Tercera Division RFEF Group 17 stats & predictions

Tactical Formations and Strategies

The tactical approaches adopted by each team will significantly impact tomorrow's outcomes. Understanding these strategies offers insights into potential match dynamics:

• Atlético Baleares: Likely to employ an attacking formation such as 4-3-3 or 4-2-3-1, focusing on wing play and quick transitions to exploit gaps in Menorca's defense.
• CD Menorca: Expected to utilize a compact 5-4-1 formation, prioritizing defensive solidity while looking for opportunities on the break through speedy wingers.
• UD Ibiza: With an emphasis on possession-based football, they may opt for a 4-2-3-1 setup that allows them to control midfield battles and dictate play from the backline.
• CF Poblense: A balanced 4-4-2 formation could be their choice, aiming for stability across the pitch while using overlapping full-backs to create width in attack.
• CE Campos: Their preference for an attacking style suggests a 4-3-3 or 4-1-4-1 formation that maximizes offensive output through dynamic midfield movement and fluid front three interplay.
• Portmany Atlètic: A disciplined approach may see them adopt a 5-3-2 setup that focuses on defensive resilience while using pacey forwards to exploit counter-attacking chances.
• UE San Rafael: Likely to stick with their preferred 4-2-3-1 formation that emphasizes strong central midfield presence and quick interchanges between wingers and forwards.
• CD Sant Jordi:A flexible approach might see them switch between formations like 4-5-1 or 4-4-2 depending on game flow and opposition tactics.rebeccagrylls/CS224n<|file_sep|>/README.md # CS224n: Natural Language Processing with Deep Learning ## Lecture Notes The lecture notes were posted online by Stanford University ([link](http://web.stanford.edu/class/cs224n/)). ## Course Project The project was designed as follows: ### Problem Definition In this project you will build an end-to-end question answering system. Your task is not only predict an answer span from within some given context but also predict if there is no answer within the context. You are provided with two datasets: * [SQuAD](https://rajpurkar.github.io/SQuAD-explorer/), which contains questions paired up with paragraphs that provide answers. * [Natural Questions](https://ai.google.com/research/NaturalQuestions/), which contains questions paired up with Wikipedia passages. In addition you are provided with an evaluation script which computes exact match accuracy. For details please refer [here](https://github.com/allenai/bilm-tf/blob/master/run_nq_eval.py). ### Submission You are asked to submit your code (in Python) together with results from your final submission. For this purpose you should run your model on both datasets provided above. We require exact-match accuracy scores for each dataset separately as well as averaged over both datasets. ### Training You can use any training procedure you wish. Please note that we do not allow external data except GloVe vectors. If you train your model on SQuAD or NQ please mention this clearly. ### Requirements * All code must run on CPU only * You may use TensorFlow or PyTorch * You may use GloVe vectors ### Grading Your grade depends on exact-match accuracy scores as well as your report. The report should contain: * The final architecture * The training procedure * Any other relevant details * Discussion of results ### Submission Instructions Please submit your code (in Python) together with results from your final submission. You can do this by uploading it onto our blackboard page. ### Evaluation Criteria Your grade depends on exact-match accuracy scores as well as your report. The report should contain: * The final architecture * The training procedure * Any other relevant details * Discussion of results ## My Notes & Code I have written up my notes & code from this course below. My code is based off of [Dongjun Lee](https://github.com/djlee05/CS224N-Spring2018)'s repo. You can find his original repo [here](https://github.com/djlee05/CS224N-Spring2018). Thanks so much Dongjun! My implementation can answer SQuAD questions without using any pre-trained word vectors. I am using PyTorch because I love it so much :) ## Dependencies `pip install -r requirements.txt` ## Usage To run my code simply do: `python main.py` To run my code with pre-trained word vectors do: `python main.py --pretrained_wv` To run my code using GloVe word vectors do: `python main.py --glove_wv` To run my code using ELMo word vectors do: `python main.py --elmo_wv` To run my code using ELMo word vectors & BERT do: `python main.py --elmo_wv --bert_wv` To run my code using all pre-trained word vectors do: `python main.py --pretrained_wv --glove_wv --elmo_wv --bert_wv` To run my code using only SQuAD do: `python main.py --train_squad_only` To run my code using only Natural Questions do: `python main.py --train_nq_only` To run my code using only test data from SQuAD do: `python main.py --test_squad_only` To run my code using only test data from Natural Questions do: `python main.py --test_nq_only` To run my code & print out intermediate outputs do: `python main.py --debug` If you want more detailed information about what each flag does then please see `main.py`. ## References [Stanford CS224n Lecture Notes](http://web.stanford.edu/class/cs224n/) [Dongjun Lee's CS224n Spring 2018 Repo](https://github.com/djlee05/CS224N-Spring2018) [Rebecca Ryder's CS224n Spring 2018 Repo](https://github.com/rebeccagrylls/CS224n) [AllenNLP WikiQA Project](https://github.com/allenai/wikiqa) [AllenNLP SQuAD Project](https://github.com/allenai/squad) [AllenNLP BiDAF Model](https://github.com/allenai/bidaf) [AllenNLP BiDAF Paper](https://arxiv.org/pdf/1611.01603.pdf) <|file_sep|># Dependencies torch==0.4.0 numpy==1.14.* scikit_learn==0.19.* pyrouge==0.1.* torchtext==0.2.* allennlp==0.* <|file_sep|># -*- coding: utf8 -*- import torch.nn.functional as F class QuestionAnswerer(nn.Module): def __init__(self, vocab_size, hidden_dim, embedding_dim, num_layers, bidirectional, dropout, pretrained_wv=None): super(QuestionAnswerer, self).__init__() self.embedding = nn.Embedding(vocab_size, embedding_dim) if pretrained_wv is not None: self.embedding.weight.data.copy_(torch.from_numpy(pretrained_wv)) self.encoder = nn.LSTM(input_size=embedding_dim, hidden_size=hidden_dim // 2 if bidirectional else hidden_dim, num_layers=num_layers, bidirectional=bidirectional, batch_first=True) self.decoder = nn.LSTM(input_size=embedding_dim + hidden_dim * (num_layers * (2 if bidirectional else 1)), hidden_size=hidden_dim // 2 if bidirectional else hidden_dim, num_layers=num_layers, bidirectional=bidirectional, batch_first=True) self.attention_layer = Attention(hidden_dim * num_layers * (2 if bidirectional else 1)) self.start_logits_layer = nn.Linear(hidden_dim * num_layers * (2 if bidirectional else 1), 1) self.end_logits_layer = nn.Linear(hidden_dim * num_layers * (2 if bidirectional else 1), 1) self.dropout = nn.Dropout(dropout) def forward(self, context_idxs, question_idxs): """ Args: context_idxs: Tensor(bsize x max_context_len) question_idxs: Tensor(bsize x max_question_len) Returns: start_logits: Tensor(bsize x max_context_len) end_logits: Tensor(bsize x max_context_len) attention_weights: Tensor(bsize x max_context_len x max_question_len) """ bsize = context_idxs.size(0) max_context_len = context_idxs.size(1) max_question_len = question_idxs.size(1) context_mask = torch.sign(context_idxs).float() question_mask = torch.sign(question_idxs).float() context_embeds = self.embedding(context_idxs) context_hiddens = self.encoder(context_embeds)[0] context_hiddens = self.dropout(context_hiddens) with torch.no_grad(): flat_context_mask = context_mask.contiguous().view(-1) flat_context_hiddens = context_hiddens.view(-1, context_hiddens.size(-1)) flat_context_hiddens.data.masked_fill_(flat_context_mask.eq(0).unsqueeze(1), -float('inf')) sorted_context_hiddens_data, sorted_context_hiddens_indices = flat_context_hiddens.sort(descending=True) sorted_context_hiddens_inverse_indices = torch.LongTensor( range(sorted_context_hiddens_data.size(0)))[sorted_context_hiddens_indices] sorted_context_hiddens_inverse_indices = Variable(sorted_context_hiddens_inverse_indices.cuda() if CUDA else sorted_context_hiddens_inverse_indices) sorted_unpacked_output_data = pack_padded_sequence(sorted_context_hiddens_data.view(bsize, -1, context_hiddens.size(-1)), flat_context_mask.view(bsize, -1).long().sum(1), batch_first=True)[0] unpacked_output_shape = sorted_unpacked_output_data.size() unpacked_output_data = torch.FloatTensor(unpacked_output_shape).zero_() unpacked_output_data.scatter_(0, sorted_context_hiddens_inverse_indices.unsqueeze(1).unsqueeze(2).expand(unpacked_output_shape), sorted_unpacked_output_data) unpacked_output_data = unpacked_output_data.contiguous().view(context_hiddens.size()) context_hiddens.data.copy_(unpacked_output_data) question_embeds = self.embedding(question_idxs) flat_question_mask = question_mask.contiguous().view(-1) flat_question_embeds = question_embeds.view(-1, question_embeds.size(-1)) flat_question_embeds.data.masked_fill_(flat_question_mask.unsqueeze(1).eq(0), -float('inf')) sorted_question_embeds_data, sorted_question_embeds_indices = flat_question_embeds.sort(descending=True) sorted_question_embeds_inverse_indices = torch.LongTensor( range(sorted_question_embeds_data.size(0)))[sorted_question_embeds_indices] sorted_question_embeds_inverse_indices = Variable(sorted_question_embeds_inverse_indices.cuda() if CUDA else sorted_question_embeds_inverse_indices) sorted_unpacked_input_data = pack_padded_sequence(sorted_question_embeds_data.view(bsize, -1, question_embeds.size(-1)), batch_first=True)[0] unpacked_input_shape = sorted_unpacked_input_data