Upload model

README.md

@@ -0,0 +1,199 @@
+---
+library_name: transformers
+tags: []
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]

config.json

@@ -0,0 +1,32 @@
+{
+  "architectures": [
+    "RZCompressionModel"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_roberta_zinc_compression_encoder.RZCompressionConfig",
+    "AutoModel": "modeling_roberta_zinc_compression_encoder.RZCompressionModel"
+  },
+  "compression_sizes": [
+    512,
+    256,
+    128,
+    64,
+    32
+  ],
+  "decoder_cosine_weight": 1.0,
+  "decoder_layers": 4,
+  "dropout": 0.1,
+  "encoder_layers": 4,
+  "input_size": 768,
+  "layer_norm_eps": 1e-12,
+  "model_type": "roberta_zinc_compression_encoder",
+  "mse_loss_weight": 0.0,
+  "pearson_loss_weight": 1.0,
+  "topk_values": [
+    10,
+    100,
+    256
+  ],
+  "torch_dtype": "float32",
+  "transformers_version": "4.51.3"
+}

configuration_roberta_zinc_compression_encoder.py

@@ -0,0 +1,49 @@
+from typing import List, Optional
+from transformers import PretrainedConfig
+
+
+class RZCompressionConfig(PretrainedConfig):
+    """
+    Configuration for the roberta_zinc embedding-compression models.
+
+    Args:
+        input_size (int):  Dimension of the input embedding.
+        compression_sizes (List[int]):  One or more output dimensions.
+        encoder_layers (int):  Number of FeedForwardLayers in the encoder path.
+        decoder_layers (int):  Number of FeedForwardLayers in the optional decoder.
+        dropout (float):  Drop-out prob in every layer except the final ones.
+        layer_norm_eps (float | None):  Epsilon for LayerNorm.
+        mse_loss_weight (float): Weight for MSE loss on base-to-compressed similarity matrices
+        pearson_loss_weight (float): Weight for Pearson loss on base-to-compressed similarity matrices
+        topk_values (List[int]): Top-k values for weighting mse/pearson loss 
+        decoder_cosine_weight (float): weight for decoder cosine similarity loss 
+    """
+    model_type = "roberta_zinc_compression_encoder"
+
+    def __init__(
+        self,
+        # ── model params ─────────────────────────────────────────────
+        input_size: int = 768,
+        compression_sizes: List[int] = (32, 64, 128, 256, 512),
+        encoder_layers: int = 2,
+        decoder_layers: int = 2,
+        dropout: float = 0.1,
+        layer_norm_eps: Optional[float] = 1e-12,
+        # ── loss knobs ───────────────────────────────────────────────
+        mse_loss_weight:          float = 0.0,
+        pearson_loss_weight:      float = 0.0,
+        topk_values:              list[int] = (10, 100),
+        decoder_cosine_weight:    float = 0.0,
+        **kwargs,
+    ):
+        self.input_size = input_size
+        self.compression_sizes = list(compression_sizes)
+        self.encoder_layers = encoder_layers
+        self.decoder_layers = decoder_layers
+        self.dropout = dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.mse_loss_weight = mse_loss_weight
+        self.topk_values = topk_values
+        self.pearson_loss_weight = pearson_loss_weight
+        self.decoder_cosine_weight = decoder_cosine_weight
+        super().__init__(**kwargs)

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:faa111596c4d62aba81b98d0f33579d7a886a2b931a72d943b28331f4c42e886
+size 244378384

modeling_roberta_zinc_compression_encoder.py

@@ -0,0 +1,266 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple
+
+from transformers import PreTrainedModel
+from transformers.utils import ModelOutput
+
+from .configuration_roberta_zinc_compression_encoder import RZCompressionConfig
+
+# pairwise cosine  ----------------------------------------------------------------
+def pairwise_cosine(x: torch.Tensor) -> torch.Tensor:
+    x = F.normalize(x, p=2, dim=-1)
+    return x @ x.t()                              # [B, B]
+
+# remove diagonal -----------------------------------------------------------------
+def drop_diag(M: torch.Tensor) -> torch.Tensor:
+    n = M.size(0)
+    return M.masked_select(~torch.eye(n, dtype=torch.bool, device=M.device)).view(n, n - 1)
+
+# pearson row-wise ----------------------------------------------------------------
+def rowwise_pearson(ref: torch.Tensor, comp: torch.Tensor, rm_diag: bool=True) -> torch.Tensor:
+    if rm_diag:
+        ref   = drop_diag(ref)
+        comp  = drop_diag(comp)
+    ref_z = F.normalize(ref  - ref.mean(dim=1, keepdim=True), p=2, dim=1)
+    cmp_z = F.normalize(comp - comp.mean(dim=1, keepdim=True), p=2, dim=1)
+    return 1 - (ref_z * cmp_z).sum(dim=1).mean()   #  0 = perfect corr
+
+# aggregate loss ------------------------------------------------------------------
+def compute_losses(
+    embedding: torch.Tensor, # (batch_size, d)
+    compressed: Dict[int, torch.Tensor], # Dict[size, (batch_size, size)]
+    recon_stack: torch.Tensor | None, # (batch_size, n_heads, d)
+    cfg,
+) -> tuple[torch.Tensor, dict[str, float]]:
+    """Return (total_loss, terms_dict)"""
+    device = embedding.device
+    loss_total = torch.zeros((), device=device)
+    terms: dict[str, float] = {}
+
+    # ---- base similarities (detach to save mem) ---------------------------
+    with torch.no_grad():
+        base_sims = pairwise_cosine(embedding)
+        ranks     = base_sims.argsort(-1, descending=True)
+
+    # ======================================================================
+    # 1) encoder / compressed losses
+    # ======================================================================
+    for size, z in compressed.items():
+        tag = f"cmp{size}"
+        comp_sims = pairwise_cosine(z)
+
+        # plain MSE --------------------------------------------------------
+        if cfg.mse_loss_weight:
+            mse = F.mse_loss(drop_diag(base_sims), drop_diag(comp_sims))
+            loss_total += cfg.mse_loss_weight * mse
+            terms[f"{tag}_mse"] = mse.detach()
+
+        # top-k MSE --------------------------------------------------------
+        if cfg.mse_loss_weight and cfg.topk_values:
+            tk_vals = []
+            for k in cfg.topk_values:
+                idx = ranks[:, 1 : k + 1]
+                ref_k = torch.gather(base_sims, 1, idx)
+                cmp_k = torch.gather(comp_sims, 1, idx)
+                tk_mse = F.mse_loss(ref_k, cmp_k)
+                tk_vals.append(tk_mse)
+                terms[f"{tag}_top{k}"] = tk_mse.detach()
+            tk_agg = torch.stack(tk_vals).mean()
+            loss_total += cfg.topk_mse_loss_weight * tk_agg
+            terms[f"{tag}_topk_mean"] = tk_agg.detach()
+
+        # Pearson ----------------------------------------------------------
+        if cfg.pearson_loss_weight:
+            pr = rowwise_pearson(base_sims, comp_sims)
+            loss_total += cfg.pearson_loss_weight * pr
+            terms[f"{tag}_pearson"] = pr.detach()
+
+        if cfg.pearson_loss_weight and cfg.topk_values:
+            pr_vals = []
+            for k in cfg.topk_values:
+                idx = ranks[:, 1 : k + 1]
+                ref_k = torch.gather(base_sims, 1, idx)
+                cmp_k = torch.gather(comp_sims, 1, idx)
+                pr = rowwise_pearson(ref_k, cmp_k, rm_diag=False)
+                pr_vals.append(pr)
+                terms[f"{tag}_pearson_top{k}"] = pr.detach()
+            pr_agg = torch.stack(pr_vals).sum()
+            loss_total += cfg.pearson_loss_weight * pr_agg
+
+    # ======================================================================
+    # 2) decoder losses
+    # ======================================================================
+    if recon_stack is not None:
+        # cosine -----------------------------------------------------------
+        if cfg.decoder_cosine_weight:
+            cos_loss = 1 - F.cosine_similarity(
+                recon_stack,
+                embedding.unsqueeze(1).expand_as(recon_stack),
+                dim=-1,
+            ).mean()
+            loss_total += cfg.decoder_cosine_weight * cos_loss
+            terms["dec_cosine"] = cos_loss.detach()
+
+    return loss_total, terms
+
+
+# ─── basic blocks ───────────────────────────────────────────────
+class FeedForward(nn.Module):
+    def __init__(self, d_in: int, d_out: int):
+        super().__init__()
+        self.fc1 = nn.Linear(d_in, d_out * 2)
+        self.fc2 = nn.Linear(d_out, d_out)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x1, x2 = x.chunk(2, dim=-1)
+        return self.fc2(F.silu(x1) * x2)
+
+class FeedForwardLayer(nn.Module):
+    def __init__(
+        self, d_in: int, d_out: int, dropout: float = 0.1, layer_norm_eps: Optional[float] = 1e-12
+    ):
+        super().__init__()
+        self.ff = FeedForward(d_in, d_out)
+        self.skip = nn.Linear(d_in, d_out) if d_in != d_out else nn.Identity()
+        self.dropout = nn.Dropout(dropout)
+        self.norm = (
+            nn.LayerNorm(d_out, eps=layer_norm_eps) 
+            if layer_norm_eps is not None else nn.Identity()
+        )
+
+    def forward(self, x):
+        y = self.ff(self.dropout(x)) + self.skip(x)
+        return self.norm(y)
+
+
+# ─── pure PyTorch compressor ────────────────────────────────────
+class CompressionModel(nn.Module):
+    """
+    Encoder → (optional) Decoder.
+    """
+
+    def __init__(
+        self,
+        d_in: int,
+        d_comp: int,
+        encoder_layers: int,
+        decoder_layers: int,
+        dropout: float,
+        layer_norm_eps: Optional[float],
+    ):
+        super().__init__()
+
+        enc_layers: List[nn.Module] = []
+        for i in range(encoder_layers):
+            last = i == encoder_layers - 1
+            enc_layers.append(
+                FeedForwardLayer(
+                    d_in,
+                    d_comp if last else d_in,
+                    dropout if not last else 0.0,
+                    None if last else layer_norm_eps,
+                )
+            )
+        self.encoder = nn.Sequential(*enc_layers)
+
+        # optional decoder
+        dec_layers: List[nn.Module] = []
+        for i in range(decoder_layers):
+            last = i == decoder_layers - 1
+            d_prev = d_comp if i==0 else d_in
+            dec_layers.append(
+                FeedForwardLayer(
+                    d_prev,
+                    d_in,
+                    dropout if not last else 0.0,
+                    None if last else layer_norm_eps,
+                )
+            )
+        self.decoder = nn.Sequential(*dec_layers) if dec_layers else None
+
+    def forward(self, x) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        z = self.encoder(x)
+        x_recon = self.decoder(z) if self.decoder is not None else None
+        return z, x_recon
+
+
+# ─── HF wrapper ─────────────────────────────────────────────────
+@dataclass
+class RZCompressionOutput(ModelOutput):
+    loss: torch.FloatTensor
+    loss_terms: Dict[str, torch.Tensor] | None = None
+    compressed: Dict[int, torch.FloatTensor] | None = None
+    reconstructed: torch.FloatTensor | None = None
+
+class RZCompressionModel(PreTrainedModel):
+    config_class = RZCompressionConfig
+
+    def __init__(self, config: RZCompressionConfig):
+        super().__init__(config)
+
+        self.compressors = nn.ModuleDict(
+            {
+                str(size): CompressionModel(
+                    d_in=config.input_size,
+                    d_comp=size,
+                    encoder_layers=config.encoder_layers,
+                    decoder_layers=config.decoder_layers,
+                    dropout=config.dropout,
+                    layer_norm_eps=config.layer_norm_eps,
+                )
+                for size in config.compression_sizes
+            }
+        )
+
+        self.post_init()
+
+    def get_encoders(self, unpack_single=False):
+        encoders = {}
+        for k,v in self.compressors.items():
+            v = v.encoder
+            if len(v)==1 and unpack_single:
+                # unpack from nn.Sequential if only a single layer
+                v = v[0]
+            encoders[k] = v
+        encoders = nn.ModuleDict(encoders)
+        return encoders 
+    
+    def save_encoders(self, path, unpack_single=False):
+        encoders = self.get_encoders(unpack_single)
+        torch.save(encoders.state_dict(), path)
+        
+    def compress(self,
+                 inputs: torch.Tensor,
+                 compression_sizes: List[int]):
+        compressed = {d: self.compressors[str(d)].encoder(inputs) for d in compression_sizes}
+        return compressed
+        
+    def forward(self, embedding, return_dict=True, compute_loss=True):
+        # ---------- forward passes ------------------------------------------------
+        compressed, recons = {}, []
+        for size, module in self.compressors.items():
+            z, rec = module(embedding)
+            compressed[int(size)] = z
+            if rec is not None:
+                recons.append(rec)
+        recon_stack = torch.stack(recons, dim=1) if recons else None
+
+        # ---------- losses --------------------------------------------------------
+        if compute_loss:
+            loss_total, terms = compute_losses(embedding, compressed, recon_stack, self.config)
+        else:
+            loss_total, terms = torch.zeros((), device=embedding.device), {}
+
+        if not return_dict:
+            return compressed, recon_stack, loss_total, terms
+
+        return RZCompressionOutput(
+            loss=loss_total,
+            loss_terms=terms,
+            compressed=compressed,
+            reconstructed=recon_stack,
+        )