden + torch.sum(curr_hidden, dim=1)
+
+            prediction = classifier(new_accumulated_hidden / (curr_length + 1 + horizon_length))
+
+            label = torch.tensor(prediction.shape[0] * [class_label], device=device, dtype=torch.long)
+            discrim_loss = ce_loss(prediction, label)
+            print(" pplm_discrim_loss:", discrim_loss.data.cpu().numpy())
+            loss += discrim_loss
+            loss_list.append(discrim_loss)
+
+        kl_loss = 0.0
+        if kl_scale > 0.0:
+            unpert_probs = F.softmax(unpert_logits[:, -1, :], dim=-1)
+            unpert_probs = unpert_probs + SMALL_CONST * (unpert_probs <= SMALL_CONST).float().to(device).detach()
+            correction = SMALL_CONST * (probs <= SMALL_CONST).float().to(device).detach()
+            corrected_probs = probs + correction.detach()
+            kl_loss = kl_scale * ((corrected_probs * (corrected_probs / unpert_probs).log()).sum())
+            print(" kl_loss", kl_loss.data.cpu().numpy())
+            loss += kl_loss
+
+        loss_per_iter.append(loss.data.cpu().numpy())
+        print(" pplm_loss", (loss - kl_loss).data.cpu().numpy())
+
+        # compute gradients
+        loss.backward()
+
+        # calculate gradient norms
+        if grad_norms is not None and loss_type == PPLM_BOW:
+            grad_norms = [
+                torch.max(grad_norms[index], torch.norm(p_.grad * window_mask))
+                for index, p_ in enumerate(curr_perturbation)
+            ]
+        else:
+            grad_norms = [
+                (torch.norm(p_.grad * window_mask) + SMALL_CONST) for index, p_ in enumerate(curr_perturbation)
+            ]
+
+        # normalize gradients
+        grad = [
+            -stepsize * (p_.grad * window_mask / grad_norms[index] ** gamma).data.cpu().numpy()
+            for index, p_ in enumerate(curr_perturbation)
+        ]
+
+        # accumulate gradient
+        grad_accumulator = list(map(add, grad, grad_accumulator))
+
+        # reset gradients, just to make sure
+        for p_ in curr_perturbation:
+            p_.grad.data.zero_()
+
+        # removing past from the graph
+        new_past = []
+        for p_ in past:
+            new_past.append(p_.detach())
+        past = new_past
+
+    # apply the accumulated perturbations to the past
+    grad_accumulator = [to_var(torch.from_numpy(p_), requires_grad=True, device=device) for p_ in grad_accumulator]
+    pert_past = list(map(add, past, grad_accumulator))
+
+    return pert_past, new_accumulated_hidden, grad_norms, loss_per_iter
+
+
+def get_classifier(
+    name: Optional[str], class_label: Union[str, int], device: str
+) -> Tuple[Optional[ClassificationHead], Optional[int]]:
+    if name is None:
+        return None, None
+
+    params = DISCRIMINATOR_MODELS_PARAMS[name]
+    classifier = ClassificationHead(class_size=params["class_size"], embed_size=params["embed_size"]).to(device)
+    if "url" in params:
+        resolved_archive_file = cached_path(params["url"])
+    elif "path" in params:
+        resolved_archive_file = params["path"]
+    else:
+        raise ValueError("Either url or path have to be specified " "in the discriminator model parameters")
+    classifier.load_state_dict(torch.load(resolved_archive_file, map_location=device))
+    classifier.eval()
+
+    if isinstance(class_label, str):
+        if class_label in params["class_vocab"]:
+            label_id = params["class_vocab"][class_label]
+        else:
+            label_id = params["default_class"]
+            print("class_label {} not in class_vocab".format(class_label))
+            print("available values are: {}".format(params["class_vocab"]))
+            print("using default class {}".format(label_id))
+
+    elif isinstance(class_label, int):
+        if class_label in set(params["class_vocab"].values()):
+            label_id = class_label
+        else:
+            label_id = params["default_class"]
+            print("class_label {} not in class_vocab".format(class_label))
+            print("available values are: {}".format(params["class_vocab"]))
+            print("using default class {}".format(label_id))
+
+    else:
+        label_id = params["default_class"]
+
+    return classifier, label_id
+
+
+def get_bag_of_words_indices(bag_of_words_ids_or_paths: List[str], tokenizer) -> List[List[List[int]]]:
+    bow_indices = []
+    for id_or_path in bag_of_words_ids_or_paths:
+        if id_or_path in BAG_OF_WORDS_ARCHIVE_MAP:
+            filepath = cached_path(BAG_OF_WORDS_ARCHIVE_MAP[id_or_path])
+        else:
+            filepath = id_or_path
+        with open(filepath, "r") as f:
+            words = f.read().strip().split("\n")
+        bow_indices.append([tokenizer.encode(word.strip(), add_prefix_space=True) for word in words])
+    return bow_indices
+
+
+def build_bows_one_hot_vectors(bow_indices, tokenizer, device="cuda"):
+    if bow_indices is None:
+        return None
+
+    one_hot_bows_vectors = []
+    for single_bow in bow_indices:
+        single_bow = list(filter(lambda x: len(x) <= 1, single_bow))
+        single_bow = torch.tensor(single_bow).to(device)
+        num_words = single_bow.shape[0]
+        one_hot_bow = torch.zeros(num_words, tokenizer.vocab_size).to(device)
+        one_hot_bow.scatter_(1, single_bow, 1)
+        one_hot_bows_vectors.append(one_hot_bow)
+    return one_hot_bows_vectors
+
+
+def full_text_generation(
+    model,
+    tokenizer,
+    context=None,
+    num_samples=1,
+    device="cuda",
+    bag_of_words=None,
+    discrim=None,
+    class_label=None,
+    length=100,
+    stepsize=0.02,
+    temperature=1.0,
+    top_k=10,
+    sample=False,
+    num_iterations=3,
+    grad_length=10000,
+    horizon_length=1,
+    window_length=0,
+    decay=False,
+    gamma=1.5,
+    gm_scale=0.9,
+    kl_scale=0.01,
+    repetition_penalty=1.0,
+    **kwargs
+):
+    classifier, class_id = get_classifier(discrim, class_label, device)
+
+    bow_indices = []
+    if bag_of_words:
+        bow_indices = get_bag_of_words_indices(bag_of_words.split(";"), tokenizer)
+
+    if bag_of_words and classifier:
+        print("Both PPLM-BoW and PPLM-Discrim are on. This is not optimized.")
+        loss_type = PPLM_BOW_DISCRIM
+
+    elif bag_of_words:
+        loss_type = PPLM_BOW
+        print("Using PPLM-BoW")
+
+    elif classifier is not None:
+        loss_type = PPLM_DISCRIM
+        print("Using PPLM-Discrim")
+
+    else:
+        raise Exception("Specify either a bag of words or a discriminator")
+
+    unpert_gen_tok_text, _, _ = generate_text_pplm(
+        model=model,
+        tokenizer=tokenizer,
+        context=context,
+        device=device,
+        length=length,
+        sample=sample,
+        perturb=False,
+        repetition_penalty=repetition_penalty,
+    )
+    if device == "cuda":
+        torch.cuda.empty_cache()
+
+    pert_gen_tok_texts = []
+    discrim_losses = []
+    losses_in_time = []
+
+    for i in range(num_samples):
+        pert_gen_tok_text, discrim_loss, loss_in_time = generate_text_pplm(
+            model=model,
+            tokenizer=tokenizer,
+            context=context,
+            device=device,
+            perturb=True,
+            bow_indices=bow_indices,
+            classifier=classifier,
+            class_label=class_id,
+            loss_type=loss_type,
+            length=length,
+            stepsize=stepsize,
+            temperature=temperature,
+            top_k=top_k,
+            sample=sample,
+            num_iterations=num_iterations,
+            grad_length=grad_length,
+            horizon_length=horizon_length,
+            window_length=window_length,
+            decay=decay,
+            gamma=gamma,
+            gm_scale=gm_scale,
+            kl_scale=kl_scale,
+            repetition_penalty=repetition_penalty,
+        )
+        pert_gen_tok_texts.append(pert_gen_tok_text)
+        if classifier is not None:
+            discrim_losses.append(discrim_loss.data.cpu().numpy())
+        losses_in_time.append(loss_in_time)
+
+    if device == "cuda":
+        torch.cuda.empty_cache()
+
+    return unpert_gen_tok_text, pert_gen_tok_texts, discrim_losses, losses_in_time
+
+
+def generate_text_pplm(
+    model,
+    tokenizer,
+    context=None,
+    past=None,
+    device="cuda",
+    perturb=True,
+    bow_indices=None,
+    classifier=None,
+    class_label=None,
+    loss_type=0,
+    length=100,
+    stepsize=0.02,
+    temperature=1.0,
+    top_k=10,
+    sample=False,
+    num_iterations=3,
+    grad_length=10000,
+    horizon_length=1,
+    window_length=0,
+    decay=False,
+    gamma=1.5,
+    gm_scale=0.9,
+    kl_scale=0.01,
+    repetition_penalty=1.0,
+):
+    output_so_far = None
+    if context:
+        context_t = torch.tensor(context, device=device, dtype=torch.long)
+        while len(context_t.shape) < 2:
+            context_t = context_t.unsqueeze(0)
+        output_so_far = context_t
+
+    # collect one hot vectors for bags of words
+    one_hot_bows_vectors = build_bows_one_hot_vectors(bow_indices, tokenizer, device)
+
+    grad_norms = None
+    last = None
+    unpert_discrim_loss = 0
+    loss_in_time = []
+    for i in trange(length, ascii=True):
+
+        # Get past/probs for current output, except for last word
+        # Note that GPT takes 2 inputs: past + current_token
+
+        # run model forward to obtain unperturbed
+        if past is None and output_so_far is not None:
+            last = output_so_far[:, -1:]
+            if output_so_far.shape[1] > 1:
+                _, past, _ = model(output_so_far[:, :-1])
+
+        unpert_logits, unpert_past, unpert_all_hidden = model(output_so_far)
+        unpert_last_hidden = unpert_all_hidden[-1]
+
+        # check if we are abowe grad max length
+        if i >= grad_length:
+            current_stepsize = stepsize * 0
+        else:
+            current_stepsize = stepsize
+
+        # modify the past if necessary
+        if not perturb or num_iterations == 0:
+            pert_past = past
+
+        else:
+            accumulated_hidden = unpert_last_hidden[:, :-1, :]
+            accumulated_hidden = torch.sum(accumulated_hidden, dim=1)
+
+            if past is not None:
+                pert_past, _, grad_norms, loss_this_iter = perturb_past(
+                    past,
+                    model,
+                    last,
+                    unpert_past=unpert_past,
+                    unpert_logits=unpert_logits,
+                    accumulated_hidden=accumulated_hidden,
+                    grad_norms=grad_norms,
+                    stepsize=current_stepsize,
+                    one_hot_bows_vectors=one_hot_bows_vectors,
+                    classifier=classifier,
+                    class_label=class_label,
+                    loss_type=loss_type,
+                    num_iterations=num_iterations,
+                    horizon_length=horizon_length,
+                    window_length=window_length,
+                    decay=decay,
+                    gamma=gamma,
+                    kl_scale=kl_scale,
+                    device=device,
+                )
+                loss_in_time.append(loss_this_iter)
+            else:
+                pert_past = past
+
+        pert_logits, past, pert_all_hidden = model(last, past=pert_past)
+        pert_logits = pert_logits[:, -1, :] / temperature  # + SMALL_CONST
+
+        for token_idx in set(output_so_far[0].tolist()):
+            if pert_logits[0, token_idx] < 0:
+                pert_logits[0, token_idx] *= repetition_penalty
+            else:
+                pert_logits[0, token_idx] /= repetition_penalty
+
+        pert_probs = F.softmax(pert_logits, dim=-1)
+
+        if classifier is not None:
+            ce_loss = torch.nn.CrossEntropyLoss()
+            prediction = classifier(torch.mean(unpert_last_hidden, dim=1))
+            label = torch.tensor([class_label], device=device, dtype=torch.long)
+            unpert_discrim_loss = ce_loss(prediction, label)
+            print("unperturbed discrim loss", unpert_discrim_loss.data.cpu().numpy())
+        else:
+            unpert_discrim_loss = 0
+
+        # Fuse the modified model and original model
+        if perturb:
+
+            unpert_probs = F.softmax(unpert_logits[:, -1, :], dim=-1)
+
+            pert_probs = (pert_probs ** gm_scale) * (unpert_probs ** (1 - gm_scale))  # + SMALL_CONST
+            pert_probs = top_k_filter(pert_probs, k=top_k, probs=True)  # + SMALL_CONST
+
+            # rescale
+            if torch.sum(pert_probs) <= 1:
+                pert_probs = pert_probs / torch.sum(pert_probs)
+
+        else:
+            pert_logits = top_k_filter(pert_logits, k=top_k)  # + SMALL_CONST
+            pert_probs = F.softmax(pert_logits, dim=-1)
+
+        # sample or greedy
+        if sample:
+            last = torch.multinomial(pert_probs, num_samples=1)
+
+        else:
+            _, last = torch.topk(pert_probs, k=1, dim=-1)
+
+        # update context/output_so_far appending the new token
+        output_so_far = last if output_so_far is None else torch.cat((output_so_far, last), dim=1)
+
+        print(tokenizer.decode(output_so_far.tolist()[0]))
+
+    return output_so_far, unpert_discrim_loss, loss_in_time
+
+
+def set_generic_model_params(discrim_weights, discrim_meta):
+    if discrim_weights is None:
+        raise ValueError("When using a generic discriminator, " "discrim_weights need to be specified")
+    if discrim_meta is None:
+        raise ValueError("When using a generic discriminator, " "discrim_meta need to be specified")
+
+    with open(discrim_meta, "r") as discrim_meta_file:
+        meta = json.load(discrim_meta_file)
+    meta["path"] = discrim_weights
+    DISCRIMINATOR_MODELS_PARAMS["generic"] = meta
+
+
+def run_pplm_example(
+    pretrained_model="gpt2-medium",
+    cond_text="",
+    uncond=False,
+    num_samples=1,
+    bag_of_words=None,
+    discrim=None,
+    discrim_weights=None,
+    discrim_meta=None,
+    class_label=-1,
+    length=100,
+    stepsize=0.02,
+    temperature=1.0,
+    top_k=10,
+    sample=False,
+    num_iterations=3,
+    grad_length=10000,
+    horizon_length=1,
+    window_length=0,
+    decay=False,
+    gamma=1.5,
+    gm_scale=0.9,
+    kl_scale=0.01,
+    seed=0,
+    no_cuda=False,
+    colorama=False,
+    repetition_penalty=1.0,
+):
+    # set Random seed
+    torch.manual_seed(seed)
+    np.random.seed(seed)
+
+    # set the device
+    device = "cuda" if torch.cuda.is_available() and not no_cuda else "cpu"
+
+    if discrim == "generic":
+        set_generic_model_params(discrim_weights, discrim_meta)
+
+    if discrim is not None:
+        pretrained_model = DISCRIMINATOR_MODELS_PARAMS[discrim]["pretrained_model"]
+        print("discrim = {}, pretrained_model set " "to discriminator's = {}".format(discrim, pretrained_model))
+
+    # load pretrained model
+    model = GPT2LMHeadModel.from_pretrained(pretrained_model, output_hidden_states=True)
+    model.to(device)
+    model.eval()
+
+    # load tokenizer
+    tokenizer = GPT2Tokenizer.from_pretrained(pretrained_model)
+
+    # Freeze GPT-2 weights
+    for param in model.parameters():
+        param.requires_grad = False
+
+    # figure out conditioning text
+    if uncond:
+        tokenized_cond_text = tokenizer.encode([tokenizer.bos_token])
+    else:
+        raw_text = cond_text
+        while not raw_text:
+            print("Did you forget to add `--cond_text`? ")
+            raw_text = input("Model prompt >>> ")
+        tokenized_cond_text = tokenizer.encode(tokenizer.bos_token + raw_text)
+
+    print("= Prefix of sentence =")
+    print(tokenizer.decode(tokenized_cond_text))
+    print()
+
+    # generate unperturbed and perturbed texts
+
+    # full_text_generation returns:
+    # unpert_gen_tok_text, pert_gen_tok_texts, discrim_losses, losses_in_time
+    unpert_gen_tok_text, pert_gen_tok_texts, _, _ = full_text_generation(
+        model=model,
+        tokenizer=tokenizer,
+        context=tokenized_cond_text,
+        device=device,
+        num_samples=num_samples,
+        bag_of_words=bag_of_words,
+        discrim=discrim,
+        class_label=class_label,
+        length=length,
+        stepsize=stepsize,
+        temperature=temperature,
+        top_k=top_k,
+        sample=sample,
+        num_iterations=num_iterations,
+        grad_length=grad_length,
+        horizon_length=horizon_length,
+        window_length=window_length,
+        decay=decay,
+        gamma=gamma,
+        gm_scale=gm_scale,
+        kl_scale=kl_scale,
+        repetition_penalty=repetition_penalty,
+    )
+
+    # untokenize unperturbed text
+    unpert_gen_text = tokenizer.decode(unpert_gen_tok_text.tolist()[0])
+
+    print("=" * 80)
+    print("= Unperturbed generated text =")
+    print(unpert_gen_text)
+    print()
+
+    generated_texts = []
+
+    bow_word_ids = set()
+    if bag_of_words and colorama:
+        bow_indices = get_bag_of_words_indices(bag_of_words.split(";"), tokenizer)
+        for single_bow_list in bow_indices:
+            # filtering all words in the list composed of more than 1 token
+            filtered = list(filter(lambda x: len(x) <= 1, single_bow_list))
+            # w[0] because we are sure w has only 1 item because previous fitler
+            bow_word_ids.update(w[0] for w in filtered)
+
+    # iterate through the perturbed texts
+    for i, pert_gen_tok_text in enumerate(pert_gen_tok_texts):
+        try:
+            # untokenize unperturbed text
+            if colorama:
+                import colorama
+
+                pert_gen_text = ""
+                for word_id in pert_gen_tok_text.tolist()[0]:
+                    if word_id in bow_word_ids:
+                        pert_gen_text += "{}{}{}".format(
+                            colorama.Fore.RED, tokenizer.decode([word_id]), colorama.Style.RESET_ALL
+                        )
+                    else:
+                        pert_gen_text += tokenizer.decode([word_id])
+            else:
+                pert_gen_text = tokenizer.decode(pert_gen_tok_text.tolist()[0])
+
+            print("= Perturbed generated text {} =".format(i + 1))
+            print(pert_gen_text)
+            print()
+        except Exception as exc:
+            print("Ignoring error while generating perturbed text:", exc)
+
+        # keep the prefix, perturbed seq, original seq for each index
+        generated_texts.append((tokenized_cond_text, pert_gen_tok_text, unpert_gen_tok_text))
+
+    return
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--pretrained_model",
+        "-M",
+        type=str,
+        default="gpt2-medium",
+        help="pretrained model name or path to local checkpoint",
+    )
+    parser.add_argument("--cond_text", type=str, default="The lake", help="Prefix texts to condition on")
+    parser.add_argument("--uncond", action="store_true", help="Generate from end-of-text as prefix")
+    parser.add_argument(
+        "--num_samples", type=int, default=1, help="Number of samples to generate from the modified latents",
+    )
+    parser.add_argument(
+        "--bag_of_words",
+        "-B",
+        type=str,
+        default=None,
+        help="Bags of words used for PPLM-BoW. "
+        "Either a BOW id (see list in code) or a filepath. "
+        "Multiple BoWs separated by ;",
+    )
+    parser.add_argument(
+        "--discrim",
+        "-D",
+        type=str,
+        default=None,
+        choices=("clickbait", "sentiment", "toxicity", "generic"),
+        help="Discriminator to use",
+    )
+    parser.add_argument("--discrim_weights", type=str, default=None, help="Weights for the generic discriminator")
+    parser.add_argument(
+        "--discrim_meta", type=str, default=None, help="Meta information for the generic discriminator"
+    )
+    parser.add_argument(
+        "--class_label", type=int, default=-1, help="Class label used for the discriminator",
+    )
+    parser.add_argument("--length", type=int, default=100)
+    parser.add_argument("--stepsize", type=float, default=0.02)
+    parser.add_argument("--temperature", type=float, default=1.0)
+    parser.add_argument("--top_k", type=int, default=10)
+    parser.add_argument("--sample", action="store_true", help="Generate from end-of-text as prefix")
+    parser.add_argument("--num_iterations", type=int, default=3)
+    parser.add_argument("--grad_length", type=int, default=10000)
+    parser.add_argument(
+        "--window_length",
+        type=int,
+        default=0,
+        help="Length of past which is being optimized; " "0 corresponds to infinite window length",
+    )
+    parser.add_argument(
+        "--horizon_length", type=int, default=1, help="Length of future to optimize over",
+    )
+    parser.add_argument("--decay", action="store_true", help="whether to decay or not")
+    parser.add_argument("--gamma", type=float, default=1.5)
+    parser.add_argument("--gm_scale", type=float, default=0.9)
+    parser.add_argument("--kl_scale", type=float, default=0.01)
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--no_cuda", action="store_true", help="no cuda")
+    parser.add_argument("--colorama", action="store_true", help="colors keywords")
+    parser.add_argument(
+        "--repetition_penalty", type=float, default=1.0, help="Penalize repetition. More than 1.0 -> less repetition",
+    )
+
+    args = parser.parse_args()
+    run_pplm_example(**vars(args))
--- a/examples/text-generation/pplm/run_pplm_discrim_train.py
+++ b/examples/text-generation/pplm/run_pplm_discrim_train.py
@@ -0,0 +1,517 @@
+#! /usr/bin/env python3
+# coding=utf-8
+
+# Copyright (c) 2019 Uber Technologies, Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import csv
+import json
+import math
+import time
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torch.optim as optim
+import torch.utils.data as data
+from nltk.tokenize.treebank import TreebankWordDetokenizer
+from torchtext import data as torchtext_data
+from torchtext import datasets
+from tqdm import tqdm, trange
+
+from pplm_classification_head import ClassificationHead
+from transformers import GPT2LMHeadModel, GPT2Tokenizer
+
+
+torch.manual_seed(0)
+np.random.seed(0)
+EPSILON = 1e-10
+example_sentence = "This is incredible! I love it, this is the best chicken I have ever had."
+max_length_seq = 100
+
+
+class Discriminator(torch.nn.Module):
+    """Transformer encoder followed by a Classification Head"""
+
+    def __init__(self, class_size, pretrained_model="gpt2-medium", cached_mode=False, device="cpu"):
+        super().__init__()
+        self.tokenizer = GPT2Tokenizer.from_pretrained(pretrained_model)
+        self.encoder = GPT2LMHeadModel.from_pretrained(pretrained_model)
+        self.embed_size = self.encoder.transformer.config.hidden_size
+        self.classifier_head = ClassificationHead(class_size=class_size, embed_size=self.embed_size)
+        self.cached_mode = cached_mode
+        self.device = device
+
+    def get_classifier(self):
+        return self.classifier_head
+
+    def train_custom(self):
+        for param in self.encoder.parameters():
+            param.requires_grad = False
+        self.classifier_head.train()
+
+    def avg_representation(self, x):
+        mask = x.ne(0).unsqueeze(2).repeat(1, 1, self.embed_size).float().to(self.device).detach()
+        hidden, _ = self.encoder.transformer(x)
+        masked_hidden = hidden * mask
+        avg_hidden = torch.sum(masked_hidden, dim=1) / (torch.sum(mask, dim=1).detach() + EPSILON)
+        return avg_hidden
+
+    def forward(self, x):
+        if self.cached_mode:
+            avg_hidden = x.to(self.device)
+        else:
+            avg_hidden = self.avg_representation(x.to(self.device))
+
+        logits = self.classifier_head(avg_hidden)
+        probs = F.log_softmax(logits, dim=-1)
+
+        return probs
+
+
+class Dataset(data.Dataset):
+    def __init__(self, X, y):
+        """Reads source and target sequences from txt files."""
+        self.X = X
+        self.y = y
+
+    def __len__(self):
+        return len(self.X)
+
+    def __getitem__(self, index):
+        """Returns one data pair (source and target)."""
+        data = {}
+        data["X"] = self.X[index]
+        data["y"] = self.y[index]
+        return data
+
+
+def collate_fn(data):
+    def pad_sequences(sequences):
+        lengths = [len(seq) for seq in sequences]
+
+        padded_sequences = torch.zeros(len(sequences), max(lengths)).long()  # padding value = 0
+
+        for i, seq in enumerate(sequences):
+            end = lengths[i]
+            padded_sequences[i, :end] = seq[:end]
+
+        return padded_sequences, lengths
+
+    item_info = {}
+    for key in data[0].keys():
+        item_info[key] = [d[key] for d in data]
+
+    x_batch, _ = pad_sequences(item_info["X"])
+    y_batch = torch.tensor(item_info["y"], dtype=torch.long)
+
+    return x_batch, y_batch
+
+
+def cached_collate_fn(data):
+    item_info = {}
+    for key in data[0].keys():
+        item_info[key] = [d[key] for d in data]
+
+    x_batch = torch.cat(item_info["X"], 0)
+    y_batch = torch.tensor(item_info["y"], dtype=torch.long)
+
+    return x_batch, y_batch
+
+
+def train_epoch(data_loader, discriminator, optimizer, epoch=0, log_interval=10, device="cpu"):
+    samples_so_far = 0
+    discriminator.train_custom()
+    for batch_idx, (input_t, target_t) in enumerate(data_loader):
+        input_t, target_t = input_t.to(device), target_t.to(device)
+
+        optimizer.zero_grad()
+
+        output_t = discriminator(input_t)
+        loss = F.nll_loss(output_t, target_t)
+        loss.backward(retain_graph=True)
+        optimizer.step()
+
+        samples_so_far += len(input_t)
+
+        if batch_idx % log_interval == 0:
+            print(
+                "Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}".format(
+                    epoch + 1,
+                    samples_so_far,
+                    len(data_loader.dataset),
+                    100 * samples_so_far / len(data_loader.dataset),
+                    loss.item(),
+                )
+            )
+
+
+def evaluate_performance(data_loader, discriminator, device="cpu"):
+    discriminator.eval()
+    test_loss = 0
+    correct = 0
+    with torch.no_grad():
+        for input_t, target_t in data_loader:
+            input_t, target_t = input_t.to(device), target_t.to(device)
+            output_t = discriminator(input_t)
+            # sum up batch loss
+            test_loss += F.nll_loss(output_t, target_t, reduction="sum").item()
+            # get the index of the max log-probability
+            pred_t = output_t.argmax(dim=1, keepdim=True)
+            correct += pred_t.eq(target_t.view_as(pred_t)).sum().item()
+
+    test_loss /= len(data_loader.dataset)
+
+    print(
+        "Performance on test set: "
+        "Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)".format(
+            test_loss, correct, len(data_loader.dataset), 100.0 * correct / len(data_loader.dataset)
+        )
+    )
+
+
+def predict(input_sentence, model, classes, cached=False, device="cpu"):
+    input_t = model.tokenizer.encode(input_sentence)
+    input_t = torch.tensor([input_t], dtype=torch.long, device=device)
+    if cached:
+        input_t = model.avg_representation(input_t)
+
+    log_probs = model(input_t).data.cpu().numpy().flatten().tolist()
+    print("Input sentence:", input_sentence)
+    print(
+        "Predictions:",
+        ", ".join("{}: {:.4f}".format(c, math.exp(log_prob)) for c, log_prob in zip(classes, log_probs)),
+    )
+
+
+def get_cached_data_loader(dataset, batch_size, discriminator, shuffle=False, device="cpu"):
+    data_loader = torch.utils.data.DataLoader(dataset=dataset, batch_size=batch_size, collate_fn=collate_fn)
+
+    xs = []
+    ys = []
+    for batch_idx, (x, y) in enumerate(tqdm(data_loader, ascii=True)):
+        with torch.no_grad():
+            x = x.to(device)
+            avg_rep = discriminator.avg_representation(x).cpu().detach()
+            avg_rep_list = torch.unbind(avg_rep.unsqueeze(1))
+            xs += avg_rep_list
+            ys += y.cpu().numpy().tolist()
+
+    data_loader = torch.utils.data.DataLoader(
+        dataset=Dataset(xs, ys), batch_size=batch_size, shuffle=shuffle, collate_fn=cached_collate_fn
+    )
+
+    return data_loader
+
+
+def train_discriminator(
+    dataset,
+    dataset_fp=None,
+    pretrained_model="gpt2-medium",
+    epochs=10,
+    batch_size=64,
+    log_interval=10,
+    save_model=False,
+    cached=False,
+    no_cuda=False,
+):
+    device = "cuda" if torch.cuda.is_available() and not no_cuda else "cpu"
+
+    print("Preprocessing {} dataset...".format(dataset))
+    start = time.time()
+
+    if dataset == "SST":
+        idx2class = ["positive", "negative", "very positive", "very negative", "neutral"]
+        class2idx = {c: i for i, c in enumerate(idx2class)}
+
+        discriminator = Discriminator(
+            class_size=len(idx2class), pretrained_model=pretrained_model, cached_mode=cached, device=device
+        ).to(device)
+
+        text = torchtext_data.Field()
+        label = torchtext_data.Field(sequential=False)
+        train_data, val_data, test_data = datasets.SST.splits(text, label, fine_grained=True, train_subtrees=True,)
+
+        x = []
+        y = []
+        for i in trange(len(train_data), ascii=True):
+            seq = TreebankWordDetokenizer().detokenize(vars(train_data[i])["text"])
+            seq = discriminator.tokenizer.encode(seq)
+            seq = torch.tensor([50256] + seq, device=device, dtype=torch.long)
+            x.append(seq)
+            y.append(class2idx[vars(train_data[i])["label"]])
+        train_dataset = Dataset(x, y)
+
+        test_x = []
+        test_y = []
+        for i in trange(len(test_data), ascii=True):
+            seq = TreebankWordDetokenizer().detokenize(vars(test_data[i])["text"])
+            seq = discriminator.tokenizer.encode(seq)
+            seq = torch.tensor([50256] + seq, device=device, dtype=torch.long)
+            test_x.append(seq)
+            test_y.append(class2idx[vars(test_data[i])["label"]])
+        test_dataset = Dataset(test_x, test_y)
+
+        discriminator_meta = {
+            "class_size": len(idx2class),
+            "embed_size": discriminator.embed_size,
+            "pretrained_model": pretrained_model,
+            "class_vocab": class2idx,
+            "default_class": 2,
+        }
+
+    elif dataset == "clickbait":
+        idx2class = ["non_clickbait", "clickbait"]
+        class2idx = {c: i for i, c in enumerate(idx2class)}
+
+        discriminator = Discriminator(
+            class_size=len(idx2class), pretrained_model=pretrained_model, cached_mode=cached, device=device
+        ).to(device)
+
+        with open("datasets/clickbait/clickbait_train_prefix.txt") as f:
+            data = []
+            for i, line in enumerate(f):
+                try:
+                    data.append(eval(line))
+                except Exception:
+                    print("Error evaluating line {}: {}".format(i, line))
+                    continue
+        x = []
+        y = []
+        with open("datasets/clickbait/clickbait_train_prefix.txt") as f:
+            for i, line in enumerate(tqdm(f, ascii=True)):
+                try:
+                    d = eval(line)
+                    seq = discriminator.tokenizer.encode(d["text"])
+
+                    if len(seq) < max_length_seq:
+                        seq = torch.tensor([50256] + seq, device=device, dtype=torch.long)
+                    else:
+                        print("Line {} is longer than maximum length {}".format(i, max_length_seq))
+                        continue
+                    x.append(seq)
+                    y.append(d["label"])
+                except Exception:
+                    print("Error evaluating / tokenizing" " line {}, skipping it".format(i))
+                    pass
+
+        full_dataset = Dataset(x, y)
+        train_size = int(0.9 * len(full_dataset))
+        test_size = len(full_dataset) - train_size
+        train_dataset, test_dataset = torch.utils.data.random_split(full_dataset, [train_size, test_size])
+
+        discriminator_meta = {
+            "class_size": len(idx2class),
+            "embed_size": discriminator.embed_size,
+            "pretrained_model": pretrained_model,
+            "class_vocab": class2idx,
+            "default_class": 1,
+        }
+
+    elif dataset == "toxic":
+        idx2class = ["non_toxic", "toxic"]
+        class2idx = {c: i for i, c in enumerate(idx2class)}
+
+        discriminator = Discriminator(
+            class_size=len(idx2class), pretrained_model=pretrained_model, cached_mode=cached, device=device
+        ).to(device)
+
+        x = []
+        y = []
+        with open("datasets/toxic/toxic_train.txt") as f:
+            for i, line in enumerate(tqdm(f, ascii=True)):
+                try:
+                    d = eval(line)
+                    seq = discriminator.tokenizer.encode(d["text"])
+
+                    if len(seq) < max_length_seq:
+                        seq = torch.tensor([50256] + seq, device=device, dtype=torch.long)
+                    else:
+                        print("Line {} is longer than maximum length {}".format(i, max_length_seq))
+                        continue
+                    x.append(seq)
+                    y.append(int(np.sum(d["label"]) > 0))
+                except Exception:
+                    print("Error evaluating / tokenizing" " line {}, skipping it".format(i))
+                    pass
+
+        full_dataset = Dataset(x, y)
+        train_size = int(0.9 * len(full_dataset))
+        test_size = len(full_dataset) - train_size
+        train_dataset, test_dataset = torch.utils.data.random_split(full_dataset, [train_size, test_size])
+
+        discriminator_meta = {
+            "class_size": len(idx2class),
+            "embed_size": discriminator.embed_size,
+            "pretrained_model": pretrained_model,
+            "class_vocab": class2idx,
+            "default_class": 0,
+        }
+
+    else:  # if dataset == "generic":
+        # This assumes the input dataset is a TSV with the following structure:
+        # class \t text
+
+        if dataset_fp is None:
+            raise ValueError("When generic dataset is selected, " "dataset_fp needs to be specified aswell.")
+
+        classes = set()
+        with open(dataset_fp) as f:
+            csv_reader = csv.reader(f, delimiter="\t")
+            for row in tqdm(csv_reader, ascii=True):
+                if row:
+                    classes.add(row[0])
+
+        idx2class = sorted(classes)
+        class2idx = {c: i for i, c in enumerate(idx2class)}
+
+        discriminator = Discriminator(
+            class_size=len(idx2class), pretrained_model=pretrained_model, cached_mode=cached, device=device
+        ).to(device)
+
+        x = []
+        y = []
+        with open(dataset_fp) as f:
+            csv_reader = csv.reader(f, delimiter="\t")
+            for i, row in enumerate(tqdm(csv_reader, ascii=True)):
+                if row:
+                    label = row[0]
+                    text = row[1]
+
+                    try:
+                        seq = discriminator.tokenizer.encode(text)
+                        if len(seq) < max_length_seq:
+                            seq = torch.tensor([50256] + seq, device=device, dtype=torch.long)
+
+                        else:
+                            print("Line {} is longer than maximum length {}".format(i, max_length_seq))
+                            continue
+
+                        x.append(seq)
+                        y.append(class2idx[label])
+
+                    except Exception:
+                        print("Error tokenizing line {}, skipping it".format(i))
+                        pass
+
+        full_dataset = Dataset(x, y)
+        train_size = int(0.9 * len(full_dataset))
+        test_size = len(full_dataset) - train_size
+        train_dataset, test_dataset = torch.utils.data.random_split(full_dataset, [train_size, test_size])
+
+        discriminator_meta = {
+            "class_size": len(idx2class),
+            "embed_size": discriminator.embed_size,
+            "pretrained_model": pretrained_model,
+            "class_vocab": class2idx,
+            "default_class": 0,
+        }
+
+    end = time.time()
+    print("Preprocessed {} data points".format(len(train_dataset) + len(test_dataset)))
+    print("Data preprocessing took: {:.3f}s".format(end - start))
+
+    if cached:
+        print("Building representation cache...")
+
+        start = time.time()
+
+        train_loader = get_cached_data_loader(train_dataset, batch_size, discriminator, shuffle=True, device=device)
+
+        test_loader = get_cached_data_loader(test_dataset, batch_size, discriminator, device=device)
+
+        end = time.time()
+        print("Building representation cache took: {:.3f}s".format(end - start))
+
+    else:
+        train_loader = torch.utils.data.DataLoader(
+            dataset=train_dataset, batch_size=batch_size, shuffle=True, collate_fn=collate_fn
+        )
+        test_loader = torch.utils.data.DataLoader(dataset=test_dataset, batch_size=batch_size, collate_fn=collate_fn)
+
+    if save_model:
+        with open("{}_classifier_head_meta.json".format(dataset), "w") as meta_file:
+            json.dump(discriminator_meta, meta_file)
+
+    optimizer = optim.Adam(discriminator.parameters(), lr=0.0001)
+
+    for epoch in range(epochs):
+        start = time.time()
+        print("\nEpoch", epoch + 1)
+
+        train_epoch(
+            discriminator=discriminator,
+            data_loader=train_loader,
+            optimizer=optimizer,
+            epoch=epoch,
+            log_interval=log_interval,
+            device=device,
+        )
+        evaluate_performance(data_loader=test_loader, discriminator=discriminator, device=device)
+
+        end = time.time()
+        print("Epoch took: {:.3f}s".format(end - start))
+
+        print("\nExample prediction")
+        predict(example_sentence, discriminator, idx2class, cached=cached, device=device)
+
+        if save_model:
+            # torch.save(discriminator.state_dict(),
+            #           "{}_discriminator_{}.pt".format(
+            #               args.dataset, epoch + 1
+            #               ))
+            torch.save(
+                discriminator.get_classifier().state_dict(),
+                "{}_classifier_head_epoch_{}.pt".format(dataset, epoch + 1),
+            )
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Train a discriminator on top of GPT-2 representations")
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default="SST",
+        choices=("SST", "clickbait", "toxic", "generic"),
+        help="dataset to train the discriminator on."
+        "In case of generic, the dataset is expected"
+        "to be a TSBV file with structure: class \\t text",
+    )
+    parser.add_argument(
+        "--dataset_fp",
+        type=str,
+        default="",
+        help="File path of the dataset to use. " "Needed only in case of generic datadset",
+    )
+    parser.add_argument(
+        "--pretrained_model", type=str, default="gpt2-medium", help="Pretrained model to use as encoder"
+    )
+    parser.add_argument("--epochs", type=int, default=10, metavar="N", help="Number of training epochs")
+    parser.add_argument(
+        "--batch_size", type=int, default=64, metavar="N", help="input batch size for training (default: 64)"
+    )
+    parser.add_argument(
+        "--log_interval",
+        type=int,
+        default=10,
+        metavar="N",
+        help="how many batches to wait before logging training status",
+    )
+    parser.add_argument("--save_model", action="store_true", help="whether to save the model")
+    parser.add_argument("--cached", action="store_true", help="whether to cache the input representations")
+    parser.add_argument("--no_cuda", action="store_true", help="use to turn off cuda")
+    args = parser.parse_args()
+
+    train_discriminator(**(vars(args)))
--- a/examples/text-generation/run_generation.py
+++ b/examples/text-generation/run_generation.py
@@ -0,0 +1,268 @@
+#!/usr/bin/env python3
+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain and Carnegie Mellon University Authors and the HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Conditional text generation with the auto-regressive models of the library (GPT/GPT-2/CTRL/Transformer-XL/XLNet)
+"""
+
+
+import argparse
+import logging
+
+import numpy as np
+import torch
+
+from transformers import (
+    CTRLLMHeadModel,
+    CTRLTokenizer,
+    GPT2LMHeadModel,
+    GPT2Tokenizer,
+    OpenAIGPTLMHeadModel,
+    OpenAIGPTTokenizer,
+    TransfoXLLMHeadModel,
+    TransfoXLTokenizer,
+    XLMTokenizer,
+    XLMWithLMHeadModel,
+    XLNetLMHeadModel,
+    XLNetTokenizer,
+)
+
+
+logging.basicConfig(
+    format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO,
+)
+logger = logging.getLogger(__name__)
+
+MAX_LENGTH = int(10000)  # Hardcoded max length to avoid infinite loop
+
+MODEL_CLASSES = {
+    "gpt2": (GPT2LMHeadModel, GPT2Tokenizer),
+    "ctrl": (CTRLLMHeadModel, CTRLTokenizer),
+    "openai-gpt": (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
+    "xlnet": (XLNetLMHeadModel, XLNetTokenizer),
+    "transfo-xl": (TransfoXLLMHeadModel, TransfoXLTokenizer),
+    "xlm": (XLMWithLMHeadModel, XLMTokenizer),
+}
+
+# Padding text to help Transformer-XL and XLNet with short prompts as proposed by Aman Rusia
+# in https://github.com/rusiaaman/XLNet-gen#methodology
+# and https://medium.com/@amanrusia/xlnet-speaks-comparison-to-gpt-2-ea1a4e9ba39e
+PADDING_TEXT = """In 1991, the remains of Russian Tsar Nicholas II and his family
+(except for Alexei and Maria) are discovered.
+The voice of Nicholas's young son, Tsarevich Alexei Nikolaevich, narrates the
+remainder of the story. 1883 Western Siberia,
+a young Grigori Rasputin is asked by his father and a group of men to perform magic.
+Rasputin has a vision and denounces one of the men as a horse thief. Although his
+father initially slaps him for making such an accusation, Rasputin watches as the
+man is chased outside and beaten. Twenty years later, Rasputin sees a vision of
+the Virgin Mary, prompting him to become a priest. Rasputin quickly becomes famous,
+with people, even a bishop, begging for his blessing. <eod> </s> <eos>"""
+
+
+def set_seed(args):
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+
+#
+# Functions to prepare models' input
+#
+
+
+def prepare_ctrl_input(args, _, tokenizer, prompt_text):
+    if args.temperature > 0.7:
+        logger.info("CTRL typically works better with lower temperatures (and lower top_k).")
+
+    encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False)
+    if not any(encoded_prompt[0] == x for x in tokenizer.control_codes.values()):
+        logger.info("WARNING! You are not starting your generation from a control code so you won't get good results")
+    return prompt_text
+
+
+def prepare_xlm_input(args, model, tokenizer, prompt_text):
+    # kwargs = {"language": None, "mask_token_id": None}
+
+    # Set the language
+    use_lang_emb = hasattr(model.config, "use_lang_emb") and model.config.use_lang_emb
+    if hasattr(model.config, "lang2id") and use_lang_emb:
+        available_languages = model.config.lang2id.keys()
+        if args.xlm_language in available_languages:
+            language = args.xlm_language
+        else:
+            language = None
+            while language not in available_languages:
+                language = input("Using XLM. Select language in " + str(list(available_languages)) + " >>> ")
+
+        model.config.lang_id = model.config.lang2id[language]
+        # kwargs["language"] = tokenizer.lang2id[language]
+
+    # TODO fix mask_token_id setup when configurations will be synchronized between models and tokenizers
+    # XLM masked-language modeling (MLM) models need masked token
+    # is_xlm_mlm = "mlm" in args.model_name_or_path
+    # if is_xlm_mlm:
+    #     kwargs["mask_token_id"] = tokenizer.mask_token_id
+
+    return prompt_text
+
+
+def prepare_xlnet_input(args, _, tokenizer, prompt_text):
+    prompt_text = (args.padding_text if args.padding_text else PADDING_TEXT) + prompt_text
+    return prompt_text
+
+
+def prepare_transfoxl_input(args, _, tokenizer, prompt_text):
+    prompt_text = (args.padding_text if args.padding_text else PADDING_TEXT) + prompt_text
+    return prompt_text
+
+
+PREPROCESSING_FUNCTIONS = {
+    "ctrl": prepare_ctrl_input,
+    "xlm": prepare_xlm_input,
+    "xlnet": prepare_xlnet_input,
+    "transfo-xl": prepare_transfoxl_input,
+}
+
+
+def adjust_length_to_model(length, max_sequence_length):
+    if length < 0 and max_sequence_length > 0:
+        length = max_sequence_length
+    elif 0 < max_sequence_length < length:
+        length = max_sequence_length  # No generation bigger than model size
+    elif length < 0:
+        length = MAX_LENGTH  # avoid infinite loop
+    return length
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model_type",
+        default=None,
+        type=str,
+        required=True,
+        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()),
+    )
+    parser.add_argument(
+        "--model_name_or_path",
+        default=None,
+        type=str,
+        required=True,
+        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(MODEL_CLASSES.keys()),
+    )
+
+    parser.add_argument("--prompt", type=str, default="")
+    parser.add_argument("--length", type=int, default=20)
+    parser.add_argument("--stop_token", type=str, default=None, help="Token at which text generation is stopped")
+
+    parser.add_argument(
+        "--temperature",
+        type=float,
+        default=1.0,
+        help="temperature of 1.0 has no effect, lower tend toward greedy sampling",
+    )
+    parser.add_argument(
+        "--repetition_penalty", type=float, default=1.0, help="primarily useful for CTRL model; in that case, use 1.2"
+    )
+    parser.add_argument("--k", type=int, default=0)
+    parser.add_argument("--p", type=float, default=0.9)
+
+    parser.add_argument("--padding_text", type=str, default="", help="Padding text for Transfo-XL and XLNet.")
+    parser.add_argument("--xlm_language", type=str, default="", help="Optional language when used with the XLM model.")
+
+    parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
+    parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
+    parser.add_argument("--num_return_sequences", type=int, default=1, help="The number of samples to generate.")
+    args = parser.parse_args()
+
+    args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+    args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
+
+    set_seed(args)
+
+    # Initialize the model and tokenizer
+    try:
+        args.model_type = args.model_type.lower()
+        model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    except KeyError:
+        raise KeyError("the model {} you specified is not supported. You are welcome to add it and open a PR :)")
+
+    tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path)
+    model = model_class.from_pretrained(args.model_name_or_path)
+    model.to(args.device)
+
+    args.length = adjust_length_to_model(args.length, max_sequence_length=model.config.max_position_embeddings)
+    logger.info(args)
+
+    prompt_text = args.prompt if args.prompt else input("Model prompt >>> ")
+
+    # Different models need different input formatting and/or extra arguments
+    requires_preprocessing = args.model_type in PREPROCESSING_FUNCTIONS.keys()
+    if requires_preprocessing:
+        prepare_input = PREPROCESSING_FUNCTIONS.get(args.model_type)
+        preprocessed_prompt_text = prepare_input(args, model, tokenizer, prompt_text)
+        encoded_prompt = tokenizer.encode(
+            preprocessed_prompt_text, add_special_tokens=False, return_tensors="pt", add_space_before_punct_symbol=True
+        )
+    else:
+        encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False, return_tensors="pt")
+    encoded_prompt = encoded_prompt.to(args.device)
+
+    if encoded_prompt.size()[-1] == 0:
+        input_ids = None
+    else:
+        input_ids = encoded_prompt
+
+    output_sequences = model.generate(
+        input_ids=input_ids,
+        max_length=args.length + len(encoded_prompt[0]),
+        temperature=args.temperature,
+        top_k=args.k,
+        top_p=args.p,
+        repetition_penalty=args.repetition_penalty,
+        do_sample=True,
+        num_return_sequences=args.num_return_sequences,
+    )
+
+    # Remove the batch dimension when returning multiple sequences
+    if len(output_sequences.shape) > 2:
+        output_sequences.squeeze_()
+
+    generated_sequences = []
+
+    for generated_sequence_idx, generated_sequence in enumerate(output_sequences):
+        print("=== GENERATED SEQUENCE {} ===".format(generated_sequence_idx + 1))
+        generated_sequence = generated_sequence.tolist()
+
+        # Decode text
+        text = tokenizer.decode(generated_sequence, clean_up_tokenization_spaces=True)
+
+        # Remove all text after the stop token
+        text = text[: text.find(args.stop_token) if args.stop_token else None]
+
+        # Add the prompt at the beginning of the sequence. Remove the excess text that was used for pre-processing
+        total_sequence = (
+            prompt_text + text[len(tokenizer.decode(encoded_prompt[0], clean_up_tokenization_spaces=True)) :]
+        )
+
+        generated_sequences.append(total_sequence)
+        print(total_sequence)
+
+    return generated_sequences
+
+
+if __name__ == "__main__":
+    main()