updated models

models/blip_override/blip.py

@@ -0,0 +1,240 @@
+'''
+ * Copyright (c) 2022, salesforce.com, inc.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ * For full license text, see LICENSE.txt file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+ * By Junnan Li
+'''
+import warnings
+
+warnings.filterwarnings("ignore")
+
+from .vit import VisionTransformer, interpolate_pos_embed
+from .med import BertModel, BertLMHeadModel
+from transformers import BertTokenizer, BertConfig
+
+import torch
+from torch import nn
+
+import os
+from urllib.parse import urlparse
+from timm.models.hub import download_cached_file
+
+
+class BLIP_Base(nn.Module):
+    def __init__(self,
+                 med_config='models/blip_override/med_config.json',
+                 image_size=224,
+                 vit='base',
+                 vit_grad_ckpt=False,
+                 vit_ckpt_layer=0,
+                 ):
+        """
+        Args:
+            med_config (str): path for the mixture of encoder-decoder model's configuration file
+            image_size (int): input image size
+            vit (str): model size of vision transformer
+        """
+        super().__init__()
+
+        self.visual_encoder, vision_width = create_vit(vit, image_size, vit_grad_ckpt, vit_ckpt_layer)
+        self.tokenizer = init_tokenizer()
+        med_config = BertConfig.from_json_file(med_config)
+        med_config.encoder_width = vision_width
+        self.text_encoder = BertModel(config=med_config, add_pooling_layer=False)
+
+    def forward(self, image, text, attention_mask, mode):
+        assert mode in ['image', 'text', 'multimodal'], "mode parameter must be image, text, or multimodal"
+        if mode == 'image':
+            # return image features
+            image_embeds = self.visual_encoder(image)
+            return image_embeds
+
+        elif mode == 'text':
+            # return text features
+            text_output = self.text_encoder(text, attention_mask=attention_mask, return_dict=True, mode='text')
+            return text_output.last_hidden_state
+
+        elif mode == 'multimodal':
+            # return multimodel features
+            image_embeds = self.visual_encoder(image)
+            image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(image.device)
+
+            text[:, 0] = self.tokenizer.enc_token_id
+            output = self.text_encoder(text,
+                                       attention_mask=attention_mask,
+                                       encoder_hidden_states=image_embeds,
+                                       encoder_attention_mask=image_atts,
+                                       return_dict=True,
+                                       )
+            return output.last_hidden_state
+
+
+class BLIP_Decoder(nn.Module):
+    def __init__(self,
+                 med_config='models/blip_override/med_config.json',
+                 image_size=384,
+                 vit='base',
+                 vit_grad_ckpt=False,
+                 vit_ckpt_layer=0,
+                 prompt='a picture of ',
+                 ):
+        """
+        Args:
+            med_config (str): path for the mixture of encoder-decoder model's configuration file
+            image_size (int): input image size
+            vit (str): model size of vision transformer
+        """
+        super().__init__()
+
+        self.visual_encoder, vision_width = create_vit(vit, image_size, vit_grad_ckpt, vit_ckpt_layer)
+        self.tokenizer = init_tokenizer()
+        med_config = BertConfig.from_json_file(med_config)
+        med_config.encoder_width = vision_width
+        self.text_decoder = BertLMHeadModel(config=med_config)
+
+        self.prompt = prompt
+        self.prompt_length = len(self.tokenizer(self.prompt).input_ids) - 1
+
+    def forward(self, image, caption):
+
+        image_embeds = self.visual_encoder(image)
+        image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(image.device)
+
+        text = self.tokenizer(caption, padding='longest', truncation=True, max_length=40, return_tensors="pt").to(
+            image.device)
+
+        text.input_ids[:, 0] = self.tokenizer.bos_token_id
+
+        decoder_targets = text.input_ids.masked_fill(text.input_ids == self.tokenizer.pad_token_id, -100)
+        decoder_targets[:, :self.prompt_length] = -100
+
+        decoder_output = self.text_decoder(text.input_ids,
+                                           attention_mask=text.attention_mask,
+                                           encoder_hidden_states=image_embeds,
+                                           encoder_attention_mask=image_atts,
+                                           labels=decoder_targets,
+                                           return_dict=True,
+                                           )
+        loss_lm = decoder_output.loss
+
+        return loss_lm
+
+    def generate(self, image, sample=False, num_beams=3, max_length=30, min_length=10, top_p=0.9,
+                 repetition_penalty=1.0):
+        image_embeds = self.visual_encoder(image)
+
+        if not sample:
+            image_embeds = image_embeds.repeat_interleave(num_beams, dim=0)
+
+        image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(image.device)
+        model_kwargs = {"encoder_hidden_states": image_embeds, "encoder_attention_mask": image_atts}
+
+        prompt = [self.prompt] * image.size(0)
+        input_ids = self.tokenizer(prompt, return_tensors="pt").input_ids.to(image.device)
+        input_ids[:, 0] = self.tokenizer.bos_token_id
+        input_ids = input_ids[:, :-1]
+
+        if sample:
+            # nucleus sampling
+            outputs = self.text_decoder.generate(input_ids=input_ids,
+                                                 max_length=max_length,
+                                                 min_length=min_length,
+                                                 do_sample=True,
+                                                 top_p=top_p,
+                                                 num_return_sequences=1,
+                                                 eos_token_id=self.tokenizer.sep_token_id,
+                                                 pad_token_id=self.tokenizer.pad_token_id,
+                                                 repetition_penalty=1.1,
+                                                 **model_kwargs)
+        else:
+            # beam search
+            outputs = self.text_decoder.generate(input_ids=input_ids,
+                                                 max_length=max_length,
+                                                 min_length=min_length,
+                                                 num_beams=num_beams,
+                                                 eos_token_id=self.tokenizer.sep_token_id,
+                                                 pad_token_id=self.tokenizer.pad_token_id,
+                                                 repetition_penalty=repetition_penalty,
+                                                 **model_kwargs)
+
+        captions = []
+        for output in outputs:
+            caption = self.tokenizer.decode(output, skip_special_tokens=True)
+            captions.append(caption[len(self.prompt):])
+        return captions
+
+
+def blip_decoder(pretrained='', **kwargs):
+    model = BLIP_Decoder(**kwargs)
+    if pretrained:
+        model, msg = load_checkpoint(model, pretrained)
+        assert (len(msg.missing_keys) == 0)
+    return model
+
+
+def blip_feature_extractor(pretrained='', **kwargs):
+    model = BLIP_Base(**kwargs)
+    if pretrained:
+        model, msg = load_checkpoint(model, pretrained)
+        assert (len(msg.missing_keys) == 0)
+    return model
+
+
+def init_tokenizer():
+    tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+    tokenizer.add_special_tokens({'bos_token': '[DEC]'})
+    tokenizer.add_special_tokens({'additional_special_tokens': ['[ENC]']})
+    tokenizer.enc_token_id = tokenizer.additional_special_tokens_ids[0]
+    return tokenizer
+
+
+def create_vit(vit, image_size, use_grad_checkpointing=False, ckpt_layer=0, drop_path_rate=0):
+    assert vit in ['base', 'large'], "vit parameter must be base or large"
+    assert use_grad_checkpointing is False, 'grad checkpointing is not supported yet'
+    if vit == 'base':
+        vision_width = 768
+        visual_encoder = VisionTransformer(img_size=image_size, patch_size=16, embed_dim=vision_width, depth=12,
+                                           num_heads=12, use_grad_checkpointing=use_grad_checkpointing,
+                                           ckpt_layer=ckpt_layer,
+                                           drop_path_rate=0 or drop_path_rate
+                                           )
+    elif vit == 'large':
+        vision_width = 1024
+        visual_encoder = VisionTransformer(img_size=image_size, patch_size=16, embed_dim=vision_width, depth=24,
+                                           num_heads=16, use_grad_checkpointing=use_grad_checkpointing,
+                                           ckpt_layer=ckpt_layer,
+                                           drop_path_rate=0.1 or drop_path_rate
+                                           )
+    return visual_encoder, vision_width
+
+
+def is_url(url_or_filename):
+    parsed = urlparse(url_or_filename)
+    return parsed.scheme in ("http", "https")
+
+
+def load_checkpoint(model, url_or_filename):
+    if is_url(url_or_filename):
+        cached_file = download_cached_file(url_or_filename, check_hash=False, progress=True)
+        checkpoint = torch.load(cached_file, map_location='cpu')
+    elif os.path.isfile(url_or_filename):
+        checkpoint = torch.load(url_or_filename, map_location='cpu')
+    else:
+        raise RuntimeError('checkpoint url or path is invalid')
+
+    state_dict = checkpoint['model']
+
+    state_dict['visual_encoder.pos_embed'] = interpolate_pos_embed(state_dict['visual_encoder.pos_embed'],
+                                                                   model.visual_encoder)
+    if 'visual_encoder_m.pos_embed' in model.state_dict().keys():
+        state_dict['visual_encoder_m.pos_embed'] = interpolate_pos_embed(state_dict['visual_encoder_m.pos_embed'],
+                                                                         model.visual_encoder_m)
+    for key in model.state_dict().keys():
+        if key in state_dict.keys():
+            if state_dict[key].shape != model.state_dict()[key].shape:
+                del state_dict[key]
+
+    msg = model.load_state_dict(state_dict, strict=False)
+    print('load checkpoint from %s' % url_or_filename)
+    return model, msg

models/blip_override/med.py

@@ -0,0 +1,955 @@
+'''
+ * Copyright (c) 2022, salesforce.com, inc.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ * For full license text, see LICENSE.txt file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+ * By Junnan Li
+ * Based on huggingface code base
+ * https://github.com/huggingface/transformers/blob/v4.15.0/src/transformers/models/bert
+'''
+
+import math
+import os
+import warnings
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import torch
+from torch import Tensor, device, dtype, nn
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+import torch.nn.functional as F
+
+from transformers.activations import ACT2FN
+from transformers.file_utils import (
+    ModelOutput,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    NextSentencePredictorOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import (
+    PreTrainedModel,
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from transformers.utils import logging
+from transformers.models.bert.configuration_bert import BertConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+class BertEmbeddings(nn.Module):
+    """Construct the embeddings from word and position embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        
+        self.config = config
+
+    def forward(
+        self, input_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
+    ):
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+
+        embeddings = inputs_embeds
+
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class BertSelfAttention(nn.Module):
+    def __init__(self, config, is_cross_attention):
+        super().__init__()
+        self.config = config
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" % (config.hidden_size, config.num_attention_heads)
+            )
+        
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        if is_cross_attention:
+            self.key = nn.Linear(config.encoder_width, self.all_head_size)
+            self.value = nn.Linear(config.encoder_width, self.all_head_size)
+        else:
+            self.key = nn.Linear(config.hidden_size, self.all_head_size)
+            self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+        self.save_attention = False   
+            
+    def save_attn_gradients(self, attn_gradients):
+        self.attn_gradients = attn_gradients
+        
+    def get_attn_gradients(self):
+        return self.attn_gradients
+    
+    def save_attention_map(self, attention_map):
+        self.attention_map = attention_map
+        
+    def get_attention_map(self):
+        return self.attention_map
+    
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+        mixed_query_layer = self.query(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
+            position_ids_r = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        
+        if is_cross_attention and self.save_attention:
+            self.save_attention_map(attention_probs)
+            attention_probs.register_hook(self.save_attn_gradients)         
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs_dropped = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs_dropped = attention_probs_dropped * head_mask
+
+        context_layer = torch.matmul(attention_probs_dropped, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        outputs = outputs + (past_key_value,)
+        return outputs
+
+
+class BertSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertAttention(nn.Module):
+    def __init__(self, config, is_cross_attention=False):
+        super().__init__()
+        self.self = BertSelfAttention(config, is_cross_attention)
+        self.output = BertSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+class BertIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+class BertOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertLayer(nn.Module):
+    def __init__(self, config, layer_num):
+        super().__init__()
+        self.config = config
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = BertAttention(config)      
+        self.layer_num = layer_num          
+        if self.config.add_cross_attention:
+            self.crossattention = BertAttention(config, is_cross_attention=self.config.add_cross_attention)
+        self.intermediate = BertIntermediate(config)
+        self.output = BertOutput(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+        mode=None,
+    ):
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+
+        outputs = self_attention_outputs[1:-1]
+        present_key_value = self_attention_outputs[-1]
+
+        if mode=='multimodal':
+            assert encoder_hidden_states is not None, "encoder_hidden_states must be given for cross-attention layers"
+
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                output_attentions=output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights                               
+        layer_output = apply_chunking_to_forward(
+            self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
+        )
+        outputs = (layer_output,) + outputs
+
+        outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+
+class BertEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([BertLayer(config,i) for i in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=False,
+        output_hidden_states=False,
+        return_dict=True,
+        mode='multimodal',
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
+
+        next_decoder_cache = () if use_cache else None
+               
+        for i in range(self.config.num_hidden_layers):
+            layer_module = self.layer[i]
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                if use_cache:
+                    logger.warn(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    mode=mode,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                    mode=mode,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+class BertPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class BertPredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+class BertLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = BertPredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+class BertOnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = BertLMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+class BertPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = BertConfig
+    base_model_prefix = "bert"
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """ Initialize the weights """
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+
+class BertModel(BertPreTrainedModel):
+    """
+    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
+    cross-attention is added between the self-attention layers, following the architecture described in `Attention is
+    all you need <https://arxiv.org/abs/1706.03762>`__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
+    Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
+    argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an
+    input to the forward pass.
+    """
+
+    def __init__(self, config, add_pooling_layer=True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = BertEmbeddings(config)
+        
+        self.encoder = BertEncoder(config)
+
+        self.pooler = BertPooler(config) if add_pooling_layer else None
+
+        self.init_weights()
+ 
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    
+    def get_extended_attention_mask(self, attention_mask: Tensor, input_shape: Tuple[int], device: device, is_decoder: bool) -> Tensor:
+        """
+        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
+
+        Arguments:
+            attention_mask (:obj:`torch.Tensor`):
+                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
+            input_shape (:obj:`Tuple[int]`):
+                The shape of the input to the model.
+            device: (:obj:`torch.device`):
+                The device of the input to the model.
+
+        Returns:
+            :obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`.
+        """
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif attention_mask.dim() == 2:
+            # Provided a padding mask of dimensions [batch_size, seq_length]
+            # - if the model is a decoder, apply a causal mask in addition to the padding mask
+            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            if is_decoder:
+                batch_size, seq_length = input_shape
+
+                seq_ids = torch.arange(seq_length, device=device)
+                causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
+                # in case past_key_values are used we need to add a prefix ones mask to the causal mask
+                # causal and attention masks must have same type with pytorch version < 1.3
+                causal_mask = causal_mask.to(attention_mask.dtype)
+   
+                if causal_mask.shape[1] < attention_mask.shape[1]:
+                    prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]
+                    causal_mask = torch.cat(
+                        [
+                            torch.ones((batch_size, seq_length, prefix_seq_len), device=device, dtype=causal_mask.dtype),
+                            causal_mask,
+                        ],
+                        axis=-1,
+                    )                     
+
+                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
+        else:
+            raise ValueError(
+                "Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(
+                    input_shape, attention_mask.shape
+                )
+            )
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        return extended_attention_mask
+    
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        is_decoder=False,
+        mode='multimodal',
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            batch_size, seq_length = input_shape
+            device = input_ids.device
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size, seq_length = input_shape
+            device = inputs_embeds.device
+        elif encoder_embeds is not None:    
+            input_shape = encoder_embeds.size()[:-1]
+            batch_size, seq_length = input_shape 
+            device = encoder_embeds.device
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds or encoder_embeds")
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
+            
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, 
+                                                                                 device, is_decoder)
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if encoder_hidden_states is not None:
+            if type(encoder_hidden_states) == list:
+                encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[0].size()
+            else:
+                encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            
+            if type(encoder_attention_mask) == list:
+                encoder_extended_attention_mask = [self.invert_attention_mask(mask) for mask in encoder_attention_mask]
+            elif encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+                encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+            else:    
+                encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+        
+        if encoder_embeds is None:
+            embedding_output = self.embeddings(
+                input_ids=input_ids,
+                position_ids=position_ids,
+                inputs_embeds=inputs_embeds,
+                past_key_values_length=past_key_values_length,
+            )
+        else:
+            embedding_output = encoder_embeds
+            
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            mode=mode,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+
+class BertLMHeadModel(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        return_logits=False,            
+        is_decoder=True,
+        reduction='mean',
+        mode='multimodal', 
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
+            ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are
+            ignored (masked), the loss is only computed for the tokens with labels n ``[0, ..., config.vocab_size]``
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        Returns:
+        Example::
+            >>> from transformers import BertTokenizer, BertLMHeadModel, BertConfig
+            >>> import torch
+            >>> tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
+            >>> config = BertConfig.from_pretrained("bert-base-cased")
+            >>> model = BertLMHeadModel.from_pretrained('bert-base-cased', config=config)
+            >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+            >>> outputs = model(**inputs)
+            >>> prediction_logits = outputs.logits
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        if labels is not None:
+            use_cache = False
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            is_decoder=is_decoder,
+            mode=mode,
+        )
+        
+        sequence_output = outputs[0]
+        prediction_scores = self.cls(sequence_output)
+        
+        if return_logits:
+            return prediction_scores[:, :-1, :].contiguous()  
+
+        lm_loss = None
+        if labels is not None:
+            # we are doing next-token prediction; shift prediction scores and input ids by one
+            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
+            labels = labels[:, 1:].contiguous()
+            loss_fct = CrossEntropyLoss(reduction=reduction, label_smoothing=0.1) 
+            lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            if reduction=='none':
+                lm_loss = lm_loss.view(prediction_scores.size(0),-1).sum(1)               
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return ((lm_loss,) + output) if lm_loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=lm_loss,
+            logits=prediction_scores,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, past=None, attention_mask=None, **model_kwargs):
+        input_shape = input_ids.shape
+        # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
+        if attention_mask is None:
+            attention_mask = input_ids.new_ones(input_shape)
+
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            input_ids = input_ids[:, -1:]
+
+        return {
+            "input_ids": input_ids, 
+            "attention_mask": attention_mask, 
+            "past_key_values": past,
+            "encoder_hidden_states": model_kwargs.get("encoder_hidden_states", None),
+            "encoder_attention_mask": model_kwargs.get("encoder_attention_mask", None),
+            "is_decoder": True,
+        }
+
+    def _reorder_cache(self, past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past

models/blip_override/med_config.json

@@ -0,0 +1,21 @@
+{
+  "architectures": [
+    "BertModel"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "type_vocab_size": 2,
+  "vocab_size": 30524,
+  "encoder_width": 768,
+  "add_cross_attention": true   
+}

models/blip_override/vit.py

@@ -0,0 +1,302 @@
+'''
+ * Copyright (c) 2022, salesforce.com, inc.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ * For full license text, see LICENSE.txt file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+ * By Junnan Li
+ * Based on timm code base
+ * https://github.com/rwightman/pytorch-image-models/tree/master/timm
+'''
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+
+from timm.models.vision_transformer import _cfg, PatchEmbed
+from timm.models.registry import register_model
+from timm.models.layers import trunc_normal_, DropPath
+from timm.models.helpers import named_apply, adapt_input_conv
+
+
+class Mlp(nn.Module):
+    """ MLP as used in Vision Transformer, MLP-Mixer and related networks
+    """
+
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
+        self.scale = qk_scale or head_dim ** -0.5
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.attn_gradients = None
+        self.attention_map = None
+
+    def save_attn_gradients(self, attn_gradients):
+        self.attn_gradients = attn_gradients
+
+    def get_attn_gradients(self):
+        return self.attn_gradients
+
+    def save_attention_map(self, attention_map):
+        self.attention_map = attention_map
+
+    def get_attention_map(self):
+        return self.attention_map
+
+    def forward(self, x, register_hook=False):
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        if register_hook:
+            self.save_attention_map(attn)
+            attn.register_hook(self.save_attn_gradients)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, use_grad_checkpointing=False):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x, register_hook=False):
+        x = x + self.drop_path(self.attn(self.norm1(x), register_hook=register_hook))
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+class VisionTransformer(nn.Module):
+    """ Vision Transformer
+    A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale`  -
+        https://arxiv.org/abs/2010.11929
+    """
+
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,
+                 num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None, representation_size=None,
+                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=None,
+                 use_grad_checkpointing=False, ckpt_layer=0):
+        """
+        Args:
+            img_size (int, tuple): input image size
+            patch_size (int, tuple): patch size
+            in_chans (int): number of input channels
+            num_classes (int): number of classes for classification head
+            embed_dim (int): embedding dimension
+            depth (int): depth of transformer
+            num_heads (int): number of attention heads
+            mlp_ratio (int): ratio of mlp hidden dim to embedding dim
+            qkv_bias (bool): enable bias for qkv if True
+            qk_scale (float): override default qk scale of head_dim ** -0.5 if set
+            representation_size (Optional[int]): enable and set representation layer (pre-logits) to this value if set
+            drop_rate (float): dropout rate
+            attn_drop_rate (float): attention dropout rate
+            drop_path_rate (float): stochastic depth rate
+            norm_layer: (nn.Module): normalization layer
+        """
+        super().__init__()
+        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
+
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
+
+        num_patches = self.patch_embed.num_patches
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+        self.blocks = nn.ModuleList([
+            Block(
+                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
+                use_grad_checkpointing=(use_grad_checkpointing and i >= depth - ckpt_layer)
+            )
+            for i in range(depth)])
+        self.norm = norm_layer(embed_dim)
+
+        trunc_normal_(self.pos_embed, std=.02)
+        trunc_normal_(self.cls_token, std=.02)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed', 'cls_token'}
+
+    def forward(self, x, register_blk=-1):
+        B = x.shape[0]
+        x = self.patch_embed(x)
+
+        cls_tokens = self.cls_token.expand(B, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
+        x = torch.cat((cls_tokens, x), dim=1)
+
+        x = x + self.pos_embed[:, :x.size(1), :]
+        x = self.pos_drop(x)
+
+        for i, blk in enumerate(self.blocks):
+            x = blk(x, register_blk == i)
+        x = self.norm(x)
+
+        return x
+
+    @torch.jit.ignore()
+    def load_pretrained(self, checkpoint_path, prefix=''):
+        _load_weights(self, checkpoint_path, prefix)
+
+
+@torch.no_grad()
+def _load_weights(model: VisionTransformer, checkpoint_path: str, prefix: str = ''):
+    """ Load weights from .npz checkpoints for official Google Brain Flax implementation
+    """
+    import numpy as np
+
+    def _n2p(w, t=True):
+        if w.ndim == 4 and w.shape[0] == w.shape[1] == w.shape[2] == 1:
+            w = w.flatten()
+        if t:
+            if w.ndim == 4:
+                w = w.transpose([3, 2, 0, 1])
+            elif w.ndim == 3:
+                w = w.transpose([2, 0, 1])
+            elif w.ndim == 2:
+                w = w.transpose([1, 0])
+        return torch.from_numpy(w)
+
+    w = np.load(checkpoint_path)
+    if not prefix and 'opt/target/embedding/kernel' in w:
+        prefix = 'opt/target/'
+
+    if hasattr(model.patch_embed, 'backbone'):
+        # hybrid
+        backbone = model.patch_embed.backbone
+        stem_only = not hasattr(backbone, 'stem')
+        stem = backbone if stem_only else backbone.stem
+        stem.conv.weight.copy_(adapt_input_conv(stem.conv.weight.shape[1], _n2p(w[f'{prefix}conv_root/kernel'])))
+        stem.norm.weight.copy_(_n2p(w[f'{prefix}gn_root/scale']))
+        stem.norm.bias.copy_(_n2p(w[f'{prefix}gn_root/bias']))
+        if not stem_only:
+            for i, stage in enumerate(backbone.stages):
+                for j, block in enumerate(stage.blocks):
+                    bp = f'{prefix}block{i + 1}/unit{j + 1}/'
+                    for r in range(3):
+                        getattr(block, f'conv{r + 1}').weight.copy_(_n2p(w[f'{bp}conv{r + 1}/kernel']))
+                        getattr(block, f'norm{r + 1}').weight.copy_(_n2p(w[f'{bp}gn{r + 1}/scale']))
+                        getattr(block, f'norm{r + 1}').bias.copy_(_n2p(w[f'{bp}gn{r + 1}/bias']))
+                    if block.downsample is not None:
+                        block.downsample.conv.weight.copy_(_n2p(w[f'{bp}conv_proj/kernel']))
+                        block.downsample.norm.weight.copy_(_n2p(w[f'{bp}gn_proj/scale']))
+                        block.downsample.norm.bias.copy_(_n2p(w[f'{bp}gn_proj/bias']))
+        embed_conv_w = _n2p(w[f'{prefix}embedding/kernel'])
+    else:
+        embed_conv_w = adapt_input_conv(
+            model.patch_embed.proj.weight.shape[1], _n2p(w[f'{prefix}embedding/kernel']))
+    model.patch_embed.proj.weight.copy_(embed_conv_w)
+    model.patch_embed.proj.bias.copy_(_n2p(w[f'{prefix}embedding/bias']))
+    model.cls_token.copy_(_n2p(w[f'{prefix}cls'], t=False))
+    pos_embed_w = _n2p(w[f'{prefix}Transformer/posembed_input/pos_embedding'], t=False)
+    if pos_embed_w.shape != model.pos_embed.shape:
+        pos_embed_w = resize_pos_embed(  # resize pos embedding when different size from pretrained weights
+            pos_embed_w, model.pos_embed, getattr(model, 'num_tokens', 1), model.patch_embed.grid_size)
+    model.pos_embed.copy_(pos_embed_w)
+    model.norm.weight.copy_(_n2p(w[f'{prefix}Transformer/encoder_norm/scale']))
+    model.norm.bias.copy_(_n2p(w[f'{prefix}Transformer/encoder_norm/bias']))
+    #     if isinstance(model.head, nn.Linear) and model.head.bias.shape[0] == w[f'{prefix}head/bias'].shape[-1]:
+    #         model.head.weight.copy_(_n2p(w[f'{prefix}head/kernel']))
+    #         model.head.bias.copy_(_n2p(w[f'{prefix}head/bias']))
+    #     if isinstance(getattr(model.pre_logits, 'fc', None), nn.Linear) and f'{prefix}pre_logits/bias' in w:
+    #         model.pre_logits.fc.weight.copy_(_n2p(w[f'{prefix}pre_logits/kernel']))
+    #         model.pre_logits.fc.bias.copy_(_n2p(w[f'{prefix}pre_logits/bias']))
+    for i, block in enumerate(model.blocks.children()):
+        block_prefix = f'{prefix}Transformer/encoderblock_{i}/'
+        mha_prefix = block_prefix + 'MultiHeadDotProductAttention_1/'
+        block.norm1.weight.copy_(_n2p(w[f'{block_prefix}LayerNorm_0/scale']))
+        block.norm1.bias.copy_(_n2p(w[f'{block_prefix}LayerNorm_0/bias']))
+        block.attn.qkv.weight.copy_(torch.cat([
+            _n2p(w[f'{mha_prefix}{n}/kernel'], t=False).flatten(1).T for n in ('query', 'key', 'value')]))
+        block.attn.qkv.bias.copy_(torch.cat([
+            _n2p(w[f'{mha_prefix}{n}/bias'], t=False).reshape(-1) for n in ('query', 'key', 'value')]))
+        block.attn.proj.weight.copy_(_n2p(w[f'{mha_prefix}out/kernel']).flatten(1))
+        block.attn.proj.bias.copy_(_n2p(w[f'{mha_prefix}out/bias']))
+        for r in range(2):
+            getattr(block.mlp, f'fc{r + 1}').weight.copy_(_n2p(w[f'{block_prefix}MlpBlock_3/Dense_{r}/kernel']))
+            getattr(block.mlp, f'fc{r + 1}').bias.copy_(_n2p(w[f'{block_prefix}MlpBlock_3/Dense_{r}/bias']))
+        block.norm2.weight.copy_(_n2p(w[f'{block_prefix}LayerNorm_2/scale']))
+        block.norm2.bias.copy_(_n2p(w[f'{block_prefix}LayerNorm_2/bias']))
+
+
+def interpolate_pos_embed(pos_embed_checkpoint, visual_encoder):
+    # interpolate position embedding
+    embedding_size = pos_embed_checkpoint.shape[-1]
+    num_patches = visual_encoder.patch_embed.num_patches
+    num_extra_tokens = visual_encoder.pos_embed.shape[-2] - num_patches
+    # height (== width) for the checkpoint position embedding
+    orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
+    # height (== width) for the new position embedding
+    new_size = int(num_patches ** 0.5)
+
+    if orig_size != new_size:
+        # class_token and dist_token are kept unchanged
+        extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
+        # only the position tokens are interpolated
+        pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
+        pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
+        pos_tokens = torch.nn.functional.interpolate(
+            pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
+        pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
+        new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
+        print('reshape position embedding from %d to %d' % (orig_size ** 2, new_size ** 2))
+
+        return new_pos_embed
+    else:
+        return pos_embed_checkpoint

models/diffusers_override/attention.py

@@ -0,0 +1,669 @@
+# Copyright 2022 The HuggingFace Team. 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.
+import math
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.modeling_utils import ModelMixin
+from diffusers.models.embeddings import ImagePositionalEmbeddings
+from diffusers.utils import BaseOutput
+from diffusers.utils.import_utils import is_xformers_available
+
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+    """
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            Hidden states conditioned on `encoder_hidden_states` input. If discrete, returns probability distributions
+            for the unnoised latent pixels.
+    """
+
+    sample: torch.FloatTensor
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+class Transformer2DModel(ModelMixin, ConfigMixin):
+    """
+    Transformer model for image-like data. Takes either discrete (classes of vector embeddings) or continuous (actual
+    embeddings) inputs.
+
+    When input is continuous: First, project the input (aka embedding) and reshape to b, t, d. Then apply standard
+    transformer action. Finally, reshape to image.
+
+    When input is discrete: First, input (classes of latent pixels) is converted to embeddings and has positional
+    embeddings applied, see `ImagePositionalEmbeddings`. Then apply standard transformer action. Finally, predict
+    classes of unnoised image.
+
+    Note that it is assumed one of the input classes is the masked latent pixel. The predicted classes of the unnoised
+    image do not contain a prediction for the masked pixel as the unnoised image cannot be masked.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            Pass if the input is continuous. The number of channels in the input and output.
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.1): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of context dimensions to use.
+        sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
+            Note that this is fixed at training time as it is used for learning a number of position embeddings. See
+            `ImagePositionalEmbeddings`.
+        num_vector_embeds (`int`, *optional*):
+            Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
+            The number of diffusion steps used during training. Note that this is fixed at training time as it is used
+            to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
+            up to but not more than steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the TransformerBlocks' attention should contain a bias parameter.
+    """
+
+    @register_to_config
+    def __init__(
+            self,
+            num_attention_heads: int = 16,
+            attention_head_dim: int = 88,
+            in_channels: Optional[int] = None,
+            num_layers: int = 1,
+            dropout: float = 0.0,
+            norm_num_groups: int = 32,
+            cross_attention_dim: Optional[int] = None,
+            attention_bias: bool = False,
+            sample_size: Optional[int] = None,
+            num_vector_embeds: Optional[int] = None,
+            activation_fn: str = "geglu",
+            num_embeds_ada_norm: Optional[int] = None,
+    ):
+        super().__init__()
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # 1. Transformer2DModel can process both standard continous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
+        # Define whether input is continuous or discrete depending on configuration
+        self.is_input_continuous = in_channels is not None
+        self.is_input_vectorized = num_vector_embeds is not None
+
+        if self.is_input_continuous and self.is_input_vectorized:
+            raise ValueError(
+                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is None."
+            )
+        elif not self.is_input_continuous and not self.is_input_vectorized:
+            raise ValueError(
+                f"Has to define either `in_channels`: {in_channels} or `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is not None."
+            )
+
+        # 2. Define input layers
+        if self.is_input_continuous:
+            self.in_channels = in_channels
+
+            self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+            self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        elif self.is_input_vectorized:
+            assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
+            assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"
+
+            self.height = sample_size
+            self.width = sample_size
+            self.num_vector_embeds = num_vector_embeds
+            self.num_latent_pixels = self.height * self.width
+
+            self.latent_image_embedding = ImagePositionalEmbeddings(
+                num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
+            )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        if self.is_input_continuous:
+            self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+        elif self.is_input_vectorized:
+            self.norm_out = nn.LayerNorm(inner_dim)
+            self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
+
+    def _set_attention_slice(self, slice_size):
+        for block in self.transformer_blocks:
+            block._set_attention_slice(slice_size)
+
+    def forward(self, hidden_states, encoder_hidden_states=None, encoder_attention_mask=None, timestep=None,
+                return_dict: bool = True):
+        """
+        Args:
+            hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`.
+                When continous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input
+                hidden_states
+            encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, context dim)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            encoder_attention_mask ( `torch.LongTensor` of shape `(batch size, context)`, *optional*):
+                Attention mask for cross attention layer.
+            timestep ( `torch.long`, *optional*):
+                Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.attention.Transformer2DModelOutput`] or `tuple`: [`~models.attention.Transformer2DModelOutput`]
+            if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample
+            tensor.
+        """
+        # 1. Input
+        if self.is_input_continuous:
+            batch, channel, height, weight = hidden_states.shape
+            residual = hidden_states
+            hidden_states = self.norm(hidden_states)
+            hidden_states = self.proj_in(hidden_states)
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
+        elif self.is_input_vectorized:
+            hidden_states = self.latent_image_embedding(hidden_states)
+
+        # 2. Blocks
+        for block in self.transformer_blocks:
+            hidden_states = block(hidden_states, context=encoder_hidden_states, mask=encoder_attention_mask,
+                                  timestep=timestep)
+
+        # 3. Output
+        if self.is_input_continuous:
+            hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2)
+            hidden_states = self.proj_out(hidden_states)
+            output = hidden_states + residual
+        elif self.is_input_vectorized:
+            hidden_states = self.norm_out(hidden_states)
+            logits = self.out(hidden_states)
+            # (batch, self.num_vector_embeds - 1, self.num_latent_pixels)
+            logits = logits.permute(0, 2, 1)
+
+            # log(p(x_0))
+            output = F.log_softmax(logits.double(), dim=1).float()
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)
+
+    def _set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        for block in self.transformer_blocks:
+            block._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
+
+
+class AttentionBlock(nn.Module):
+    """
+    An attention block that allows spatial positions to attend to each other. Originally ported from here, but adapted
+    to the N-d case.
+    https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
+    Uses three q, k, v linear layers to compute attention.
+
+    Parameters:
+        channels (`int`): The number of channels in the input and output.
+        num_head_channels (`int`, *optional*):
+            The number of channels in each head. If None, then `num_heads` = 1.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for group norm.
+        rescale_output_factor (`float`, *optional*, defaults to 1.0): The factor to rescale the output by.
+        eps (`float`, *optional*, defaults to 1e-5): The epsilon value to use for group norm.
+    """
+
+    def __init__(
+            self,
+            channels: int,
+            num_head_channels: Optional[int] = None,
+            norm_num_groups: int = 32,
+            rescale_output_factor: float = 1.0,
+            eps: float = 1e-5,
+    ):
+        super().__init__()
+        self.channels = channels
+
+        self.num_heads = channels // num_head_channels if num_head_channels is not None else 1
+        self.num_head_size = num_head_channels
+        self.group_norm = nn.GroupNorm(num_channels=channels, num_groups=norm_num_groups, eps=eps, affine=True)
+
+        # define q,k,v as linear layers
+        self.query = nn.Linear(channels, channels)
+        self.key = nn.Linear(channels, channels)
+        self.value = nn.Linear(channels, channels)
+
+        self.rescale_output_factor = rescale_output_factor
+        self.proj_attn = nn.Linear(channels, channels, 1)
+
+    def transpose_for_scores(self, projection: torch.Tensor) -> torch.Tensor:
+        new_projection_shape = projection.size()[:-1] + (self.num_heads, -1)
+        # move heads to 2nd position (B, T, H * D) -> (B, T, H, D) -> (B, H, T, D)
+        new_projection = projection.view(new_projection_shape).permute(0, 2, 1, 3)
+        return new_projection
+
+    def forward(self, hidden_states):
+        residual = hidden_states
+        batch, channel, height, width = hidden_states.shape
+
+        # norm
+        hidden_states = self.group_norm(hidden_states)
+
+        hidden_states = hidden_states.view(batch, channel, height * width).transpose(1, 2)
+
+        # proj to q, k, v
+        query_proj = self.query(hidden_states)
+        key_proj = self.key(hidden_states)
+        value_proj = self.value(hidden_states)
+
+        # transpose
+        query_states = self.transpose_for_scores(query_proj)
+        key_states = self.transpose_for_scores(key_proj)
+        value_states = self.transpose_for_scores(value_proj)
+
+        # get scores
+        scale = 1 / math.sqrt(math.sqrt(self.channels / self.num_heads))
+        attention_scores = torch.matmul(query_states * scale, key_states.transpose(-1, -2) * scale)  # TODO: use baddmm
+        attention_probs = torch.softmax(attention_scores.float(), dim=-1).type(attention_scores.dtype)
+
+        # compute attention output
+        hidden_states = torch.matmul(attention_probs, value_states)
+
+        hidden_states = hidden_states.permute(0, 2, 1, 3).contiguous()
+        new_hidden_states_shape = hidden_states.size()[:-2] + (self.channels,)
+        hidden_states = hidden_states.view(new_hidden_states_shape)
+
+        # compute next hidden_states
+        hidden_states = self.proj_attn(hidden_states)
+        hidden_states = hidden_states.transpose(-1, -2).reshape(batch, channel, height, width)
+
+        # res connect and rescale
+        hidden_states = (hidden_states + residual) / self.rescale_output_factor
+        return hidden_states
+
+
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the context vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+    """
+
+    def __init__(
+            self,
+            dim: int,
+            num_attention_heads: int,
+            attention_head_dim: int,
+            dropout=0.0,
+            cross_attention_dim: Optional[int] = None,
+            activation_fn: str = "geglu",
+            num_embeds_ada_norm: Optional[int] = None,
+            attention_bias: bool = False,
+    ):
+        super().__init__()
+        self.attn1 = CrossAttention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+        )  # is a self-attention
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.attn2 = CrossAttention(
+            query_dim=dim,
+            cross_attention_dim=cross_attention_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+        )  # is self-attn if context is none
+
+        # layer norms
+        self.use_ada_layer_norm = num_embeds_ada_norm is not None
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+            self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        else:
+            self.norm1 = nn.LayerNorm(dim)
+            self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+
+    def _set_attention_slice(self, slice_size):
+        self.attn1._slice_size = slice_size
+        self.attn2._slice_size = slice_size
+
+    def _set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        if not is_xformers_available():
+            print("Here is how to install it")
+            raise ModuleNotFoundError(
+                "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
+                " xformers",
+                name="xformers",
+            )
+        elif not torch.cuda.is_available():
+            raise ValueError(
+                "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is only"
+                " available for GPU "
+            )
+        else:
+            try:
+                # Make sure we can run the memory efficient attention
+                _ = xformers.ops.memory_efficient_attention(
+                    torch.randn((1, 2, 40), device="cuda"),
+                    torch.randn((1, 2, 40), device="cuda"),
+                    torch.randn((1, 2, 40), device="cuda"),
+                )
+            except Exception as e:
+                raise e
+            self.attn1._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+            self.attn2._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+
+    def forward(self, hidden_states, context=None, mask=None, timestep=None):
+        # 1. Self-Attention
+        norm_hidden_states = (
+            self.norm1(hidden_states, timestep) if self.use_ada_layer_norm else self.norm1(hidden_states)
+        )
+        hidden_states = self.attn1(norm_hidden_states) + hidden_states
+
+        # 2. Cross-Attention
+        norm_hidden_states = (
+            self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
+        )
+        hidden_states = self.attn2(norm_hidden_states, context=context, mask=mask) + hidden_states
+
+        # 3. Feed-forward
+        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+        return hidden_states
+
+
+class CrossAttention(nn.Module):
+    r"""
+    A cross attention layer.
+
+    Parameters:
+        query_dim (`int`): The number of channels in the query.
+        cross_attention_dim (`int`, *optional*):
+            The number of channels in the context. If not given, defaults to `query_dim`.
+        heads (`int`,  *optional*, defaults to 8): The number of heads to use for multi-head attention.
+        dim_head (`int`,  *optional*, defaults to 64): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        bias (`bool`, *optional*, defaults to False):
+            Set to `True` for the query, key, and value linear layers to contain a bias parameter.
+    """
+
+    def __init__(
+            self,
+            query_dim: int,
+            cross_attention_dim: Optional[int] = None,
+            heads: int = 8,
+            dim_head: int = 64,
+            dropout: float = 0.0,
+            bias=False,
+    ):
+        super().__init__()
+        inner_dim = dim_head * heads
+        cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
+
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+        # for slice_size > 0 the attention score computation
+        # is split across the batch axis to save memory
+        # You can set slice_size with `set_attention_slice`
+        self._slice_size = None
+        self._use_memory_efficient_attention_xformers = False
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=bias)
+        self.to_k = nn.Linear(cross_attention_dim, inner_dim, bias=bias)
+        self.to_v = nn.Linear(cross_attention_dim, inner_dim, bias=bias)
+
+        self.to_out = nn.ModuleList([])
+        self.to_out.append(nn.Linear(inner_dim, query_dim))
+        self.to_out.append(nn.Dropout(dropout))
+
+    def reshape_heads_to_batch_dim(self, tensor):
+        batch_size, seq_len, dim = tensor.shape
+        head_size = self.heads
+        tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
+        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size * head_size, seq_len, dim // head_size)
+        return tensor
+
+    def reshape_batch_dim_to_heads(self, tensor):
+        batch_size, seq_len, dim = tensor.shape
+        head_size = self.heads
+        tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
+        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
+        return tensor
+
+    def forward(self, hidden_states, context=None, mask=None):
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        query = self.to_q(hidden_states)
+        context = context if context is not None else hidden_states
+        key = self.to_k(context)
+        value = self.to_v(context)
+
+        dim = query.shape[-1]
+
+        query = self.reshape_heads_to_batch_dim(query)
+        key = self.reshape_heads_to_batch_dim(key)
+        value = self.reshape_heads_to_batch_dim(value)
+        mask = mask.repeat_interleave(self.heads, dim=0).unsqueeze(1) if mask is not None else None
+
+        # attention, what we cannot get enough of
+        if self._use_memory_efficient_attention_xformers:
+            hidden_states = self._memory_efficient_attention_xformers(query, key, value)
+        else:
+            if self._slice_size is None or query.shape[0] // self._slice_size == 1:
+                hidden_states = self._attention(query, key, value, mask)
+            else:
+                assert mask is None, "masking is not supported for sliced attention"
+                hidden_states = self._sliced_attention(query, key, value, sequence_length, dim)
+
+        # linear proj
+        hidden_states = self.to_out[0](hidden_states)
+        # dropout
+        hidden_states = self.to_out[1](hidden_states)
+        return hidden_states
+
+    def _attention(self, query, key, value, mask):
+        # TODO: use baddbmm for better performance
+        if query.device.type == "mps":
+            # Better performance on mps (~20-25%)
+            attention_scores = torch.einsum("b i d, b j d -> b i j", query, key) * self.scale
+        else:
+            attention_scores = torch.matmul(query, key.transpose(-1, -2)) * self.scale
+        attention_scores = attention_scores.masked_fill(mask.expand(attention_scores.shape), value=float("-inf")) \
+            if mask is not None else attention_scores
+        attention_probs = attention_scores.softmax(dim=-1)
+        # compute attention output
+
+        if query.device.type == "mps":
+            hidden_states = torch.einsum("b i j, b j d -> b i d", attention_probs, value)
+        else:
+            hidden_states = torch.matmul(attention_probs, value)
+
+        # reshape hidden_states
+        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
+        return hidden_states
+
+    def _sliced_attention(self, query, key, value, sequence_length, dim):
+        batch_size_attention = query.shape[0]
+        hidden_states = torch.zeros(
+            (batch_size_attention, sequence_length, dim // self.heads), device=query.device, dtype=query.dtype
+        )
+        slice_size = self._slice_size if self._slice_size is not None else hidden_states.shape[0]
+        for i in range(hidden_states.shape[0] // slice_size):
+            start_idx = i * slice_size
+            end_idx = (i + 1) * slice_size
+            if query.device.type == "mps":
+                # Better performance on mps (~20-25%)
+                attn_slice = (
+                        torch.einsum("b i d, b j d -> b i j", query[start_idx:end_idx], key[start_idx:end_idx])
+                        * self.scale
+                )
+            else:
+                attn_slice = (
+                        torch.matmul(query[start_idx:end_idx], key[start_idx:end_idx].transpose(1, 2)) * self.scale
+                )  # TODO: use baddbmm for better performance
+            attn_slice = attn_slice.softmax(dim=-1)
+            if query.device.type == "mps":
+                attn_slice = torch.einsum("b i j, b j d -> b i d", attn_slice, value[start_idx:end_idx])
+            else:
+                attn_slice = torch.matmul(attn_slice, value[start_idx:end_idx])
+
+            hidden_states[start_idx:end_idx] = attn_slice
+
+        # reshape hidden_states
+        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
+        return hidden_states
+
+    def _memory_efficient_attention_xformers(self, query, key, value):
+        hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=None)
+        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
+        return hidden_states
+
+
+class FeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        dim (`int`): The number of channels in the input.
+        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
+        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+    """
+
+    def __init__(
+            self,
+            dim: int,
+            dim_out: Optional[int] = None,
+            mult: int = 4,
+            dropout: float = 0.0,
+            activation_fn: str = "geglu",
+    ):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+
+        if activation_fn == "geglu":
+            geglu = GEGLU(dim, inner_dim)
+        elif activation_fn == "geglu-approximate":
+            geglu = ApproximateGELU(dim, inner_dim)
+
+        self.net = nn.ModuleList([])
+        # project in
+        self.net.append(geglu)
+        # project dropout
+        self.net.append(nn.Dropout(dropout))
+        # project out
+        self.net.append(nn.Linear(inner_dim, dim_out))
+
+    def forward(self, hidden_states):
+        for module in self.net:
+            hidden_states = module(hidden_states)
+        return hidden_states
+
+
+# feedforward
+class GEGLU(nn.Module):
+    r"""
+    A variant of the gated linear unit activation function from https://arxiv.org/abs/2002.05202.
+
+    Parameters:
+        dim_in (`int`): The number of channels in the input.
+        dim_out (`int`): The number of channels in the output.
+    """
+
+    def __init__(self, dim_in: int, dim_out: int):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def gelu(self, gate):
+        if gate.device.type != "mps":
+            return F.gelu(gate)
+        # mps: gelu is not implemented for float16
+        return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
+
+    def forward(self, hidden_states):
+        hidden_states, gate = self.proj(hidden_states).chunk(2, dim=-1)
+        return hidden_states * self.gelu(gate)
+
+
+class ApproximateGELU(nn.Module):
+    """
+    The approximate form of Gaussian Error Linear Unit (GELU)
+
+    For more details, see section 2: https://arxiv.org/abs/1606.08415
+    """
+
+    def __init__(self, dim_in: int, dim_out: int):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out)
+
+    def forward(self, x):
+        x = self.proj(x)
+        return x * torch.sigmoid(1.702 * x)
+
+
+class AdaLayerNorm(nn.Module):
+    """
+    Norm layer modified to incorporate timestep embeddings.
+    """
+
+    def __init__(self, embedding_dim, num_embeddings):
+        super().__init__()
+        self.emb = nn.Embedding(num_embeddings, embedding_dim)
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(embedding_dim, embedding_dim * 2)
+        self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False)
+
+    def forward(self, x, timestep):
+        emb = self.linear(self.silu(self.emb(timestep)))
+        scale, shift = torch.chunk(emb, 2)
+        x = self.norm(x) * (1 + scale) + shift
+        return x

models/diffusers_override/unet_2d_blocks.py

@@ -0,0 +1,1602 @@
+# Copyright 2022 The HuggingFace Team. 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.
+import numpy as np
+import torch
+from torch import nn
+
+from .attention import AttentionBlock, Transformer2DModel
+from diffusers.models.resnet import Downsample2D, FirDownsample2D, FirUpsample2D, ResnetBlock2D, Upsample2D
+
+
+def get_down_block(
+        down_block_type,
+        num_layers,
+        in_channels,
+        out_channels,
+        temb_channels,
+        add_downsample,
+        resnet_eps,
+        resnet_act_fn,
+        attn_num_head_channels,
+        resnet_groups=None,
+        cross_attention_dim=None,
+        downsample_padding=None,
+):
+    down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
+    if down_block_type == "DownBlock2D":
+        return DownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+        )
+    elif down_block_type == "AttnDownBlock2D":
+        return AttnDownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            attn_num_head_channels=attn_num_head_channels,
+        )
+    elif down_block_type == "CrossAttnDownBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
+        return CrossAttnDownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            attn_num_head_channels=attn_num_head_channels,
+        )
+    elif down_block_type == "SkipDownBlock2D":
+        return SkipDownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            downsample_padding=downsample_padding,
+        )
+    elif down_block_type == "AttnSkipDownBlock2D":
+        return AttnSkipDownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            downsample_padding=downsample_padding,
+            attn_num_head_channels=attn_num_head_channels,
+        )
+    elif down_block_type == "DownEncoderBlock2D":
+        return DownEncoderBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+        )
+    elif down_block_type == "AttnDownEncoderBlock2D":
+        return AttnDownEncoderBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            attn_num_head_channels=attn_num_head_channels,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+        up_block_type,
+        num_layers,
+        in_channels,
+        out_channels,
+        prev_output_channel,
+        temb_channels,
+        add_upsample,
+        resnet_eps,
+        resnet_act_fn,
+        attn_num_head_channels,
+        resnet_groups=None,
+        cross_attention_dim=None,
+):
+    up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+    if up_block_type == "UpBlock2D":
+        return UpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+        )
+    elif up_block_type == "CrossAttnUpBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
+        return CrossAttnUpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            attn_num_head_channels=attn_num_head_channels,
+        )
+    elif up_block_type == "AttnUpBlock2D":
+        return AttnUpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            attn_num_head_channels=attn_num_head_channels,
+        )
+    elif up_block_type == "SkipUpBlock2D":
+        return SkipUpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+        )
+    elif up_block_type == "AttnSkipUpBlock2D":
+        return AttnSkipUpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            attn_num_head_channels=attn_num_head_channels,
+        )
+    elif up_block_type == "UpDecoderBlock2D":
+        return UpDecoderBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+        )
+    elif up_block_type == "AttnUpDecoderBlock2D":
+        return AttnUpDecoderBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            attn_num_head_channels=attn_num_head_channels,
+        )
+    raise ValueError(f"{up_block_type} does not exist.")
+
+
+class UNetMidBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            temb_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_groups: int = 32,
+            resnet_pre_norm: bool = True,
+            attn_num_head_channels=1,
+            attention_type="default",
+            output_scale_factor=1.0,
+            **kwargs,
+    ):
+        super().__init__()
+
+        self.attention_type = attention_type
+        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        for _ in range(num_layers):
+            attentions.append(
+                AttentionBlock(
+                    in_channels,
+                    num_head_channels=attn_num_head_channels,
+                    rescale_output_factor=output_scale_factor,
+                    eps=resnet_eps,
+                    norm_num_groups=resnet_groups,
+                )
+            )
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def forward(self, hidden_states, temb=None, encoder_states=None):
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if self.attention_type == "default":
+                hidden_states = attn(hidden_states)
+            else:
+                hidden_states = attn(hidden_states, encoder_states)
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class UNetMidBlock2DCrossAttn(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            temb_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_groups: int = 32,
+            resnet_pre_norm: bool = True,
+            attn_num_head_channels=1,
+            attention_type="default",
+            output_scale_factor=1.0,
+            cross_attention_dim=1280,
+            **kwargs,
+    ):
+        super().__init__()
+
+        self.attention_type = attention_type
+        self.attn_num_head_channels = attn_num_head_channels
+        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        for _ in range(num_layers):
+            attentions.append(
+                Transformer2DModel(
+                    attn_num_head_channels,
+                    in_channels // attn_num_head_channels,
+                    in_channels=in_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                )
+            )
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def set_attention_slice(self, slice_size):
+        if slice_size is not None and self.attn_num_head_channels % slice_size != 0:
+            raise ValueError(
+                f"Make sure slice_size {slice_size} is a divisor of "
+                f"the number of heads used in cross_attention {self.attn_num_head_channels}"
+            )
+        if slice_size is not None and slice_size > self.attn_num_head_channels:
+            raise ValueError(
+                f"Chunk_size {slice_size} has to be smaller or equal to "
+                f"the number of heads used in cross_attention {self.attn_num_head_channels}"
+            )
+
+        for attn in self.attentions:
+            attn._set_attention_slice(slice_size)
+
+    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        for attn in self.attentions:
+            attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
+
+    def forward(self, hidden_states, temb=None, encoder_hidden_states=None, encoder_attention_mask=None):
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            hidden_states = attn(hidden_states, encoder_hidden_states, encoder_attention_mask).sample
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class AttnDownBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            out_channels: int,
+            temb_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_groups: int = 32,
+            resnet_pre_norm: bool = True,
+            attn_num_head_channels=1,
+            attention_type="default",
+            output_scale_factor=1.0,
+            downsample_padding=1,
+            add_downsample=True,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.attention_type = attention_type
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            attentions.append(
+                AttentionBlock(
+                    out_channels,
+                    num_head_channels=attn_num_head_channels,
+                    rescale_output_factor=output_scale_factor,
+                    eps=resnet_eps,
+                    norm_num_groups=resnet_groups,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        in_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+    def forward(self, hidden_states, temb=None):
+        output_states = ()
+
+        for resnet, attn in zip(self.resnets, self.attentions):
+            hidden_states = resnet(hidden_states, temb)
+            hidden_states = attn(hidden_states)
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class CrossAttnDownBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            out_channels: int,
+            temb_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_groups: int = 32,
+            resnet_pre_norm: bool = True,
+            attn_num_head_channels=1,
+            cross_attention_dim=1280,
+            attention_type="default",
+            output_scale_factor=1.0,
+            downsample_padding=1,
+            add_downsample=True,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.attention_type = attention_type
+        self.attn_num_head_channels = attn_num_head_channels
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            attentions.append(
+                Transformer2DModel(
+                    attn_num_head_channels,
+                    out_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                )
+            )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        in_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def set_attention_slice(self, slice_size):
+        if slice_size is not None and self.attn_num_head_channels % slice_size != 0:
+            raise ValueError(
+                f"Make sure slice_size {slice_size} is a divisor of "
+                f"the number of heads used in cross_attention {self.attn_num_head_channels}"
+            )
+        if slice_size is not None and slice_size > self.attn_num_head_channels:
+            raise ValueError(
+                f"Chunk_size {slice_size} has to be smaller or equal to "
+                f"the number of heads used in cross_attention {self.attn_num_head_channels}"
+            )
+
+        for attn in self.attentions:
+            attn._set_attention_slice(slice_size)
+
+    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        for attn in self.attentions:
+            attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
+
+    def forward(self, hidden_states, temb=None, encoder_hidden_states=None, encoder_attention_mask=None):
+        output_states = ()
+
+        for resnet, attn in zip(self.resnets, self.attentions):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states,
+                    encoder_attention_mask
+                )[0]
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states,
+                                     encoder_attention_mask=encoder_attention_mask).sample
+
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class DownBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            out_channels: int,
+            temb_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_groups: int = 32,
+            resnet_pre_norm: bool = True,
+            output_scale_factor=1.0,
+            add_downsample=True,
+            downsample_padding=1,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        in_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(self, hidden_states, temb=None):
+        output_states = ()
+
+        for resnet in self.resnets:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class DownEncoderBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            out_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_groups: int = 32,
+            resnet_pre_norm: bool = True,
+            output_scale_factor=1.0,
+            add_downsample=True,
+            downsample_padding=1,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=None,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        in_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+    def forward(self, hidden_states):
+        for resnet in self.resnets:
+            hidden_states = resnet(hidden_states, temb=None)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+        return hidden_states
+
+
+class AttnDownEncoderBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            out_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_groups: int = 32,
+            resnet_pre_norm: bool = True,
+            attn_num_head_channels=1,
+            output_scale_factor=1.0,
+            add_downsample=True,
+            downsample_padding=1,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=None,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            attentions.append(
+                AttentionBlock(
+                    out_channels,
+                    num_head_channels=attn_num_head_channels,
+                    rescale_output_factor=output_scale_factor,
+                    eps=resnet_eps,
+                    norm_num_groups=resnet_groups,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        in_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+    def forward(self, hidden_states):
+        for resnet, attn in zip(self.resnets, self.attentions):
+            hidden_states = resnet(hidden_states, temb=None)
+            hidden_states = attn(hidden_states)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+        return hidden_states
+
+
+class AttnSkipDownBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            out_channels: int,
+            temb_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_pre_norm: bool = True,
+            attn_num_head_channels=1,
+            attention_type="default",
+            output_scale_factor=np.sqrt(2.0),
+            downsample_padding=1,
+            add_downsample=True,
+    ):
+        super().__init__()
+        self.attentions = nn.ModuleList([])
+        self.resnets = nn.ModuleList([])
+
+        self.attention_type = attention_type
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            self.resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=min(in_channels // 4, 32),
+                    groups_out=min(out_channels // 4, 32),
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            self.attentions.append(
+                AttentionBlock(
+                    out_channels,
+                    num_head_channels=attn_num_head_channels,
+                    rescale_output_factor=output_scale_factor,
+                    eps=resnet_eps,
+                )
+            )
+
+        if add_downsample:
+            self.resnet_down = ResnetBlock2D(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=min(out_channels // 4, 32),
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+                use_in_shortcut=True,
+                down=True,
+                kernel="fir",
+            )
+            self.downsamplers = nn.ModuleList([FirDownsample2D(in_channels, out_channels=out_channels)])
+            self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
+        else:
+            self.resnet_down = None
+            self.downsamplers = None
+            self.skip_conv = None
+
+    def forward(self, hidden_states, temb=None, skip_sample=None):
+        output_states = ()
+
+        for resnet, attn in zip(self.resnets, self.attentions):
+            hidden_states = resnet(hidden_states, temb)
+            hidden_states = attn(hidden_states)
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            hidden_states = self.resnet_down(hidden_states, temb)
+            for downsampler in self.downsamplers:
+                skip_sample = downsampler(skip_sample)
+
+            hidden_states = self.skip_conv(skip_sample) + hidden_states
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states, skip_sample
+
+
+class SkipDownBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            out_channels: int,
+            temb_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_pre_norm: bool = True,
+            output_scale_factor=np.sqrt(2.0),
+            add_downsample=True,
+            downsample_padding=1,
+    ):
+        super().__init__()
+        self.resnets = nn.ModuleList([])
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            self.resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=min(in_channels // 4, 32),
+                    groups_out=min(out_channels // 4, 32),
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        if add_downsample:
+            self.resnet_down = ResnetBlock2D(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=min(out_channels // 4, 32),
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+                use_in_shortcut=True,
+                down=True,
+                kernel="fir",
+            )
+            self.downsamplers = nn.ModuleList([FirDownsample2D(in_channels, out_channels=out_channels)])
+            self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
+        else:
+            self.resnet_down = None
+            self.downsamplers = None
+            self.skip_conv = None
+
+    def forward(self, hidden_states, temb=None, skip_sample=None):
+        output_states = ()
+
+        for resnet in self.resnets:
+            hidden_states = resnet(hidden_states, temb)
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            hidden_states = self.resnet_down(hidden_states, temb)
+            for downsampler in self.downsamplers:
+                skip_sample = downsampler(skip_sample)
+
+            hidden_states = self.skip_conv(skip_sample) + hidden_states
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states, skip_sample
+
+
+class AttnUpBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            prev_output_channel: int,
+            out_channels: int,
+            temb_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_groups: int = 32,
+            resnet_pre_norm: bool = True,
+            attention_type="default",
+            attn_num_head_channels=1,
+            output_scale_factor=1.0,
+            add_upsample=True,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.attention_type = attention_type
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            attentions.append(
+                AttentionBlock(
+                    out_channels,
+                    num_head_channels=attn_num_head_channels,
+                    rescale_output_factor=output_scale_factor,
+                    eps=resnet_eps,
+                    norm_num_groups=resnet_groups,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+        else:
+            self.upsamplers = None
+
+    def forward(self, hidden_states, res_hidden_states_tuple, temb=None):
+        for resnet, attn in zip(self.resnets, self.attentions):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            hidden_states = resnet(hidden_states, temb)
+            hidden_states = attn(hidden_states)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states)
+
+        return hidden_states
+
+
+class CrossAttnUpBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            out_channels: int,
+            prev_output_channel: int,
+            temb_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_groups: int = 32,
+            resnet_pre_norm: bool = True,
+            attn_num_head_channels=1,
+            cross_attention_dim=1280,
+            attention_type="default",
+            output_scale_factor=1.0,
+            add_upsample=True,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.attention_type = attention_type
+        self.attn_num_head_channels = attn_num_head_channels
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            attentions.append(
+                Transformer2DModel(
+                    attn_num_head_channels,
+                    out_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                )
+            )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def set_attention_slice(self, slice_size):
+        if slice_size is not None and self.attn_num_head_channels % slice_size != 0:
+            raise ValueError(
+                f"Make sure slice_size {slice_size} is a divisor of "
+                f"the number of heads used in cross_attention {self.attn_num_head_channels}"
+            )
+        if slice_size is not None and slice_size > self.attn_num_head_channels:
+            raise ValueError(
+                f"Chunk_size {slice_size} has to be smaller or equal to "
+                f"the number of heads used in cross_attention {self.attn_num_head_channels}"
+            )
+
+        for attn in self.attentions:
+            attn._set_attention_slice(slice_size)
+
+        self.gradient_checkpointing = False
+
+    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        for attn in self.attentions:
+            attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
+
+    def forward(
+            self,
+            hidden_states,
+            res_hidden_states_tuple,
+            temb=None,
+            encoder_hidden_states=None,
+            encoder_attention_mask=None,
+            upsample_size=None,
+    ):
+        for resnet, attn in zip(self.resnets, self.attentions):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states,
+                    encoder_attention_mask
+                )[0]
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states,
+                                     encoder_attention_mask=encoder_attention_mask).sample
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states
+
+
+class UpBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            prev_output_channel: int,
+            out_channels: int,
+            temb_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_groups: int = 32,
+            resnet_pre_norm: bool = True,
+            output_scale_factor=1.0,
+            add_upsample=True,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
+        for resnet in self.resnets:
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states
+
+
+class UpDecoderBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            out_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_groups: int = 32,
+            resnet_pre_norm: bool = True,
+            output_scale_factor=1.0,
+            add_upsample=True,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            input_channels = in_channels if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=input_channels,
+                    out_channels=out_channels,
+                    temb_channels=None,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+        else:
+            self.upsamplers = None
+
+    def forward(self, hidden_states):
+        for resnet in self.resnets:
+            hidden_states = resnet(hidden_states, temb=None)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states)
+
+        return hidden_states
+
+
+class AttnUpDecoderBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            out_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_groups: int = 32,
+            resnet_pre_norm: bool = True,
+            attn_num_head_channels=1,
+            output_scale_factor=1.0,
+            add_upsample=True,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        for i in range(num_layers):
+            input_channels = in_channels if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=input_channels,
+                    out_channels=out_channels,
+                    temb_channels=None,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            attentions.append(
+                AttentionBlock(
+                    out_channels,
+                    num_head_channels=attn_num_head_channels,
+                    rescale_output_factor=output_scale_factor,
+                    eps=resnet_eps,
+                    norm_num_groups=resnet_groups,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+        else:
+            self.upsamplers = None
+
+    def forward(self, hidden_states):
+        for resnet, attn in zip(self.resnets, self.attentions):
+            hidden_states = resnet(hidden_states, temb=None)
+            hidden_states = attn(hidden_states)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states)
+
+        return hidden_states
+
+
+class AttnSkipUpBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            prev_output_channel: int,
+            out_channels: int,
+            temb_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_pre_norm: bool = True,
+            attn_num_head_channels=1,
+            attention_type="default",
+            output_scale_factor=np.sqrt(2.0),
+            upsample_padding=1,
+            add_upsample=True,
+    ):
+        super().__init__()
+        self.attentions = nn.ModuleList([])
+        self.resnets = nn.ModuleList([])
+
+        self.attention_type = attention_type
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            self.resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=min(resnet_in_channels + res_skip_channels // 4, 32),
+                    groups_out=min(out_channels // 4, 32),
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions.append(
+            AttentionBlock(
+                out_channels,
+                num_head_channels=attn_num_head_channels,
+                rescale_output_factor=output_scale_factor,
+                eps=resnet_eps,
+            )
+        )
+
+        self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
+        if add_upsample:
+            self.resnet_up = ResnetBlock2D(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=min(out_channels // 4, 32),
+                groups_out=min(out_channels // 4, 32),
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+                use_in_shortcut=True,
+                up=True,
+                kernel="fir",
+            )
+            self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+            self.skip_norm = torch.nn.GroupNorm(
+                num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
+            )
+            self.act = nn.SiLU()
+        else:
+            self.resnet_up = None
+            self.skip_conv = None
+            self.skip_norm = None
+            self.act = None
+
+    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, skip_sample=None):
+        for resnet in self.resnets:
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            hidden_states = resnet(hidden_states, temb)
+
+        hidden_states = self.attentions[0](hidden_states)
+
+        if skip_sample is not None:
+            skip_sample = self.upsampler(skip_sample)
+        else:
+            skip_sample = 0
+
+        if self.resnet_up is not None:
+            skip_sample_states = self.skip_norm(hidden_states)
+            skip_sample_states = self.act(skip_sample_states)
+            skip_sample_states = self.skip_conv(skip_sample_states)
+
+            skip_sample = skip_sample + skip_sample_states
+
+            hidden_states = self.resnet_up(hidden_states, temb)
+
+        return hidden_states, skip_sample
+
+
+class SkipUpBlock2D(nn.Module):
+    def __init__(
+            self,
+            in_channels: int,
+            prev_output_channel: int,
+            out_channels: int,
+            temb_channels: int,
+            dropout: float = 0.0,
+            num_layers: int = 1,
+            resnet_eps: float = 1e-6,
+            resnet_time_scale_shift: str = "default",
+            resnet_act_fn: str = "swish",
+            resnet_pre_norm: bool = True,
+            output_scale_factor=np.sqrt(2.0),
+            add_upsample=True,
+            upsample_padding=1,
+    ):
+        super().__init__()
+        self.resnets = nn.ModuleList([])
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            self.resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=min((resnet_in_channels + res_skip_channels) // 4, 32),
+                    groups_out=min(out_channels // 4, 32),
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
+        if add_upsample:
+            self.resnet_up = ResnetBlock2D(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=min(out_channels // 4, 32),
+                groups_out=min(out_channels // 4, 32),
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+                use_in_shortcut=True,
+                up=True,
+                kernel="fir",
+            )
+            self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+            self.skip_norm = torch.nn.GroupNorm(
+                num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
+            )
+            self.act = nn.SiLU()
+        else:
+            self.resnet_up = None
+            self.skip_conv = None
+            self.skip_norm = None
+            self.act = None
+
+    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, skip_sample=None):
+        for resnet in self.resnets:
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            hidden_states = resnet(hidden_states, temb)
+
+        if skip_sample is not None:
+            skip_sample = self.upsampler(skip_sample)
+        else:
+            skip_sample = 0
+
+        if self.resnet_up is not None:
+            skip_sample_states = self.skip_norm(hidden_states)
+            skip_sample_states = self.act(skip_sample_states)
+            skip_sample_states = self.skip_conv(skip_sample_states)
+
+            skip_sample = skip_sample + skip_sample_states
+
+            hidden_states = self.resnet_up(hidden_states, temb)
+
+        return hidden_states, skip_sample

models/diffusers_override/unet_2d_condition.py

@@ -0,0 +1,359 @@
+# Copyright 2022 The HuggingFace Team. 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.
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.modeling_utils import ModelMixin
+from diffusers.utils import BaseOutput, logging
+from diffusers.models.embeddings import TimestepEmbedding, Timesteps
+from .unet_2d_blocks import (
+    CrossAttnDownBlock2D,
+    CrossAttnUpBlock2D,
+    DownBlock2D,
+    UNetMidBlock2DCrossAttn,
+    UpBlock2D,
+    get_down_block,
+    get_up_block,
+)
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Hidden states conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.FloatTensor
+
+
+class UNet2DConditionModel(ModelMixin, ConfigMixin):
+    r"""
+    UNet2DConditionModel is a conditional 2D UNet model that takes in a noisy sample, conditional state, and a timestep
+    and returns sample shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
+    implements for all the models (such as downloading or saving, etc.)
+
+    Parameters:
+        sample_size (`int`, *optional*): The size of the input sample.
+        in_channels (`int`, *optional*, defaults to 4): The number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): The number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `True`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D",)`):
+            The tuple of upsample blocks to use.
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int`, *optional*, defaults to 1280): The dimension of the cross attention features.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+            self,
+            sample_size: Optional[int] = None,
+            in_channels: int = 4,
+            out_channels: int = 4,
+            center_input_sample: bool = False,
+            flip_sin_to_cos: bool = True,
+            freq_shift: int = 0,
+            down_block_types: Tuple[str] = (
+                    "CrossAttnDownBlock2D",
+                    "CrossAttnDownBlock2D",
+                    "CrossAttnDownBlock2D",
+                    "DownBlock2D",
+            ),
+            up_block_types: Tuple[str] = (
+                    "UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+            block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+            layers_per_block: int = 2,
+            downsample_padding: int = 1,
+            mid_block_scale_factor: float = 1,
+            act_fn: str = "silu",
+            norm_num_groups: int = 32,
+            norm_eps: float = 1e-5,
+            cross_attention_dim: int = 1280,
+            attention_head_dim: int = 8,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+        time_embed_dim = block_out_channels[0] * 4
+
+        # input
+        self.conv_in = nn.Conv2d(in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1))
+
+        # time
+        self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+        timestep_input_dim = block_out_channels[0]
+
+        self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+
+        self.down_blocks = nn.ModuleList([])
+        self.mid_block = None
+        self.up_blocks = nn.ModuleList([])
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=attention_head_dim,
+                downsample_padding=downsample_padding,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        self.mid_block = UNetMidBlock2DCrossAttn(
+            in_channels=block_out_channels[-1],
+            temb_channels=time_embed_dim,
+            resnet_eps=norm_eps,
+            resnet_act_fn=act_fn,
+            output_scale_factor=mid_block_scale_factor,
+            resnet_time_scale_shift="default",
+            cross_attention_dim=cross_attention_dim,
+            attn_num_head_channels=attention_head_dim,
+            resnet_groups=norm_num_groups,
+        )
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=layers_per_block + 1,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=attention_head_dim,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps)
+        self.conv_act = nn.SiLU()
+        self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1)
+
+    def set_attention_slice(self, slice_size):
+        if slice_size is not None and self.config.attention_head_dim % slice_size != 0:
+            raise ValueError(
+                f"Make sure slice_size {slice_size} is a divisor of "
+                f"the number of heads used in cross_attention {self.config.attention_head_dim}"
+            )
+        if slice_size is not None and slice_size > self.config.attention_head_dim:
+            raise ValueError(
+                f"Chunk_size {slice_size} has to be smaller or equal to "
+                f"the number of heads used in cross_attention {self.config.attention_head_dim}"
+            )
+
+        for block in self.down_blocks:
+            if hasattr(block, "attentions") and block.attentions is not None:
+                block.set_attention_slice(slice_size)
+
+        self.mid_block.set_attention_slice(slice_size)
+
+        for block in self.up_blocks:
+            if hasattr(block, "attentions") and block.attentions is not None:
+                block.set_attention_slice(slice_size)
+
+    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        for block in self.down_blocks:
+            if hasattr(block, "attentions") and block.attentions is not None:
+                block.set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
+
+        self.mid_block.set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
+
+        for block in self.up_blocks:
+            if hasattr(block, "attentions") and block.attentions is not None:
+                block.set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D, CrossAttnUpBlock2D, UpBlock2D)):
+            module.gradient_checkpointing = value
+
+    def forward(
+            self,
+            sample: torch.FloatTensor,
+            timestep: Union[torch.Tensor, float, int],
+            encoder_hidden_states: torch.Tensor,
+            encoder_attention_mask: torch.Tensor,
+            return_dict: bool = True,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        Args:
+            sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor
+            timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
+            encoder_hidden_states (`torch.FloatTensor`):
+                (batch_size, sequence_length, hidden_size) encoder hidden states
+            encoder_attention_mask (`torch.FloatTensor`):
+                (batch_size, sequence_length) encoder attention mask
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2 ** self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            timesteps = torch.tensor([timesteps], dtype=torch.long, device=sample.device)
+        elif torch.is_tensor(timesteps) and len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # timesteps does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=self.dtype)
+        emb = self.time_embedding(t_emb)
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # 3. down
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "attentions") and downsample_block.attentions is not None:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+            down_block_res_samples += res_samples
+
+        # 4. mid
+        sample = self.mid_block(sample, emb, encoder_hidden_states=encoder_hidden_states,
+                                encoder_attention_mask=encoder_attention_mask)
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets):]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "attentions") and upsample_block.attentions is not None:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    upsample_size=upsample_size,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
+                )
+        # 6. post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)

models/inception.py

@@ -0,0 +1,314 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision import models
+
+try:
+    from torchvision.models.utils import load_state_dict_from_url
+except ImportError:
+    from torch.utils.model_zoo import load_url as load_state_dict_from_url
+
+# Inception weights ported to Pytorch from
+# http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz
+FID_WEIGHTS_URL = 'https://github.com/mseitzer/pytorch-fid/releases/download/fid_weights/pt_inception-2015-12-05-6726825d.pth'
+
+
+class InceptionV3(nn.Module):
+    """Pretrained InceptionV3 network returning feature maps"""
+
+    # Index of default block of inception to return,
+    # corresponds to output of final average pooling
+    DEFAULT_BLOCK_INDEX = 3
+
+    # Maps feature dimensionality to their output blocks indices
+    BLOCK_INDEX_BY_DIM = {
+        64: 0,  # First max pooling features
+        192: 1,  # Second max pooling featurs
+        768: 2,  # Pre-aux classifier features
+        2048: 3  # Final average pooling features
+    }
+
+    def __init__(self,
+                 output_blocks=[DEFAULT_BLOCK_INDEX],
+                 resize_input=True,
+                 normalize_input=True,
+                 requires_grad=False,
+                 use_fid_inception=True):
+        """Build pretrained InceptionV3
+
+        Parameters
+        ----------
+        output_blocks : list of int
+            Indices of blocks to return features of. Possible values are:
+                - 0: corresponds to output of first max pooling
+                - 1: corresponds to output of second max pooling
+                - 2: corresponds to output which is fed to aux classifier
+                - 3: corresponds to output of final average pooling
+        resize_input : bool
+            If true, bilinearly resizes input to width and height 299 before
+            feeding input to model. As the network without fully connected
+            layers is fully convolutional, it should be able to handle inputs
+            of arbitrary size, so resizing might not be strictly needed
+        normalize_input : bool
+            If true, scales the input from range (0, 1) to the range the
+            pretrained Inception network expects, namely (-1, 1)
+        requires_grad : bool
+            If true, parameters of the model require gradients. Possibly useful
+            for finetuning the network
+        use_fid_inception : bool
+            If true, uses the pretrained Inception model used in Tensorflow's
+            FID implementation. If false, uses the pretrained Inception model
+            available in torchvision. The FID Inception model has different
+            weights and a slightly different structure from torchvision's
+            Inception model. If you want to compute FID scores, you are
+            strongly advised to set this parameter to true to get comparable
+            results.
+        """
+        super(InceptionV3, self).__init__()
+
+        self.resize_input = resize_input
+        self.normalize_input = normalize_input
+        self.output_blocks = sorted(output_blocks)
+        self.last_needed_block = max(output_blocks)
+
+        assert self.last_needed_block <= 3, \
+            'Last possible output block index is 3'
+
+        self.blocks = nn.ModuleList()
+
+        if use_fid_inception:
+            inception = fid_inception_v3()
+        else:
+            inception = models.inception_v3(pretrained=True)
+
+        # Block 0: input to maxpool1
+        block0 = [
+            inception.Conv2d_1a_3x3,
+            inception.Conv2d_2a_3x3,
+            inception.Conv2d_2b_3x3,
+            nn.MaxPool2d(kernel_size=3, stride=2)
+        ]
+        self.blocks.append(nn.Sequential(*block0))
+
+        # Block 1: maxpool1 to maxpool2
+        if self.last_needed_block >= 1:
+            block1 = [
+                inception.Conv2d_3b_1x1,
+                inception.Conv2d_4a_3x3,
+                nn.MaxPool2d(kernel_size=3, stride=2)
+            ]
+            self.blocks.append(nn.Sequential(*block1))
+
+        # Block 2: maxpool2 to aux classifier
+        if self.last_needed_block >= 2:
+            block2 = [
+                inception.Mixed_5b,
+                inception.Mixed_5c,
+                inception.Mixed_5d,
+                inception.Mixed_6a,
+                inception.Mixed_6b,
+                inception.Mixed_6c,
+                inception.Mixed_6d,
+                inception.Mixed_6e,
+            ]
+            self.blocks.append(nn.Sequential(*block2))
+
+        # Block 3: aux classifier to final avgpool
+        if self.last_needed_block >= 3:
+            block3 = [
+                inception.Mixed_7a,
+                inception.Mixed_7b,
+                inception.Mixed_7c,
+                nn.AdaptiveAvgPool2d(output_size=(1, 1))
+            ]
+            self.blocks.append(nn.Sequential(*block3))
+
+        for param in self.parameters():
+            param.requires_grad = requires_grad
+
+    def forward(self, inp):
+        """Get Inception feature maps
+
+        Parameters
+        ----------
+        inp : torch.autograd.Variable
+            Input tensor of shape Bx3xHxW. Values are expected to be in
+            range (0, 1)
+
+        Returns
+        -------
+        List of torch.autograd.Variable, corresponding to the selected output
+        block, sorted ascending by index
+        """
+        outp = []
+        x = inp
+
+        if self.resize_input:
+            x = F.interpolate(x,
+                              size=(299, 299),
+                              mode='bilinear',
+                              align_corners=False)
+
+        if self.normalize_input:
+            x = 2 * x - 1  # Scale from range (0, 1) to range (-1, 1)
+
+        for idx, block in enumerate(self.blocks):
+            x = block(x)
+            if idx in self.output_blocks:
+                outp.append(x)
+
+            if idx == self.last_needed_block:
+                break
+
+        return outp
+
+
+def fid_inception_v3():
+    """Build pretrained Inception model for FID computation
+
+    The Inception model for FID computation uses a different set of weights
+    and has a slightly different structure than torchvision's Inception.
+
+    This method first constructs torchvision's Inception and then patches the
+    necessary parts that are different in the FID Inception model.
+    """
+    inception = models.inception_v3(num_classes=1008,
+                                    aux_logits=False,
+                                    pretrained=False)
+    inception.Mixed_5b = FIDInceptionA(192, pool_features=32)
+    inception.Mixed_5c = FIDInceptionA(256, pool_features=64)
+    inception.Mixed_5d = FIDInceptionA(288, pool_features=64)
+    inception.Mixed_6b = FIDInceptionC(768, channels_7x7=128)
+    inception.Mixed_6c = FIDInceptionC(768, channels_7x7=160)
+    inception.Mixed_6d = FIDInceptionC(768, channels_7x7=160)
+    inception.Mixed_6e = FIDInceptionC(768, channels_7x7=192)
+    inception.Mixed_7b = FIDInceptionE_1(1280)
+    inception.Mixed_7c = FIDInceptionE_2(2048)
+
+    state_dict = load_state_dict_from_url(FID_WEIGHTS_URL, progress=True)
+    inception.load_state_dict(state_dict)
+    return inception
+
+
+class FIDInceptionA(models.inception.InceptionA):
+    """InceptionA block patched for FID computation"""
+
+    def __init__(self, in_channels, pool_features):
+        super(FIDInceptionA, self).__init__(in_channels, pool_features)
+
+    def forward(self, x):
+        branch1x1 = self.branch1x1(x)
+
+        branch5x5 = self.branch5x5_1(x)
+        branch5x5 = self.branch5x5_2(branch5x5)
+
+        branch3x3dbl = self.branch3x3dbl_1(x)
+        branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl)
+        branch3x3dbl = self.branch3x3dbl_3(branch3x3dbl)
+
+        # Patch: Tensorflow's average pool does not use the padded zero's in
+        # its average calculation
+        branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1,
+                                   count_include_pad=False)
+        branch_pool = self.branch_pool(branch_pool)
+
+        outputs = [branch1x1, branch5x5, branch3x3dbl, branch_pool]
+        return torch.cat(outputs, 1)
+
+
+class FIDInceptionC(models.inception.InceptionC):
+    """InceptionC block patched for FID computation"""
+
+    def __init__(self, in_channels, channels_7x7):
+        super(FIDInceptionC, self).__init__(in_channels, channels_7x7)
+
+    def forward(self, x):
+        branch1x1 = self.branch1x1(x)
+
+        branch7x7 = self.branch7x7_1(x)
+        branch7x7 = self.branch7x7_2(branch7x7)
+        branch7x7 = self.branch7x7_3(branch7x7)
+
+        branch7x7dbl = self.branch7x7dbl_1(x)
+        branch7x7dbl = self.branch7x7dbl_2(branch7x7dbl)
+        branch7x7dbl = self.branch7x7dbl_3(branch7x7dbl)
+        branch7x7dbl = self.branch7x7dbl_4(branch7x7dbl)
+        branch7x7dbl = self.branch7x7dbl_5(branch7x7dbl)
+
+        # Patch: Tensorflow's average pool does not use the padded zero's in
+        # its average calculation
+        branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1,
+                                   count_include_pad=False)
+        branch_pool = self.branch_pool(branch_pool)
+
+        outputs = [branch1x1, branch7x7, branch7x7dbl, branch_pool]
+        return torch.cat(outputs, 1)
+
+
+class FIDInceptionE_1(models.inception.InceptionE):
+    """First InceptionE block patched for FID computation"""
+
+    def __init__(self, in_channels):
+        super(FIDInceptionE_1, self).__init__(in_channels)
+
+    def forward(self, x):
+        branch1x1 = self.branch1x1(x)
+
+        branch3x3 = self.branch3x3_1(x)
+        branch3x3 = [
+            self.branch3x3_2a(branch3x3),
+            self.branch3x3_2b(branch3x3),
+        ]
+        branch3x3 = torch.cat(branch3x3, 1)
+
+        branch3x3dbl = self.branch3x3dbl_1(x)
+        branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl)
+        branch3x3dbl = [
+            self.branch3x3dbl_3a(branch3x3dbl),
+            self.branch3x3dbl_3b(branch3x3dbl),
+        ]
+        branch3x3dbl = torch.cat(branch3x3dbl, 1)
+
+        # Patch: Tensorflow's average pool does not use the padded zero's in
+        # its average calculation
+        branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1,
+                                   count_include_pad=False)
+        branch_pool = self.branch_pool(branch_pool)
+
+        outputs = [branch1x1, branch3x3, branch3x3dbl, branch_pool]
+        return torch.cat(outputs, 1)
+
+
+class FIDInceptionE_2(models.inception.InceptionE):
+    """Second InceptionE block patched for FID computation"""
+
+    def __init__(self, in_channels):
+        super(FIDInceptionE_2, self).__init__(in_channels)
+
+    def forward(self, x):
+        branch1x1 = self.branch1x1(x)
+
+        branch3x3 = self.branch3x3_1(x)
+        branch3x3 = [
+            self.branch3x3_2a(branch3x3),
+            self.branch3x3_2b(branch3x3),
+        ]
+        branch3x3 = torch.cat(branch3x3, 1)
+
+        branch3x3dbl = self.branch3x3dbl_1(x)
+        branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl)
+        branch3x3dbl = [
+            self.branch3x3dbl_3a(branch3x3dbl),
+            self.branch3x3dbl_3b(branch3x3dbl),
+        ]
+        branch3x3dbl = torch.cat(branch3x3dbl, 1)
+
+        # Patch: The FID Inception model uses max pooling instead of average
+        # pooling. This is likely an error in this specific Inception
+        # implementation, as other Inception models use average pooling here
+        # (which matches the description in the paper).
+        branch_pool = F.max_pool2d(x, kernel_size=3, stride=1, padding=1)
+        branch_pool = self.branch_pool(branch_pool)
+
+        outputs = [branch1x1, branch3x3, branch3x3dbl, branch_pool]
+        return torch.cat(outputs, 1)